From: Cordell Bloor Date: Thu, 13 Apr 2023 17:21:20 -0600 Subject: remove references to dfsg-violating kernels The DGEMM_Aldebaran_PKFixedAtomic512Latest and DGEMM_Aldebaran_PKFixedAtomic512_104 kernels were removed for dfsg reasons, and references to those kernels must be removed to fix the build. This will result in a performance drop on MI200 GPUs because the tuned assembly kernels will be replaced with fallback implementations for these problems. This problem has been reported upstream and they intend to supply a better fix. Forwarded: not-needed --- .../aldebaran/aldebaran_Cijk_Ailk_Bjlk_DB.yaml | 169126 ------------------ .../aldebaran/aldebaran_Cijk_Ailk_Bjlk_DB_GB.yaml | 152505 ---------------- .../aldebaran_Cijk_Ailk_Bjlk_DB.yaml | 155832 ---------------- .../aldebaran_Cijk_Ailk_Bjlk_DB.yaml | 144526 --------------- 4 files changed, 621989 deletions(-) delete mode 100644 library/src/blas3/Tensile/Logic/asm_full/aldebaran/aldebaran_Cijk_Ailk_Bjlk_DB.yaml delete mode 100644 library/src/blas3/Tensile/Logic/asm_full/aldebaran/aldebaran_Cijk_Ailk_Bjlk_DB_GB.yaml delete mode 100644 library/src/blas3/Tensile/Logic/asm_full/aldebaran_104cu/aldebaran_Cijk_Ailk_Bjlk_DB.yaml delete mode 100644 library/src/blas3/Tensile/Logic/nonMFMA_legacy/aldebaran_Cijk_Ailk_Bjlk_DB.yaml diff --git a/library/src/blas3/Tensile/Logic/asm_full/aldebaran/aldebaran_Cijk_Ailk_Bjlk_DB.yaml b/library/src/blas3/Tensile/Logic/asm_full/aldebaran/aldebaran_Cijk_Ailk_Bjlk_DB.yaml deleted file mode 100644 index 20f3a53..0000000 --- a/library/src/blas3/Tensile/Logic/asm_full/aldebaran/aldebaran_Cijk_Ailk_Bjlk_DB.yaml +++ /dev/null @@ -1,169126 +0,0 @@ -- {MinimumRequiredVersion: 4.8.1} -- aldebaran -- gfx90a -- [Device 0050, Device 0051, Device 0052, Device 0054, Device 0062, Device 7400, Device - 740c] -- AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false -- - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 0 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x064x04_PLR1_TT04_04_WG16_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: 1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 512 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 0 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 1 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 1 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 2 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 4 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 4 - NumGlobalWriteVectorsPerThread: 2 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 4 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 2 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 2] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: 1 - _VectorStore: 1 - _staggerStrideShift: 1 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 16 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsOffsetA: 0 - LdsOffsetB: 512 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 16 - LoopTail: true - LoopUnroll: 16 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 3 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 16 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsOffsetA: 0 - LdsOffsetB: 512 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 4 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 4 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 16 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsOffsetA: 0 - LdsOffsetB: 512 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 5 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 512 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 6 - StaggerU: 32 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 1 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 512 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 4 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 0 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 7 - StaggerU: 32 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 1 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 8 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_SE_FL0_WGM11 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 9 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_SE_FL0_WGM8 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - FractionalLoad: 1 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 10 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_SE_FL1_WGM8 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 11 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x4_SE_FL0_WGM11 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 3 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 12 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG16_16_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 13 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG8_32_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 4 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 14 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SE_1LDSB1_EPS1_IU4_NLCA1_PGR1_SIA2_TT8_32_WG16_16_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 15 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 1 - LSPB: 2 - LVCA: 64 - LVCB: 32 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 16 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_64_WG64_4_1_WGM11 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 17 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 18 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 19 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM10 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 10 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: -1 - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {3: 512} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {0: 128, 1: 128} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: DGEMM_Aldebaran_PKFixedAtomic512Latest - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: Branch - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 1 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 20 - SolutionNameMin: DGEMM_Aldebaran_PKFixedAtomic512Latest - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: -1 - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {3: 512} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {0: 128, 1: 128} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: DGEMM_Aldebaran_PKFixedAtomic512_104 - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: Branch - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 1 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 21 - SolutionNameMin: DGEMM_Aldebaran_PKFixedAtomic512_104 - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 22 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 23 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 24 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 25 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 26 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 27 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 28 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 29 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 30 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 31 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 32 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 33 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 34 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 35 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 36 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 37 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 38 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 39 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 40 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 41 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 42 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 43 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 44 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 45 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 46 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 47 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 48 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 49 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 50 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 51 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 52 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 53 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 54 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 55 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 56 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 57 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 58 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 59 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 60 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 61 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 62 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 63 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 64 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 65 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 66 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 67 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 68 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 69 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 70 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 71 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 72 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 73 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 74 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 75 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 76 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 77 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 78 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 79 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 80 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 81 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 82 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 83 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 84 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 85 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 86 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 87 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 88 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 89 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 90 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 91 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 92 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 93 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 94 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 95 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 96 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 97 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 98 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 99 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 100 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 101 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 102 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 103 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 104 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT8_32_WG16_16_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 105 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 106 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 107 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 108 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 109 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 110 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 111 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 112 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 113 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 114 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 115 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 116 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 117 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 118 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 119 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 120 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 121 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 122 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 123 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 124 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 125 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 126 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 127 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 128 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 129 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 130 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 131 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 132 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 133 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 134 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 135 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS128_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 136 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 137 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 138 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 139 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 140 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 141 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 142 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 143 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 144 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 145 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 146 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 147 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 148 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 149 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 150 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 151 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 152 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 153 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 154 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 155 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 156 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 157 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU32_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 158 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 159 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 160 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 161 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 162 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 163 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 164 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 165 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 166 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 167 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 168 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 169 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 170 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 171 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 172 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 173 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 174 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 175 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 176 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 177 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 178 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 179 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 180 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 181 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 182 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 183 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 184 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 185 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 186 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 187 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 188 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 189 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 190 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 191 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 192 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 193 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 194 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 195 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 196 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 197 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 198 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 199 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 200 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 201 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT4_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 202 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT4_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 203 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 204 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS128_TT4_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 16 - LSCB: 16 - LSPA: 4 - LSPB: 4 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 256 - LdsNumElementsAlignedA: 128 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 256 - LdsOffsetB: 128 - LdsOffsetB_Blk: 384 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 1] - MIWaveTile: [1, 1] - MIWaveTileA: 1 - MIWaveTileB: 1 - MacroTile0: 16 - MacroTile1: 16 - MacroTileA: 16 - MacroTileB: 16 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 4 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 64 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 205 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT16x16x8_MI16x16x4x1_SN_TT1_16_WG16_4_1 - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 16 - SubGroupA: 4 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [1, 16] - ThreadTile0: 4 - ThreadTile1: 1 - ThreadTileA: 4 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 206 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 207 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 208 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 209 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 210 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 211 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 212 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 213 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 214 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 215 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 216 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 217 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 218 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 219 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 220 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 221 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 222 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 223 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 224 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_PLR3_SU0_SUS0_SSO8_TT4_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 225 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_PLR3_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 226 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_PLR5_SU0_SUS0_SSO8_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 1 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 227 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_PLR3_SU0_SUS0_SSO4_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 228 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 8 - LSPB: 2 - LVCA: 32 - LVCB: 128 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 768 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 768 - LdsOffsetB_Blk: 2816 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 229 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x8_MI16x16x4x1_SN_PLR3_SU0_SUS0_SSO4_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 2] - MIWaveTileA: 3 - MIWaveTileB: 2 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 3 - NumLoadsB: 2 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 230 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 32] - ThreadTile0: 12 - ThreadTile1: 2 - ThreadTileA: 12 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 231 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_PLR5_SU32_SUS128_SSO4_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 232 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: 1 - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 233 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 234 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO8_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 235 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 236 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 237 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: 1 - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 238 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 239 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_PLR5_SU32_SUS128_SSO8_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: 1 - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 240 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 241 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU0_SUS0_SSO4_TT2_96_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 242 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW2_NLCA3_NLCB1_PLR5_SU32_SUS128_SSO4_TT3_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 243 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 244 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT2_64_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 245 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT4_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 246 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 247 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 1 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 248 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 249 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 1 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 250 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT2_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 2 - LSPB: 4 - LVCA: 128 - LVCB: 64 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 8 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 251 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 252 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 253 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 254 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 255 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 256 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW2_NLCA3_NLCB1_PLR5_SU0_SUS0_SSO4_TT3_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 257 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU0_SUS0_SSO4_TT2_96_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 258 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW2_NLCA3_NLCB1_PLR5_SU0_SUS0_SSO8_TT3_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 259 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO8_TT4_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 260 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 261 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_32_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 262 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO8_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 263 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 264 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO8_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 265 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT2_64_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 266 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT4_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 267 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 2 - LVCA: 64 - LVCB: 128 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 268 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT4_32_WG16_16_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 16 - LSPB: 16 - LVCA: 16 - LVCB: 16 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2560 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2560 - LdsOffsetB_Blk: 6656 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [5, 3] - MIWaveTileA: 5 - MIWaveTileB: 3 - MacroTile0: 160 - MacroTile1: 96 - MacroTileA: 160 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 60 - NumGlobalWriteVectorsPerThread: 60 - NumLoadsA: 5 - NumLoadsB: 3 - NumLoadsCoalescedA: 5 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 269 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT160x96x16_MI16x16x4x1_SN_GRVW2_NLCA5_NLCB3_PLR5_SU32_SUS128_SSO8_TT5_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [5, 48] - ThreadTile0: 20 - ThreadTile1: 3 - ThreadTileA: 20 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 270 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_32_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 271 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT2_32_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 272 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU32_SUS256_SSO8_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 273 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS256_SSO4_TT2_48_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 274 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO4_TT4_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 275 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT2_32_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 276 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 277 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU32_SUS256_SSO4_TT4_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 278 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO4_TT4_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 279 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS256_SSO4_TT4_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 2 - LSPB: 4 - LVCA: 128 - LVCB: 64 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 8 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 280 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT2_64_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 281 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO4_TT4_48_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2560 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2560 - LdsOffsetB_Blk: 6656 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [5, 3] - MIWaveTileA: 5 - MIWaveTileB: 3 - MacroTile0: 160 - MacroTile1: 96 - MacroTileA: 160 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 60 - NumGlobalWriteVectorsPerThread: 60 - NumLoadsA: 10 - NumLoadsB: 6 - NumLoadsCoalescedA: 5 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 282 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT160x96x16_MI16x16x4x1_SN_GRVW1_NLCA5_NLCB3_PLR5_SU0_SUS0_SSO4_TT5_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [5, 48] - ThreadTile0: 20 - ThreadTile1: 3 - ThreadTileA: 20 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 283 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT4_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 1 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 284 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1280 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1280 - LdsOffsetB_Blk: 3328 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [5, 3] - MIWaveTileA: 5 - MIWaveTileB: 3 - MacroTile0: 160 - MacroTile1: 96 - MacroTileA: 160 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 60 - NumGlobalWriteVectorsPerThread: 60 - NumLoadsA: 5 - NumLoadsB: 3 - NumLoadsCoalescedA: 5 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 285 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT160x96x8_MI16x16x4x1_SN_GRVW1_NLCA5_NLCB3_PLR3_SU32_SUS128_SSO8_TT5_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [5, 48] - ThreadTile0: 20 - ThreadTile1: 3 - ThreadTileA: 20 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 286 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 287 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 8 - LSPB: 2 - LVCA: 32 - LVCB: 128 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 768 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 768 - LdsOffsetB_Blk: 2816 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 288 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x8_MI16x16x4x1_SN_GRVW1_NLCA3_NLCB1_PLR3_SU0_SUS0_SSO4_TT3_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 289 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT4_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 290 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO8_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 2 - LSPB: 4 - LVCA: 128 - LVCB: 64 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 8 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 291 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 292 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 293 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT8_32_WG16_16_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 294 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 295 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT8_32_WG16_16_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 296 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU32_SUS128_SSO8_TT8_32_WG16_16_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 2 - LVCA: 64 - LVCB: 128 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 297 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT2_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 298 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB3_PLR3_SU32_SUS128_SSO8_TT4_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [6, 1] - MIWaveTileA: 6 - MIWaveTileB: 1 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 6 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 299 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x16_MI16x16x4x1_SN_GRVW1_NLCA3_NLCB1_PLR5_SU0_SUS0_SSO4_TT6_16_WG16_16_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [6, 16] - ThreadTile0: 24 - ThreadTile1: 1 - ThreadTileA: 24 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 300 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_64_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 2 - LVCA: 64 - LVCB: 128 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 301 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 302 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 8 - LSPB: 2 - LVCA: 32 - LVCB: 128 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 6 - NumLoadsB: 8 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 303 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW1_NLCA3_NLCB1_PLR5_SU0_SUS0_SSO8_TT3_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 304 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT8_32_WG16_16_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2560 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2560 - LdsOffsetB_Blk: 6656 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [5, 3] - MIWaveTileA: 5 - MIWaveTileB: 3 - MacroTile0: 160 - MacroTile1: 96 - MacroTileA: 160 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 60 - NumGlobalWriteVectorsPerThread: 60 - NumLoadsA: 10 - NumLoadsB: 6 - NumLoadsCoalescedA: 5 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 305 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT160x96x16_MI16x16x4x1_SN_GRVW1_NLCA5_NLCB3_PLR5_SU32_SUS256_SSO4_TT5_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [5, 48] - ThreadTile0: 20 - ThreadTile1: 3 - ThreadTileA: 20 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 306 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT8_32_WG16_16_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 307 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_32_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 308 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT4_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 309 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 310 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS256_SSO4_TT4_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2560 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2560 - LdsOffsetB_Blk: 6656 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [5, 3] - MIWaveTileA: 5 - MIWaveTileB: 3 - MacroTile0: 160 - MacroTile1: 96 - MacroTileA: 160 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 60 - NumGlobalWriteVectorsPerThread: 60 - NumLoadsA: 10 - NumLoadsB: 6 - NumLoadsCoalescedA: 5 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 311 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT160x96x16_MI16x16x4x1_SN_GRVW1_NLCA5_NLCB3_PLR5_SU32_SUS128_SSO8_TT5_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [5, 48] - ThreadTile0: 20 - ThreadTile1: 3 - ThreadTileA: 20 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 2 - LVCA: 64 - LVCB: 128 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 312 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT4_32_WG16_16_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 6 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 313 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO8_TT2_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 314 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 2 - LSPB: 4 - LVCA: 128 - LVCB: 64 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 8 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 315 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO8_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [1, 6] - MIWaveTileA: 1 - MIWaveTileB: 6 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 6 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 316 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB3_PLR5_SU32_SUS256_SSO4_TT1_96_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [1, 96] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 2 - LVCA: 64 - LVCB: 128 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 317 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU32_SUS128_SSO4_TT2_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 318 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT8_32_WG16_16_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 319 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT8_32_WG16_16_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 320 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT8_32_WG16_16_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 8 - NumLoadsB: 6 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 321 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO4_TT2_96_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 322 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT2_32_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 8 - NumLoadsB: 6 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 323 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO4_TT4_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 8 - LSPB: 2 - LVCA: 32 - LVCB: 128 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 6 - NumLoadsB: 8 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 324 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW1_NLCA3_NLCB1_PLR5_SU32_SUS128_SSO4_TT3_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 325 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_32_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 326 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO8_TT2_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 327 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT2_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 328 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO8_TT2_32_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 329 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT2_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 330 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_ - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 0 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 331 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 332 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 333 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 334 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 335 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_48_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 336 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_48_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 337 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 338 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 339 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_48_VW2_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 340 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 341 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 342 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 343 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 344 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 345 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 346 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_48_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 347 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 348 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 349 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 350 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 351 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 352 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 353 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 354 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 355 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 356 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 357 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 358 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 359 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 360 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 361 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 362 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 363 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 364 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 365 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 366 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 367 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 368 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 369 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 370 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 371 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 372 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 373 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 374 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 375 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 376 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 377 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 378 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 379 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 380 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 381 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 382 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 383 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 384 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 385 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 386 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 387 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 388 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 389 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 390 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB2_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 391 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 392 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 393 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 394 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 395 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 396 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 397 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 398 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 399 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 400 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 401 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 402 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 403 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 404 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 405 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 406 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 407 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 408 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 409 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 410 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 411 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 412 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB4_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 413 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB4_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 414 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB4_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 415 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 416 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB4_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 417 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 418 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 419 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_32_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 420 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_48_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 421 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS256_SVW2_TT2_48_VW2_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 422 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 423 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 424 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 425 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 426 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 2] - MIWaveTileA: 3 - MIWaveTileB: 2 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 3 - NumLoadsB: 2 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 427 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU0_SUS0_SVW1_TT3_32_VW1_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 32] - ThreadTile0: 12 - ThreadTile1: 2 - ThreadTileA: 12 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 428 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 429 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 430 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 2] - MIWaveTileA: 3 - MIWaveTileB: 2 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 3 - NumLoadsB: 2 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 431 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU0_SUS0_SVW1_TT3_32_VW1_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 32] - ThreadTile0: 12 - ThreadTile1: 2 - ThreadTileA: 12 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 2] - MIWaveTileA: 3 - MIWaveTileB: 2 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 3 - NumLoadsB: 2 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 432 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU32_SUS128_SVW1_TT3_32_VW1_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 32] - ThreadTile0: 12 - ThreadTile1: 2 - ThreadTileA: 12 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 433 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 434 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 8 - LSPB: 2 - LVCA: 32 - LVCB: 128 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 768 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 768 - LdsOffsetB_Blk: 2816 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 435 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x8_MI16x16x4x1_SN_AMAS0_GRVW1_NEPBS2_NLCA3_NLCB1_PLR3_SU0_SUS0_SVW1_TT3_64_VW1_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 436 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU32_SUS256_SVW1_TT3_64_VW1_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 437 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 438 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 439 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 440 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 441 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 442 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 443 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU32_SUS256_SVW1_TT3_64_VW1_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 444 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 445 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 446 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 447 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 448 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 449 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 450 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 451 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 452 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB4_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 453 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 454 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 455 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB4_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 456 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB4_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 457 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB4_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 458 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 459 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_48_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 460 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 461 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_48_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 462 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 463 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 464 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 465 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 466 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 467 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 468 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 469 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 470 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 471 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 472 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 473 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 474 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 475 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 476 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 477 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 478 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 479 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 480 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 481 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 482 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 483 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 484 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 485 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 486 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 487 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 488 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 489 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 490 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 491 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 492 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 493 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 494 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 495 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 496 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 497 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 498 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 499 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 500 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 501 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 502 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 503 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU0_SUS0_SSO4_TT2_96_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollInterval: 2 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 504 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO8_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollInterval: 2 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 505 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU0_SUS0_SSO4_TT2_96_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollInterval: 2 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 506 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollInterval: 2 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 507 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollInterval: 2 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 508 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollInterval: 2 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 509 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU0_SUS0_SSO4_TT2_96_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollInterval: 2 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 510 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW2_NLCA3_NLCB1_PLR5_SU0_SUS0_SSO4_TT3_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollInterval: 2 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 511 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU0_SUS0_SSO4_TT2_96_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollInterval: 2 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 512 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO8_TT2_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollInterval: 2 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 513 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT2_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollInterval: 2 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 514 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollInterval: 2 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 32 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 16 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 16 - LSPB: 16 - LVCA: 16 - LVCB: 16 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 32 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [1, 1] - MIWaveTileA: 1 - MIWaveTileB: 1 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 4 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 515 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x32x32_MI16x16x4x1_SN_AF0EM1_AMAS0_ETSP_EPS1_GRVW2_GSU16_PGR2_SUS256_SVW1_TT1_16_VW1_WG32_8_1 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [1, 16] - ThreadTile0: 4 - ThreadTile1: 1 - ThreadTileA: 4 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 32 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 32 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: Branch - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 8 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 16 - LSPB: 16 - LVCA: 16 - LVCB: 16 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 32 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [1, 1] - MIWaveTileA: 1 - MIWaveTileB: 1 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 4 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 516 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x32x32_MI16x16x4x1_SN_AF0EM1_AMAS0_ETB_EPS1_GRVW2_GSU8_PGR2_SUS256_SVW1_TT1_16_VW1_WG32_8_1 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [1, 16] - ThreadTile0: 4 - ThreadTile1: 1 - ThreadTileA: 4 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 32 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 64 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: Branch - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 16 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 16 - LSPA: 16 - LSPB: 32 - LVCA: 16 - LVCB: 8 - LVPA: 8 - LVPB: 16 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 7168 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 16 - LoopTail: true - LoopUnroll: 64 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [4, 4, 4, 4, 4, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 32 - MacroTile1: 16 - MacroTileA: 32 - MacroTileB: 16 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 4 - MatrixInstBM: 4 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 4 - MatrixInstN: 4 - MatrixInstruction: [4, 4, 4, 4] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 2 - NumGlobalWriteVectorsPerThread: 2 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 517 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x16x64_MI4x4x4x4_SN_AF0EM2_AMAS0_ETB_GRVW2_GSU16_PGR1_SUS512_SVW1_VW1 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 512 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 1 - ThreadTileA: 2 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 64 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 64 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: None - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 16 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 16 - LSPA: 16 - LSPB: 32 - LVCA: 16 - LVCB: 8 - LVPA: 8 - LVPB: 16 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 7168 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 16 - LoopTail: true - LoopUnroll: 64 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [4, 4, 4, 4, 4, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 32 - MacroTile1: 16 - MacroTileA: 32 - MacroTileB: 16 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 4 - MatrixInstBM: 4 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 4 - MatrixInstN: 4 - MatrixInstruction: [4, 4, 4, 4] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 2 - NumGlobalWriteVectorsPerThread: 2 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 518 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x16x64_MI4x4x4x4_SN_AF0EM2_AMAS0_ETN_GRVW2_GSU16_PGR1_SUS512_SVW1_VW1 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 512 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 1 - ThreadTileA: 2 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 64 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 0 - LdsOffsetB_Blk: 1024 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 519 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_PLR3_SU32_SUS128_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 0 - LdsOffsetB_Blk: 1024 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 520 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_PLR3_SU8_SUS128_WGM11 - SourceSwap: true - StaggerU: 8 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 521 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_PLR5_SU8_SUS128_WGM11 - SourceSwap: true - StaggerU: 8 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 0 - LdsOffsetB_Blk: 1024 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 522 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_PLR3_SU8_SUS128_WGM8 - SourceSwap: true - StaggerU: 8 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 523 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 0 - LdsOffsetB_Blk: 1024 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 524 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_PLR3_SU16_SUS128_WGM11 - SourceSwap: true - StaggerU: 16 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 525 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM6 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 6 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 526 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM7 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 7 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 527 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM13 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 13 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 528 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM9 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 9 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 529 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM8 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 530 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM10 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 10 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 531 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM5 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 532 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM12 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 12 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 533 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM11 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 534 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM14 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 14 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 535 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM4 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false -- [2, 3, 0, 1] -- - - [38144, 38144, 1, 256] - - [22, 74.893] - - - [29568, 128, 1, 384] - - [34, 47.913] - - - [30848, 128, 1, 256] - - [23, 41.422] - - - [25728, 128, 1, 384] - - [40, 57.278] - - - [32256, 32256, 1, 256] - - [24, 75.584] - - - [7680, 7680, 1, 256] - - [57, 73.456] - - - [41984, 41984, 1, 256] - - [25, 75.197] - - - [40448, 40448, 1, 256] - - [24, 74.937] - - - [25728, 128, 1, 256] - - [26, 49.099] - - - [64, 64, 1, 64] - - [205, 0.117] - - - [15104, 15104, 1, 256] - - [27, 75.487] - - - [17280, 17280, 1, 384] - - [27, 90.896] - - - [34688, 128, 1, 384] - - [23, 54.573] - - - [27392, 27392, 1, 256] - - [28, 74.926] - - - [6528, 128, 1, 256] - - [108, 37.791] - - - [35328, 35328, 1, 256] - - [51, 75.386] - - - [18432, 18432, 1, 384] - - [29, 89.766] - - - [31232, 31232, 1, 256] - - [48, 75.543] - - - [7808, 128, 1, 256] - - [109, 35.357] - - - [38400, 38400, 1, 384] - - [25, 90.583] - - - [16128, 16128, 1, 256] - - [37, 75.578] - - - [9472, 9472, 1, 256] - - [27, 74.227] - - - [21888, 21888, 1, 384] - - [28, 89.215] - - - [38656, 38656, 1, 256] - - [30, 74.685] - - - [20224, 20224, 1, 256] - - [37, 75.749] - - - [8960, 8960, 1, 256] - - [31, 73.31] - - - [29952, 29952, 1, 384] - - [27, 90.917] - - - [36864, 36864, 1, 384] - - [32, 87.926] - - - [33408, 33408, 1, 384] - - [29, 90.847] - - - [20608, 128, 1, 384] - - [33, 48.729] - - - [23424, 23424, 1, 384] - - [27, 91.11] - - - [4864, 4864, 1, 256] - - [34, 68.048] - - - [21504, 21504, 1, 384] - - [41, 89.214] - - - [25600, 25600, 1, 256] - - [29, 75.967] - - - [40960, 40960, 1, 256] - - [54, 66.623] - - - [19200, 19200, 1, 384] - - [29, 90.96] - - - [64, 1, 1, 64] - - [205, 0.002] - - - [25088, 25088, 1, 256] - - [42, 75.905] - - - [41728, 41728, 1, 256] - - [28, 73.787] - - - [35840, 35840, 1, 256] - - [37, 75.487] - - - [34560, 34560, 1, 256] - - [27, 75.1] - - - [26368, 26368, 1, 256] - - [25, 75.32] - - - [5888, 5888, 1, 256] - - [35, 70.595] - - - [28032, 28032, 1, 384] - - [27, 91.053] - - - [42496, 42496, 1, 256] - - [28, 75.137] - - - [27008, 128, 1, 256] - - [33, 50.615] - - - [38400, 38400, 1, 256] - - [25, 75.384] - - - [11008, 11008, 1, 256] - - [85, 73.892] - - - [32000, 32000, 1, 256] - - [27, 75.005] - - - [37248, 37248, 1, 384] - - [25, 90.692] - - - [10496, 10496, 1, 256] - - [25, 74.439] - - - [16640, 16640, 1, 256] - - [25, 75.73] - - - [24960, 24960, 1, 384] - - [25, 91.209] - - - [18688, 18688, 1, 256] - - [27, 75.593] - - - [22272, 22272, 1, 384] - - [36, 91.067] - - - [15488, 128, 1, 256] - - [59, 32.909] - - - [28416, 28416, 1, 384] - - [25, 90.867] - - - [3840, 3840, 1, 256] - - [34, 64.173] - - - [19968, 19968, 1, 384] - - [25, 90.738] - - - [43776, 43776, 1, 256] - - [28, 74.28] - - - [35072, 35072, 1, 256] - - [25, 75.071] - - - [20736, 20736, 1, 256] - - [27, 75.611] - - - [7168, 7168, 1, 256] - - [37, 72.544] - - - [18432, 18432, 1, 256] - - [27, 76.049] - - - [38016, 38016, 1, 384] - - [25, 91.017] - - - [35328, 35328, 1, 384] - - [25, 90.526] - - - [38784, 38784, 1, 384] - - [25, 90.979] - - - [26112, 26112, 1, 384] - - [29, 91.076] - - - [27264, 27264, 1, 384] - - [29, 90.776] - - - [44928, 44928, 1, 384] - - [27, 90.696] - - - [41088, 128, 1, 384] - - [36, 60.328] - - - [42368, 128, 1, 256] - - [23, 44.691] - - - [10752, 10752, 1, 256] - - [27, 75.284] - - - [9088, 128, 1, 384] - - [109, 46.368] - - - [17152, 17152, 1, 256] - - [47, 75.562] - - - [44928, 128, 1, 384] - - [36, 54.276] - - - [7808, 128, 1, 384] - - [110, 40.491] - - - [29184, 29184, 1, 256] - - [37, 75.558] - - - [11776, 11776, 1, 256] - - [25, 75.132] - - - [1, 64, 1, 64] - - [205, 0.002] - - - [27136, 27136, 1, 256] - - [42, 75.911] - - - [33408, 128, 1, 256] - - [35, 44.177] - - - [33792, 33792, 1, 384] - - [41, 89.232] - - - [43520, 43520, 1, 256] - - [38, 75.193] - - - [14592, 14592, 1, 384] - - [37, 90.309] - - - [41472, 41472, 1, 256] - - [25, 74.99] - - - [14080, 14080, 1, 256] - - [57, 74.446] - - - [34688, 128, 1, 256] - - [39, 45.463] - - - [16896, 16896, 1, 256] - - [37, 76.106] - - - [15744, 15744, 1, 384] - - [25, 90.68] - - - [28416, 28416, 1, 256] - - [25, 74.869] - - - [23808, 23808, 1, 256] - - [27, 75.555] - - - [27648, 27648, 1, 256] - - [25, 75.552] - - - [1152, 3072, 1, 384] - - [40, 63.809] - - - [21888, 128, 1, 256] - - [35, 43.504] - - - [34816, 34816, 1, 256] - - [29, 75.387] - - - [43776, 43776, 1, 384] - - [41, 89.916] - - - [36096, 36096, 1, 256] - - [41, 74.203] - - - [24320, 24320, 1, 256] - - [24, 75.666] - - - [12544, 12544, 1, 256] - - [27, 74.955] - - - [29184, 29184, 1, 384] - - [29, 90.781] - - - [29568, 29568, 1, 384] - - [24, 90.297] - - - [12928, 128, 1, 384] - - [111, 54.014] - - - [36480, 36480, 1, 384] - - [25, 90.923] - - - [30720, 30720, 1, 256] - - [29, 75.552] - - - [25728, 25728, 1, 384] - - [58, 90.933] - - - [34048, 34048, 1, 256] - - [42, 74.525] - - - [12928, 128, 1, 256] - - [112, 45.988] - - - [9728, 9728, 1, 256] - - [29, 74.669] - - - [128, 128, 1, 256] - - [113, 1.074] - - - [33024, 33024, 1, 256] - - [43, 75.213] - - - [15488, 128, 1, 384] - - [57, 37.868] - - - [39808, 128, 1, 384] - - [44, 59.233] - - - [18176, 18176, 1, 256] - - [27, 75.705] - - - [21504, 21504, 1, 256] - - [27, 76.075] - - - [16384, 16384, 1, 256] - - [54, 62.756] - - - [27008, 128, 1, 384] - - [45, 59.129] - - - [27904, 27904, 1, 256] - - [38, 75.163] - - - [24448, 128, 1, 384] - - [46, 55.216] - - - [35968, 128, 1, 384] - - [53, 55.639] - - - [37632, 37632, 1, 256] - - [29, 74.955] - - - [14848, 14848, 1, 256] - - [28, 75.624] - - - [23552, 23552, 1, 256] - - [29, 76.111] - - - [4608, 4608, 1, 50000] - - [53, 97.942] - - - [13056, 13056, 1, 256] - - [29, 75.313] - - - [38528, 128, 1, 256] - - [58, 48.177] - - - [19584, 19584, 1, 384] - - [37, 91.095] - - - [16768, 128, 1, 384] - - [57, 40.639] - - - [22784, 22784, 1, 256] - - [47, 75.492] - - - [44160, 44160, 1, 384] - - [29, 90.838] - - - [28160, 28160, 1, 256] - - [48, 75.558] - - - [14592, 14592, 1, 256] - - [68, 74.291] - - - [20992, 20992, 1, 256] - - [37, 76.172] - - - [41216, 41216, 1, 256] - - [25, 74.854] - - - [21760, 21760, 1, 256] - - [27, 75.734] - - - [25344, 25344, 1, 256] - - [41, 74.966] - - - [4608, 4608, 1, 256] - - [49, 69.948] - - - [2560, 2048, 1, 256] - - [49, 56.06] - - - [30464, 30464, 1, 256] - - [50, 74.371] - - - [19200, 19200, 1, 256] - - [37, 75.675] - - - [22272, 22272, 1, 256] - - [51, 75.594] - - - [29952, 29952, 1, 256] - - [47, 75.275] - - - [20480, 20480, 1, 256] - - [29, 75.812] - - - [17408, 17408, 1, 256] - - [25, 75.599] - - - [32768, 32768, 1, 256] - - [52, 57.634] - - - [18816, 18816, 1, 384] - - [58, 90.869] - - - [34944, 34944, 1, 384] - - [37, 90.977] - - - [18048, 18048, 1, 384] - - [58, 90.958] - - - [34560, 34560, 1, 384] - - [37, 90.826] - - - [9088, 128, 1, 256] - - [114, 40.41] - - - [24576, 24576, 1, 256] - - [54, 70.192] - - - [32128, 128, 1, 384] - - [49, 51.321] - - - [8448, 8448, 1, 256] - - [57, 73.715] - - - [42752, 42752, 1, 256] - - [27, 74.735] - - - [5376, 5376, 1, 256] - - [49, 69.366] - - - [18048, 128, 1, 256] - - [39, 37.6] - - - [3584, 3584, 1, 256] - - [35, 62.742] - - - [37120, 37120, 1, 256] - - [27, 74.921] - - - [39936, 39936, 1, 384] - - [61, 89.019] - - - [20736, 20736, 1, 384] - - [53, 91.075] - - - [35584, 35584, 1, 256] - - [37, 74.893] - - - [26112, 26112, 1, 256] - - [48, 76.0] - - - [16896, 16896, 1, 384] - - [37, 90.639] - - - [40704, 40704, 1, 384] - - [25, 90.834] - - - [33280, 33280, 1, 256] - - [24, 75.635] - - - [5632, 5632, 1, 256] - - [34, 71.658] - - - [19456, 19456, 1, 256] - - [27, 75.886] - - - [22016, 22016, 1, 256] - - [30, 76.148] - - - [14208, 128, 1, 256] - - [33, 30.277] - - - [13568, 13568, 1, 256] - - [25, 75.309] - - - [30848, 128, 1, 384] - - [60, 49.625] - - - [1408, 128, 1, 384] - - [115, 13.439] - - - [5760, 5760, 1, 5760] - - [239, 98.181] - - - [39936, 39936, 1, 256] - - [25, 75.103] - - - [1920, 3072, 1, 384] - - [35, 62.558] - - - [9984, 9984, 1, 256] - - [35, 74.421] - - - [2816, 2048, 1, 256] - - [35, 51.272] - - - [23168, 128, 1, 256] - - [47, 45.196] - - - [19968, 19968, 1, 256] - - [28, 76.018] - - - [44800, 44800, 1, 256] - - [25, 74.449] - - - [14976, 14976, 1, 384] - - [44, 90.548] - - - [35712, 35712, 1, 384] - - [25, 90.953] - - - [43008, 43008, 1, 384] - - [54, 89.03] - - - [41088, 41088, 1, 384] - - [41, 90.052] - - - [16128, 16128, 1, 384] - - [25, 90.362] - - - [5120, 5120, 1, 256] - - [35, 70.401] - - - [25856, 25856, 1, 256] - - [29, 75.426] - - - [12288, 12288, 1, 256] - - [27, 75.012] - - - [6400, 6400, 1, 256] - - [55, 72.272] - - - [2688, 128, 1, 256] - - [116, 20.246] - - - [11648, 128, 1, 256] - - [117, 42.668] - - - [43264, 43264, 1, 256] - - [25, 74.81] - - - [19712, 19712, 1, 256] - - [32, 75.0] - - - [34176, 34176, 1, 384] - - [27, 90.755] - - - [31104, 31104, 1, 384] - - [27, 90.907] - - - [36608, 36608, 1, 256] - - [38, 74.827] - - - [39808, 128, 1, 256] - - [40, 49.04] - - - [13824, 13824, 1, 384] - - [29, 90.095] - - - [42624, 42624, 1, 384] - - [28, 87.986] - - - [21120, 21120, 1, 384] - - [37, 91.055] - - - [23296, 23296, 1, 256] - - [25, 75.634] - - - [42240, 42240, 1, 256] - - [27, 74.739] - - - [33408, 128, 1, 384] - - [35, 52.984] - - - [43648, 128, 1, 256] - - [40, 45.265] - - - [19328, 128, 1, 384] - - [56, 46.245] - - - [33792, 33792, 1, 256] - - [25, 75.516] - - - [31488, 31488, 1, 256] - - [37, 74.936] - - - [768, 3072, 1, 384] - - [57, 44.769] - - - [6144, 6144, 1, 256] - - [37, 71.929] - - - [20352, 20352, 1, 384] - - [27, 91.117] - - - [23168, 128, 1, 384] - - [39, 53.248] - - - [33536, 33536, 1, 256] - - [50, 75.178] - - - [32640, 32640, 1, 384] - - [29, 90.087] - - - [1536, 3072, 1, 384] - - [35, 63.449] - - - [19328, 128, 1, 256] - - [40, 40.04] - - - [2688, 3072, 1, 384] - - [49, 72.101] - - - [24192, 24192, 1, 384] - - [27, 91.241] - - - [6912, 6912, 1, 256] - - [34, 72.526] - - - [15360, 15360, 1, 256] - - [25, 76.063] - - - [18944, 18944, 1, 256] - - [24, 76.049] - - - [37376, 37376, 1, 256] - - [41, 75.399] - - - [31488, 31488, 1, 384] - - [29, 90.749] - - - [26880, 26880, 1, 256] - - [25, 75.31] - - - [44928, 128, 1, 128] - - [40, 31.959] - - - [24448, 128, 1, 256] - - [57, 46.959] - - - [31872, 31872, 1, 384] - - [29, 90.846] - - - [1408, 128, 1, 256] - - [118, 11.358] - - - [38528, 128, 1, 384] - - [58, 58.149] - - - [15616, 15616, 1, 256] - - [25, 75.505] - - - [39552, 39552, 1, 384] - - [25, 90.918] - - - [4352, 4352, 1, 256] - - [34, 67.359] - - - [28288, 128, 1, 384] - - [34, 46.221] - - - [10368, 128, 1, 256] - - [109, 43.982] - - - [32128, 128, 1, 256] - - [40, 42.898] - - - [4608, 4608, 1, 4608] - - [233, 97.048] - - - [8704, 8704, 1, 256] - - [29, 74.167] - - - [17664, 17664, 1, 256] - - [30, 74.999] - - - [24576, 24576, 1, 384] - - [32, 83.07] - - - [37248, 128, 1, 384] - - [55, 56.876] - - - [34304, 34304, 1, 256] - - [43, 75.525] - - - [42368, 128, 1, 384] - - [27, 52.092] - - - [17664, 17664, 1, 384] - - [27, 90.703] - - - [12800, 12800, 1, 256] - - [27, 75.345] - - - [26624, 26624, 1, 256] - - [37, 75.83] - - - [36864, 36864, 1, 256] - - [29, 74.612] - - - [40704, 40704, 1, 256] - - [27, 74.86] - - - [12032, 12032, 1, 256] - - [27, 74.709] - - - [33024, 33024, 1, 384] - - [25, 90.647] - - - [28800, 28800, 1, 384] - - [53, 91.023] - - - [22656, 22656, 1, 384] - - [25, 91.179] - - - [41472, 41472, 1, 384] - - [27, 90.471] - - - [39680, 39680, 1, 256] - - [29, 74.613] - - - [44032, 44032, 1, 256] - - [37, 75.129] - - - [43392, 43392, 1, 384] - - [25, 90.699] - - - [42240, 42240, 1, 384] - - [27, 90.742] - - - [38912, 38912, 1, 256] - - [29, 75.302] - - - [23040, 23040, 1, 384] - - [27, 91.024] - - - [13312, 13312, 1, 256] - - [25, 75.812] - - - [128, 128, 1, 384] - - [119, 1.211] - - - [39168, 39168, 1, 256] - - [41, 74.682] - - - [25344, 25344, 1, 384] - - [27, 91.121] - - - [5248, 128, 1, 256] - - [109, 31.75] - - - [30208, 30208, 1, 256] - - [42, 75.621] - - - [40192, 40192, 1, 256] - - [29, 74.915] - - - [15872, 15872, 1, 256] - - [25, 76.103] - - - [44544, 44544, 1, 256] - - [22, 74.966] - - - [11520, 11520, 1, 256] - - [29, 74.674] - - - [15360, 15360, 1, 384] - - [47, 89.355] - - - [23040, 23040, 1, 256] - - [27, 76.194] - - - [26496, 26496, 1, 384] - - [29, 91.077] - - - [11264, 11264, 1, 256] - - [25, 75.369] - - - [18048, 128, 1, 384] - - [59, 43.408] - - - [30976, 30976, 1, 256] - - [28, 74.406] - - - [11648, 128, 1, 384] - - [120, 50.278] - - - [2304, 3072, 1, 384] - - [35, 73.922] - - - [28928, 28928, 1, 256] - - [30, 74.911] - - - [43008, 43008, 1, 256] - - [25, 74.922] - - - [29440, 29440, 1, 256] - - [30, 75.408] - - - [36352, 36352, 1, 256] - - [41, 75.406] - - - [32256, 32256, 1, 384] - - [37, 90.64] - - - [23808, 23808, 1, 384] - - [53, 91.162] - - - [37248, 128, 1, 256] - - [33, 47.064] - - - [1, 1, 1, 64] - - [0, 0.0] - - - [37888, 37888, 1, 256] - - [29, 75.212] - - - [35968, 128, 1, 256] - - [59, 46.038] - - - [13824, 13824, 1, 256] - - [51, 75.675] - - - [39168, 39168, 1, 384] - - [25, 90.701] - - - [37632, 37632, 1, 384] - - [25, 90.859] - - - [29568, 128, 1, 256] - - [60, 39.885] - - - [14336, 14336, 1, 256] - - [29, 75.668] - - - [28288, 128, 1, 256] - - [60, 38.51] - - - [16512, 16512, 1, 384] - - [54, 89.482] - - - [30720, 30720, 1, 384] - - [32, 89.477] - - - [21248, 21248, 1, 256] - - [37, 75.578] - - - [29696, 29696, 1, 256] - - [37, 75.577] - - - [384, 3072, 1, 384] - - [109, 48.474] - - - [28672, 28672, 1, 256] - - [27, 75.173] - - - [32512, 32512, 1, 256] - - [24, 75.093] - - - [9216, 9216, 1, 256] - - [27, 74.101] - - - [6656, 6656, 1, 256] - - [34, 72.68] - - - [30336, 30336, 1, 384] - - [25, 90.946] - - - [20608, 128, 1, 256] - - [47, 41.527] - - - [7936, 7936, 1, 256] - - [35, 73.5] - - - [41856, 41856, 1, 384] - - [29, 90.931] - - - [44288, 44288, 1, 256] - - [37, 74.601] - - - [7744, 7744, 1, 7744] - - [25, 97.287] - - - [7424, 7424, 1, 256] - - [49, 73.14] - - - [39424, 39424, 1, 256] - - [25, 75.347] - - - [43648, 128, 1, 384] - - [36, 52.576] - - - [14208, 14208, 1, 384] - - [37, 89.943] - - - [36096, 36096, 1, 384] - - [28, 89.977] - - - [44544, 44544, 1, 384] - - [37, 90.389] - - - [22528, 22528, 1, 256] - - [29, 75.715] - - - [4096, 4096, 1, 256] - - [35, 66.031] - - - [31744, 31744, 1, 256] - - [37, 75.52] - - - [3968, 128, 1, 384] - - [108, 28.344] - - - [17920, 17920, 1, 256] - - [29, 76.073] - - - [5248, 128, 1, 384] - - [108, 36.66] - - - [26880, 26880, 1, 384] - - [37, 91.155] - - - [8192, 8192, 1, 256] - - [25, 72.824] - - - [3968, 128, 1, 256] - - [110, 24.407] - - - [41088, 128, 1, 256] - - [59, 49.94] - - - [21888, 128, 1, 384] - - [47, 50.833] - - - [16768, 128, 1, 256] - - [59, 35.074] - - - [24064, 24064, 1, 256] - - [24, 76.017] - - - [44928, 128, 1, 256] - - [27, 46.495] - - - [27648, 27648, 1, 384] - - [61, 89.139] - - - [24832, 24832, 1, 256] - - [24, 75.662] - - - [10240, 10240, 1, 256] - - [27, 75.096] - - - [40320, 40320, 1, 384] - - [29, 90.989] - - - [18432, 2688, 1, 384] - - [27, 85.949] - - - [43008, 2304, 1, 384] - - [29, 89.185] - - - [3840, 3072, 1, 384] - - [55, 76.792] - - - [33408, 1920, 1, 384] - - [29, 87.863] - - - [33792, 2688, 1, 384] - - [37, 87.542] - - - [8064, 2688, 1, 384] - - [37, 79.14] - - - [33408, 2304, 1, 384] - - [29, 88.628] - - - [31872, 1536, 1, 384] - - [25, 85.612] - - - [41088, 1920, 1, 384] - - [61, 86.524] - - - [41088, 2304, 1, 384] - - [74, 87.329] - - - [5376, 1536, 1, 384] - - [34, 71.681] - - - [16128, 1536, 1, 384] - - [29, 83.019] - - - [36480, 2688, 1, 384] - - [36, 88.977] - - - [15360, 768, 1, 384] - - [60, 75.527] - - - [42624, 768, 1, 384] - - [53, 82.371] - - - [4992, 1536, 1, 384] - - [35, 67.772] - - - [29952, 1536, 1, 384] - - [25, 86.156] - - - [10752, 2688, 1, 384] - - [36, 82.264] - - - [42240, 2688, 1, 384] - - [53, 89.835] - - - [36096, 1536, 1, 384] - - [24, 85.555] - - - [26496, 1536, 1, 384] - - [25, 85.909] - - - [42624, 2688, 1, 384] - - [27, 87.217] - - - [17664, 2688, 1, 384] - - [36, 86.621] - - - [37248, 1536, 1, 384] - - [37, 87.151] - - - [16896, 2304, 1, 384] - - [37, 86.299] - - - [22272, 1920, 1, 384] - - [25, 87.094] - - - [26880, 2688, 1, 384] - - [36, 87.772] - - - [384, 768, 1, 384] - - [110, 21.71] - - - [16896, 1920, 1, 384] - - [37, 86.53] - - - [32640, 2304, 1, 384] - - [25, 85.785] - - - [5760, 2304, 1, 384] - - [49, 76.76] - - - [11904, 2304, 1, 384] - - [27, 82.809] - - - [24576, 2304, 1, 384] - - [37, 85.039] - - - [33024, 1536, 1, 384] - - [37, 84.384] - - - [36096, 2304, 1, 384] - - [24, 86.916] - - - [20352, 2688, 1, 384] - - [44, 87.058] - - - [14592, 2304, 1, 384] - - [37, 85.738] - - - [16128, 1920, 1, 384] - - [27, 82.946] - - - [16512, 1920, 1, 384] - - [41, 79.688] - - - [35712, 1920, 1, 384] - - [37, 87.275] - - - [9216, 2688, 1, 384] - - [27, 83.651] - - - [23808, 2688, 1, 384] - - [53, 88.104] - - - [18048, 768, 1, 384] - - [44, 78.673] - - - [14592, 2688, 1, 384] - - [25, 86.006] - - - [14208, 1920, 1, 384] - - [29, 82.418] - - - [14976, 2688, 1, 384] - - [36, 85.749] - - - [17280, 2304, 1, 384] - - [29, 84.863] - - - [11520, 2304, 1, 384] - - [25, 84.796] - - - [18432, 768, 1, 384] - - [44, 79.512] - - - [4608, 768, 1, 384] - - [35, 65.306] - - - [34944, 1920, 1, 384] - - [37, 87.524] - - - [13824, 2688, 1, 384] - - [53, 86.081] - - - [39936, 2304, 1, 384] - - [25, 88.199] - - - [7680, 2688, 1, 384] - - [27, 81.415] - - - [19968, 2304, 1, 384] - - [27, 86.824] - - - [27648, 2688, 1, 384] - - [37, 86.702] - - - [4224, 768, 1, 384] - - [34, 59.788] - - - [24192, 1920, 1, 384] - - [27, 87.591] - - - [32640, 1920, 1, 384] - - [25, 84.928] - - - [34176, 2688, 1, 384] - - [56, 88.958] - - - [35328, 1536, 1, 384] - - [27, 85.698] - - - [8832, 2688, 1, 384] - - [53, 81.421] - - - [18048, 1920, 1, 384] - - [27, 84.439] - - - [31488, 768, 1, 384] - - [53, 82.286] - - - [21504, 2304, 1, 384] - - [25, 86.898] - - - [11136, 2688, 1, 384] - - [53, 84.703] - - - [768, 1152, 1, 384] - - [109, 46.95] - - - [29184, 2688, 1, 384] - - [44, 88.494] - - - [4608, 2688, 1, 384] - - [49, 79.851] - - - [21504, 2688, 1, 384] - - [27, 85.527] - - - [34176, 768, 1, 384] - - [53, 83.561] - - - [23808, 1536, 1, 384] - - [27, 84.573] - - - [43392, 1536, 1, 384] - - [25, 87.946] - - - [13824, 768, 1, 384] - - [36, 77.721] - - - [38016, 1536, 1, 384] - - [29, 86.692] - - - [20736, 2688, 1, 384] - - [58, 88.463] - - - [15744, 1536, 1, 384] - - [25, 82.237] - - - [16512, 1536, 1, 384] - - [38, 75.676] - - - [3072, 2304, 1, 384] - - [49, 73.318] - - - [5760, 2688, 1, 384] - - [464, 81.364] - - - [38400, 2304, 1, 384] - - [25, 88.2] - - - [15360, 2688, 1, 384] - - [27, 86.326] - - - [29952, 2688, 1, 384] - - [58, 88.86] - - - [43008, 2688, 1, 384] - - [29, 87.686] - - - [13440, 1920, 1, 384] - - [27, 82.655] - - - [6528, 2688, 1, 384] - - [25, 81.588] - - - [2304, 1536, 1, 384] - - [49, 64.822] - - - [40320, 1536, 1, 384] - - [27, 86.967] - - - [13440, 1536, 1, 384] - - [27, 81.341] - - - [40320, 2688, 1, 384] - - [53, 88.955] - - - [30336, 2304, 1, 384] - - [36, 88.619] - - - [24192, 2688, 1, 384] - - [44, 87.204] - - - [35328, 768, 1, 384] - - [53, 81.961] - - - [23040, 768, 1, 384] - - [58, 81.55] - - - [29952, 2304, 1, 384] - - [27, 87.87] - - - [33024, 1920, 1, 384] - - [37, 86.647] - - - [14976, 768, 1, 384] - - [58, 74.424] - - - [42624, 1920, 1, 384] - - [37, 87.484] - - - [32640, 2688, 1, 384] - - [28, 85.433] - - - [11520, 1536, 1, 384] - - [29, 82.185] - - - [6912, 768, 1, 384] - - [60, 69.191] - - - [39552, 1920, 1, 384] - - [27, 87.96] - - - [32256, 1920, 1, 384] - - [27, 87.539] - - - [10752, 1536, 1, 384] - - [25, 77.35] - - - [24576, 2688, 1, 384] - - [37, 84.613] - - - [12672, 2688, 1, 384] - - [53, 86.586] - - - [10752, 1920, 1, 384] - - [27, 81.453] - - - [40704, 1536, 1, 384] - - [25, 87.483] - - - [32256, 768, 1, 384] - - [53, 83.799] - - - [18816, 2688, 1, 384] - - [44, 85.896] - - - [11520, 2688, 1, 384] - - [36, 83.45] - - - [35712, 2688, 1, 384] - - [53, 88.852] - - - [29952, 1920, 1, 384] - - [25, 88.372] - - - [26880, 1920, 1, 384] - - [25, 87.563] - - - [33408, 2688, 1, 384] - - [58, 89.067] - - - [35328, 2688, 1, 384] - - [44, 88.976] - - - [21120, 2688, 1, 384] - - [62, 86.9] - - - [19584, 1920, 1, 384] - - [27, 86.96] - - - [17664, 1536, 1, 384] - - [29, 81.463] - - - [36864, 768, 1, 384] - - [44, 83.653] - - - [14592, 1536, 1, 384] - - [25, 81.728] - - - [11136, 2304, 1, 384] - - [25, 82.333] - - - [9600, 2688, 1, 384] - - [29, 82.609] - - - [9216, 2304, 1, 384] - - [29, 83.156] - - - [21120, 768, 1, 384] - - [36, 81.735] - - - [4992, 2688, 1, 384] - - [464, 82.226] - - - [41472, 768, 1, 384] - - [44, 85.184] - - - [37632, 1536, 1, 384] - - [27, 86.027] - - - [38784, 2304, 1, 384] - - [37, 88.806] - - - [8448, 2688, 1, 384] - - [44, 82.716] - - - [36864, 2304, 1, 384] - - [37, 87.729] - - - [40704, 1920, 1, 384] - - [29, 88.34] - - - [39552, 2688, 1, 384] - - [53, 89.664] - - - [26112, 768, 1, 384] - - [44, 79.562] - - - [29184, 1536, 1, 384] - - [25, 86.987] - - - [32640, 1536, 1, 384] - - [29, 83.298] - - - [5376, 2688, 1, 384] - - [464, 81.429] - - - [13056, 768, 1, 384] - - [53, 73.998] - - - [13824, 2304, 1, 384] - - [27, 85.541] - - - [16896, 768, 1, 384] - - [36, 74.655] - - - [30336, 1920, 1, 384] - - [29, 87.334] - - - [27264, 2304, 1, 384] - - [29, 88.195] - - - [7680, 1536, 1, 384] - - [25, 76.276] - - - [30720, 2688, 1, 384] - - [27, 87.806] - - - [36096, 2688, 1, 384] - - [28, 87.086] - - - [5760, 1920, 1, 384] - - [35, 72.496] - - - [42240, 1536, 1, 384] - - [27, 87.779] - - - [8448, 1920, 1, 384] - - [34, 82.584] - - - [32256, 1536, 1, 384] - - [37, 86.303] - - - [44160, 2304, 1, 384] - - [58, 89.108] - - - [30336, 2688, 1, 384] - - [44, 88.332] - - - [6144, 2688, 1, 384] - - [468, 81.03] - - - [39168, 1536, 1, 384] - - [25, 86.962] - - - [11904, 1920, 1, 384] - - [58, 83.359] - - - [8064, 1536, 1, 384] - - [55, 78.707] - - - [21120, 1920, 1, 384] - - [27, 86.0] - - - [22656, 2304, 1, 384] - - [44, 88.383] - - - [19968, 2688, 1, 384] - - [44, 87.953] - - - [10752, 768, 1, 384] - - [63, 71.538] - - - [18432, 2304, 1, 384] - - [29, 86.351] - - - [14976, 1920, 1, 384] - - [58, 86.02] - - - [33024, 2688, 1, 384] - - [58, 87.96] - - - [1536, 768, 1, 384] - - [420, 46.928] - - - [33024, 2304, 1, 384] - - [44, 87.324] - - - [14208, 2688, 1, 384] - - [44, 84.865] - - - [38016, 2304, 1, 384] - - [53, 88.859] - - - [16896, 2688, 1, 384] - - [36, 85.692] - - - [31104, 768, 1, 384] - - [53, 81.555] - - - [41472, 2304, 1, 384] - - [25, 88.447] - - - [23424, 2688, 1, 384] - - [29, 87.378] - - - [26496, 2688, 1, 384] - - [36, 88.317] - - - [16512, 2304, 1, 384] - - [28, 80.998] - - - [11520, 1920, 1, 384] - - [25, 81.321] - - - [39552, 768, 1, 384] - - [53, 85.633] - - - [6144, 2304, 1, 384] - - [34, 80.665] - - - [14208, 2304, 1, 384] - - [27, 83.895] - - - [19584, 2304, 1, 384] - - [29, 85.331] - - - [36480, 768, 1, 384] - - [58, 83.786] - - - [15744, 2688, 1, 384] - - [53, 86.233] - - - [34560, 1536, 1, 384] - - [29, 86.554] - - - [8448, 2304, 1, 384] - - [25, 82.987] - - - [26112, 2688, 1, 384] - - [53, 88.934] - - - [39936, 768, 1, 384] - - [36, 82.561] - - - [19200, 1920, 1, 384] - - [27, 85.398] - - - [38400, 768, 1, 384] - - [53, 83.34] - - - [8448, 1536, 1, 384] - - [34, 74.895] - - - [13824, 1536, 1, 384] - - [25, 82.987] - - - [9600, 768, 1, 384] - - [23, 64.683] - - - [10368, 768, 1, 384] - - [63, 69.246] - - - [20736, 1536, 1, 384] - - [25, 85.105] - - - [28800, 768, 1, 384] - - [36, 80.331] - - - [10368, 1536, 1, 384] - - [27, 81.171] - - - [21888, 1536, 1, 384] - - [27, 82.833] - - - [38784, 2688, 1, 384] - - [36, 89.453] - - - [27648, 2304, 1, 384] - - [29, 87.437] - - - [11136, 1920, 1, 384] - - [27, 83.769] - - - [37248, 768, 1, 384] - - [44, 85.091] - - - [23040, 2688, 1, 384] - - [36, 87.36] - - - [37632, 1920, 1, 384] - - [25, 87.723] - - - [7680, 768, 1, 384] - - [23, 62.177] - - - [38016, 1920, 1, 384] - - [25, 88.34] - - - [35712, 2304, 1, 384] - - [44, 88.776] - - - [37248, 2688, 1, 384] - - [44, 88.92] - - - [29568, 1920, 1, 384] - - [51, 86.846] - - - [38400, 2688, 1, 384] - - [53, 88.725] - - - [25728, 768, 1, 384] - - [58, 83.582] - - - [8832, 1920, 1, 384] - - [53, 79.336] - - - [43776, 1920, 1, 384] - - [30, 86.47] - - - [15744, 768, 1, 384] - - [53, 77.357] - - - [27264, 1920, 1, 384] - - [29, 85.812] - - - [33792, 2304, 1, 384] - - [27, 87.915] - - - [8832, 2304, 1, 384] - - [27, 80.583] - - - [39168, 2688, 1, 384] - - [36, 89.136] - - - [35328, 1920, 1, 384] - - [27, 88.337] - - - [35328, 2304, 1, 384] - - [27, 88.11] - - - [29184, 768, 1, 384] - - [53, 81.9] - - - [18048, 2688, 1, 384] - - [58, 87.839] - - - [32256, 2688, 1, 384] - - [36, 88.127] - - - [18816, 1536, 1, 384] - - [27, 81.82] - - - [13056, 1536, 1, 384] - - [25, 78.981] - - - [34944, 1536, 1, 384] - - [27, 87.087] - - - [38400, 1920, 1, 384] - - [37, 88.679] - - - [15360, 2304, 1, 384] - - [25, 85.764] - - - [27264, 2688, 1, 384] - - [27, 87.083] - - - [11136, 1536, 1, 384] - - [27, 79.702] - - - [30720, 2304, 1, 384] - - [37, 87.779] - - - [24960, 2688, 1, 384] - - [58, 87.696] - - - [13824, 1920, 1, 384] - - [25, 84.703] - - - [17280, 2688, 1, 384] - - [44, 87.383] - - - [31872, 768, 1, 384] - - [36, 83.012] - - - [11904, 2688, 1, 384] - - [36, 85.826] - - - [7296, 768, 1, 384] - - [63, 58.889] - - - [19200, 1536, 1, 384] - - [29, 83.176] - - - [12288, 768, 1, 384] - - [36, 69.334] - - - [33792, 768, 1, 384] - - [53, 82.575] - - - [21888, 2688, 1, 384] - - [64, 83.957] - - - [2688, 1920, 1, 384] - - [35, 67.714] - - - [19968, 768, 1, 384] - - [44, 78.744] - - - [12288, 2688, 1, 384] - - [27, 83.987] - - - [12288, 2304, 1, 384] - - [25, 84.682] - - - [28416, 768, 1, 384] - - [44, 80.118] - - - [34560, 768, 1, 384] - - [53, 84.329] - - - [39936, 2688, 1, 384] - - [27, 88.07] - - - [8064, 1920, 1, 384] - - [37, 79.126] - - - [26880, 1536, 1, 384] - - [37, 86.713] - - - [28032, 2688, 1, 384] - - [36, 89.022] - - - [41472, 2688, 1, 384] - - [58, 89.01] - - - [29568, 2688, 1, 384] - - [53, 87.448] - - - [31104, 2688, 1, 384] - - [53, 88.587] - - - [5376, 1920, 1, 384] - - [49, 76.125] - - - [41856, 2688, 1, 384] - - [36, 89.376] - - - [9984, 768, 1, 384] - - [60, 67.304] - - - [3456, 2688, 1, 384] - - [35, 70.02] - - - [43392, 2688, 1, 384] - - [44, 89.587] - - - [36480, 1920, 1, 384] - - [44, 88.451] - - - [29568, 1536, 1, 384] - - [47, 84.24] - - - [36864, 2688, 1, 384] - - [37, 87.725] - - - [12672, 768, 1, 384] - - [23, 72.052] - - - [24064, 3072, 1, 256] - - [31, 72.89] - - - [256, 512, 1, 256] - - [122, 8.398] - - - [40960, 27648, 1, 256] - - [22, 67.36] - - - [31744, 3072, 1, 256] - - [29, 73.477] - - - [13056, 1792, 1, 256] - - [55, 68.961] - - - [35328, 22785, 1, 256] - - [73, 73.723] - - - [28160, 15872, 1, 256] - - [65, 75.611] - - - [39168, 1792, 1, 256] - - [59, 71.707] - - - [23808, 11265, 1, 256] - - [32, 71.85] - - - [16640, 4353, 1, 256] - - [73, 69.432] - - - [38912, 26624, 1, 256] - - [37, 75.182] - - - [6912, 3585, 1, 256] - - [29, 65.964] - - - [32768, 1792, 1, 256] - - [22, 58.928] - - - [30976, 18688, 1, 256] - - [28, 74.426] - - - [512, 2048, 1, 256] - - [121, 37.096] - - - [15872, 3584, 1, 256] - - [31, 72.985] - - - [6400, 1792, 1, 256] - - [35, 62.034] - - - [39680, 27393, 1, 256] - - [54, 72.507] - - - [36864, 24577, 1, 256] - - [32, 71.805] - - - [26112, 1536, 1, 256] - - [49, 70.393] - - - [26368, 1536, 1, 256] - - [49, 69.857] - - - [16896, 4353, 1, 256] - - [30, 71.01] - - - [14336, 1793, 1, 256] - - [25, 64.85] - - - [3840, 3072, 1, 256] - - [34, 63.197] - - - [2560, 3072, 1, 256] - - [35, 56.708] - - - [6656, 1536, 1, 256] - - [49, 61.995] - - - [27136, 1792, 1, 256] - - [34, 71.697] - - - [43776, 3072, 1, 256] - - [28, 71.088] - - - [23296, 1792, 1, 256] - - [34, 70.456] - - - [11264, 7937, 1, 256] - - [32, 72.469] - - - [768, 3072, 1, 256] - - [57, 38.337] - - - [6912, 3841, 1, 256] - - [34, 66.587] - - - [40960, 769, 1, 256] - - [29, 55.161] - - - [40448, 9216, 1, 256] - - [42, 74.661] - - - [7680, 4353, 1, 256] - - [25, 67.882] - - - [23296, 3072, 1, 256] - - [25, 72.423] - - - [7936, 4609, 1, 256] - - [57, 69.169] - - - [20736, 8448, 1, 256] - - [29, 74.596] - - - [768, 1024, 1, 256] - - [123, 36.135] - - - [38656, 3072, 1, 256] - - [30, 72.725] - - - [28160, 1792, 1, 256] - - [31, 71.494] - - - [13824, 3072, 1, 256] - - [27, 71.6] - - - [42752, 1792, 1, 256] - - [29, 72.151] - - - [35584, 23041, 1, 256] - - [28, 73.114] - - - [13056, 3072, 1, 256] - - [27, 71.139] - - - [37888, 768, 1, 256] - - [44, 68.261] - - - [19456, 3072, 1, 256] - - [27, 72.97] - - - [15872, 9216, 1, 256] - - [37, 74.823] - - - [30976, 1792, 1, 256] - - [57, 71.133] - - - [26368, 14081, 1, 256] - - [54, 72.876] - - - [35328, 23041, 1, 256] - - [43, 73.686] - - - [27648, 15105, 1, 256] - - [25, 74.161] - - - [25856, 13568, 1, 256] - - [37, 75.047] - - - [23296, 9216, 1, 256] - - [29, 74.259] - - - [2048, 1024, 1, 256] - - [59, 34.871] - - - [12032, 1792, 1, 256] - - [49, 69.388] - - - [11520, 1536, 1, 256] - - [55, 67.434] - - - [16128, 768, 1, 256] - - [60, 62.842] - - - [15360, 3072, 1, 256] - - [27, 72.025] - - - [38912, 26369, 1, 256] - - [54, 73.922] - - - [25344, 13056, 1, 256] - - [41, 74.792] - - - [39168, 26880, 1, 256] - - [38, 74.625] - - - [39424, 768, 1, 256] - - [63, 68.846] - - - [10496, 1792, 1, 256] - - [35, 66.402] - - - [28672, 3072, 1, 256] - - [25, 73.07] - - - [27392, 768, 1, 256] - - [39, 63.787] - - - [39680, 768, 1, 256] - - [44, 68.165] - - - [11520, 8193, 1, 256] - - [54, 70.412] - - - [17408, 4865, 1, 256] - - [25, 71.47] - - - [14080, 1537, 1, 256] - - [27, 63.305] - - - [29184, 768, 1, 256] - - [39, 67.284] - - - [19200, 6913, 1, 256] - - [31, 71.548] - - - [33536, 9216, 1, 256] - - [41, 74.284] - - - [5632, 3072, 1, 256] - - [55, 66.804] - - - [32768, 20480, 1, 256] - - [66, 58.823] - - - [29440, 9216, 1, 256] - - [47, 74.429] - - - [40960, 1792, 1, 256] - - [38, 66.694] - - - [10240, 3072, 1, 256] - - [34, 70.816] - - - [20992, 1792, 1, 256] - - [27, 71.07] - - - [42240, 9216, 1, 256] - - [28, 74.271] - - - [19200, 6912, 1, 256] - - [27, 74.327] - - - [27392, 1792, 1, 256] - - [86, 68.339] - - - [42496, 1536, 1, 256] - - [37, 71.485] - - - [29440, 16897, 1, 256] - - [28, 73.235] - - - [20480, 8192, 1, 256] - - [25, 74.921] - - - [11264, 8193, 1, 256] - - [29, 70.854] - - - [26880, 14337, 1, 256] - - [54, 72.433] - - - [28928, 16641, 1, 256] - - [73, 73.123] - - - [15360, 2817, 1, 256] - - [27, 67.942] - - - [44288, 1536, 1, 256] - - [29, 70.471] - - - [7936, 1536, 1, 256] - - [34, 63.735] - - - [18176, 5633, 1, 256] - - [25, 71.129] - - - [8448, 3072, 1, 256] - - [49, 68.909] - - - [17920, 5632, 1, 256] - - [24, 74.331] - - - [1792, 2048, 1, 256] - - [35, 40.334] - - - [39936, 3072, 1, 256] - - [37, 73.824] - - - [20480, 3072, 1, 256] - - [29, 72.732] - - - [24832, 1792, 1, 256] - - [67, 71.196] - - - [37376, 25088, 1, 256] - - [37, 75.229] - - - [7168, 4097, 1, 256] - - [25, 67.186] - - - [21504, 768, 1, 256] - - [53, 64.581] - - - [13312, 3072, 1, 256] - - [27, 71.791] - - - [40960, 1025, 1, 256] - - [29, 57.587] - - - [12032, 1536, 1, 256] - - [34, 65.445] - - - [9216, 768, 1, 256] - - [60, 59.444] - - - [44288, 27648, 1, 256] - - [28, 74.594] - - - [32512, 1792, 1, 256] - - [34, 71.363] - - - [23808, 11520, 1, 256] - - [25, 75.256] - - - [25600, 13057, 1, 256] - - [54, 73.924] - - - [40448, 1792, 1, 256] - - [29, 72.271] - - - [25088, 12800, 1, 256] - - [47, 75.683] - - - [22784, 10496, 1, 256] - - [67, 74.874] - - - [38400, 26113, 1, 256] - - [28, 73.709] - - - [9728, 3072, 1, 256] - - [27, 70.812] - - - [20736, 1792, 1, 256] - - [55, 70.596] - - - [7680, 3072, 1, 256] - - [35, 67.931] - - - [5376, 2305, 1, 256] - - [49, 58.991] - - - [12800, 3072, 1, 256] - - [31, 71.387] - - - [43520, 3584, 1, 256] - - [37, 74.091] - - - [12288, 3072, 1, 256] - - [27, 71.431] - - - [12800, 1536, 1, 256] - - [67, 67.084] - - - [21504, 8961, 1, 256] - - [25, 73.17] - - - [39680, 9216, 1, 256] - - [27, 74.129] - - - [3584, 513, 1, 256] - - [33, 30.521] - - - [1280, 3072, 1, 256] - - [40, 43.215] - - - [13056, 9216, 1, 256] - - [37, 74.371] - - - [22016, 768, 1, 256] - - [58, 64.586] - - - [33024, 1536, 1, 256] - - [55, 69.126] - - - [26880, 9216, 1, 256] - - [27, 74.278] - - - [44032, 27648, 1, 256] - - [27, 74.877] - - - [7680, 768, 1, 256] - - [40, 51.917] - - - [32000, 19712, 1, 256] - - [37, 75.0] - - - [26880, 14593, 1, 256] - - [54, 72.979] - - - [24064, 9216, 1, 256] - - [43, 74.935] - - - [39424, 26881, 1, 256] - - [54, 73.642] - - - [27392, 3072, 1, 256] - - [43, 70.269] - - - [10752, 1792, 1, 256] - - [55, 67.663] - - - [8960, 5633, 1, 256] - - [29, 70.165] - - - [34560, 3072, 1, 256] - - [25, 72.869] - - - [23808, 9216, 1, 256] - - [25, 74.572] - - - [29696, 17153, 1, 256] - - [29, 74.326] - - - [11776, 1536, 1, 256] - - [34, 64.33] - - - [13568, 1536, 1, 256] - - [34, 67.459] - - - [30208, 9216, 1, 256] - - [42, 75.008] - - - [36608, 1536, 1, 256] - - [55, 70.529] - - - [12800, 513, 1, 256] - - [37, 48.009] - - - [7680, 1792, 1, 256] - - [35, 65.222] - - - [42496, 2305, 1, 256] - - [37, 68.926] - - - [37376, 1536, 1, 256] - - [29, 70.914] - - - [20224, 1792, 1, 256] - - [34, 70.026] - - - [43520, 1536, 1, 256] - - [25, 71.431] - - - [26368, 768, 1, 256] - - [40, 66.092] - - - [18176, 3072, 1, 256] - - [27, 72.178] - - - [24320, 12033, 1, 256] - - [73, 73.218] - - - [17408, 9216, 1, 256] - - [27, 74.837] - - - [36352, 1792, 1, 256] - - [29, 72.37] - - - [20992, 8705, 1, 256] - - [32, 72.643] - - - [19712, 7424, 1, 256] - - [59, 73.379] - - - [38144, 768, 1, 256] - - [23, 67.922] - - - [10752, 1536, 1, 256] - - [29, 64.306] - - - [4096, 3072, 1, 256] - - [51, 64.649] - - - [29696, 17409, 1, 256] - - [54, 72.74] - - - [10240, 6913, 1, 256] - - [32, 71.858] - - - [18944, 1536, 1, 256] - - [49, 68.276] - - - [38656, 26113, 1, 256] - - [41, 72.926] - - - [37376, 25089, 1, 256] - - [41, 73.77] - - - [38400, 1536, 1, 256] - - [25, 70.995] - - - [8448, 1792, 1, 256] - - [55, 64.45] - - - [13056, 769, 1, 256] - - [44, 54.241] - - - [24320, 11777, 1, 256] - - [73, 73.035] - - - [17664, 9216, 1, 256] - - [24, 73.709] - - - [8192, 4865, 1, 256] - - [37, 69.019] - - - [17920, 1792, 1, 256] - - [59, 70.638] - - - [32000, 19713, 1, 256] - - [54, 73.259] - - - [8960, 768, 1, 256] - - [60, 58.995] - - - [31232, 3072, 1, 256] - - [67, 73.089] - - - [12544, 257, 1, 256] - - [36, 35.771] - - - [43776, 3585, 1, 256] - - [28, 68.553] - - - [11008, 1792, 1, 256] - - [68, 65.563] - - - [29696, 17408, 1, 256] - - [27, 75.352] - - - [34560, 22272, 1, 256] - - [25, 74.996] - - - [256, 2048, 1, 256] - - [114, 25.406] - - - [32768, 20481, 1, 256] - - [66, 55.844] - - - [14336, 3072, 1, 256] - - [55, 71.953] - - - [19456, 7168, 1, 256] - - [27, 74.12] - - - [13312, 9216, 1, 256] - - [37, 74.894] - - - [22272, 768, 1, 256] - - [60, 65.541] - - - [24064, 1792, 1, 256] - - [31, 71.538] - - - [16896, 1792, 1, 256] - - [55, 70.328] - - - [27904, 15616, 1, 256] - - [27, 75.078] - - - [37888, 3072, 1, 256] - - [25, 73.795] - - - [13056, 513, 1, 256] - - [36, 48.579] - - - [36608, 24065, 1, 256] - - [28, 72.88] - - - [40704, 3072, 1, 256] - - [27, 73.142] - - - [28928, 16640, 1, 256] - - [47, 75.132] - - - [24576, 12288, 1, 256] - - [54, 69.898] - - - [17152, 3072, 1, 256] - - [49, 71.874] - - - [17152, 4864, 1, 256] - - [35, 73.546] - - - [42496, 9216, 1, 256] - - [28, 74.809] - - - [32256, 768, 1, 256] - - [39, 68.635] - - - [4352, 1792, 1, 256] - - [34, 55.99] - - - [5632, 768, 1, 256] - - [39, 46.604] - - - [40704, 513, 1, 256] - - [58, 56.078] - - - [19712, 768, 1, 256] - - [69, 61.993] - - - [33536, 20993, 1, 256] - - [73, 73.277] - - - [2816, 3072, 1, 256] - - [55, 61.168] - - - [3584, 3072, 1, 256] - - [25, 59.974] - - - [4608, 1537, 1, 256] - - [35, 51.08] - - - [44032, 9216, 1, 256] - - [25, 74.553] - - - [33792, 21249, 1, 256] - - [32, 74.025] - - - [32512, 20225, 1, 256] - - [73, 73.399] - - - [38656, 9216, 1, 256] - - [41, 74.324] - - - [17664, 5377, 1, 256] - - [30, 70.624] - - - [19456, 7169, 1, 256] - - [32, 71.093] - - - [8448, 5121, 1, 256] - - [29, 68.835] - - - [29440, 17152, 1, 256] - - [30, 75.486] - - - [40448, 513, 1, 256] - - [36, 56.249] - - - [41472, 1792, 1, 256] - - [29, 72.408] - - - [17920, 3072, 1, 256] - - [57, 72.264] - - - [35072, 9216, 1, 256] - - [25, 74.214] - - - [34816, 22273, 1, 256] - - [54, 74.005] - - - [35072, 22785, 1, 256] - - [32, 73.107] - - - [39168, 9216, 1, 256] - - [28, 74.256] - - - [42752, 2817, 1, 256] - - [54, 69.038] - - - [11776, 3072, 1, 256] - - [25, 70.075] - - - [24832, 12289, 1, 256] - - [43, 72.414] - - - [24576, 12033, 1, 256] - - [32, 68.22] - - - [6400, 1536, 1, 256] - - [34, 60.956] - - - [32512, 3072, 1, 256] - - [57, 72.723] - - - [30976, 3072, 1, 256] - - [31, 71.129] - - - [22016, 9473, 1, 256] - - [24, 73.241] - - - [19968, 1792, 1, 256] - - [55, 71.033] - - - [29440, 3072, 1, 256] - - [30, 73.013] - - - [43776, 3840, 1, 256] - - [28, 72.616] - - - [41472, 768, 1, 256] - - [53, 68.878] - - - [8192, 1792, 1, 256] - - [35, 63.092] - - - [35840, 3072, 1, 256] - - [27, 73.788] - - - [8704, 3072, 1, 256] - - [34, 70.142] - - - [9728, 1792, 1, 256] - - [49, 67.491] - - - [22272, 9729, 1, 256] - - [70, 72.315] - - - [32768, 3072, 1, 256] - - [22, 59.221] - - - [3072, 2048, 1, 256] - - [35, 54.959] - - - [36864, 24576, 1, 256] - - [25, 74.304] - - - [9984, 1536, 1, 256] - - [55, 64.234] - - - [12032, 8961, 1, 256] - - [47, 71.392] - - - [38400, 25857, 1, 256] - - [28, 73.786] - - - [20224, 7937, 1, 256] - - [54, 72.001] - - - [34304, 21761, 1, 256] - - [43, 73.859] - - - [30720, 18432, 1, 256] - - [27, 75.598] - - - [31744, 9216, 1, 256] - - [27, 74.725] - - - [27136, 14848, 1, 256] - - [47, 75.662] - - - [34048, 9216, 1, 256] - - [43, 73.879] - - - [3584, 257, 1, 256] - - [114, 33.197] - - - [18688, 6145, 1, 256] - - [32, 70.227] - - - [36096, 768, 1, 256] - - [39, 66.163] - - - [36608, 9216, 1, 256] - - [28, 74.237] - - - [35584, 9216, 1, 256] - - [28, 74.453] - - - [29952, 17664, 1, 256] - - [24, 75.301] - - - [34816, 1792, 1, 256] - - [29, 72.664] - - - [24064, 11776, 1, 256] - - [24, 75.569] - - - [40448, 3072, 1, 256] - - [51, 73.246] - - - [18688, 6401, 1, 256] - - [54, 70.993] - - - [20480, 1536, 1, 256] - - [27, 69.236] - - - [18432, 3072, 1, 256] - - [25, 72.829] - - - [20224, 768, 1, 256] - - [60, 64.775] - - - [25344, 768, 1, 256] - - [69, 66.478] - - - [36608, 24320, 1, 256] - - [38, 74.884] - - - [34816, 9216, 1, 256] - - [25, 74.783] - - - [41216, 27648, 1, 256] - - [28, 74.634] - - - [30464, 9216, 1, 256] - - [41, 73.654] - - - [7424, 3072, 1, 256] - - [35, 69.039] - - - [20480, 1792, 1, 256] - - [49, 70.652] - - - [41984, 1793, 1, 256] - - [29, 67.812] - - - [18688, 1792, 1, 256] - - [49, 70.3] - - - [13824, 1792, 1, 256] - - [35, 70.014] - - - [38144, 3072, 1, 256] - - [37, 73.144] - - - [33280, 3072, 1, 256] - - [51, 73.418] - - - [35584, 23296, 1, 256] - - [50, 74.79] - - - [43520, 768, 1, 256] - - [36, 69.457] - - - [40704, 1536, 1, 256] - - [29, 71.038] - - - [29696, 3072, 1, 256] - - [37, 73.642] - - - [32256, 19969, 1, 256] - - [73, 73.669] - - - [40960, 9216, 1, 256] - - [38, 67.099] - - - [37632, 9216, 1, 256] - - [29, 73.983] - - - [42240, 2305, 1, 256] - - [32, 68.068] - - - [17920, 5377, 1, 256] - - [24, 71.454] - - - [27904, 9216, 1, 256] - - [41, 74.405] - - - [34304, 22016, 1, 256] - - [42, 75.492] - - - [11776, 8705, 1, 256] - - [51, 72.055] - - - [22272, 1536, 1, 256] - - [31, 69.158] - - - [25856, 9216, 1, 256] - - [25, 74.358] - - - [19712, 3072, 1, 256] - - [51, 70.534] - - - [41472, 9216, 1, 256] - - [42, 74.605] - - - [42496, 27648, 1, 256] - - [28, 75.01] - - - [44288, 4352, 1, 256] - - [38, 73.828] - - - [42496, 2561, 1, 256] - - [25, 68.824] - - - [9984, 6657, 1, 256] - - [29, 71.101] - - - [43008, 3073, 1, 256] - - [54, 69.489] - - - [36352, 24065, 1, 256] - - [41, 73.734] - - - [24832, 3072, 1, 256] - - [57, 72.622] - - - [29184, 16641, 1, 256] - - [32, 73.633] - - - [1024, 2048, 1, 256] - - [72, 34.714] - - - [42240, 27648, 1, 256] - - [28, 74.587] - - - [9984, 1792, 1, 256] - - [55, 67.945] - - - [44288, 3072, 1, 256] - - [54, 72.705] - - - [11008, 768, 1, 256] - - [72, 56.84] - - - [28672, 16129, 1, 256] - - [54, 73.603] - - - [17920, 9216, 1, 256] - - [74, 74.862] - - - [25088, 12801, 1, 256] - - [43, 73.417] - - - [19712, 9216, 1, 256] - - [41, 74.04] - - - [31744, 19457, 1, 256] - - [25, 72.863] - - - [36864, 1792, 1, 256] - - [27, 72.242] - - - [42496, 1792, 1, 256] - - [25, 72.873] - - - [39936, 9216, 1, 256] - - [25, 74.655] - - - [8960, 1792, 1, 256] - - [34, 66.553] - - - [17664, 5121, 1, 256] - - [51, 69.096] - - - [38144, 25601, 1, 256] - - [28, 72.464] - - - [27136, 14849, 1, 256] - - [43, 73.926] - - - [31744, 19456, 1, 256] - - [27, 75.314] - - - [33024, 3072, 1, 256] - - [24, 72.339] - - - [37888, 9216, 1, 256] - - [27, 74.723] - - - [6912, 1792, 1, 256] - - [35, 64.897] - - - [42240, 2049, 1, 256] - - [58, 66.103] - - - [34048, 3072, 1, 256] - - [31, 72.477] - - - [37120, 9216, 1, 256] - - [41, 74.336] - - - [14080, 9216, 1, 256] - - [51, 73.043] - - - [38400, 1792, 1, 256] - - [37, 72.333] - - - [43776, 9216, 1, 256] - - [41, 74.086] - - - [14336, 2049, 1, 256] - - [29, 63.371] - - - [37120, 24577, 1, 256] - - [41, 72.438] - - - [30976, 18433, 1, 256] - - [28, 72.375] - - - [37632, 3072, 1, 256] - - [25, 72.923] - - - [34560, 1792, 1, 256] - - [55, 71.797] - - - [5120, 3072, 1, 256] - - [35, 66.228] - - - [21760, 9217, 1, 256] - - [32, 71.598] - - - [24064, 11521, 1, 256] - - [30, 73.393] - - - [7936, 3072, 1, 256] - - [49, 68.835] - - - [21760, 9472, 1, 256] - - [29, 75.121] - - - [9216, 6145, 1, 256] - - [27, 69.56] - - - [8192, 1536, 1, 256] - - [27, 63.981] - - - [39936, 27648, 1, 256] - - [25, 74.967] - - - [21248, 9216, 1, 256] - - [27, 74.493] - - - [5376, 2049, 1, 256] - - [27, 56.9] - - - [35072, 22529, 1, 256] - - [54, 72.351] - - - [13312, 769, 1, 256] - - [58, 55.124] - - - [35840, 9216, 1, 256] - - [25, 74.674] - - - [39424, 27136, 1, 256] - - [38, 75.313] - - - [26368, 9216, 1, 256] - - [29, 74.3] - - - [34048, 21505, 1, 256] - - [28, 72.296] - - - [26112, 1792, 1, 256] - - [59, 71.895] - - - [23296, 768, 1, 256] - - [63, 67.076] - - - [43264, 27648, 1, 256] - - [27, 74.283] - - - [18432, 9216, 1, 256] - - [37, 75.027] - - - [38912, 3072, 1, 256] - - [27, 73.845] - - - [30464, 17921, 1, 256] - - [41, 72.452] - - - [37376, 9216, 1, 256] - - [41, 74.853] - - - [256, 3072, 1, 256] - - [123, 36.28] - - - [9472, 3072, 1, 256] - - [29, 69.062] - - - [35840, 23552, 1, 256] - - [29, 75.421] - - - [8960, 3072, 1, 256] - - [34, 68.93] - - - [34816, 3072, 1, 256] - - [25, 73.722] - - - [11008, 3072, 1, 256] - - [68, 69.116] - - - [36864, 1536, 1, 256] - - [27, 70.739] - - - [23552, 9216, 1, 256] - - [25, 74.931] - - - [31232, 18945, 1, 256] - - [41, 73.789] - - - [27136, 9216, 1, 256] - - [43, 74.972] - - - [19968, 7681, 1, 256] - - [41, 72.249] - - - [31488, 18945, 1, 256] - - [32, 73.064] - - - [33280, 1792, 1, 256] - - [31, 72.275] - - - [14592, 3072, 1, 256] - - [59, 70.802] - - - [30976, 18689, 1, 256] - - [41, 72.645] - - - [4096, 769, 1, 256] - - [45, 35.775] - - - [31488, 3072, 1, 256] - - [37, 72.752] - - - [33024, 1792, 1, 256] - - [59, 70.613] - - - [11520, 8449, 1, 256] - - [30, 71.344] - - - [44544, 4353, 1, 256] - - [32, 71.136] - - - [18176, 5889, 1, 256] - - [25, 70.944] - - - [5632, 2305, 1, 256] - - [34, 60.921] - - - [39936, 27393, 1, 256] - - [54, 73.828] - - - [10240, 7169, 1, 256] - - [37, 70.424] - - - [39168, 26625, 1, 256] - - [28, 72.658] - - - [10752, 7681, 1, 256] - - [27, 71.73] - - - [13824, 1536, 1, 256] - - [27, 68.41] - - - [14336, 9216, 1, 256] - - [37, 74.678] - - - [37632, 25345, 1, 256] - - [54, 73.074] - - - [35840, 23553, 1, 256] - - [32, 73.019] - - - [23552, 3072, 1, 256] - - [25, 73.223] - - - [19712, 7169, 1, 256] - - [54, 69.862] - - - [5888, 2561, 1, 256] - - [34, 63.656] - - - [27136, 768, 1, 256] - - [39, 67.459] - - - [22272, 1792, 1, 256] - - [34, 70.322] - - - [15616, 1536, 1, 256] - - [29, 67.46] - - - [3840, 769, 1, 256] - - [34, 46.289] - - - [42240, 2304, 1, 256] - - [37, 72.628] - - - [24576, 3072, 1, 256] - - [22, 68.65] - - - [27136, 1536, 1, 256] - - [31, 70.093] - - - [25344, 12801, 1, 256] - - [28, 72.186] - - - [32512, 20224, 1, 256] - - [47, 75.143] - - - [17664, 3072, 1, 256] - - [57, 71.451] - - - [28160, 15873, 1, 256] - - [32, 73.603] - - - [40960, 3072, 1, 256] - - [38, 66.819] - - - [14592, 9216, 1, 256] - - [71, 72.836] - - - [22784, 10497, 1, 256] - - [41, 72.571] - - - [22272, 3072, 1, 256] - - [57, 72.388] - - - [39680, 27137, 1, 256] - - [32, 72.405] - - - [20992, 8704, 1, 256] - - [37, 75.323] - - - [24320, 1536, 1, 256] - - [57, 69.964] - - - [7936, 4865, 1, 256] - - [57, 69.997] - - - [17664, 5376, 1, 256] - - [57, 73.382] - - - [37888, 25345, 1, 256] - - [54, 73.968] - - - [23296, 10753, 1, 256] - - [32, 72.528] - - - [28416, 15873, 1, 256] - - [70, 73.027] - - - [27648, 15361, 1, 256] - - [32, 72.759] - - - [39424, 1536, 1, 256] - - [31, 70.892] - - - [15104, 2817, 1, 256] - - [29, 68.216] - - - [19456, 9216, 1, 256] - - [25, 74.881] - - - [24064, 11777, 1, 256] - - [43, 73.397] - - - [40448, 1536, 1, 256] - - [30, 71.11] - - - [512, 3072, 1, 256] - - [117, 44.68] - - - [38912, 9216, 1, 256] - - [29, 74.717] - - - [19456, 6913, 1, 256] - - [27, 72.86] - - - [29440, 1792, 1, 256] - - [55, 71.532] - - - [41984, 9216, 1, 256] - - [29, 74.61] - - - [14080, 1793, 1, 256] - - [27, 65.232] - - - [20992, 8449, 1, 256] - - [32, 72.683] - - - [17920, 768, 1, 256] - - [40, 64.388] - - - [10496, 7169, 1, 256] - - [54, 69.631] - - - [40704, 27648, 1, 256] - - [27, 74.406] - - - [13568, 1025, 1, 256] - - [40, 59.023] - - - [38144, 9216, 1, 256] - - [41, 74.19] - - - [27392, 15104, 1, 256] - - [43, 74.726] - - - [2304, 3072, 1, 256] - - [49, 60.894] - - - [9472, 6401, 1, 256] - - [25, 70.416] - - - [39424, 1792, 1, 256] - - [59, 72.285] - - - [41728, 768, 1, 256] - - [72, 67.967] - - - [11264, 3072, 1, 256] - - [29, 70.634] - - - [25344, 3072, 1, 256] - - [57, 72.1] - - - [24576, 1792, 1, 256] - - [25, 67.901] - - - [27392, 14849, 1, 256] - - [41, 72.861] - - - [14848, 2561, 1, 256] - - [27, 66.512] - - - [28160, 3072, 1, 256] - - [27, 73.026] - - - [23552, 11009, 1, 256] - - [54, 73.699] - - - [11776, 8449, 1, 256] - - [54, 72.092] - - - [16640, 1792, 1, 256] - - [55, 67.869] - - - [24576, 12289, 1, 256] - - [32, 67.331] - - - [38656, 26369, 1, 256] - - [28, 72.925] - - - [13824, 9216, 1, 256] - - [61, 74.326] - - - [28928, 1792, 1, 256] - - [27, 71.293] - - - [27904, 15361, 1, 256] - - [41, 72.692] - - - [3840, 1792, 1, 256] - - [34, 60.099] - - - [14848, 3072, 1, 256] - - [37, 71.769] - - - [27904, 1536, 1, 256] - - [67, 70.041] - - - [34816, 1536, 1, 256] - - [27, 71.375] - - - [14592, 2305, 1, 256] - - [31, 66.345] - - - [22528, 9985, 1, 256] - - [29, 73.686] - - - [26368, 13825, 1, 256] - - [73, 72.918] - - - [4096, 1792, 1, 256] - - [34, 53.676] - - - [30720, 18177, 1, 256] - - [54, 74.079] - - - [37120, 24833, 1, 256] - - [41, 73.124] - - - [24320, 3072, 1, 256] - - [25, 72.794] - - - [2560, 1536, 1, 256] - - [55, 43.756] - - - [44032, 4097, 1, 256] - - [54, 70.115] - - - [44544, 27648, 1, 256] - - [28, 74.734] - - - [34048, 21761, 1, 256] - - [28, 72.608] - - - [24064, 1536, 1, 256] - - [49, 69.843] - - - [24832, 12545, 1, 256] - - [73, 73.308] - - - [44032, 3841, 1, 256] - - [25, 70.954] - - - [40448, 257, 1, 256] - - [36, 45.308] - - - [26624, 14337, 1, 256] - - [32, 72.796] - - - [8192, 5121, 1, 256] - - [25, 68.016] - - - [42240, 1536, 1, 256] - - [27, 70.765] - - - [5888, 2817, 1, 256] - - [55, 63.618] - - - [6144, 1792, 1, 256] - - [25, 60.128] - - - [16384, 1792, 1, 256] - - [22, 60.792] - - - [35584, 23297, 1, 256] - - [32, 73.151] - - - [36352, 24064, 1, 256] - - [28, 75.214] - - - [23040, 1536, 1, 256] - - [67, 69.287] - - - [8704, 1536, 1, 256] - - [55, 63.632] - - - [18432, 6145, 1, 256] - - [37, 70.988] - - - [12032, 3072, 1, 256] - - [27, 70.991] - - - [39168, 3072, 1, 256] - - [30, 72.818] - - - [28160, 1536, 1, 256] - - [54, 70.098] - - - [41728, 27648, 1, 256] - - [41, 73.797] - - - [28416, 1792, 1, 256] - - [31, 70.953] - - - [24320, 12032, 1, 256] - - [24, 75.433] - - - [28928, 16385, 1, 256] - - [54, 72.397] - - - [34816, 22528, 1, 256] - - [25, 75.054] - - - [26368, 1792, 1, 256] - - [55, 71.252] - - - [25856, 13569, 1, 256] - - [28, 72.951] - - - [25600, 13312, 1, 256] - - [27, 75.771] - - - [31232, 18689, 1, 256] - - [28, 73.883] - - - [20736, 9216, 1, 256] - - [37, 74.492] - - - [34304, 9216, 1, 256] - - [42, 75.021] - - - [43264, 3073, 1, 256] - - [54, 69.053] - - - [8704, 5633, 1, 256] - - [37, 70.581] - - - [4864, 1793, 1, 256] - - [34, 53.98] - - - [41984, 3072, 1, 256] - - [29, 73.928] - - - [20992, 3072, 1, 256] - - [25, 72.977] - - - [9728, 6401, 1, 256] - - [29, 70.891] - - - [16640, 4097, 1, 256] - - [54, 67.999] - - - [38400, 9216, 1, 256] - - [28, 74.802] - - - [38656, 1536, 1, 256] - - [31, 70.607] - - - [1536, 3072, 1, 256] - - [35, 50.841] - - - [12544, 1792, 1, 256] - - [34, 68.266] - - - [37632, 1792, 1, 256] - - [67, 71.968] - - - [17152, 4609, 1, 256] - - [27, 70.191] - - - [18944, 6656, 1, 256] - - [30, 74.667] - - - [34560, 22017, 1, 256] - - [32, 73.029] - - - [23296, 11008, 1, 256] - - [27, 74.913] - - - [14848, 768, 1, 256] - - [40, 60.74] - - - [38656, 1792, 1, 256] - - [31, 71.937] - - - [8448, 5377, 1, 256] - - [31, 70.346] - - - [29952, 17665, 1, 256] - - [73, 73.321] - - - [33792, 21504, 1, 256] - - [27, 75.406] - - - [24576, 1536, 1, 256] - - [38, 66.688] - - - [37376, 1792, 1, 256] - - [25, 72.422] - - - [42752, 768, 1, 256] - - [36, 68.776] - - - [4096, 1025, 1, 256] - - [35, 45.482] - - - [35840, 768, 1, 256] - - [60, 68.381] - - - [19200, 3072, 1, 256] - - [49, 71.676] - - - [33536, 1792, 1, 256] - - [34, 71.782] - - - [36864, 9216, 1, 256] - - [22, 74.017] - - - [38656, 26368, 1, 256] - - [28, 74.56] - - - [44288, 9216, 1, 256] - - [41, 74.399] - - - [44288, 4097, 1, 256] - - [54, 69.82] - - - [26112, 3072, 1, 256] - - [27, 73.247] - - - [512, 768, 1, 256] - - [108, 22.789] - - - [36096, 3072, 1, 256] - - [24, 71.493] - - - [4864, 1537, 1, 256] - - [55, 53.44] - - - [31232, 18944, 1, 256] - - [30, 75.689] - - - [20224, 7681, 1, 256] - - [28, 71.553] - - - [26112, 9216, 1, 256] - - [41, 74.995] - - - [21504, 3072, 1, 256] - - [25, 73.129] - - - [12544, 3072, 1, 256] - - [35, 71.024] - - - [32256, 19713, 1, 256] - - [70, 73.799] - - - [40704, 1792, 1, 256] - - [57, 72.126] - - - [18176, 5888, 1, 256] - - [34, 73.909] - - - [33792, 9216, 1, 256] - - [27, 74.743] - - - [26624, 14336, 1, 256] - - [37, 75.385] - - - [38912, 1792, 1, 256] - - [29, 72.82] - - - [7936, 1792, 1, 256] - - [34, 66.373] - - - [28672, 16385, 1, 256] - - [54, 72.361] - - - [18944, 3072, 1, 256] - - [59, 72.445] - - - [33280, 20993, 1, 256] - - [43, 73.865] - - - [37120, 24832, 1, 256] - - [27, 74.866] - - - [43520, 1792, 1, 256] - - [27, 72.876] - - - [16896, 4609, 1, 256] - - [30, 70.699] - - - [41472, 1536, 1, 256] - - [37, 71.132] - - - [39936, 768, 1, 256] - - [40, 68.67] - - - [23296, 11009, 1, 256] - - [54, 72.525] - - - [26624, 9216, 1, 256] - - [37, 74.827] - - - [29184, 9216, 1, 256] - - [74, 74.75] - - - [36352, 9216, 1, 256] - - [28, 74.823] - - - [37632, 25344, 1, 256] - - [29, 74.856] - - - [37888, 25600, 1, 256] - - [27, 75.356] - - - [16640, 9216, 1, 256] - - [41, 74.324] - - - [44544, 9216, 1, 256] - - [28, 74.614] - - - [14080, 1792, 1, 256] - - [31, 68.979] - - - [33536, 21249, 1, 256] - - [70, 73.389] - - - [34048, 21760, 1, 256] - - [64, 74.55] - - - [9984, 768, 1, 256] - - [23, 54.574] - - - [40192, 1536, 1, 256] - - [25, 71.09] - - - [41728, 3072, 1, 256] - - [30, 71.64] - - - [35328, 9216, 1, 256] - - [43, 74.937] - - - [32512, 768, 1, 256] - - [39, 67.213] - - - [14592, 2049, 1, 256] - - [39, 63.564] - - - [14848, 9216, 1, 256] - - [41, 74.684] - - - [23808, 3072, 1, 256] - - [25, 72.789] - - - [13568, 9216, 1, 256] - - [29, 74.107] - - - [42496, 2560, 1, 256] - - [25, 73.708] - - - [42752, 3072, 1, 256] - - [25, 73.224] - - - [39680, 27392, 1, 256] - - [37, 74.582] - - - [14592, 1792, 1, 256] - - [57, 68.208] - - - [25600, 13313, 1, 256] - - [32, 72.762] - - - [26624, 1792, 1, 256] - - [27, 72.142] - - - [20480, 8193, 1, 256] - - [54, 71.705] - - - [36096, 23808, 1, 256] - - [41, 74.247] - - - [15104, 2561, 1, 256] - - [40, 67.093] - - - [43520, 3072, 1, 256] - - [27, 73.788] - - - [1280, 2048, 1, 256] - - [57, 42.367] - - - [43008, 1792, 1, 256] - - [25, 72.967] - - - [18688, 3072, 1, 256] - - [25, 72.245] - - - [35328, 23040, 1, 256] - - [51, 75.432] - - - [18944, 6401, 1, 256] - - [73, 71.873] - - - [16128, 3585, 1, 256] - - [25, 69.134] - - - [29952, 1536, 1, 256] - - [37, 70.233] - - - [17408, 5121, 1, 256] - - [54, 70.042] - - - [36608, 1792, 1, 256] - - [34, 71.769] - - - [13056, 768, 1, 256] - - [53, 61.047] - - - [26112, 13824, 1, 256] - - [65, 75.824] - - - [43520, 3585, 1, 256] - - [54, 70.545] - - - [40704, 9216, 1, 256] - - [37, 74.131] - - - [27904, 15617, 1, 256] - - [42, 73.165] - - - [21248, 3072, 1, 256] - - [49, 72.225] - - - [38912, 1536, 1, 256] - - [29, 71.532] - - - [28672, 1792, 1, 256] - - [27, 71.832] - - - [18432, 1792, 1, 256] - - [25, 70.635] - - - [29952, 9216, 1, 256] - - [47, 74.232] - - - [4352, 1025, 1, 256] - - [23, 47.919] - - - [34304, 22017, 1, 256] - - [43, 73.821] - - - [28160, 15617, 1, 256] - - [54, 73.595] - - - [19968, 9216, 1, 256] - - [28, 74.953] - - - [7424, 4353, 1, 256] - - [34, 68.112] - - - [19200, 1792, 1, 256] - - [59, 69.134] - - - [27648, 15360, 1, 256] - - [29, 75.75] - - - [23040, 10497, 1, 256] - - [70, 73.426] - - - [21248, 8961, 1, 256] - - [70, 72.223] - - - [32256, 1792, 1, 256] - - [57, 72.227] - - - [26112, 13569, 1, 256] - - [43, 73.707] - - - [12288, 8961, 1, 256] - - [37, 72.875] - - - [6656, 3585, 1, 256] - - [59, 66.739] - - - [19968, 7425, 1, 256] - - [54, 72.526] - - - [9472, 768, 1, 256] - - [53, 52.718] - - - [33792, 3072, 1, 256] - - [29, 73.595] - - - [15616, 3072, 1, 256] - - [37, 71.795] - - - [8704, 5377, 1, 256] - - [27, 70.194] - - - [11520, 3072, 1, 256] - - [67, 70.214] - - - [25856, 1536, 1, 256] - - [35, 69.702] - - - [28416, 768, 1, 256] - - [40, 66.259] - - - [32256, 3072, 1, 256] - - [25, 73.426] - - - [20736, 1536, 1, 256] - - [55, 69.294] - - - [22784, 10241, 1, 256] - - [41, 71.799] - - - [36608, 24321, 1, 256] - - [41, 72.971] - - - [36096, 9216, 1, 256] - - [28, 73.759] - - - [10752, 768, 1, 256] - - [40, 58.055] - - - [38400, 26112, 1, 256] - - [27, 75.376] - - - [9216, 5889, 1, 256] - - [55, 70.705] - - - [41472, 27648, 1, 256] - - [43, 74.643] - - - [38144, 25856, 1, 256] - - [37, 74.767] - - - [15360, 3073, 1, 256] - - [27, 67.924] - - - [29184, 16896, 1, 256] - - [25, 75.487] - - - [16128, 1792, 1, 256] - - [34, 68.836] - - - [32768, 20225, 1, 256] - - [95, 57.927] - - - [23040, 10752, 1, 256] - - [47, 75.65] - - - [15872, 3585, 1, 256] - - [25, 69.965] - - - [11008, 7681, 1, 256] - - [75, 70.59] - - - [15360, 9216, 1, 256] - - [37, 74.876] - - - [28416, 16128, 1, 256] - - [51, 74.891] - - - [30208, 1792, 1, 256] - - [29, 71.803] - - - [41728, 1792, 1, 256] - - [64, 71.002] - - - [32256, 19968, 1, 256] - - [51, 75.446] - - - [18944, 1792, 1, 256] - - [59, 70.622] - - - [41728, 1793, 1, 256] - - [25, 66.538] - - - [31488, 19201, 1, 256] - - [32, 73.037] - - - [40192, 257, 1, 256] - - [25, 44.954] - - - [42752, 27648, 1, 256] - - [29, 74.278] - - - [40704, 768, 1, 256] - - [33, 68.398] - - - [25088, 12545, 1, 256] - - [24, 73.619] - - - [24576, 9216, 1, 256] - - [38, 70.049] - - - [33024, 20737, 1, 256] - - [73, 73.586] - - - [29696, 9216, 1, 256] - - [27, 74.863] - - - [31232, 1536, 1, 256] - - [25, 70.863] - - - [30208, 17920, 1, 256] - - [24, 75.773] - - - [44544, 4609, 1, 256] - - [54, 71.087] - - - [22016, 9728, 1, 256] - - [37, 75.247] - - - [30208, 17921, 1, 256] - - [42, 73.75] - - - [19200, 6657, 1, 256] - - [70, 71.173] - - - [22016, 9729, 1, 256] - - [24, 73.212] - - - [18176, 768, 1, 256] - - [23, 64.992] - - - [29184, 1792, 1, 256] - - [59, 71.58] - - - [12288, 1792, 1, 256] - - [35, 67.603] - - - [22528, 1536, 1, 256] - - [34, 69.808] - - - [14848, 2305, 1, 256] - - [25, 66.318] - - - [41216, 1025, 1, 256] - - [69, 63.189] - - - [8192, 3072, 1, 256] - - [27, 69.015] - - - [5888, 1792, 1, 256] - - [34, 64.218] - - - [21760, 3072, 1, 256] - - [35, 72.375] - - - [22272, 9985, 1, 256] - - [73, 72.608] - - - [29184, 1536, 1, 256] - - [31, 70.393] - - - [22016, 3072, 1, 256] - - [47, 72.804] - - - [30720, 9216, 1, 256] - - [27, 74.894] - - - [39680, 1792, 1, 256] - - [29, 72.047] - - - [9728, 1536, 1, 256] - - [35, 63.991] - - - [34560, 9216, 1, 256] - - [37, 74.255] - - - [12032, 8705, 1, 256] - - [37, 71.355] - - - [10752, 7425, 1, 256] - - [51, 71.821] - - - [18688, 1536, 1, 256] - - [55, 69.385] - - - [16128, 3840, 1, 256] - - [55, 72.828] - - - [38656, 768, 1, 256] - - [33, 68.285] - - - [21248, 1792, 1, 256] - - [34, 70.279] - - - [36352, 3072, 1, 256] - - [27, 73.493] - - - [19968, 7680, 1, 256] - - [29, 75.125] - - - [3840, 513, 1, 256] - - [34, 32.748] - - - [38400, 3072, 1, 256] - - [37, 73.324] - - - [5376, 768, 1, 256] - - [49, 44.727] - - - [20224, 9216, 1, 256] - - [37, 74.585] - - - [17408, 5120, 1, 256] - - [27, 74.321] - - - [28928, 9216, 1, 256] - - [25, 74.117] - - - [35072, 1792, 1, 256] - - [29, 71.863] - - - [31488, 19200, 1, 256] - - [27, 75.016] - - - [11008, 7937, 1, 256] - - [71, 70.797] - - - [21248, 8705, 1, 256] - - [73, 72.196] - - - [13568, 3072, 1, 256] - - [34, 71.084] - - - [34560, 22273, 1, 256] - - [32, 73.128] - - - [34048, 768, 1, 256] - - [40, 67.567] - - - [40448, 27648, 1, 256] - - [42, 74.661] - - - [28416, 16129, 1, 256] - - [54, 73.072] - - - [34816, 22529, 1, 256] - - [32, 72.704] - - - [22528, 3072, 1, 256] - - [27, 73.326] - - - [27136, 14593, 1, 256] - - [70, 73.833] - - - [35584, 3072, 1, 256] - - [29, 72.72] - - - [43008, 3072, 1, 256] - - [27, 73.831] - - - [30464, 1792, 1, 256] - - [31, 70.013] - - - [16384, 4097, 1, 256] - - [27, 59.138] - - - [20992, 9216, 1, 256] - - [41, 75.069] - - - [31488, 1792, 1, 256] - - [34, 71.727] - - - [31488, 9216, 1, 256] - - [27, 74.117] - - - [22272, 9984, 1, 256] - - [57, 75.013] - - - [41728, 1537, 1, 256] - - [27, 65.171] - - - [26880, 1792, 1, 256] - - [27, 71.327] - - - [30464, 768, 1, 256] - - [33, 66.682] - - - [2816, 1792, 1, 256] - - [35, 54.473] - - - [41472, 1537, 1, 256] - - [37, 66.289] - - - [43008, 27648, 1, 256] - - [25, 74.892] - - - [39424, 27137, 1, 256] - - [41, 73.628] - - - [24320, 1792, 1, 256] - - [31, 70.912] - - - [32000, 3072, 1, 256] - - [25, 72.768] - - - [12800, 1792, 1, 256] - - [31, 68.573] - - - [15872, 3072, 1, 256] - - [25, 72.001] - - - [15872, 1792, 1, 256] - - [35, 70.342] - - - [10496, 7425, 1, 256] - - [54, 71.106] - - - [16896, 4608, 1, 256] - - [25, 73.401] - - - [9984, 6913, 1, 256] - - [25, 71.091] - - - [21248, 8960, 1, 256] - - [27, 74.939] - - - [14336, 1792, 1, 256] - - [49, 69.041] - - - [24832, 12544, 1, 256] - - [31, 75.359] - - - [30464, 18176, 1, 256] - - [28, 74.481] - - - [31744, 19201, 1, 256] - - [27, 74.065] - - - [1792, 768, 1, 256] - - [124, 40.595] - - - [1536, 2048, 1, 256] - - [57, 49.311] - - - [40192, 3072, 1, 256] - - [37, 73.329] - - - [42240, 3072, 1, 256] - - [29, 73.139] - - - [32256, 9216, 1, 256] - - [42, 74.881] - - - [41984, 2049, 1, 256] - - [74, 66.57] - - - [6656, 1792, 1, 256] - - [59, 63.224] - - - [13824, 1537, 1, 256] - - [27, 63.771] - - - [20736, 3072, 1, 256] - - [27, 72.45] - - - [36096, 23809, 1, 256] - - [41, 72.71] - - - [41728, 9216, 1, 256] - - [28, 73.108] - - - [25600, 768, 1, 256] - - [60, 67.051] - - - [37632, 768, 1, 256] - - [23, 68.035] - - - [25600, 9216, 1, 256] - - [29, 75.002] - - - [19968, 3072, 1, 256] - - [29, 72.482] - - - [15616, 9216, 1, 256] - - [29, 74.411] - - - [29184, 16897, 1, 256] - - [32, 73.625] - - - [7168, 3841, 1, 256] - - [29, 67.57] - - - [40704, 769, 1, 256] - - [33, 60.707] - - - [6144, 3073, 1, 256] - - [37, 64.657] - - - [34304, 1792, 1, 256] - - [59, 72.252] - - - [18688, 6400, 1, 256] - - [37, 74.512] - - - [20992, 1536, 1, 256] - - [27, 69.992] - - - [21760, 768, 1, 256] - - [60, 64.205] - - - [43264, 3072, 1, 256] - - [25, 73.263] - - - [21760, 9216, 1, 256] - - [25, 74.502] - - - [11264, 768, 1, 256] - - [53, 60.064] - - - [42496, 3072, 1, 256] - - [37, 73.716] - - - [30208, 17665, 1, 256] - - [73, 73.922] - - - [27392, 15105, 1, 256] - - [28, 72.88] - - - [29952, 17409, 1, 256] - - [73, 72.467] - - - [44032, 3072, 1, 256] - - [37, 73.868] - - - [41216, 9216, 1, 256] - - [41, 74.596] - - - [8448, 1536, 1, 256] - - [49, 61.889] - - - [36352, 768, 1, 256] - - [63, 68.459] - - - [23552, 768, 1, 256] - - [58, 64.869] - - - [7168, 3072, 1, 256] - - [35, 67.869] - - - [44288, 4353, 1, 256] - - [54, 70.362] - - - [36608, 768, 1, 256] - - [63, 68.316] - - - [15616, 3073, 1, 256] - - [27, 68.049] - - - [37376, 24833, 1, 256] - - [28, 73.814] - - - [38144, 25857, 1, 256] - - [28, 72.949] - - - [26880, 14592, 1, 256] - - [29, 75.253] - - - [6144, 2817, 1, 256] - - [37, 62.577] - - - [23808, 768, 1, 256] - - [63, 64.822] - - - [39168, 26881, 1, 256] - - [41, 72.928] - - - [5120, 1793, 1, 256] - - [49, 55.962] - - - [32512, 19969, 1, 256] - - [43, 73.225] - - - [43008, 2817, 1, 256] - - [25, 69.9] - - - [26112, 13825, 1, 256] - - [43, 73.821] - - - [33536, 3072, 1, 256] - - [51, 72.881] - - - [9728, 6657, 1, 256] - - [27, 71.388] - - - [2048, 3072, 1, 256] - - [49, 55.21] - - - [24832, 9216, 1, 256] - - [29, 74.486] - - - [5632, 2561, 1, 256] - - [27, 61.714] - - - [33280, 20992, 1, 256] - - [51, 75.582] - - - [20224, 7936, 1, 256] - - [27, 74.62] - - - [28672, 16384, 1, 256] - - [27, 74.944] - - - [28416, 9216, 1, 256] - - [32, 73.936] - - - [7936, 768, 1, 256] - - [27, 52.95] - - - [23552, 11265, 1, 256] - - [25, 72.46] - - - [25088, 3072, 1, 256] - - [29, 72.818] - - - [32000, 19457, 1, 256] - - [32, 72.725] - - - [44800, 3072, 1, 256] - - [25, 72.719] - - - [37120, 1792, 1, 256] - - [27, 71.806] - - - [30464, 18177, 1, 256] - - [28, 72.706] - - - [44544, 4608, 1, 256] - - [28, 73.899] - - - [7168, 768, 1, 256] - - [27, 49.311] - - - [18944, 9216, 1, 256] - - [42, 74.773] - - - [33280, 20737, 1, 256] - - [70, 73.951] - - - [25856, 3072, 1, 256] - - [49, 72.641] - - - [27648, 9216, 1, 256] - - [37, 74.837] - - - [5120, 2049, 1, 256] - - [76, 55.943] - - - [28160, 9216, 1, 256] - - [48, 74.804] - - - [37632, 25089, 1, 256] - - [32, 73.058] - - - [22016, 1792, 1, 256] - - [29, 71.088] - - - [16384, 9216, 1, 256] - - [22, 63.2] - - - [21504, 9217, 1, 256] - - [32, 72.143] - - - [20480, 7937, 1, 256] - - [32, 72.788] - - - [33536, 21248, 1, 256] - - [51, 75.086] - - - [12800, 768, 1, 256] - - [40, 60.244] - - - [28672, 9216, 1, 256] - - [37, 74.277] - - - [32000, 9216, 1, 256] - - [27, 74.163] - - - [44544, 3072, 1, 256] - - [25, 73.355] - - - [5376, 3072, 1, 256] - - [25, 64.581] - - - [35840, 23297, 1, 256] - - [32, 74.027] - - - [23808, 11521, 1, 256] - - [32, 72.524] - - - [13312, 1025, 1, 256] - - [36, 58.916] - - - [18176, 9216, 1, 256] - - [37, 74.483] - - - [17920, 5633, 1, 256] - - [59, 71.536] - - - [27648, 3072, 1, 256] - - [25, 73.527] - - - [1024, 3072, 1, 256] - - [72, 49.177] - - - [22016, 9216, 1, 256] - - [41, 75.131] - - - [21760, 9473, 1, 256] - - [32, 72.475] - - - [6144, 1536, 1, 256] - - [34, 58.455] - - - [16896, 1536, 1, 256] - - [55, 68.548] - - - [19968, 768, 1, 256] - - [40, 64.87] - - - [23552, 11264, 1, 256] - - [37, 75.562] - - - [27904, 3072, 1, 256] - - [59, 72.351] - - - [19712, 7425, 1, 256] - - [73, 71.126] - - - [26624, 14081, 1, 256] - - [37, 73.889] - - - [3328, 257, 1, 256] - - [110, 31.019] - - - [24320, 9216, 1, 256] - - [47, 74.48] - - - [14080, 3072, 1, 256] - - [59, 70.439] - - - [17408, 3072, 1, 256] - - [37, 72.646] - - - [21504, 9216, 1, 256] - - [27, 74.988] - - - [14848, 2560, 1, 256] - - [34, 71.053] - - - [34304, 3072, 1, 256] - - [47, 73.279] - - - [15104, 9216, 1, 256] - - [25, 74.312] - - - [17152, 4865, 1, 256] - - [25, 70.61] - - - [38912, 26625, 1, 256] - - [54, 72.727] - - - [41216, 1792, 1, 256] - - [25, 71.942] - - - [39424, 3072, 1, 256] - - [29, 73.417] - - - [30720, 18433, 1, 256] - - [32, 72.945] - - - [18944, 6657, 1, 256] - - [24, 72.217] - - - [5632, 1792, 1, 256] - - [34, 61.719] - - - [18176, 1792, 1, 256] - - [55, 69.577] - - - [31232, 9216, 1, 256] - - [28, 74.854] - - - [42752, 2561, 1, 256] - - [32, 68.13] - - - [18688, 9216, 1, 256] - - [37, 74.553] - - - [43776, 1792, 1, 256] - - [41, 69.792] - - - [10240, 1792, 1, 256] - - [35, 66.547] - - - [33792, 21505, 1, 256] - - [32, 72.863] - - - [25856, 13313, 1, 256] - - [28, 72.38] - - - [29952, 3072, 1, 256] - - [30, 72.875] - - - [5888, 768, 1, 256] - - [55, 48.723] - - - [20480, 9216, 1, 256] - - [37, 74.621] - - - [17152, 1792, 1, 256] - - [34, 69.112] - - - [23040, 10753, 1, 256] - - [73, 73.355] - - - [8960, 5889, 1, 256] - - [57, 70.343] - - - [16640, 4352, 1, 256] - - [34, 72.785] - - - [30464, 3072, 1, 256] - - [57, 71.363] - - - [16128, 9216, 1, 256] - - [25, 74.406] - - - [25344, 13057, 1, 256] - - [41, 72.612] - - - [39424, 9216, 1, 256] - - [41, 74.936] - - - [25600, 3072, 1, 256] - - [29, 73.346] - - - [28416, 3072, 1, 256] - - [47, 72.357] - - - [12800, 257, 1, 256] - - [60, 36.642] - - - [43264, 1792, 1, 256] - - [25, 72.181] - - - [20736, 8193, 1, 256] - - [54, 71.591] - - - [30976, 9216, 1, 256] - - [41, 73.289] - - - [40192, 27648, 1, 256] - - [25, 74.477] - - - [31232, 1792, 1, 256] - - [35, 71.89] - - - [36352, 23809, 1, 256] - - [28, 73.868] - - - [9984, 3072, 1, 256] - - [25, 69.262] - - - [11776, 1792, 1, 256] - - [57, 68.461] - - - [37120, 1536, 1, 256] - - [49, 70.444] - - - [14592, 2304, 1, 256] - - [59, 69.777] - - - [7424, 768, 1, 256] - - [60, 50.368] - - - [10240, 1536, 1, 256] - - [49, 65.762] - - - [27392, 9216, 1, 256] - - [41, 74.083] - - - [15104, 3072, 1, 256] - - [34, 71.582] - - - [19200, 9216, 1, 256] - - [28, 74.336] - - - [36096, 23553, 1, 256] - - [28, 72.256] - - - [16128, 3841, 1, 256] - - [27, 69.422] - - - [18432, 5889, 1, 256] - - [25, 72.062] - - - [43776, 3841, 1, 256] - - [28, 69.142] - - - [22528, 10241, 1, 256] - - [54, 72.313] - - - [20224, 3072, 1, 256] - - [29, 72.17] - - - [39680, 3072, 1, 256] - - [37, 72.868] - - - [20736, 8449, 1, 256] - - [32, 72.238] - - - [30720, 1792, 1, 256] - - [55, 72.468] - - - [36864, 24321, 1, 256] - - [32, 73.196] - - - [22784, 1536, 1, 256] - - [34, 69.341] - - - [7424, 4097, 1, 256] - - [25, 66.537] - - - [7680, 4609, 1, 256] - - [25, 68.021] - - - [12032, 768, 1, 256] - - [63, 57.752] - - - [1792, 3072, 1, 256] - - [55, 49.169] - - - [6400, 3073, 1, 256] - - [25, 63.363] - - - [29440, 17153, 1, 256] - - [73, 73.313] - - - [8704, 1792, 1, 256] - - [27, 65.844] - - - [30720, 3072, 1, 256] - - [29, 73.577] - - - [16384, 3841, 1, 256] - - [38, 58.468] - - - [40192, 9216, 1, 256] - - [29, 74.318] - - - [23040, 1792, 1, 256] - - [49, 71.001] - - - [37888, 25601, 1, 256] - - [54, 72.878] - - - [26368, 14080, 1, 256] - - [25, 75.251] - - - [30208, 3072, 1, 256] - - [29, 72.997] - - - [33024, 20736, 1, 256] - - [43, 75.079] - - - [35072, 22784, 1, 256] - - [25, 74.903] - - - [9472, 6145, 1, 256] - - [27, 69.301] - - - [22784, 1792, 1, 256] - - [34, 70.849] - - - [768, 2048, 1, 256] - - [112, 44.352] - - - [1024, 1280, 1, 256] - - [125, 43.314] - - - [41984, 27648, 1, 256] - - [37, 74.925] - - - [33024, 20481, 1, 256] - - [42, 72.827] - - - [33280, 1536, 1, 256] - - [49, 71.098] - - - [9216, 3072, 1, 256] - - [34, 70.425] - - - [22528, 1792, 1, 256] - - [34, 71.567] - - - [25088, 768, 1, 256] - - [23, 66.789] - - - [13825, 128, 1, 128] - - [126, 26.713] - - - [20609, 128, 1, 256] - - [206, 33.58] - - - [6017, 128, 1, 256] - - [127, 30.524] - - - [2305, 128, 1, 128] - - [128, 10.976] - - - [15745, 128, 1, 256] - - [206, 27.345] - - - [8833, 128, 1, 128] - - [129, 23.452] - - - [641, 128, 1, 128] - - [130, 3.197] - - - [9217, 128, 1, 128] - - [131, 23.704] - - - [15361, 128, 1, 256] - - [219, 26.647] - - - [22913, 128, 1, 256] - - [207, 36.177] - - - [2177, 128, 1, 128] - - [118, 10.367] - - - [19073, 128, 1, 256] - - [214, 31.601] - - - [28289, 128, 1, 128] - - [206, 20.694] - - - [13057, 128, 1, 256] - - [120, 37.574] - - - [1793, 128, 1, 128] - - [132, 8.538] - - - [16769, 128, 1, 128] - - [215, 18.474] - - - [23681, 128, 1, 256] - - [208, 36.993] - - - [14593, 128, 1, 256] - - [211, 25.525] - - - [24449, 128, 1, 128] - - [207, 24.486] - - - [4609, 128, 1, 256] - - [133, 24.923] - - - [10625, 128, 1, 128] - - [134, 23.134] - - - [12545, 128, 1, 256] - - [135, 36.242] - - - [5633, 128, 1, 128] - - [136, 18.479] - - - [641, 128, 1, 256] - - [133, 4.731] - - - [18305, 128, 1, 256] - - [207, 30.812] - - - [23297, 128, 1, 256] - - [218, 36.822] - - - [21377, 128, 1, 256] - - [208, 34.302] - - - [9601, 128, 1, 128] - - [137, 24.866] - - - [13697, 128, 1, 256] - - [138, 39.415] - - - [23681, 128, 1, 128] - - [210, 23.717] - - - [24833, 128, 1, 256] - - [209, 38.387] - - - [25985, 128, 1, 128] - - [215, 25.37] - - - [9601, 128, 1, 256] - - [139, 35.612] - - - [17153, 128, 1, 128] - - [210, 18.729] - - - [9985, 128, 1, 128] - - [140, 25.327] - - - [23297, 128, 1, 128] - - [215, 23.719] - - - [19073, 128, 1, 128] - - [211, 20.373] - - - [2689, 128, 1, 256] - - [118, 18.463] - - - [4993, 128, 1, 128] - - [118, 19.814] - - - [6913, 128, 1, 256] - - [141, 33.681] - - - [6785, 128, 1, 128] - - [132, 22.258] - - - [27905, 128, 1, 128] - - [206, 26.646] - - - [7169, 128, 1, 256] - - [142, 27.705] - - - [11905, 128, 1, 256] - - [137, 35.504] - - - [1409, 128, 1, 128] - - [143, 7.027] - - - [12673, 128, 1, 128] - - [139, 26.276] - - - [27521, 128, 1, 256] - - [212, 41.204] - - - [1409, 128, 1, 256] - - [132, 10.4] - - - [25217, 128, 1, 128] - - [215, 24.917] - - - [7297, 128, 1, 128] - - [137, 20.023] - - - [14081, 128, 1, 128] - - [206, 15.796] - - - [22913, 128, 1, 128] - - [211, 23.168] - - - [10753, 128, 1, 256] - - [144, 32.662] - - - [7937, 128, 1, 128] - - [136, 21.46] - - - [11393, 128, 1, 128] - - [145, 24.445] - - - [26369, 128, 1, 128] - - [206, 25.951] - - - [12161, 128, 1, 256] - - [137, 36.415] - - - [8449, 128, 1, 128] - - [129, 22.514] - - - [22145, 128, 1, 256] - - [213, 35.265] - - - [20225, 128, 1, 256] - - [213, 33.174] - - - [10241, 128, 1, 256] - - [146, 35.572] - - - [6913, 128, 1, 128] - - [132, 21.806] - - - [4993, 128, 1, 256] - - [147, 26.999] - - - [6401, 128, 1, 256] - - [148, 32.584] - - - [13057, 128, 1, 128] - - [149, 26.404] - - - [2945, 128, 1, 128] - - [150, 13.846] - - - [3713, 128, 1, 256] - - [118, 20.452] - - - [10753, 128, 1, 128] - - [151, 23.073] - - - [14849, 128, 1, 256] - - [207, 26.003] - - - [3841, 128, 1, 128] - - [118, 15.578] - - - [28289, 128, 1, 256] - - [214, 32.474] - - - [12929, 128, 1, 128] - - [152, 26.806] - - - [14081, 128, 1, 256] - - [210, 24.426] - - - [14977, 128, 1, 256] - - [207, 26.165] - - - [12545, 128, 1, 128] - - [149, 25.865] - - - [16129, 128, 1, 256] - - [207, 27.59] - - - [11777, 128, 1, 256] - - [144, 34.359] - - - [11777, 128, 1, 128] - - [153, 24.695] - - - [17537, 128, 1, 256] - - [206, 29.553] - - - [5377, 128, 1, 128] - - [154, 17.639] - - - [8065, 128, 1, 256] - - [139, 31.167] - - - [6145, 128, 1, 128] - - [148, 19.981] - - - [20993, 128, 1, 128] - - [214, 21.979] - - - [15617, 128, 1, 128] - - [206, 17.518] - - - [5633, 128, 1, 256] - - [155, 28.675] - - - [4865, 128, 1, 128] - - [133, 19.203] - - - [385, 128, 1, 256] - - [128, 2.856] - - - [3841, 128, 1, 256] - - [147, 21.236] - - - [8833, 128, 1, 256] - - [152, 33.694] - - - [4225, 128, 1, 128] - - [128, 17.135] - - - [11009, 128, 1, 256] - - [152, 33.235] - - - [385, 128, 1, 128] - - [128, 1.973] - - - [9473, 128, 1, 256] - - [152, 35.137] - - - [5761, 128, 1, 128] - - [132, 19.155] - - - [11905, 128, 1, 128] - - [149, 25.324] - - - [4097, 128, 1, 256] - - [156, 22.568] - - - [25217, 128, 1, 256] - - [208, 38.617] - - - [9089, 128, 1, 256] - - [157, 34.055] - - - [10369, 128, 1, 256] - - [152, 37.425] - - - [14209, 128, 1, 256] - - [213, 24.794] - - - [6401, 128, 1, 128] - - [158, 20.814] - - - [27137, 128, 1, 256] - - [208, 41.004] - - - [16385, 128, 1, 256] - - [210, 28.591] - - - [24833, 128, 1, 128] - - [207, 24.538] - - - [18689, 128, 1, 128] - - [219, 19.906] - - - [7553, 128, 1, 256] - - [155, 29.265] - - - [8321, 128, 1, 128] - - [151, 22.334] - - - [15361, 128, 1, 128] - - [206, 17.337] - - - [1153, 128, 1, 128] - - [132, 5.75] - - - [1025, 128, 1, 128] - - [128, 5.182] - - - [19841, 128, 1, 256] - - [216, 32.4] - - - [15233, 128, 1, 128] - - [206, 17.088] - - - [21761, 128, 1, 256] - - [207, 34.842] - - - [17153, 128, 1, 256] - - [211, 29.478] - - - [15617, 128, 1, 256] - - [214, 27.123] - - - [4865, 128, 1, 256] - - [132, 26.404] - - - [14209, 128, 1, 128] - - [215, 16.135] - - - [19457, 128, 1, 256] - - [206, 32.31] - - - [9857, 128, 1, 256] - - [159, 36.107] - - - [11521, 128, 1, 128] - - [148, 24.366] - - - [8449, 128, 1, 256] - - [140, 32.146] - - - [4097, 128, 1, 128] - - [118, 16.708] - - - [28673, 128, 1, 256] - - [211, 31.898] - - - [12161, 128, 1, 128] - - [151, 25.868] - - - [1921, 128, 1, 256] - - [128, 13.314] - - - [9985, 128, 1, 256] - - [140, 36.306] - - - [7937, 128, 1, 256] - - [151, 30.434] - - - [9857, 128, 1, 128] - - [131, 25.528] - - - [13825, 128, 1, 256] - - [138, 39.476] - - - [9089, 128, 1, 128] - - [131, 24.22] - - - [6785, 128, 1, 256] - - [141, 33.387] - - - [5249, 128, 1, 256] - - [151, 27.094] - - - [7681, 128, 1, 256] - - [139, 29.376] - - - [3329, 128, 1, 128] - - [128, 15.075] - - - [14465, 128, 1, 128] - - [215, 16.375] - - - [11137, 128, 1, 256] - - [151, 34.039] - - - [1153, 128, 1, 256] - - [132, 8.596] - - - [16001, 128, 1, 128] - - [210, 17.707] - - - [26753, 128, 1, 128] - - [206, 25.881] - - - [13697, 128, 1, 128] - - [149, 27.251] - - - [3073, 128, 1, 128] - - [150, 14.356] - - - [22529, 128, 1, 256] - - [207, 35.8] - - - [18689, 128, 1, 256] - - [220, 30.86] - - - [257, 128, 1, 128] - - [143, 1.282] - - - [15233, 128, 1, 256] - - [211, 26.581] - - - [27521, 128, 1, 128] - - [214, 26.45] - - - [16385, 128, 1, 128] - - [211, 18.492] - - - [4481, 128, 1, 256] - - [160, 24.683] - - - [6017, 128, 1, 128] - - [161, 19.916] - - - [7297, 128, 1, 256] - - [127, 28.273] - - - [7553, 128, 1, 128] - - [162, 20.572] - - - [21761, 128, 1, 128] - - [214, 22.216] - - - [11393, 128, 1, 256] - - [137, 34.324] - - - [11521, 128, 1, 256] - - [163, 34.428] - - - [12929, 128, 1, 256] - - [159, 37.424] - - - [20225, 128, 1, 128] - - [210, 21.175] - - - [13313, 128, 1, 128] - - [164, 25.928] - - - [2561, 128, 1, 128] - - [165, 12.04] - - - [1537, 128, 1, 128] - - [133, 7.718] - - - [24449, 128, 1, 256] - - [223, 37.636] - - - [12289, 128, 1, 256] - - [144, 35.432] - - - [4225, 128, 1, 256] - - [118, 23.101] - - - [26369, 128, 1, 256] - - [217, 40.172] - - - [17921, 128, 1, 256] - - [215, 30.409] - - - [2945, 128, 1, 256] - - [118, 20.127] - - - [24065, 128, 1, 128] - - [207, 24.069] - - - [6529, 128, 1, 128] - - [133, 21.514] - - - [6145, 128, 1, 256] - - [151, 30.96] - - - [25985, 128, 1, 256] - - [213, 39.546] - - - [8705, 128, 1, 256] - - [140, 32.533] - - - [384, 128, 1, 256] - - [166, 3.132] - - - [25601, 128, 1, 256] - - [218, 38.882] - - - [28673, 128, 1, 128] - - [207, 20.508] - - - [20609, 128, 1, 128] - - [210, 21.547] - - - [19457, 128, 1, 128] - - [219, 20.546] - - - [16769, 128, 1, 256] - - [213, 28.553] - - - [12673, 128, 1, 256] - - [111, 36.9] - - - [8321, 128, 1, 256] - - [163, 31.823] - - - [5249, 128, 1, 128] - - [132, 17.854] - - - [16129, 128, 1, 128] - - [206, 18.011] - - - [13441, 128, 1, 256] - - [112, 38.454] - - - [5377, 128, 1, 256] - - [163, 27.466] - - - [21377, 128, 1, 128] - - [207, 21.885] - - - [14465, 128, 1, 256] - - [214, 25.211] - - - [11137, 128, 1, 128] - - [164, 24.107] - - - [7681, 128, 1, 128] - - [131, 20.692] - - - [7169, 128, 1, 128] - - [127, 19.671] - - - [22145, 128, 1, 128] - - [206, 22.957] - - - [11009, 128, 1, 128] - - [167, 23.76] - - - [20993, 128, 1, 256] - - [214, 34.018] - - - [13313, 128, 1, 256] - - [168, 37.292] - - - [25601, 128, 1, 128] - - [215, 24.669] - - - [4609, 128, 1, 128] - - [132, 18.29] - - - [5761, 128, 1, 256] - - [148, 29.529] - - - [17921, 128, 1, 128] - - [206, 19.773] - - - [2689, 128, 1, 128] - - [143, 12.405] - - - [8705, 128, 1, 128] - - [131, 22.785] - - - [10241, 128, 1, 128] - - [153, 25.624] - - - [14977, 128, 1, 128] - - [207, 16.955] - - - [18305, 128, 1, 128] - - [215, 19.753] - - - [3457, 128, 1, 128] - - [128, 15.372] - - - [24065, 128, 1, 256] - - [207, 37.356] - - - [12289, 128, 1, 128] - - [149, 25.127] - - - [14593, 128, 1, 128] - - [206, 16.546] - - - [2177, 128, 1, 256] - - [169, 15.018] - - - [4481, 128, 1, 128] - - [133, 18.173] - - - [8065, 128, 1, 128] - - [127, 21.886] - - - [3457, 128, 1, 256] - - [128, 22.482] - - - [6529, 128, 1, 256] - - [137, 32.895] - - - [26753, 128, 1, 256] - - [221, 40.673] - - - [17537, 128, 1, 128] - - [214, 19.148] - - - [22529, 128, 1, 128] - - [207, 22.748] - - - [10625, 128, 1, 256] - - [163, 32.474] - - - [14849, 128, 1, 128] - - [206, 17.019] - - - [9217, 128, 1, 256] - - [153, 33.105] - - - [19841, 128, 1, 128] - - [206, 21.072] - - - [15745, 128, 1, 128] - - [214, 17.609] - - - [13441, 128, 1, 128] - - [152, 27.181] - - - [3713, 128, 1, 128] - - [118, 15.058] - - - [27137, 128, 1, 128] - - [220, 26.286] - - - [16001, 128, 1, 256] - - [215, 27.499] - - - [10369, 128, 1, 128] - - [131, 26.301] - - - [1921, 128, 1, 128] - - [118, 8.975] - - - [9473, 128, 1, 128] - - [140, 24.62] - - - [27905, 128, 1, 256] - - [222, 41.737] - - - [30976, 1024, 1, 128] - - [27, 42.513] - - - [42240, 26369, 1, 128] - - [22, 45.705] - - - [33024, 17025, 1, 128] - - [27, 46.355] - - - [39168, 512, 1, 128] - - [49, 41.06] - - - [30848, 1024, 1, 128] - - [60, 42.967] - - - [41728, 8192, 1, 128] - - [64, 45.865] - - - [39552, 23553, 1, 128] - - [22, 45.133] - - - [35072, 512, 1, 128] - - [58, 41.862] - - - [29952, 14081, 1, 128] - - [25, 45.934] - - - [33280, 2048, 1, 128] - - [39, 44.567] - - - [40320, 128, 1, 128] - - [34, 32.824] - - - [35456, 1024, 1, 128] - - [36, 42.971] - - - [36096, 1024, 1, 128] - - [59, 40.444] - - - [36992, 20993, 1, 128] - - [22, 45.407] - - - [36096, 20097, 1, 128] - - [64, 44.703] - - - [31488, 15489, 1, 128] - - [22, 45.828] - - - [39552, 23681, 1, 128] - - [22, 45.361] - - - [36864, 128, 1, 128] - - [53, 30.903] - - - [40320, 4096, 1, 128] - - [27, 45.636] - - - [35200, 2048, 1, 128] - - [53, 44.094] - - - [29824, 2048, 1, 128] - - [67, 43.127] - - - [34688, 2048, 1, 128] - - [40, 44.433] - - - [42752, 26753, 1, 128] - - [22, 45.655] - - - [34304, 4096, 1, 128] - - [67, 46.126] - - - [36480, 20481, 1, 128] - - [50, 45.313] - - - [33408, 128, 1, 128] - - [40, 28.689] - - - [38784, 4096, 1, 128] - - [34, 45.248] - - - [43264, 27393, 1, 128] - - [38, 45.698] - - - [34560, 128, 1, 128] - - [34, 29.604] - - - [30336, 4096, 1, 128] - - [67, 45.255] - - - [29056, 2048, 1, 128] - - [35, 43.507] - - - [34816, 512, 1, 128] - - [23, 42.118] - - - [38272, 2048, 1, 128] - - [96, 36.638] - - - [39808, 23937, 1, 128] - - [64, 43.271] - - - [30848, 512, 1, 128] - - [34, 40.458] - - - [40448, 512, 1, 128] - - [60, 41.853] - - - [40448, 24577, 1, 128] - - [64, 45.618] - - - [44544, 28545, 1, 128] - - [50, 45.737] - - - [30208, 14209, 1, 128] - - [30, 46.282] - - - [34688, 18689, 1, 128] - - [29, 45.523] - - - [31360, 512, 1, 128] - - [29, 40.429] - - - [38912, 512, 1, 128] - - [34, 41.394] - - - [39680, 1024, 1, 128] - - [34, 42.937] - - - [34048, 1024, 1, 128] - - [34, 43.28] - - - [39552, 4096, 1, 128] - - [38, 45.488] - - - [40320, 24321, 1, 128] - - [22, 45.408] - - - [40832, 24833, 1, 128] - - [38, 45.035] - - - [36736, 1024, 1, 128] - - [35, 42.751] - - - [44672, 1024, 1, 128] - - [60, 43.451] - - - [32000, 128, 1, 128] - - [40, 27.968] - - - [40704, 4096, 1, 128] - - [27, 45.772] - - - [38144, 1024, 1, 128] - - [29, 43.025] - - - [30720, 14849, 1, 128] - - [38, 46.757] - - - [38144, 8192, 1, 128] - - [37, 46.834] - - - [30208, 1024, 1, 128] - - [35, 42.874] - - - [43136, 1024, 1, 128] - - [34, 42.913] - - - [38528, 1024, 1, 128] - - [36, 42.892] - - - [43264, 2048, 1, 128] - - [58, 44.133] - - - [38400, 22529, 1, 128] - - [64, 45.897] - - - [37120, 128, 1, 128] - - [31, 31.012] - - - [32256, 128, 1, 128] - - [58, 28.12] - - - [29952, 13953, 1, 128] - - [37, 45.983] - - - [34560, 8192, 1, 128] - - [27, 46.641] - - - [37504, 21505, 1, 128] - - [51, 44.546] - - - [33536, 128, 1, 128] - - [49, 28.894] - - - [41856, 2048, 1, 128] - - [60, 44.478] - - - [32896, 4096, 1, 128] - - [74, 43.557] - - - [41856, 8192, 1, 128] - - [37, 46.566] - - - [29440, 4096, 1, 128] - - [38, 45.779] - - - [33664, 8192, 1, 128] - - [47, 46.232] - - - [36992, 512, 1, 128] - - [72, 39.86] - - - [33280, 512, 1, 128] - - [31, 40.592] - - - [41728, 128, 1, 128] - - [59, 33.563] - - - [31744, 128, 1, 128] - - [34, 27.614] - - - [31360, 1024, 1, 128] - - [60, 42.98] - - - [29952, 8192, 1, 128] - - [51, 46.599] - - - [38016, 2048, 1, 128] - - [63, 44.223] - - - [34176, 8192, 1, 128] - - [47, 46.412] - - - [30464, 512, 1, 128] - - [60, 39.585] - - - [41984, 2048, 1, 128] - - [61, 44.722] - - - [40448, 4096, 1, 128] - - [51, 46.116] - - - [33920, 4096, 1, 128] - - [27, 45.581] - - - [41088, 8192, 1, 128] - - [61, 45.778] - - - [39808, 8192, 1, 128] - - [22, 44.228] - - - [40832, 4096, 1, 128] - - [25, 45.606] - - - [30592, 2048, 1, 128] - - [60, 43.833] - - - [36352, 1024, 1, 128] - - [40, 43.102] - - - [30336, 2048, 1, 128] - - [40, 43.833] - - - [30976, 512, 1, 128] - - [40, 40.197] - - - [42368, 1024, 1, 128] - - [40, 43.084] - - - [29056, 1024, 1, 128] - - [40, 42.348] - - - [38784, 22913, 1, 128] - - [22, 45.544] - - - [28928, 512, 1, 128] - - [34, 39.211] - - - [40576, 512, 1, 128] - - [27, 41.519] - - - [34816, 4096, 1, 128] - - [27, 46.696] - - - [41600, 2048, 1, 128] - - [22, 43.878] - - - [29696, 8192, 1, 128] - - [25, 47.304] - - - [41856, 4096, 1, 128] - - [27, 45.762] - - - [35584, 2048, 1, 128] - - [63, 44.122] - - - [30848, 14849, 1, 128] - - [37, 45.765] - - - [33280, 17281, 1, 128] - - [47, 46.514] - - - [43776, 2048, 1, 128] - - [60, 43.542] - - - [42112, 8192, 1, 128] - - [25, 46.462] - - - [37376, 128, 1, 128] - - [63, 31.226] - - - [41600, 4096, 1, 128] - - [38, 45.603] - - - [36224, 20353, 1, 128] - - [38, 45.672] - - - [29952, 1024, 1, 128] - - [35, 42.84] - - - [34176, 1024, 1, 128] - - [34, 42.871] - - - [31744, 512, 1, 128] - - [23, 40.693] - - - [42624, 8192, 1, 128] - - [27, 41.712] - - - [41216, 128, 1, 128] - - [44, 33.152] - - - [42624, 26753, 1, 128] - - [77, 40.481] - - - [32512, 2048, 1, 128] - - [44, 44.312] - - - [40064, 4096, 1, 128] - - [22, 45.152] - - - [32640, 4096, 1, 128] - - [27, 45.254] - - - [42112, 26241, 1, 128] - - [22, 45.551] - - - [32256, 512, 1, 128] - - [57, 40.905] - - - [40960, 1024, 1, 128] - - [27, 40.673] - - - [35968, 128, 1, 128] - - [40, 30.387] - - - [32384, 8192, 1, 128] - - [29, 46.529] - - - [42880, 512, 1, 128] - - [55, 40.788] - - - [33024, 8192, 1, 128] - - [29, 47.007] - - - [43904, 1024, 1, 128] - - [40, 42.98] - - - [33664, 17665, 1, 128] - - [38, 45.702] - - - [41856, 512, 1, 128] - - [53, 42.279] - - - [40704, 128, 1, 128] - - [33, 33.254] - - - [33408, 17537, 1, 128] - - [38, 45.966] - - - [37120, 512, 1, 128] - - [60, 40.735] - - - [41216, 25345, 1, 128] - - [64, 45.806] - - - [39680, 8192, 1, 128] - - [37, 46.567] - - - [40192, 24193, 1, 128] - - [38, 45.525] - - - [33024, 17153, 1, 128] - - [22, 46.343] - - - [38272, 1024, 1, 128] - - [30, 38.221] - - - [35328, 1024, 1, 128] - - [35, 43.01] - - - [31104, 8192, 1, 128] - - [27, 46.434] - - - [40320, 8192, 1, 128] - - [25, 46.483] - - - [29312, 2048, 1, 128] - - [58, 43.896] - - - [36608, 20737, 1, 128] - - [22, 45.83] - - - [42240, 4096, 1, 128] - - [25, 45.827] - - - [43520, 2048, 1, 128] - - [33, 44.598] - - - [29056, 512, 1, 128] - - [60, 39.186] - - - [35328, 19329, 1, 128] - - [50, 46.24] - - - [30464, 128, 1, 128] - - [23, 26.971] - - - [29696, 13697, 1, 128] - - [27, 46.562] - - - [43904, 28033, 1, 128] - - [27, 45.375] - - - [35584, 19713, 1, 128] - - [22, 45.7] - - - [41088, 4096, 1, 128] - - [36, 45.105] - - - [42368, 2048, 1, 128] - - [58, 44.078] - - - [36736, 128, 1, 128] - - [35, 30.789] - - - [30336, 8192, 1, 128] - - [35, 46.227] - - - [43008, 128, 1, 128] - - [34, 31.442] - - - [37120, 1024, 1, 128] - - [40, 42.986] - - - [31104, 2048, 1, 128] - - [60, 44.22] - - - [33152, 4096, 1, 128] - - [34, 45.818] - - - [43392, 27521, 1, 128] - - [38, 45.358] - - - [37248, 21249, 1, 128] - - [38, 45.316] - - - [33920, 17921, 1, 128] - - [38, 45.594] - - - [39680, 4096, 1, 128] - - [25, 45.776] - - - [43264, 512, 1, 128] - - [55, 40.916] - - - [35712, 8192, 1, 128] - - [25, 46.366] - - - [31616, 2048, 1, 128] - - [38, 43.609] - - - [35328, 512, 1, 128] - - [34, 40.342] - - - [43136, 27265, 1, 128] - - [22, 45.393] - - - [30208, 128, 1, 128] - - [60, 26.654] - - - [40320, 24449, 1, 128] - - [38, 45.352] - - - [44288, 2048, 1, 128] - - [36, 44.467] - - - [35072, 1024, 1, 128] - - [27, 43.152] - - - [30464, 14465, 1, 128] - - [74, 43.511] - - - [44160, 8192, 1, 128] - - [25, 46.21] - - - [33792, 17793, 1, 128] - - [29, 46.448] - - - [37632, 1024, 1, 128] - - [34, 43.181] - - - [35968, 2048, 1, 128] - - [60, 44.208] - - - [38400, 8192, 1, 128] - - [30, 46.925] - - - [32512, 4096, 1, 128] - - [27, 45.842] - - - [32512, 16641, 1, 128] - - [51, 46.246] - - - [39424, 128, 1, 128] - - [57, 32.351] - - - [30976, 8192, 1, 128] - - [76, 45.168] - - - [35968, 20097, 1, 128] - - [25, 45.717] - - - [38656, 512, 1, 128] - - [58, 40.966] - - - [34944, 18945, 1, 128] - - [27, 45.624] - - - [33664, 17793, 1, 128] - - [22, 45.754] - - - [38656, 22657, 1, 128] - - [50, 45.798] - - - [34944, 1024, 1, 128] - - [34, 43.102] - - - [31872, 16001, 1, 128] - - [38, 45.704] - - - [43392, 8192, 1, 128] - - [37, 46.144] - - - [38016, 512, 1, 128] - - [40, 40.69] - - - [29440, 8192, 1, 128] - - [22, 46.686] - - - [35200, 1024, 1, 128] - - [40, 42.748] - - - [34304, 18433, 1, 128] - - [50, 46.162] - - - [44672, 28801, 1, 128] - - [38, 45.545] - - - [29184, 4096, 1, 128] - - [57, 45.73] - - - [33408, 8192, 1, 128] - - [27, 46.532] - - - [39040, 128, 1, 128] - - [39, 31.994] - - - [39680, 23681, 1, 128] - - [38, 45.518] - - - [38144, 4096, 1, 128] - - [25, 45.883] - - - [42368, 26497, 1, 128] - - [51, 45.346] - - - [42368, 4096, 1, 128] - - [67, 45.365] - - - [31872, 128, 1, 128] - - [63, 27.758] - - - [41984, 512, 1, 128] - - [44, 42.307] - - - [39296, 2048, 1, 128] - - [38, 43.934] - - - [33920, 2048, 1, 128] - - [36, 44.095] - - - [36736, 20865, 1, 128] - - [22, 45.681] - - - [34432, 8192, 1, 128] - - [30, 44.91] - - - [30848, 14977, 1, 128] - - [37, 45.811] - - - [31744, 15873, 1, 128] - - [29, 46.488] - - - [42880, 27009, 1, 128] - - [64, 44.895] - - - [42240, 26241, 1, 128] - - [38, 45.662] - - - [38400, 4096, 1, 128] - - [37, 45.966] - - - [42624, 26625, 1, 128] - - [27, 40.471] - - - [35072, 4096, 1, 128] - - [37, 45.875] - - - [40576, 4096, 1, 128] - - [25, 45.622] - - - [39296, 8192, 1, 128] - - [29, 46.525] - - - [42624, 512, 1, 128] - - [61, 40.003] - - - [32768, 8192, 1, 128] - - [37, 35.565] - - - [36864, 1024, 1, 128] - - [35, 42.756] - - - [43392, 128, 1, 128] - - [59, 31.053] - - - [41344, 2048, 1, 128] - - [40, 44.187] - - - [35584, 4096, 1, 128] - - [27, 45.526] - - - [40064, 2048, 1, 128] - - [22, 42.994] - - - [40576, 24705, 1, 128] - - [38, 45.497] - - - [39808, 1024, 1, 128] - - [27, 42.27] - - - [36992, 1024, 1, 128] - - [72, 42.438] - - - [42496, 1024, 1, 128] - - [63, 43.847] - - - [43904, 128, 1, 128] - - [47, 31.389] - - - [31232, 512, 1, 128] - - [49, 40.907] - - - [42112, 128, 1, 128] - - [40, 33.659] - - - [37376, 2048, 1, 128] - - [58, 44.486] - - - [38016, 128, 1, 128] - - [60, 31.37] - - - [42368, 8192, 1, 128] - - [30, 46.148] - - - [43392, 512, 1, 128] - - [23, 41.032] - - - [41984, 1024, 1, 128] - - [58, 43.616] - - - [42240, 2048, 1, 128] - - [53, 44.379] - - - [29952, 128, 1, 128] - - [23, 26.55] - - - [36608, 8192, 1, 128] - - [38, 46.612] - - - [32512, 16513, 1, 128] - - [64, 46.298] - - - [29568, 512, 1, 128] - - [31, 38.598] - - - [34304, 1024, 1, 128] - - [57, 43.493] - - - [41984, 4096, 1, 128] - - [25, 46.4] - - - [30464, 4096, 1, 128] - - [76, 43.433] - - - [41216, 2048, 1, 128] - - [36, 44.506] - - - [36480, 20609, 1, 128] - - [38, 45.594] - - - [44800, 4096, 1, 128] - - [29, 45.493] - - - [36864, 512, 1, 128] - - [23, 40.566] - - - [39680, 2048, 1, 128] - - [29, 43.919] - - - [43648, 4096, 1, 128] - - [25, 45.327] - - - [33664, 128, 1, 128] - - [23, 29.078] - - - [41600, 512, 1, 128] - - [74, 42.222] - - - [43776, 1024, 1, 128] - - [63, 41.689] - - - [37632, 512, 1, 128] - - [35, 40.728] - - - [44160, 128, 1, 128] - - [39, 31.541] - - - [37248, 8192, 1, 128] - - [37, 46.313] - - - [34816, 18817, 1, 128] - - [29, 46.801] - - - [38528, 22529, 1, 128] - - [38, 45.314] - - - [40192, 24321, 1, 128] - - [38, 45.492] - - - [40832, 128, 1, 128] - - [59, 32.914] - - - [29312, 8192, 1, 128] - - [27, 46.211] - - - [43776, 27777, 1, 128] - - [61, 43.983] - - - [37632, 21633, 1, 128] - - [38, 45.804] - - - [33792, 4096, 1, 128] - - [25, 46.375] - - - [35968, 1024, 1, 128] - - [23, 43.0] - - - [37888, 512, 1, 128] - - [34, 41.05] - - - [35968, 512, 1, 128] - - [55, 40.217] - - - [30592, 1024, 1, 128] - - [53, 42.842] - - - [38400, 512, 1, 128] - - [39, 40.674] - - - [43264, 1024, 1, 128] - - [25, 43.316] - - - [38528, 4096, 1, 128] - - [49, 45.377] - - - [28928, 1024, 1, 128] - - [35, 42.433] - - - [33152, 1024, 1, 128] - - [23, 42.341] - - - [41344, 1024, 1, 128] - - [40, 43.141] - - - [30848, 8192, 1, 128] - - [37, 46.564] - - - [41344, 4096, 1, 128] - - [29, 45.428] - - - [38912, 2048, 1, 128] - - [29, 44.728] - - - [38272, 128, 1, 128] - - [84, 31.539] - - - [31488, 4096, 1, 128] - - [25, 45.665] - - - [44416, 4096, 1, 128] - - [25, 45.725] - - - [39552, 2048, 1, 128] - - [40, 44.397] - - - [37760, 1024, 1, 128] - - [40, 43.235] - - - [34304, 18305, 1, 128] - - [50, 46.325] - - - [44544, 28673, 1, 128] - - [64, 45.502] - - - [44416, 8192, 1, 128] - - [25, 46.505] - - - [38144, 512, 1, 128] - - [27, 40.804] - - - [30208, 14337, 1, 128] - - [24, 46.144] - - - [38144, 2048, 1, 128] - - [37, 44.289] - - - [40448, 128, 1, 128] - - [78, 33.044] - - - [42240, 8192, 1, 128] - - [25, 46.544] - - - [39424, 2048, 1, 128] - - [36, 44.584] - - - [41088, 512, 1, 128] - - [57, 39.121] - - - [36224, 2048, 1, 128] - - [60, 43.974] - - - [31744, 4096, 1, 128] - - [27, 46.397] - - - [44160, 512, 1, 128] - - [49, 40.882] - - - [32000, 1024, 1, 128] - - [34, 42.737] - - - [42752, 1024, 1, 128] - - [40, 43.363] - - - [42496, 2048, 1, 128] - - [39, 44.786] - - - [32640, 2048, 1, 128] - - [25, 43.065] - - - [42752, 26881, 1, 128] - - [38, 45.625] - - - [32256, 8192, 1, 128] - - [51, 47.12] - - - [44800, 512, 1, 128] - - [37, 41.246] - - - [34816, 128, 1, 128] - - [63, 29.795] - - - [38272, 8192, 1, 128] - - [61, 42.399] - - - [44800, 28929, 1, 128] - - [22, 45.322] - - - [37120, 8192, 1, 128] - - [22, 46.547] - - - [43776, 512, 1, 128] - - [33, 37.791] - - - [43008, 1024, 1, 128] - - [35, 43.999] - - - [34432, 18561, 1, 128] - - [74, 44.536] - - - [36736, 4096, 1, 128] - - [27, 45.56] - - - [36224, 512, 1, 128] - - [40, 40.29] - - - [32768, 512, 1, 128] - - [36, 36.086] - - - [30592, 128, 1, 128] - - [40, 26.897] - - - [43008, 27137, 1, 128] - - [22, 46.575] - - - [34048, 18177, 1, 128] - - [64, 45.732] - - - [43136, 2048, 1, 128] - - [37, 43.944] - - - [29184, 13313, 1, 128] - - [22, 45.766] - - - [40064, 24193, 1, 128] - - [64, 44.519] - - - [40960, 128, 1, 128] - - [53, 32.703] - - - [29184, 2048, 1, 128] - - [39, 44.175] - - - [37248, 128, 1, 128] - - [34, 31.154] - - - [35328, 128, 1, 128] - - [34, 30.192] - - - [43264, 128, 1, 128] - - [57, 31.142] - - - [29952, 4096, 1, 128] - - [57, 45.58] - - - [36736, 20737, 1, 128] - - [38, 45.629] - - - [34176, 4096, 1, 128] - - [31, 45.382] - - - [32768, 1024, 1, 128] - - [25, 36.303] - - - [44160, 4096, 1, 128] - - [25, 45.544] - - - [31104, 1024, 1, 128] - - [49, 43.24] - - - [33792, 512, 1, 128] - - [60, 41.448] - - - [41216, 25217, 1, 128] - - [64, 45.833] - - - [31872, 1024, 1, 128] - - [25, 42.624] - - - [38528, 8192, 1, 128] - - [24, 46.201] - - - [44672, 4096, 1, 128] - - [29, 45.698] - - - [32512, 1024, 1, 128] - - [23, 42.329] - - - [39168, 8192, 1, 128] - - [64, 46.6] - - - [31360, 15361, 1, 128] - - [38, 45.526] - - - [38016, 22145, 1, 128] - - [38, 45.578] - - - [35712, 128, 1, 128] - - [34, 30.005] - - - [30208, 4096, 1, 128] - - [59, 45.872] - - - [33920, 128, 1, 128] - - [35, 29.163] - - - [30336, 128, 1, 128] - - [40, 26.666] - - - [42368, 128, 1, 128] - - [26, 30.858] - - - [38912, 4096, 1, 128] - - [37, 46.69] - - - [34176, 512, 1, 128] - - [25, 41.65] - - - [42752, 8192, 1, 128] - - [27, 46.612] - - - [31488, 1024, 1, 128] - - [34, 43.165] - - - [36608, 1024, 1, 128] - - [60, 42.897] - - - [41856, 128, 1, 128] - - [40, 33.455] - - - [29312, 13441, 1, 128] - - [25, 45.652] - - - [43520, 128, 1, 128] - - [36, 31.296] - - - [31616, 8192, 1, 128] - - [25, 45.923] - - - [40448, 2048, 1, 128] - - [33, 44.652] - - - [35328, 2048, 1, 128] - - [44, 44.595] - - - [36864, 20865, 1, 128] - - [38, 46.566] - - - [32000, 2048, 1, 128] - - [27, 44.031] - - - [34176, 18177, 1, 128] - - [64, 45.62] - - - [37504, 128, 1, 128] - - [40, 31.255] - - - [33792, 1024, 1, 128] - - [35, 43.346] - - - [31872, 8192, 1, 128] - - [25, 46.308] - - - [40704, 512, 1, 128] - - [34, 41.912] - - - [37632, 128, 1, 128] - - [34, 31.221] - - - [32640, 1024, 1, 128] - - [25, 41.437] - - - [44544, 8192, 1, 128] - - [22, 46.449] - - - [39424, 8192, 1, 128] - - [25, 46.96] - - - [39296, 512, 1, 128] - - [34, 41.084] - - - [35840, 128, 1, 128] - - [27, 30.25] - - - [39168, 1024, 1, 128] - - [49, 43.49] - - - [35712, 19841, 1, 128] - - [38, 45.634] - - - [29568, 13569, 1, 128] - - [22, 45.271] - - - [34944, 4096, 1, 128] - - [27, 45.826] - - - [32768, 2048, 1, 128] - - [29, 34.701] - - - [39296, 128, 1, 128] - - [27, 32.103] - - - [29568, 4096, 1, 128] - - [61, 44.664] - - - [39040, 1024, 1, 128] - - [58, 43.154] - - - [37376, 1024, 1, 128] - - [34, 43.367] - - - [33536, 2048, 1, 128] - - [25, 44.356] - - - [31488, 8192, 1, 128] - - [27, 46.506] - - - [37888, 1024, 1, 128] - - [40, 43.644] - - - [41472, 4096, 1, 128] - - [30, 45.992] - - - [30592, 512, 1, 128] - - [34, 40.13] - - - [34560, 18561, 1, 128] - - [29, 45.867] - - - [29184, 512, 1, 128] - - [34, 39.349] - - - [32256, 16257, 1, 128] - - [24, 46.612] - - - [43392, 27393, 1, 128] - - [38, 45.405] - - - [29312, 4096, 1, 128] - - [35, 45.362] - - - [43648, 2048, 1, 128] - - [29, 43.517] - - - [44288, 1024, 1, 128] - - [60, 43.32] - - - [35456, 128, 1, 128] - - [67, 30.098] - - - [44160, 28289, 1, 128] - - [37, 45.17] - - - [40320, 1024, 1, 128] - - [40, 43.195] - - - [37888, 22017, 1, 128] - - [38, 46.388] - - - [29696, 512, 1, 128] - - [34, 39.605] - - - [35840, 2048, 1, 128] - - [44, 44.646] - - - [37504, 2048, 1, 128] - - [63, 43.735] - - - [41728, 4096, 1, 128] - - [67, 45.048] - - - [42752, 4096, 1, 128] - - [37, 45.734] - - - [29824, 4096, 1, 128] - - [61, 44.859] - - - [44800, 1024, 1, 128] - - [35, 43.392] - - - [30592, 4096, 1, 128] - - [25, 45.582] - - - [43904, 4096, 1, 128] - - [27, 45.454] - - - [39552, 8192, 1, 128] - - [38, 46.1] - - - [37632, 2048, 1, 128] - - [60, 44.242] - - - [29312, 128, 1, 128] - - [23, 25.614] - - - [30080, 512, 1, 128] - - [39, 39.423] - - - [33664, 2048, 1, 128] - - [53, 44.161] - - - [43520, 27521, 1, 128] - - [64, 46.018] - - - [36224, 128, 1, 128] - - [34, 30.533] - - - [28928, 12929, 1, 128] - - [37, 45.873] - - - [29440, 1024, 1, 128] - - [35, 42.53] - - - [35840, 19969, 1, 128] - - [38, 46.357] - - - [42880, 4096, 1, 128] - - [57, 44.933] - - - [42496, 8192, 1, 128] - - [25, 46.936] - - - [39936, 24065, 1, 128] - - [38, 46.243] - - - [33408, 1024, 1, 128] - - [25, 42.705] - - - [32256, 2048, 1, 128] - - [36, 44.555] - - - [35712, 19713, 1, 128] - - [38, 45.682] - - - [40192, 4096, 1, 128] - - [37, 45.816] - - - [32000, 16129, 1, 128] - - [27, 46.012] - - - [44032, 512, 1, 128] - - [23, 41.391] - - - [35584, 128, 1, 128] - - [40, 30.103] - - - [35584, 8192, 1, 128] - - [29, 46.454] - - - [37888, 21889, 1, 128] - - [38, 46.38] - - - [37504, 1024, 1, 128] - - [59, 42.416] - - - [33664, 512, 1, 128] - - [60, 41.523] - - - [32384, 1024, 1, 128] - - [94, 42.482] - - - [38400, 1024, 1, 128] - - [72, 43.355] - - - [35200, 128, 1, 128] - - [63, 29.841] - - - [43648, 1024, 1, 128] - - [35, 42.242] - - - [36608, 128, 1, 128] - - [23, 30.717] - - - [32768, 128, 1, 128] - - [40, 28.982] - - - [28928, 4096, 1, 128] - - [29, 45.55] - - - [35200, 19329, 1, 128] - - [22, 45.615] - - - [41216, 8192, 1, 128] - - [25, 46.646] - - - [36864, 8192, 1, 128] - - [29, 47.323] - - - [40064, 128, 1, 128] - - [57, 32.473] - - - [42624, 1024, 1, 128] - - [37, 41.156] - - - [34688, 128, 1, 128] - - [34, 29.748] - - - [43648, 27777, 1, 128] - - [22, 44.988] - - - [37888, 8192, 1, 128] - - [37, 47.162] - - - [41472, 25601, 1, 128] - - [64, 45.782] - - - [38272, 512, 1, 128] - - [27, 37.579] - - - [35456, 4096, 1, 128] - - [49, 45.661] - - - [42496, 26625, 1, 128] - - [50, 45.874] - - - [43136, 4096, 1, 128] - - [37, 45.517] - - - [44800, 8192, 1, 128] - - [29, 46.255] - - - [36480, 8192, 1, 128] - - [29, 46.305] - - - [37504, 4096, 1, 128] - - [63, 44.983] - - - [39040, 8192, 1, 128] - - [27, 46.156] - - - [31104, 512, 1, 128] - - [23, 40.757] - - - [34176, 2048, 1, 128] - - [40, 44.049] - - - [31616, 512, 1, 128] - - [63, 40.723] - - - [35456, 2048, 1, 128] - - [23, 44.265] - - - [43136, 8192, 1, 128] - - [37, 46.264] - - - [33024, 128, 1, 128] - - [60, 28.36] - - - [38656, 4096, 1, 128] - - [51, 45.653] - - - [33408, 17409, 1, 128] - - [38, 45.715] - - - [39424, 1024, 1, 128] - - [39, 43.586] - - - [29312, 13313, 1, 128] - - [27, 45.515] - - - [35840, 4096, 1, 128] - - [29, 46.364] - - - [42496, 512, 1, 128] - - [35, 41.174] - - - [37632, 8192, 1, 128] - - [37, 46.625] - - - [41088, 2048, 1, 128] - - [58, 43.811] - - - [38528, 512, 1, 128] - - [25, 40.582] - - - [35072, 2048, 1, 128] - - [53, 44.333] - - - [31104, 4096, 1, 128] - - [38, 45.51] - - - [33280, 4096, 1, 128] - - [67, 46.23] - - - [43904, 8192, 1, 128] - - [37, 46.33] - - - [34816, 8192, 1, 128] - - [27, 47.493] - - - [38016, 1024, 1, 128] - - [55, 43.185] - - - [33152, 128, 1, 128] - - [59, 28.403] - - - [42496, 128, 1, 128] - - [42, 31.099] - - - [40832, 24961, 1, 128] - - [38, 45.111] - - - [41728, 1024, 1, 128] - - [67, 42.868] - - - [41472, 25473, 1, 128] - - [50, 46.035] - - - [34560, 2048, 1, 128] - - [63, 44.545] - - - [31616, 15617, 1, 128] - - [22, 45.318] - - - [33664, 4096, 1, 128] - - [38, 45.445] - - - [35328, 8192, 1, 128] - - [30, 47.021] - - - [39808, 4096, 1, 128] - - [37, 43.349] - - - [37248, 512, 1, 128] - - [40, 40.754] - - - [31360, 4096, 1, 128] - - [27, 45.616] - - - [41344, 8192, 1, 128] - - [37, 46.237] - - - [32000, 512, 1, 128] - - [49, 40.442] - - - [35968, 19969, 1, 128] - - [27, 45.615] - - - [30080, 14081, 1, 128] - - [47, 43.41] - - - [35840, 8192, 1, 128] - - [37, 47.132] - - - [44672, 2048, 1, 128] - - [63, 44.597] - - - [31872, 2048, 1, 128] - - [37, 43.727] - - - [42496, 4096, 1, 128] - - [22, 46.1] - - - [43776, 128, 1, 128] - - [58, 31.207] - - - [40704, 2048, 1, 128] - - [44, 44.366] - - - [34432, 128, 1, 128] - - [35, 29.398] - - - [44544, 2048, 1, 128] - - [69, 44.537] - - - [32384, 16385, 1, 128] - - [25, 45.9] - - - [43776, 27905, 1, 128] - - [61, 44.016] - - - [44032, 4096, 1, 128] - - [37, 46.373] - - - [36480, 512, 1, 128] - - [39, 40.173] - - - [44160, 1024, 1, 128] - - [63, 43.268] - - - [41216, 4096, 1, 128] - - [37, 45.814] - - - [44032, 2048, 1, 128] - - [58, 44.796] - - - [33152, 2048, 1, 128] - - [63, 44.512] - - - [41984, 25985, 1, 128] - - [38, 46.352] - - - [39552, 512, 1, 128] - - [35, 41.308] - - - [41344, 25473, 1, 128] - - [29, 45.432] - - - [40960, 4096, 1, 128] - - [22, 41.231] - - - [32640, 128, 1, 128] - - [60, 28.095] - - - [35968, 4096, 1, 128] - - [35, 45.611] - - - [33536, 4096, 1, 128] - - [29, 45.86] - - - [30976, 15105, 1, 128] - - [76, 44.642] - - - [35072, 8192, 1, 128] - - [29, 46.675] - - - [39424, 23425, 1, 128] - - [64, 45.929] - - - [43520, 1024, 1, 128] - - [33, 43.744] - - - [44288, 28417, 1, 128] - - [50, 45.373] - - - [30848, 128, 1, 128] - - [40, 27.219] - - - [35712, 512, 1, 128] - - [27, 40.09] - - - [44160, 2048, 1, 128] - - [53, 44.248] - - - [34048, 8192, 1, 128] - - [64, 46.442] - - - [40448, 24449, 1, 128] - - [50, 46.019] - - - [39168, 23297, 1, 128] - - [50, 45.842] - - - [32128, 1024, 1, 128] - - [60, 42.82] - - - [36864, 20993, 1, 128] - - [38, 46.547] - - - [40064, 1024, 1, 128] - - [37, 42.512] - - - [38784, 8192, 1, 128] - - [37, 46.238] - - - [37248, 2048, 1, 128] - - [37, 43.739] - - - [34560, 4096, 1, 128] - - [27, 45.799] - - - [39040, 23041, 1, 128] - - [22, 45.415] - - - [36480, 1024, 1, 128] - - [60, 42.697] - - - [39040, 2048, 1, 128] - - [53, 44.221] - - - [39808, 23809, 1, 128] - - [64, 43.154] - - - [36992, 4096, 1, 128] - - [37, 45.424] - - - [32768, 16897, 1, 128] - - [37, 34.596] - - - [30976, 2048, 1, 128] - - [31, 42.926] - - - [32640, 16769, 1, 128] - - [29, 45.395] - - - [29824, 13953, 1, 128] - - [64, 45.133] - - - [29184, 128, 1, 128] - - [40, 25.776] - - - [30720, 8192, 1, 128] - - [29, 47.523] - - - [30848, 2048, 1, 128] - - [40, 43.898] - - - [38016, 4096, 1, 128] - - [22, 45.516] - - - [35456, 8192, 1, 128] - - [27, 46.552] - - - [36992, 21121, 1, 128] - - [22, 45.411] - - - [36736, 2048, 1, 128] - - [53, 44.039] - - - [37888, 128, 1, 128] - - [60, 31.44] - - - [39808, 2048, 1, 128] - - [27, 42.312] - - - [41856, 25985, 1, 128] - - [22, 45.662] - - - [34688, 4096, 1, 128] - - [49, 45.486] - - - [38784, 1024, 1, 128] - - [40, 42.968] - - - [40960, 25089, 1, 128] - - [22, 40.357] - - - [32000, 4096, 1, 128] - - [25, 45.811] - - - [41600, 25601, 1, 128] - - [22, 45.352] - - - [37504, 512, 1, 128] - - [57, 40.171] - - - [32128, 16129, 1, 128] - - [29, 45.863] - - - [37248, 21377, 1, 128] - - [38, 45.342] - - - [35840, 512, 1, 128] - - [49, 40.542] - - - [36096, 128, 1, 128] - - [36, 30.362] - - - [32512, 8192, 1, 128] - - [25, 46.813] - - - [36736, 8192, 1, 128] - - [29, 46.469] - - - [42880, 1024, 1, 128] - - [39, 43.034] - - - [44288, 8192, 1, 128] - - [27, 46.275] - - - [36224, 1024, 1, 128] - - [40, 42.95] - - - [41344, 25345, 1, 128] - - [29, 45.426] - - - [32384, 512, 1, 128] - - [35, 39.622] - - - [38272, 4096, 1, 128] - - [61, 41.461] - - - [37120, 2048, 1, 128] - - [23, 44.324] - - - [33152, 8192, 1, 128] - - [29, 46.687] - - - [36096, 4096, 1, 128] - - [76, 44.072] - - - [34560, 18689, 1, 128] - - [38, 45.867] - - - [36864, 4096, 1, 128] - - [25, 46.428] - - - [34944, 512, 1, 128] - - [37, 41.997] - - - [37760, 128, 1, 128] - - [34, 31.327] - - - [31616, 128, 1, 128] - - [39, 27.502] - - - [36224, 4096, 1, 128] - - [37, 45.663] - - - [40576, 24577, 1, 128] - - [22, 45.04] - - - [34688, 1024, 1, 128] - - [63, 43.274] - - - [40192, 1024, 1, 128] - - [23, 43.302] - - - [44672, 512, 1, 128] - - [27, 41.103] - - - [33664, 1024, 1, 128] - - [34, 43.004] - - - [39424, 512, 1, 128] - - [57, 41.0] - - - [44416, 1024, 1, 128] - - [40, 43.311] - - - [33408, 2048, 1, 128] - - [40, 44.337] - - - [43648, 8192, 1, 128] - - [37, 46.058] - - - [43520, 27649, 1, 128] - - [64, 45.797] - - - [40448, 1024, 1, 128] - - [63, 43.246] - - - [33152, 17153, 1, 128] - - [37, 46.124] - - - [33024, 512, 1, 128] - - [40, 40.032] - - - [39680, 128, 1, 128] - - [57, 32.482] - - - [29696, 4096, 1, 128] - - [37, 46.382] - - - [42112, 2048, 1, 128] - - [27, 44.11] - - - [38016, 8192, 1, 128] - - [25, 46.349] - - - [30464, 8192, 1, 128] - - [78, 44.92] - - - [43648, 128, 1, 128] - - [53, 31.115] - - - [32896, 16897, 1, 128] - - [74, 44.774] - - - [43008, 8192, 1, 128] - - [27, 47.387] - - - [34304, 512, 1, 128] - - [69, 41.954] - - - [38528, 128, 1, 128] - - [40, 31.686] - - - [41216, 1024, 1, 128] - - [72, 43.046] - - - [38272, 22401, 1, 128] - - [74, 40.701] - - - [34048, 4096, 1, 128] - - [35, 45.503] - - - [30720, 512, 1, 128] - - [60, 40.496] - - - [41728, 512, 1, 128] - - [29, 42.02] - - - [43136, 512, 1, 128] - - [60, 40.856] - - - [41088, 1024, 1, 128] - - [84, 42.006] - - - [33536, 1024, 1, 128] - - [27, 43.074] - - - [41088, 25089, 1, 128] - - [74, 44.401] - - - [36352, 20353, 1, 128] - - [22, 46.128] - - - [29184, 1024, 1, 128] - - [59, 42.608] - - - [44800, 128, 1, 128] - - [59, 32.154] - - - [41600, 8192, 1, 128] - - [22, 46.389] - - - [44416, 28545, 1, 128] - - [22, 45.558] - - - [34048, 512, 1, 128] - - [34, 41.814] - - - [32128, 16257, 1, 128] - - [37, 45.982] - - - [44288, 4096, 1, 128] - - [29, 45.46] - - - [34432, 18433, 1, 128] - - [74, 44.145] - - - [41856, 25857, 1, 128] - - [22, 45.643] - - - [32128, 2048, 1, 128] - - [58, 44.016] - - - [34688, 512, 1, 128] - - [63, 41.886] - - - [39936, 4096, 1, 128] - - [37, 46.381] - - - [38656, 1024, 1, 128] - - [37, 42.922] - - - [37760, 512, 1, 128] - - [34, 40.533] - - - [30336, 512, 1, 128] - - [35, 40.215] - - - [38016, 22017, 1, 128] - - [38, 45.593] - - - [44544, 4096, 1, 128] - - [50, 45.793] - - - [38912, 8192, 1, 128] - - [27, 47.448] - - - [39936, 128, 1, 128] - - [60, 32.476] - - - [36480, 2048, 1, 128] - - [40, 44.212] - - - [35200, 4096, 1, 128] - - [25, 45.526] - - - [30976, 14977, 1, 128] - - [74, 44.608] - - - [31104, 15105, 1, 128] - - [29, 45.786] - - - [40832, 1024, 1, 128] - - [37, 42.57] - - - [32384, 16513, 1, 128] - - [25, 46.085] - - - [43392, 4096, 1, 128] - - [38, 45.466] - - - [32768, 4096, 1, 128] - - [25, 35.663] - - - [38272, 22273, 1, 128] - - [78, 40.665] - - - [32128, 512, 1, 128] - - [34, 40.604] - - - [32896, 2048, 1, 128] - - [28, 41.278] - - - [37376, 21505, 1, 128] - - [37, 45.849] - - - [41856, 1024, 1, 128] - - [63, 43.584] - - - [33536, 8192, 1, 128] - - [64, 46.748] - - - [29568, 1024, 1, 128] - - [58, 41.624] - - - [44032, 28033, 1, 128] - - [38, 46.336] - - - [33280, 8192, 1, 128] - - [51, 47.149] - - - [39296, 4096, 1, 128] - - [25, 45.716] - - - [30592, 14593, 1, 128] - - [29, 45.761] - - - [37504, 8192, 1, 128] - - [61, 45.639] - - - [30336, 14465, 1, 128] - - [29, 45.698] - - - [29952, 2048, 1, 128] - - [23, 44.232] - - - [40832, 512, 1, 128] - - [44, 40.792] - - - [44672, 28673, 1, 128] - - [22, 45.372] - - - [30080, 4096, 1, 128] - - [78, 43.362] - - - [37888, 2048, 1, 128] - - [61, 44.739] - - - [37632, 21761, 1, 128] - - [22, 45.775] - - - [29824, 8192, 1, 128] - - [74, 45.814] - - - [35328, 19457, 1, 128] - - [64, 46.076] - - - [37376, 4096, 1, 128] - - [35, 45.946] - - - [33792, 17921, 1, 128] - - [37, 46.337] - - - [34304, 8192, 1, 128] - - [47, 47.014] - - - [42752, 512, 1, 128] - - [40, 40.648] - - - [36992, 2048, 1, 128] - - [39, 43.993] - - - [39168, 4096, 1, 128] - - [64, 45.615] - - - [31360, 15489, 1, 128] - - [37, 45.835] - - - [43520, 8192, 1, 128] - - [25, 46.897] - - - [30080, 2048, 1, 128] - - [73, 41.788] - - - [30720, 4096, 1, 128] - - [25, 46.713] - - - [34176, 128, 1, 128] - - [63, 29.275] - - - [32768, 16769, 1, 128] - - [37, 34.684] - - - [35072, 128, 1, 128] - - [35, 29.938] - - - [35712, 4096, 1, 128] - - [38, 45.542] - - - [36480, 4096, 1, 128] - - [37, 45.435] - - - [39424, 4096, 1, 128] - - [27, 45.949] - - - [38400, 128, 1, 128] - - [25, 31.61] - - - [34432, 2048, 1, 128] - - [59, 42.941] - - - [41344, 512, 1, 128] - - [49, 41.676] - - - [35200, 512, 1, 128] - - [23, 41.728] - - - [39936, 8192, 1, 128] - - [29, 47.064] - - - [31488, 128, 1, 128] - - [35, 27.718] - - - [43008, 512, 1, 128] - - [60, 41.471] - - - [33024, 4096, 1, 128] - - [25, 46.064] - - - [36608, 512, 1, 128] - - [49, 40.617] - - - [37376, 8192, 1, 128] - - [37, 46.86] - - - [29824, 13825, 1, 128] - - [50, 45.018] - - - [36352, 2048, 1, 128] - - [40, 44.472] - - - [30336, 1024, 1, 128] - - [35, 42.605] - - - [44416, 28417, 1, 128] - - [38, 45.544] - - - [38144, 22273, 1, 128] - - [38, 45.787] - - - [28928, 2048, 1, 128] - - [60, 43.673] - - - [29568, 13697, 1, 128] - - [38, 45.321] - - - [43136, 27137, 1, 128] - - [22, 45.344] - - - [42112, 4096, 1, 128] - - [37, 45.686] - - - [40960, 512, 1, 128] - - [44, 40.455] - - - [35584, 1024, 1, 128] - - [63, 43.056] - - - [31232, 15361, 1, 128] - - [25, 45.928] - - - [40960, 8192, 1, 128] - - [25, 41.571] - - - [31232, 1024, 1, 128] - - [34, 43.531] - - - [29312, 512, 1, 128] - - [40, 39.334] - - - [44416, 512, 1, 128] - - [35, 41.196] - - - [42240, 512, 1, 128] - - [53, 41.921] - - - [31232, 8192, 1, 128] - - [27, 46.787] - - - [35072, 19201, 1, 128] - - [37, 45.897] - - - [29568, 128, 1, 128] - - [35, 26.027] - - - [33792, 2048, 1, 128] - - [53, 44.646] - - - [35712, 2048, 1, 128] - - [63, 44.276] - - - [40576, 128, 1, 128] - - [63, 32.708] - - - [40704, 1024, 1, 128] - - [34, 43.238] - - - [29824, 1024, 1, 128] - - [31, 42.016] - - - [33536, 17665, 1, 128] - - [64, 46.186] - - - [43008, 27009, 1, 128] - - [22, 46.557] - - - [34304, 2048, 1, 128] - - [72, 44.427] - - - [37120, 21249, 1, 128] - - [29, 45.742] - - - [41600, 1024, 1, 128] - - [25, 43.256] - - - [33024, 1024, 1, 128] - - [34, 42.574] - - - [42368, 512, 1, 128] - - [57, 40.466] - - - [30592, 14721, 1, 128] - - [38, 45.745] - - - [29696, 2048, 1, 128] - - [44, 44.525] - - - [31232, 128, 1, 128] - - [40, 27.421] - - - [38784, 22785, 1, 128] - - [22, 45.55] - - - [32896, 1024, 1, 128] - - [97, 37.593] - - - [32128, 128, 1, 128] - - [35, 27.849] - - - [35968, 8192, 1, 128] - - [37, 46.444] - - - [38400, 2048, 1, 128] - - [56, 44.604] - - - [36864, 2048, 1, 128] - - [27, 44.373] - - - [31616, 4096, 1, 128] - - [64, 45.071] - - - [34688, 18817, 1, 128] - - [22, 45.573] - - - [42624, 4096, 1, 128] - - [29, 41.862] - - - [29312, 1024, 1, 128] - - [34, 42.221] - - - [37760, 2048, 1, 128] - - [58, 44.239] - - - [39808, 512, 1, 128] - - [44, 40.895] - - - [41472, 128, 1, 128] - - [40, 33.401] - - - [32128, 4096, 1, 128] - - [37, 45.8] - - - [43520, 4096, 1, 128] - - [27, 46.119] - - - [41472, 512, 1, 128] - - [25, 42.267] - - - [38912, 22913, 1, 128] - - [38, 46.708] - - - [30464, 1024, 1, 128] - - [27, 42.124] - - - [33280, 128, 1, 128] - - [40, 28.747] - - - [31872, 15873, 1, 128] - - [38, 45.676] - - - [36352, 4096, 1, 128] - - [25, 45.942] - - - [30720, 2048, 1, 128] - - [37, 44.652] - - - [33792, 128, 1, 128] - - [63, 29.019] - - - [36096, 8192, 1, 128] - - [59, 45.087] - - - [38784, 128, 1, 128] - - [23, 32.111] - - - [30208, 2048, 1, 128] - - [39, 44.35] - - - [34432, 4096, 1, 128] - - [50, 44.362] - - - [42880, 128, 1, 128] - - [31, 30.95] - - - [31616, 15745, 1, 128] - - [22, 45.341] - - - [40960, 2048, 1, 128] - - [22, 39.906] - - - [41344, 128, 1, 128] - - [23, 33.298] - - - [41728, 25857, 1, 128] - - [64, 45.145] - - - [32896, 512, 1, 128] - - [71, 32.523] - - - [41728, 2048, 1, 128] - - [59, 43.251] - - - [42368, 26369, 1, 128] - - [47, 45.288] - - - [30720, 14721, 1, 128] - - [29, 46.809] - - - [37376, 512, 1, 128] - - [40, 41.061] - - - [35456, 19457, 1, 128] - - [27, 45.548] - - - [29184, 13185, 1, 128] - - [22, 45.997] - - - [34944, 128, 1, 128] - - [72, 29.692] - - - [36608, 20609, 1, 128] - - [27, 45.82] - - - [35584, 19585, 1, 128] - - [38, 45.685] - - - [42880, 8192, 1, 128] - - [47, 45.728] - - - [39936, 1024, 1, 128] - - [63, 43.517] - - - [34944, 19073, 1, 128] - - [29, 45.75] - - - [32512, 128, 1, 128] - - [63, 28.309] - - - [40064, 512, 1, 128] - - [27, 41.616] - - - [30464, 2048, 1, 128] - - [51, 42.134] - - - [30592, 8192, 1, 128] - - [27, 46.473] - - - [39040, 512, 1, 128] - - [34, 40.947] - - - [41088, 128, 1, 128] - - [79, 33.019] - - - [29824, 128, 1, 128] - - [60, 26.31] - - - [32384, 128, 1, 128] - - [39, 28.038] - - - [41728, 25729, 1, 128] - - [50, 45.1] - - - [30976, 4096, 1, 128] - - [51, 44.37] - - - [42624, 128, 1, 128] - - [79, 30.795] - - - [42112, 512, 1, 128] - - [29, 41.873] - - - [38784, 2048, 1, 128] - - [60, 44.143] - - - [35200, 8192, 1, 128] - - [27, 46.357] - - - [30976, 128, 1, 128] - - [35, 26.972] - - - [32640, 16641, 1, 128] - - [27, 45.367] - - - [41984, 8192, 1, 128] - - [27, 47.088] - - - [30080, 128, 1, 128] - - [40, 26.605] - - - [35584, 512, 1, 128] - - [69, 39.6] - - - [44800, 2048, 1, 128] - - [38, 44.034] - - - [34048, 128, 1, 128] - - [40, 29.335] - - - [35712, 1024, 1, 128] - - [49, 42.79] - - - [43136, 128, 1, 128] - - [33, 31.165] - - - [33280, 1024, 1, 128] - - [31, 42.848] - - - [34816, 18945, 1, 128] - - [22, 46.764] - - - [40704, 8192, 1, 128] - - [37, 46.688] - - - [34304, 128, 1, 128] - - [44, 29.487] - - - [39936, 512, 1, 128] - - [34, 41.614] - - - [36096, 2048, 1, 128] - - [78, 42.932] - - - [40832, 8192, 1, 128] - - [29, 46.421] - - - [37760, 4096, 1, 128] - - [37, 45.537] - - - [36736, 512, 1, 128] - - [55, 40.645] - - - [31744, 8192, 1, 128] - - [27, 47.209] - - - [33920, 1024, 1, 128] - - [63, 42.96] - - - [39808, 128, 1, 128] - - [33, 32.336] - - - [36608, 2048, 1, 128] - - [60, 44.236] - - - [30464, 14593, 1, 128] - - [78, 42.867] - - - [35200, 19201, 1, 128] - - [38, 45.538] - - - [41472, 1024, 1, 128] - - [27, 43.428] - - - [30720, 128, 1, 128] - - [40, 26.818] - - - [41600, 128, 1, 128] - - [59, 33.352] - - - [38144, 22145, 1, 128] - - [22, 45.754] - - - [37120, 4096, 1, 128] - - [38, 45.686] - - - [40704, 24705, 1, 128] - - [22, 45.747] - - - [41088, 25217, 1, 128] - - [38, 44.43] - - - [43776, 8192, 1, 128] - - [61, 45.549] - - - [38912, 1024, 1, 128] - - [34, 43.79] - - - [43008, 2048, 1, 128] - - [25, 44.887] - - - [42496, 26497, 1, 128] - - [64, 46.082] - - - [33536, 512, 1, 128] - - [55, 41.487] - - - [43520, 512, 1, 128] - - [35, 41.357] - - - [39040, 23169, 1, 128] - - [22, 45.446] - - - [29568, 2048, 1, 128] - - [40, 43.692] - - - [44672, 8192, 1, 128] - - [27, 46.483] - - - [29824, 512, 1, 128] - - [60, 39.696] - - - [34944, 2048, 1, 128] - - [40, 44.314] - - - [33408, 4096, 1, 128] - - [25, 45.857] - - - [41600, 25729, 1, 128] - - [38, 45.529] - - - [40832, 2048, 1, 128] - - [27, 44.139] - - - [38912, 128, 1, 128] - - [40, 32.145] - - - [34048, 2048, 1, 128] - - [22, 43.336] - - - [43904, 2048, 1, 128] - - [53, 44.341] - - - [39296, 23297, 1, 128] - - [22, 45.524] - - - [31232, 4096, 1, 128] - - [37, 45.845] - - - [35840, 1024, 1, 128] - - [23, 43.597] - - - [28928, 128, 1, 128] - - [40, 25.434] - - - [42752, 2048, 1, 128] - - [38, 44.089] - - - [44032, 1024, 1, 128] - - [23, 43.694] - - - [29440, 13569, 1, 128] - - [25, 45.99] - - - [35456, 19585, 1, 128] - - [29, 45.703] - - - [35840, 19841, 1, 128] - - [37, 46.395] - - - [31360, 128, 1, 128] - - [36, 27.409] - - - [40192, 2048, 1, 128] - - [60, 44.606] - - - [33920, 8192, 1, 128] - - [27, 46.43] - - - [43648, 512, 1, 128] - - [25, 40.198] - - - [30080, 14209, 1, 128] - - [45, 43.283] - - - [39680, 23809, 1, 128] - - [22, 45.422] - - - [32512, 512, 1, 128] - - [72, 39.476] - - - [34816, 2048, 1, 128] - - [25, 44.568] - - - [43392, 1024, 1, 128] - - [55, 42.75] - - - [39040, 4096, 1, 128] - - [27, 45.454] - - - [43264, 4096, 1, 128] - - [29, 45.807] - - - [44416, 2048, 1, 128] - - [36, 44.36] - - - [31488, 512, 1, 128] - - [37, 40.673] - - - [31616, 1024, 1, 128] - - [27, 43.131] - - - [44032, 8192, 1, 128] - - [27, 47.089] - - - [39424, 23553, 1, 128] - - [38, 45.76] - - - [31360, 8192, 1, 128] - - [27, 46.446] - - - [42752, 128, 1, 128] - - [23, 31.193] - - - [40192, 512, 1, 128] - - [27, 41.544] - - - [36096, 20225, 1, 128] - - [47, 44.722] - - - [41984, 26113, 1, 128] - - [22, 46.343] - - - [39936, 2048, 1, 128] - - [58, 44.806] - - - [42880, 2048, 1, 128] - - [33, 43.617] - - - [29440, 128, 1, 128] - - [34, 26.096] - - - [40192, 128, 1, 128] - - [35, 32.619] - - - [36608, 4096, 1, 128] - - [37, 45.643] - - - [37760, 21761, 1, 128] - - [22, 45.57] - - - [44160, 28161, 1, 128] - - [29, 45.034] - - - [44288, 512, 1, 128] - - [49, 40.758] - - - [29056, 13185, 1, 128] - - [29, 45.782] - - - [43904, 512, 1, 128] - - [55, 40.392] - - - [29696, 128, 1, 128] - - [35, 26.197] - - - [36224, 8192, 1, 128] - - [27, 46.49] - - - [33024, 2048, 1, 128] - - [40, 44.345] - - - [44032, 28161, 1, 128] - - [38, 46.258] - - - [44032, 128, 1, 128] - - [37, 31.413] - - - [38784, 512, 1, 128] - - [72, 40.788] - - - [29056, 8192, 1, 128] - - [37, 46.411] - - - [33920, 18049, 1, 128] - - [38, 45.703] - - - [34816, 1024, 1, 128] - - [35, 43.646] - - - [29056, 128, 1, 128] - - [40, 25.546] - - - [39552, 1024, 1, 128] - - [23, 43.178] - - - [36992, 8192, 1, 128] - - [37, 46.344] - - - [44544, 1024, 1, 128] - - [40, 43.492] - - - [43904, 27905, 1, 128] - - [25, 45.391] - - - [29440, 512, 1, 128] - - [35, 39.452] - - - [29568, 8192, 1, 128] - - [29, 46.022] - - - [41472, 2048, 1, 128] - - [72, 44.493] - - - [29184, 8192, 1, 128] - - [35, 46.616] - - - [33408, 512, 1, 128] - - [60, 40.96] - - - [38656, 22785, 1, 128] - - [64, 45.782] - - - [31744, 15745, 1, 128] - - [37, 46.536] - - - [38656, 2048, 1, 128] - - [63, 44.176] - - - [30080, 8192, 1, 128] - - [74, 44.644] - - - [44672, 128, 1, 128] - - [25, 31.601] - - - [40704, 24833, 1, 128] - - [50, 45.754] - - - [33792, 8192, 1, 128] - - [27, 47.181] - - - [33920, 512, 1, 128] - - [35, 41.639] - - - [40576, 1024, 1, 128] - - [53, 43.164] - - - [36224, 20225, 1, 128] - - [38, 45.701] - - - [34432, 1024, 1, 128] - - [27, 42.711] - - - [31488, 15617, 1, 128] - - [27, 45.941] - - - [40576, 2048, 1, 128] - - [53, 44.334] - - - [30208, 512, 1, 128] - - [25, 40.117] - - - [36480, 128, 1, 128] - - [60, 30.685] - - - [37504, 21633, 1, 128] - - [30, 44.759] - - - [32896, 17025, 1, 128] - - [74, 44.944] - - - [39168, 2048, 1, 128] - - [23, 44.037] - - - [29440, 2048, 1, 128] - - [55, 43.835] - - - [29440, 13441, 1, 128] - - [27, 45.952] - - - [32640, 8192, 1, 128] - - [37, 46.225] - - - [35072, 19073, 1, 128] - - [38, 45.943] - - - [33152, 512, 1, 128] - - [40, 40.22] - - - [40576, 8192, 1, 128] - - [29, 46.372] - - - [34944, 8192, 1, 128] - - [27, 46.479] - - - [38656, 128, 1, 128] - - [89, 31.82] - - - [33536, 17537, 1, 128] - - [50, 46.122] - - - [29952, 512, 1, 128] - - [34, 39.867] - - - [31488, 2048, 1, 128] - - [40, 44.242] - - - [31872, 4096, 1, 128] - - [38, 45.435] - - - [31232, 15233, 1, 128] - - [50, 46.235] - - - [38912, 23041, 1, 128] - - [22, 46.582] - - - [31232, 2048, 1, 128] - - [69, 44.357] - - - [40448, 8192, 1, 128] - - [47, 46.84] - - - [36352, 128, 1, 128] - - [36, 30.612] - - - [43776, 4096, 1, 128] - - [61, 44.709] - - - [32000, 8192, 1, 128] - - [29, 46.687] - - - [37760, 8192, 1, 128] - - [25, 46.459] - - - [30080, 1024, 1, 128] - - [31, 41.65] - - - [44544, 128, 1, 128] - - [57, 32.088] - - - [29696, 1024, 1, 128] - - [80, 42.71] - - - [32640, 512, 1, 128] - - [25, 39.477] - - - [44416, 128, 1, 128] - - [57, 31.571] - - - [41216, 512, 1, 128] - - [59, 41.238] - - - [31872, 512, 1, 128] - - [40, 40.288] - - - [34432, 512, 1, 128] - - [27, 41.713] - - - [34560, 1024, 1, 128] - - [40, 43.417] - - - [42240, 128, 1, 128] - - [72, 33.682] - - - [44288, 28289, 1, 128] - - [22, 45.366] - - - [30336, 14337, 1, 128] - - [47, 45.421] - - - [32384, 2048, 1, 128] - - [63, 44.101] - - - [38400, 22401, 1, 128] - - [64, 46.128] - - - [39296, 1024, 1, 128] - - [34, 43.051] - - - [28928, 8192, 1, 128] - - [27, 46.561] - - - [40320, 2048, 1, 128] - - [63, 44.54] - - - [31104, 15233, 1, 128] - - [25, 45.866] - - - [39680, 512, 1, 128] - - [44, 41.26] - - - [34048, 18049, 1, 128] - - [64, 45.789] - - - [30720, 1024, 1, 128] - - [35, 43.411] - - - [42880, 26881, 1, 128] - - [64, 44.892] - - - [32896, 8192, 1, 128] - - [74, 45.437] - - - [43264, 8192, 1, 128] - - [37, 46.531] - - - [37632, 4096, 1, 128] - - [22, 45.643] - - - [32256, 4096, 1, 128] - - [57, 46.088] - - - [37248, 4096, 1, 128] - - [27, 45.415] - - - [33280, 17409, 1, 128] - - [64, 46.304] - - - [36096, 512, 1, 128] - - [44, 39.045] - - - [37120, 21121, 1, 128] - - [27, 45.711] - - - [32896, 128, 1, 128] - - [23, 27.925] - - - [36352, 20481, 1, 128] - - [38, 45.734] - - - [43392, 2048, 1, 128] - - [37, 43.776] - - - [36352, 512, 1, 128] - - [60, 40.948] - - - [29056, 13057, 1, 128] - - [22, 45.668] - - - [29056, 4096, 1, 128] - - [55, 45.545] - - - [37888, 4096, 1, 128] - - [37, 46.418] - - - [40320, 512, 1, 128] - - [34, 41.566] - - - [39168, 128, 1, 128] - - [34, 32.284] - - - [41472, 8192, 1, 128] - - [51, 46.812] - - - [34560, 512, 1, 128] - - [60, 42.049] - - - [34176, 18305, 1, 128] - - [51, 45.623] - - - [34688, 8192, 1, 128] - - [29, 46.367] - - - [29696, 13825, 1, 128] - - [37, 46.514] - - - [33152, 17281, 1, 128] - - [37, 46.072] - - - [30208, 8192, 1, 128] - - [51, 46.932] - - - [43648, 27649, 1, 128] - - [38, 44.631] - - - [31360, 2048, 1, 128] - - [60, 43.773] - - - [41984, 128, 1, 128] - - [25, 33.557] - - - [38528, 2048, 1, 128] - - [23, 43.961] - - - [32256, 16385, 1, 128] - - [30, 46.375] - - - [42240, 1024, 1, 128] - - [60, 43.243] - - - [32000, 16001, 1, 128] - - [22, 46.01] - - - [37248, 1024, 1, 128] - - [27, 42.966] - - - [32256, 1024, 1, 128] - - [25, 43.039] - - - [39296, 23425, 1, 128] - - [22, 45.506] - - - [43008, 4096, 1, 128] - - [25, 46.736] - - - [31104, 128, 1, 128] - - [60, 27.18] - - - [38656, 8192, 1, 128] - - [64, 46.561] - - - [44288, 128, 1, 128] - - [51, 31.718] - - - [38528, 22657, 1, 128] - - [22, 45.514] - - - [39552, 128, 1, 128] - - [23, 32.199] - - - [37376, 21377, 1, 128] - - [38, 46.05] - - - [28928, 13057, 1, 128] - - [27, 45.96] - - - [43264, 27265, 1, 128] - - [22, 45.734] - - - [35328, 4096, 1, 128] - - [64, 46.012] - - - [30848, 4096, 1, 128] - - [35, 45.587] - - - [44800, 28801, 1, 128] - - [38, 45.329] - - - [35456, 512, 1, 128] - - [80, 39.962] - - - [40960, 24961, 1, 128] - - [22, 40.336] - - - [39936, 23937, 1, 128] - - [22, 46.264] - - - [31744, 1024, 1, 128] - - [44, 43.136] - - - [32128, 8192, 1, 128] - - [22, 46.465] - - - [42112, 26113, 1, 128] - - [22, 45.455] - - - [31744, 2048, 1, 128] - - [36, 44.56] - - - [42112, 1024, 1, 128] - - [23, 43.21] - - - [40064, 8192, 1, 128] - - [22, 45.948] - - - [38144, 128, 1, 128] - - [40, 31.54] - - - [42624, 2048, 1, 128] - - [27, 40.989] - - - [36992, 128, 1, 128] - - [35, 30.794] - - - [40192, 8192, 1, 128] - - [29, 46.619] - - - [40064, 24065, 1, 128] - - [50, 44.533] - - - [37760, 21889, 1, 128] - - [22, 45.545] - - - [36352, 8192, 1, 128] - - [25, 46.787] - - - [44544, 512, 1, 128] - - [57, 41.173] - - - [32384, 4096, 1, 128] - - [27, 45.721] - - - [39168, 23169, 1, 128] - - [50, 45.761] - - - [1408, 897, 1, 128] - - [123, 29.971] - - - [16512, 512, 1, 128] - - [98, 30.672] - - - [20480, 12673, 1, 128] - - [29, 47.23] - - - [20992, 512, 1, 128] - - [27, 41.196] - - - [9344, 512, 1, 128] - - [34, 32.285] - - - [18048, 2048, 1, 128] - - [25, 42.348] - - - [20352, 12673, 1, 128] - - [29, 45.916] - - - [640, 128, 1, 128] - - [128, 3.662] - - - [28160, 512, 1, 128] - - [67, 41.669] - - - [20608, 4096, 1, 128] - - [27, 44.987] - - - [19328, 1024, 1, 128] - - [34, 41.669] - - - [26496, 4096, 1, 128] - - [67, 44.965] - - - [10624, 512, 1, 128] - - [49, 32.382] - - - [20352, 1024, 1, 128] - - [29, 42.206] - - - [10240, 6529, 1, 128] - - [37, 45.357] - - - [22144, 14465, 1, 128] - - [27, 45.898] - - - [13184, 2048, 1, 128] - - [40, 41.458] - - - [14720, 6913, 1, 128] - - [25, 44.874] - - - [21248, 512, 1, 128] - - [55, 36.614] - - - [10496, 128, 1, 128] - - [114, 31.751] - - - [13056, 5377, 1, 128] - - [25, 44.678] - - - [10880, 128, 1, 128] - - [114, 29.524] - - - [18688, 512, 1, 128] - - [34, 37.935] - - - [22656, 4096, 1, 128] - - [55, 45.081] - - - [15232, 1024, 1, 128] - - [37, 40.493] - - - [20224, 4096, 1, 128] - - [35, 45.205] - - - [6016, 2305, 1, 128] - - [29, 37.786] - - - [13184, 4096, 1, 128] - - [36, 43.952] - - - [256, 129, 1, 128] - - [170, 1.465] - - - [11264, 7553, 1, 128] - - [37, 45.632] - - - [18176, 128, 1, 128] - - [33, 26.582] - - - [15872, 8193, 1, 128] - - [37, 45.629] - - - [26112, 4096, 1, 128] - - [57, 45.885] - - - [22784, 2048, 1, 128] - - [53, 43.471] - - - [10880, 7297, 1, 128] - - [37, 44.655] - - - [14720, 2048, 1, 128] - - [23, 42.96] - - - [9216, 5633, 1, 128] - - [27, 44.514] - - - [23040, 15233, 1, 128] - - [37, 46.577] - - - [8832, 5121, 1, 128] - - [29, 43.392] - - - [18816, 1024, 1, 128] - - [31, 40.938] - - - [128, 129, 1, 128] - - [119, 0.738] - - - [15488, 512, 1, 128] - - [35, 35.007] - - - [18176, 1024, 1, 128] - - [29, 40.995] - - - [16128, 8449, 1, 128] - - [27, 45.56] - - - [16000, 2048, 1, 128] - - [35, 42.288] - - - [24960, 9089, 1, 128] - - [37, 45.317] - - - [14336, 1024, 1, 128] - - [29, 39.739] - - - [25472, 8192, 1, 128] - - [25, 46.584] - - - [23040, 128, 1, 128] - - [23, 31.621] - - - [9472, 512, 1, 128] - - [27, 33.193] - - - [19072, 128, 1, 128] - - [57, 27.345] - - - [10624, 6913, 1, 128] - - [37, 44.484] - - - [7808, 1024, 1, 128] - - [29, 35.397] - - - [27008, 11137, 1, 128] - - [50, 45.459] - - - [21504, 4096, 1, 128] - - [29, 46.162] - - - [7936, 1024, 1, 128] - - [55, 36.263] - - - [12928, 5121, 1, 128] - - [53, 43.389] - - - [26240, 8192, 1, 128] - - [29, 42.362] - - - [18304, 2048, 1, 128] - - [63, 43.223] - - - [24576, 1024, 1, 128] - - [27, 41.641] - - - [10624, 128, 1, 128] - - [109, 28.619] - - - [24576, 128, 1, 128] - - [34, 33.175] - - - [25600, 9601, 1, 128] - - [38, 46.181] - - - [5248, 128, 1, 128] - - [165, 22.135] - - - [24448, 4096, 1, 128] - - [35, 45.432] - - - [19328, 128, 1, 128] - - [34, 27.717] - - - [24064, 512, 1, 128] - - [60, 39.773] - - - [11136, 512, 1, 128] - - [36, 33.263] - - - [14592, 1024, 1, 128] - - [29, 39.377] - - - [12544, 4737, 1, 128] - - [27, 44.272] - - - [17280, 128, 1, 128] - - [63, 25.266] - - - [25344, 8192, 1, 128] - - [64, 46.021] - - - [4608, 512, 1, 128] - - [39, 26.791] - - - [4608, 128, 1, 128] - - [116, 20.246] - - - [21760, 512, 1, 128] - - [34, 37.072] - - - [7936, 128, 1, 128] - - [108, 26.151] - - - [11008, 7425, 1, 128] - - [85, 44.734] - - - [13824, 2048, 1, 128] - - [40, 43.654] - - - [18048, 512, 1, 128] - - [27, 37.199] - - - [19584, 11905, 1, 128] - - [25, 45.771] - - - [22656, 512, 1, 128] - - [40, 38.172] - - - [4608, 3073, 1, 128] - - [35, 38.869] - - - [5504, 128, 1, 128] - - [123, 22.823] - - - [4864, 1024, 1, 128] - - [27, 33.684] - - - [17664, 1024, 1, 128] - - [27, 40.529] - - - [18176, 2048, 1, 128] - - [40, 43.601] - - - [2048, 1537, 1, 128] - - [59, 33.885] - - - [22528, 128, 1, 128] - - [23, 30.976] - - - [21760, 13953, 1, 128] - - [22, 46.306] - - - [7040, 128, 1, 128] - - [110, 27.349] - - - [3328, 1665, 1, 128] - - [27, 32.65] - - - [768, 512, 1, 128] - - [119, 16.196] - - - [21504, 13697, 1, 128] - - [22, 46.635] - - - [18560, 10881, 1, 128] - - [29, 45.473] - - - [2560, 128, 1, 128] - - [116, 13.997] - - - [15616, 1024, 1, 128] - - [49, 40.824] - - - [19456, 4096, 1, 128] - - [27, 45.916] - - - [25600, 2048, 1, 128] - - [28, 44.245] - - - [2304, 128, 1, 128] - - [150, 12.504] - - - [1664, 1025, 1, 128] - - [34, 19.797] - - - [23168, 15361, 1, 128] - - [25, 45.518] - - - [9856, 128, 1, 128] - - [108, 30.957] - - - [13312, 2048, 1, 128] - - [25, 42.713] - - - [19200, 512, 1, 128] - - [35, 38.611] - - - [19200, 2048, 1, 128] - - [23, 43.454] - - - [23168, 2048, 1, 128] - - [53, 43.985] - - - [18688, 128, 1, 128] - - [44, 27.163] - - - [13568, 1024, 1, 128] - - [27, 41.514] - - - [17792, 9985, 1, 128] - - [50, 45.306] - - - [20608, 1024, 1, 128] - - [47, 42.014] - - - [11648, 8065, 1, 128] - - [37, 44.991] - - - [1280, 128, 1, 128] - - [113, 7.324] - - - [16256, 4096, 1, 128] - - [29, 44.813] - - - [17024, 1024, 1, 128] - - [25, 41.744] - - - [19456, 128, 1, 128] - - [34, 28.119] - - - [20736, 512, 1, 128] - - [53, 40.399] - - - [14464, 6785, 1, 128] - - [25, 44.794] - - - [20736, 13057, 1, 128] - - [37, 46.071] - - - [8704, 2048, 1, 128] - - [27, 42.448] - - - [640, 512, 1, 128] - - [113, 13.792] - - - [768, 129, 1, 128] - - [170, 4.395] - - - [27776, 1024, 1, 128] - - [49, 42.75] - - - [19200, 11521, 1, 128] - - [27, 45.806] - - - [6400, 2048, 1, 128] - - [27, 40.695] - - - [14976, 7297, 1, 128] - - [27, 44.964] - - - [7040, 2048, 1, 128] - - [27, 42.133] - - - [25984, 128, 1, 128] - - [34, 34.207] - - - [13696, 128, 1, 128] - - [114, 33.364] - - - [2688, 1153, 1, 128] - - [80, 32.625] - - - [15232, 2048, 1, 128] - - [29, 42.804] - - - [11776, 128, 1, 128] - - [171, 29.869] - - - [3328, 512, 1, 128] - - [123, 32.644] - - - [11648, 7937, 1, 128] - - [25, 44.948] - - - [19456, 2048, 1, 128] - - [23, 43.9] - - - [11008, 128, 1, 128] - - [109, 29.228] - - - [9984, 6401, 1, 128] - - [27, 44.572] - - - [25856, 9857, 1, 128] - - [37, 45.618] - - - [4224, 512, 1, 128] - - [57, 24.753] - - - [13568, 5761, 1, 128] - - [25, 44.662] - - - [5632, 2049, 1, 128] - - [59, 37.024] - - - [8832, 2048, 1, 128] - - [25, 40.482] - - - [5632, 3969, 1, 128] - - [59, 41.798] - - - [25856, 2048, 1, 128] - - [40, 43.873] - - - [25472, 2048, 1, 128] - - [23, 43.855] - - - [20736, 12929, 1, 128] - - [27, 46.039] - - - [14592, 128, 1, 128] - - [33, 22.261] - - - [1792, 512, 1, 128] - - [109, 24.27] - - - [14208, 2048, 1, 128] - - [40, 42.396] - - - [15360, 7681, 1, 128] - - [37, 45.649] - - - [5760, 2048, 1, 128] - - [37, 38.982] - - - [6400, 512, 1, 128] - - [35, 34.368] - - - [5248, 3713, 1, 128] - - [34, 40.632] - - - [16768, 1024, 1, 128] - - [53, 40.445] - - - [10752, 512, 1, 128] - - [60, 32.745] - - - [26624, 2048, 1, 128] - - [38, 44.32] - - - [384, 128, 1, 128] - - [116, 2.232] - - - [27392, 8192, 1, 128] - - [74, 45.808] - - - [24448, 512, 1, 128] - - [35, 38.266] - - - [11136, 7553, 1, 128] - - [29, 44.898] - - - [17024, 9345, 1, 128] - - [27, 45.64] - - - [16000, 8193, 1, 128] - - [27, 45.091] - - - [5888, 2048, 1, 128] - - [36, 39.591] - - - [18304, 10497, 1, 128] - - [25, 45.496] - - - [3968, 128, 1, 128] - - [113, 17.751] - - - [14336, 6529, 1, 128] - - [29, 45.676] - - - [19840, 128, 1, 128] - - [57, 28.341] - - - [25600, 8192, 1, 128] - - [29, 47.314] - - - [18688, 11009, 1, 128] - - [25, 45.921] - - - [7680, 1024, 1, 128] - - [25, 35.722] - - - [7168, 128, 1, 128] - - [172, 24.049] - - - [1664, 512, 1, 128] - - [108, 26.701] - - - [12544, 1024, 1, 128] - - [27, 39.443] - - - [6528, 2048, 1, 128] - - [29, 40.762] - - - [19072, 4096, 1, 128] - - [34, 45.091] - - - [2048, 512, 1, 128] - - [173, 27.115] - - - [13568, 5889, 1, 128] - - [27, 44.448] - - - [23680, 16001, 1, 128] - - [64, 45.198] - - - [26112, 10113, 1, 128] - - [50, 46.101] - - - [15872, 128, 1, 128] - - [57, 23.922] - - - [16384, 512, 1, 128] - - [29, 33.327] - - - [9856, 6273, 1, 128] - - [37, 42.549] - - - [26368, 1024, 1, 128] - - [60, 42.303] - - - [16256, 2048, 1, 128] - - [27, 42.32] - - - [3968, 2305, 1, 128] - - [34, 36.793] - - - [28672, 8192, 1, 128] - - [25, 47.54] - - - [10368, 1024, 1, 128] - - [34, 40.775] - - - [11008, 1024, 1, 128] - - [68, 35.775] - - - [11776, 4097, 1, 128] - - [31, 43.237] - - - [26496, 2048, 1, 128] - - [39, 43.274] - - - [17792, 4096, 1, 128] - - [57, 45.005] - - - [2304, 512, 1, 128] - - [109, 29.196] - - - [9216, 2048, 1, 128] - - [27, 41.282] - - - [12416, 512, 1, 128] - - [80, 36.201] - - - [18048, 128, 1, 128] - - [59, 26.342] - - - [21888, 14209, 1, 128] - - [76, 41.982] - - - [9344, 5761, 1, 128] - - [25, 44.358] - - - [19712, 2048, 1, 128] - - [41, 41.116] - - - [12288, 1024, 1, 128] - - [25, 39.436] - - - [3584, 1921, 1, 128] - - [49, 33.069] - - - [22784, 128, 1, 128] - - [58, 31.335] - - - [26880, 128, 1, 128] - - [63, 35.57] - - - [17408, 1024, 1, 128] - - [27, 41.997] - - - [15488, 4096, 1, 128] - - [25, 44.919] - - - [13312, 5633, 1, 128] - - [29, 45.186] - - - [22016, 14337, 1, 128] - - [30, 46.472] - - - [19328, 2048, 1, 128] - - [63, 43.728] - - - [25600, 128, 1, 128] - - [60, 33.998] - - - [22784, 15105, 1, 128] - - [64, 46.187] - - - [5376, 3713, 1, 128] - - [34, 41.376] - - - [14208, 512, 1, 128] - - [53, 32.868] - - - [12928, 4096, 1, 128] - - [34, 44.633] - - - [768, 257, 1, 128] - - [119, 8.688] - - - [27776, 11777, 1, 128] - - [37, 45.416] - - - [12032, 1024, 1, 128] - - [34, 40.109] - - - [14208, 4096, 1, 128] - - [25, 44.998] - - - [19840, 12161, 1, 128] - - [37, 45.78] - - - [17536, 512, 1, 128] - - [53, 38.659] - - - [19840, 4096, 1, 128] - - [34, 45.206] - - - [26624, 512, 1, 128] - - [34, 40.688] - - - [27136, 11137, 1, 128] - - [30, 46.29] - - - [11008, 512, 1, 128] - - [34, 32.65] - - - [1024, 513, 1, 128] - - [166, 18.031] - - - [15744, 512, 1, 128] - - [53, 35.825] - - - [22016, 128, 1, 128] - - [81, 30.621] - - - [9344, 1024, 1, 128] - - [27, 38.357] - - - [28544, 1024, 1, 128] - - [29, 42.309] - - - [13440, 5633, 1, 128] - - [50, 43.699] - - - [21632, 13825, 1, 128] - - [38, 45.844] - - - [24064, 4096, 1, 128] - - [30, 45.78] - - - [24192, 512, 1, 128] - - [23, 39.572] - - - [22912, 15233, 1, 128] - - [27, 46.056] - - - [20864, 13185, 1, 128] - - [29, 45.554] - - - [8064, 4353, 1, 128] - - [27, 42.586] - - - [8704, 5121, 1, 128] - - [37, 43.442] - - - [19840, 1024, 1, 128] - - [25, 41.559] - - - [15616, 128, 1, 128] - - [40, 23.393] - - - [21632, 512, 1, 128] - - [35, 36.974] - - - [13440, 512, 1, 128] - - [34, 38.697] - - - [23936, 128, 1, 128] - - [34, 32.37] - - - [8960, 5377, 1, 128] - - [27, 43.666] - - - [27008, 512, 1, 128] - - [34, 40.97] - - - [13440, 5761, 1, 128] - - [57, 43.687] - - - [3072, 512, 1, 128] - - [174, 30.642] - - - [4096, 1024, 1, 128] - - [57, 29.651] - - - [7296, 3585, 1, 128] - - [35, 41.586] - - - [12416, 4737, 1, 128] - - [31, 43.615] - - - [6912, 512, 1, 128] - - [34, 36.398] - - - [11136, 2048, 1, 128] - - [40, 41.832] - - - [18176, 10369, 1, 128] - - [25, 45.848] - - - [14976, 4096, 1, 128] - - [25, 44.963] - - - [19712, 4096, 1, 128] - - [40, 43.38] - - - [8064, 1024, 1, 128] - - [25, 34.668] - - - [9600, 128, 1, 128] - - [123, 30.281] - - - [26240, 1024, 1, 128] - - [27, 41.079] - - - [5248, 3585, 1, 128] - - [29, 40.924] - - - [16768, 2048, 1, 128] - - [40, 42.748] - - - [13184, 128, 1, 128] - - [114, 32.655] - - - [19328, 11521, 1, 128] - - [37, 45.479] - - - [4864, 512, 1, 128] - - [39, 27.847] - - - [3584, 2049, 1, 128] - - [27, 33.674] - - - [18560, 128, 1, 128] - - [34, 26.93] - - - [27392, 11393, 1, 128] - - [25, 44.508] - - - [27520, 512, 1, 128] - - [72, 40.386] - - - [18176, 4096, 1, 128] - - [37, 45.274] - - - [7808, 4225, 1, 128] - - [49, 42.956] - - - [15232, 128, 1, 128] - - [72, 22.86] - - - [25728, 1024, 1, 128] - - [34, 42.049] - - - [23936, 512, 1, 128] - - [34, 39.341] - - - [23424, 2048, 1, 128] - - [32, 42.136] - - - [28032, 12161, 1, 128] - - [25, 45.603] - - - [27136, 512, 1, 128] - - [31, 41.025] - - - [14336, 6657, 1, 128] - - [22, 45.789] - - - [15616, 4096, 1, 128] - - [37, 45.338] - - - [3328, 1793, 1, 128] - - [25, 34.56] - - - [28416, 512, 1, 128] - - [76, 38.809] - - - [16384, 8705, 1, 128] - - [29, 37.424] - - - [3200, 1537, 1, 128] - - [37, 32.96] - - - [26368, 128, 1, 128] - - [63, 35.08] - - - [16000, 512, 1, 128] - - [69, 34.729] - - - [25216, 9345, 1, 128] - - [27, 45.479] - - - [28288, 4096, 1, 128] - - [22, 45.457] - - - [24832, 512, 1, 128] - - [69, 38.5] - - - [18048, 10369, 1, 128] - - [24, 44.334] - - - [20480, 4096, 1, 128] - - [34, 46.324] - - - [17792, 10113, 1, 128] - - [24, 45.432] - - - [13312, 5505, 1, 128] - - [29, 45.025] - - - [17024, 2048, 1, 128] - - [44, 43.476] - - - [20608, 12929, 1, 128] - - [29, 45.702] - - - [16896, 4096, 1, 128] - - [57, 45.64] - - - [27776, 2048, 1, 128] - - [63, 43.876] - - - [6912, 3201, 1, 128] - - [49, 41.76] - - - [15744, 2048, 1, 128] - - [60, 43.428] - - - [24448, 128, 1, 128] - - [40, 32.828] - - - [2688, 128, 1, 128] - - [116, 14.172] - - - [7808, 2048, 1, 128] - - [27, 40.69] - - - [1408, 512, 1, 128] - - [123, 23.224] - - - [12032, 512, 1, 128] - - [34, 35.58] - - - [26752, 512, 1, 128] - - [34, 40.792] - - - [16128, 8321, 1, 128] - - [29, 45.623] - - - [25856, 128, 1, 128] - - [59, 34.214] - - - [24064, 8192, 1, 128] - - [57, 47.018] - - - [28160, 4096, 1, 128] - - [57, 45.418] - - - [13312, 128, 1, 128] - - [114, 31.595] - - - [10112, 6401, 1, 128] - - [31, 43.844] - - - [16384, 4096, 1, 128] - - [37, 38.021] - - - [16512, 2048, 1, 128] - - [82, 39.086] - - - [27520, 11521, 1, 128] - - [27, 45.627] - - - [8192, 4481, 1, 128] - - [35, 42.962] - - - [16768, 512, 1, 128] - - [27, 35.966] - - - [6144, 128, 1, 128] - - [108, 25.055] - - - [13568, 512, 1, 128] - - [34, 38.952] - - - [9344, 5633, 1, 128] - - [34, 44.287] - - - [13440, 4096, 1, 128] - - [25, 44.134] - - - [2176, 1665, 1, 128] - - [25, 25.457] - - - [28288, 128, 1, 128] - - [40, 25.262] - - - [11776, 4096, 1, 128] - - [59, 44.562] - - - [17280, 512, 1, 128] - - [35, 38.266] - - - [5504, 3841, 1, 128] - - [99, 36.733] - - - [14848, 7041, 1, 128] - - [25, 45.163] - - - [3584, 128, 1, 128] - - [113, 16.958] - - - [26880, 8192, 1, 128] - - [37, 46.482] - - - [2944, 1409, 1, 128] - - [35, 28.732] - - - [26368, 10369, 1, 128] - - [37, 45.792] - - - [21888, 512, 1, 128] - - [58, 35.508] - - - [15872, 2048, 1, 128] - - [44, 43.266] - - - [20224, 512, 1, 128] - - [27, 40.45] - - - [24320, 8449, 1, 128] - - [37, 45.716] - - - [5632, 1024, 1, 128] - - [27, 34.129] - - - [17152, 9473, 1, 128] - - [27, 45.713] - - - [4096, 128, 1, 128] - - [166, 18.547] - - - [8832, 128, 1, 128] - - [171, 28.344] - - - [2048, 1409, 1, 128] - - [57, 31.533] - - - [28160, 12289, 1, 128] - - [51, 45.864] - - - [9088, 5505, 1, 128] - - [37, 43.99] - - - [19200, 1024, 1, 128] - - [27, 41.453] - - - [18048, 4096, 1, 128] - - [37, 44.168] - - - [12928, 512, 1, 128] - - [44, 37.398] - - - [20864, 4096, 1, 128] - - [55, 45.018] - - - [27008, 2048, 1, 128] - - [40, 43.353] - - - [16640, 128, 1, 128] - - [72, 24.927] - - - [24960, 8192, 1, 128] - - [27, 46.265] - - - [24320, 1024, 1, 128] - - [37, 42.944] - - - [23552, 15873, 1, 128] - - [25, 46.69] - - - [26240, 4096, 1, 128] - - [27, 42.717] - - - [24320, 128, 1, 128] - - [57, 32.949] - - - [26240, 128, 1, 128] - - [23, 34.73] - - - [3200, 1665, 1, 128] - - [27, 31.363] - - - [11776, 2048, 1, 128] - - [37, 42.593] - - - [6144, 512, 1, 128] - - [34, 33.481] - - - [24960, 128, 1, 128] - - [23, 33.334] - - - [23424, 128, 1, 128] - - [72, 31.852] - - - [11776, 8065, 1, 128] - - [67, 45.272] - - - [19072, 11265, 1, 128] - - [25, 45.403] - - - [8192, 4609, 1, 128] - - [34, 43.552] - - - [21888, 4096, 1, 128] - - [74, 40.382] - - - [14976, 2048, 1, 128] - - [44, 42.838] - - - [23680, 4096, 1, 128] - - [50, 44.576] - - - [14080, 1024, 1, 128] - - [27, 42.357] - - - [19968, 4096, 1, 128] - - [55, 45.484] - - - [8704, 128, 1, 128] - - [110, 28.178] - - - [23424, 15745, 1, 128] - - [29, 43.83] - - - [8320, 2048, 1, 128] - - [36, 40.584] - - - [6144, 2433, 1, 128] - - [29, 39.949] - - - [19200, 11393, 1, 128] - - [37, 45.778] - - - [28416, 128, 1, 128] - - [63, 25.438] - - - [14080, 2048, 1, 128] - - [27, 42.359] - - - [12544, 4096, 1, 128] - - [25, 45.013] - - - [17024, 128, 1, 128] - - [47, 24.991] - - - [23936, 16257, 1, 128] - - [27, 46.159] - - - [12288, 128, 1, 128] - - [123, 30.85] - - - [28800, 1024, 1, 128] - - [67, 41.985] - - - [13824, 6017, 1, 128] - - [37, 45.055] - - - [23040, 2048, 1, 128] - - [36, 43.938] - - - [9984, 6273, 1, 128] - - [27, 44.695] - - - [23680, 512, 1, 128] - - [34, 39.204] - - - [7936, 4353, 1, 128] - - [25, 43.529] - - - [24192, 2048, 1, 128] - - [36, 44.024] - - - [8448, 512, 1, 128] - - [31, 29.71] - - - [5760, 2177, 1, 128] - - [35, 38.6] - - - [22656, 14977, 1, 128] - - [22, 45.731] - - - [17024, 4096, 1, 128] - - [29, 45.364] - - - [24960, 8961, 1, 128] - - [25, 45.169] - - - [5888, 1024, 1, 128] - - [40, 34.886] - - - [9344, 2048, 1, 128] - - [27, 41.138] - - - [11520, 1024, 1, 128] - - [27, 38.952] - - - [17024, 9217, 1, 128] - - [29, 45.351] - - - [10368, 6657, 1, 128] - - [27, 44.392] - - - [21632, 2048, 1, 128] - - [53, 43.672] - - - [26880, 2048, 1, 128] - - [53, 43.869] - - - [20736, 4096, 1, 128] - - [27, 45.481] - - - [26624, 8192, 1, 128] - - [25, 47.587] - - - [26752, 2048, 1, 128] - - [40, 43.698] - - - [24192, 8321, 1, 128] - - [37, 45.195] - - - [4736, 1024, 1, 128] - - [27, 33.114] - - - [27648, 8192, 1, 128] - - [27, 47.306] - - - [27392, 11521, 1, 128] - - [74, 44.51] - - - [27776, 4096, 1, 128] - - [35, 45.284] - - - [28672, 12801, 1, 128] - - [25, 46.648] - - - [13056, 512, 1, 128] - - [63, 37.842] - - - [25088, 2048, 1, 128] - - [69, 43.926] - - - [17408, 9601, 1, 128] - - [37, 46.144] - - - [5120, 3585, 1, 128] - - [34, 40.938] - - - [13824, 512, 1, 128] - - [34, 39.335] - - - [8576, 1024, 1, 128] - - [83, 38.069] - - - [16768, 4096, 1, 128] - - [35, 45.208] - - - [25728, 9729, 1, 128] - - [27, 45.32] - - - [27392, 512, 1, 128] - - [69, 38.691] - - - [13824, 128, 1, 128] - - [114, 32.393] - - - [27264, 1024, 1, 128] - - [35, 41.537] - - - [22272, 14465, 1, 128] - - [30, 46.223] - - - [19840, 2048, 1, 128] - - [53, 43.568] - - - [18176, 10497, 1, 128] - - [25, 45.819] - - - [4992, 3329, 1, 128] - - [49, 41.312] - - - [14976, 7169, 1, 128] - - [29, 44.407] - - - [10112, 512, 1, 128] - - [58, 34.534] - - - [24704, 128, 1, 128] - - [27, 33.286] - - - [16896, 128, 1, 128] - - [69, 25.26] - - - [10880, 7169, 1, 128] - - [27, 44.091] - - - [9600, 512, 1, 128] - - [49, 33.131] - - - [22528, 1024, 1, 128] - - [27, 42.125] - - - [27008, 128, 1, 128] - - [34, 35.055] - - - [4480, 2945, 1, 128] - - [34, 40.291] - - - [15872, 8065, 1, 128] - - [25, 45.84] - - - [28672, 128, 1, 128] - - [40, 25.507] - - - [9344, 128, 1, 128] - - [110, 29.6] - - - [15360, 2048, 1, 128] - - [53, 43.411] - - - [11392, 512, 1, 128] - - [34, 33.918] - - - [9216, 128, 1, 128] - - [172, 28.702] - - - [8192, 2048, 1, 128] - - [37, 40.844] - - - [14464, 1024, 1, 128] - - [25, 39.175] - - - [4096, 2433, 1, 128] - - [59, 39.249] - - - [6528, 2945, 1, 128] - - [37, 40.713] - - - [12672, 512, 1, 128] - - [78, 36.955] - - - [26624, 128, 1, 128] - - [23, 35.168] - - - [19712, 1024, 1, 128] - - [49, 40.481] - - - [4480, 2817, 1, 128] - - [35, 39.015] - - - [13440, 2048, 1, 128] - - [55, 42.115] - - - [256, 257, 1, 128] - - [166, 2.987] - - - [16000, 128, 1, 128] - - [57, 23.867] - - - [7552, 3969, 1, 128] - - [49, 42.069] - - - [12416, 2048, 1, 128] - - [27, 41.554] - - - [18432, 512, 1, 128] - - [49, 37.578] - - - [14464, 512, 1, 128] - - [34, 33.638] - - - [1280, 769, 1, 128] - - [123, 25.116] - - - [14976, 512, 1, 128] - - [36, 34.617] - - - [28032, 4096, 1, 128] - - [49, 45.397] - - - [27904, 128, 1, 128] - - [34, 36.155] - - - [20224, 12545, 1, 128] - - [27, 45.854] - - - [15872, 4096, 1, 128] - - [37, 45.739] - - - [3456, 1793, 1, 128] - - [34, 35.322] - - - [14336, 128, 1, 128] - - [31, 21.74] - - - [21248, 2048, 1, 128] - - [60, 43.649] - - - [23040, 1024, 1, 128] - - [31, 41.885] - - - [15232, 7425, 1, 128] - - [22, 44.631] - - - [14592, 512, 1, 128] - - [34, 33.889] - - - [22912, 15105, 1, 128] - - [29, 45.856] - - - [22528, 2048, 1, 128] - - [27, 44.374] - - - [3072, 1024, 1, 128] - - [57, 32.874] - - - [17536, 4096, 1, 128] - - [25, 45.101] - - - [384, 257, 1, 128] - - [122, 4.378] - - - [14464, 6657, 1, 128] - - [29, 44.75] - - - [20096, 1024, 1, 128] - - [27, 41.66] - - - [26880, 4096, 1, 128] - - [25, 45.686] - - - [18816, 2048, 1, 128] - - [58, 43.333] - - - [17152, 512, 1, 128] - - [25, 38.061] - - - [18432, 4096, 1, 128] - - [27, 46.42] - - - [10368, 2048, 1, 128] - - [40, 42.664] - - - [1408, 769, 1, 128] - - [114, 27.266] - - - [7168, 2048, 1, 128] - - [55, 39.766] - - - [17664, 128, 1, 128] - - [59, 26.138] - - - [1152, 513, 1, 128] - - [115, 19.814] - - - [7296, 3713, 1, 128] - - [25, 42.233] - - - [24064, 2048, 1, 128] - - [72, 43.879] - - - [8576, 2048, 1, 128] - - [40, 42.014] - - - [23168, 15489, 1, 128] - - [29, 45.702] - - - [14848, 7169, 1, 128] - - [27, 44.814] - - - [2432, 512, 1, 128] - - [175, 30.686] - - - [19712, 12033, 1, 128] - - [50, 44.208] - - - [25856, 4096, 1, 128] - - [25, 45.649] - - - [17152, 9345, 1, 128] - - [30, 45.62] - - - [3712, 128, 1, 128] - - [165, 16.913] - - - [22272, 128, 1, 128] - - [59, 30.86] - - - [25600, 9729, 1, 128] - - [27, 46.199] - - - [6016, 2433, 1, 128] - - [34, 39.116] - - - [12928, 128, 1, 128] - - [114, 32.348] - - - [25088, 8192, 1, 128] - - [64, 46.92] - - - [7040, 1024, 1, 128] - - [35, 39.056] - - - [4736, 3201, 1, 128] - - [57, 39.823] - - - [16000, 1024, 1, 128] - - [37, 39.786] - - - [1920, 512, 1, 128] - - [123, 25.535] - - - [8192, 1024, 1, 128] - - [34, 36.996] - - - [8448, 4865, 1, 128] - - [27, 43.535] - - - [11136, 7425, 1, 128] - - [37, 45.079] - - - [23296, 4096, 1, 128] - - [27, 45.463] - - - [27904, 2048, 1, 128] - - [58, 43.926] - - - [23552, 4096, 1, 128] - - [37, 46.138] - - - [24960, 2048, 1, 128] - - [58, 44.025] - - - [2816, 128, 1, 128] - - [128, 14.953] - - - [7424, 3841, 1, 128] - - [37, 42.525] - - - [20480, 128, 1, 128] - - [23, 29.193] - - - [18816, 11137, 1, 128] - - [25, 45.576] - - - [26496, 128, 1, 128] - - [34, 34.688] - - - [16896, 9217, 1, 128] - - [25, 45.874] - - - [23296, 512, 1, 128] - - [35, 38.895] - - - [8064, 2048, 1, 128] - - [25, 40.104] - - - [19968, 128, 1, 128] - - [57, 28.741] - - - [8320, 4737, 1, 128] - - [75, 42.111] - - - [27648, 1024, 1, 128] - - [35, 43.172] - - - [3712, 512, 1, 128] - - [35, 22.02] - - - [256, 128, 1, 128] - - [166, 1.5] - - - [3072, 1537, 1, 128] - - [35, 31.708] - - - [5504, 1024, 1, 128] - - [27, 32.134] - - - [20992, 2048, 1, 128] - - [63, 44.058] - - - [20480, 1024, 1, 128] - - [34, 41.912] - - - [20864, 128, 1, 128] - - [72, 29.177] - - - [28544, 12545, 1, 128] - - [37, 45.641] - - - [1152, 512, 1, 128] - - [166, 20.49] - - - [24320, 8321, 1, 128] - - [27, 45.636] - - - [2688, 512, 1, 128] - - [125, 29.178] - - - [27904, 8192, 1, 128] - - [50, 46.305] - - - [3840, 2177, 1, 128] - - [37, 36.982] - - - [25344, 128, 1, 128] - - [69, 33.786] - - - [13184, 512, 1, 128] - - [40, 38.064] - - - [7680, 512, 1, 128] - - [27, 27.702] - - - [11904, 2048, 1, 128] - - [44, 42.903] - - - [12544, 512, 1, 128] - - [36, 36.617] - - - [8448, 4737, 1, 128] - - [27, 43.198] - - - [28544, 128, 1, 128] - - [63, 25.216] - - - [21760, 14081, 1, 128] - - [27, 46.293] - - - [12800, 128, 1, 128] - - [121, 31.388] - - - [17664, 4096, 1, 128] - - [34, 45.009] - - - [2432, 1793, 1, 128] - - [49, 29.769] - - - [16384, 8577, 1, 128] - - [29, 37.574] - - - [28544, 512, 1, 128] - - [59, 38.922] - - - [28032, 12033, 1, 128] - - [37, 45.692] - - - [4864, 3329, 1, 128] - - [29, 40.718] - - - [12928, 5249, 1, 128] - - [37, 44.052] - - - [4736, 512, 1, 128] - - [33, 27.214] - - - [27264, 2048, 1, 128] - - [37, 42.91] - - - [19840, 12033, 1, 128] - - [27, 45.797] - - - [19584, 4096, 1, 128] - - [27, 45.273] - - - [21376, 4096, 1, 128] - - [37, 45.323] - - - [20352, 4096, 1, 128] - - [55, 45.354] - - - [6400, 2689, 1, 128] - - [35, 41.995] - - - [24704, 8192, 1, 128] - - [61, 45.87] - - - [22528, 14849, 1, 128] - - [27, 46.927] - - - [18304, 512, 1, 128] - - [35, 37.603] - - - [6656, 1024, 1, 128] - - [34, 38.433] - - - [13568, 4096, 1, 128] - - [29, 45.258] - - - [6016, 512, 1, 128] - - [84, 32.844] - - - [17664, 2048, 1, 128] - - [29, 42.904] - - - [17408, 512, 1, 128] - - [60, 38.114] - - - [24960, 4096, 1, 128] - - [37, 45.456] - - - [20608, 12801, 1, 128] - - [37, 45.592] - - - [27648, 11649, 1, 128] - - [25, 46.394] - - - [5760, 128, 1, 128] - - [109, 23.233] - - - [17792, 512, 1, 128] - - [40, 36.774] - - - [17664, 512, 1, 128] - - [36, 37.061] - - - [19968, 12161, 1, 128] - - [25, 46.15] - - - [19840, 512, 1, 128] - - [37, 39.521] - - - [12032, 4353, 1, 128] - - [27, 43.908] - - - [25984, 512, 1, 128] - - [78, 40.063] - - - [27648, 4096, 1, 128] - - [37, 46.345] - - - [10752, 7041, 1, 128] - - [25, 45.241] - - - [28544, 2048, 1, 128] - - [44, 43.938] - - - [7680, 2048, 1, 128] - - [37, 41.027] - - - [13184, 5377, 1, 128] - - [31, 43.357] - - - [6784, 3201, 1, 128] - - [35, 40.973] - - - [16384, 2048, 1, 128] - - [25, 36.708] - - - [22656, 1024, 1, 128] - - [27, 41.7] - - - [12800, 512, 1, 128] - - [40, 37.247] - - - [23936, 1024, 1, 128] - - [27, 42.686] - - - [15360, 1024, 1, 128] - - [27, 40.76] - - - [15488, 2048, 1, 128] - - [60, 43.276] - - - [11392, 1024, 1, 128] - - [27, 38.453] - - - [15744, 1024, 1, 128] - - [34, 41.287] - - - [9856, 2048, 1, 128] - - [25, 41.072] - - - [5888, 2305, 1, 128] - - [55, 40.7] - - - [10496, 512, 1, 128] - - [84, 35.51] - - - [1664, 1153, 1, 128] - - [59, 22.184] - - - [3456, 1024, 1, 128] - - [35, 36.005] - - - [20992, 13313, 1, 128] - - [29, 46.058] - - - [11904, 4096, 1, 128] - - [27, 44.407] - - - [13056, 1024, 1, 128] - - [27, 40.26] - - - [12800, 2048, 1, 128] - - [69, 42.43] - - - [12160, 512, 1, 128] - - [53, 36.03] - - - [5760, 2049, 1, 128] - - [23, 38.035] - - - [11392, 128, 1, 128] - - [114, 30.138] - - - [5632, 128, 1, 128] - - [109, 22.716] - - - [11520, 2048, 1, 128] - - [37, 41.898] - - - [11648, 2048, 1, 128] - - [53, 42.355] - - - [28544, 8192, 1, 128] - - [37, 46.393] - - - [22912, 1024, 1, 128] - - [27, 42.084] - - - [10752, 7169, 1, 128] - - [25, 45.01] - - - [8320, 128, 1, 128] - - [109, 26.935] - - - [23808, 1024, 1, 128] - - [55, 42.655] - - - [25984, 8192, 1, 128] - - [24, 46.384] - - - [22656, 2048, 1, 128] - - [58, 43.771] - - - [7296, 1024, 1, 128] - - [25, 33.942] - - - [28032, 512, 1, 128] - - [29, 41.815] - - - [22400, 2048, 1, 128] - - [44, 43.599] - - - [22144, 512, 1, 128] - - [23, 37.331] - - - [13312, 4096, 1, 128] - - [25, 45.594] - - - [10240, 2048, 1, 128] - - [37, 42.314] - - - [12672, 128, 1, 128] - - [114, 32.477] - - - [10752, 2048, 1, 128] - - [60, 42.028] - - - [1152, 128, 1, 128] - - [119, 6.643] - - - [13696, 5889, 1, 128] - - [49, 42.592] - - - [9216, 1024, 1, 128] - - [35, 37.865] - - - [17152, 128, 1, 128] - - [40, 25.431] - - - [24320, 2048, 1, 128] - - [40, 44.178] - - - [16512, 8705, 1, 128] - - [27, 44.284] - - - [3072, 1409, 1, 128] - - [29, 29.52] - - - [1024, 128, 1, 128] - - [116, 5.905] - - - [22400, 14593, 1, 128] - - [22, 45.652] - - - [4096, 512, 1, 128] - - [33, 24.897] - - - [4992, 128, 1, 128] - - [113, 21.548] - - - [9472, 5889, 1, 128] - - [37, 44.292] - - - [9472, 5761, 1, 128] - - [25, 43.936] - - - [27136, 1024, 1, 128] - - [31, 43.113] - - - [6528, 1024, 1, 128] - - [35, 37.555] - - - [25472, 1024, 1, 128] - - [40, 42.1] - - - [5120, 512, 1, 128] - - [76, 29.255] - - - [5504, 512, 1, 128] - - [34, 29.719] - - - [21120, 13441, 1, 128] - - [50, 44.906] - - - [4352, 128, 1, 128] - - [113, 19.236] - - - [8832, 5249, 1, 128] - - [25, 43.648] - - - [1536, 1025, 1, 128] - - [172, 29.099] - - - [11520, 512, 1, 128] - - [84, 34.41] - - - [5632, 2048, 1, 128] - - [24, 38.174] - - - [7424, 128, 1, 128] - - [110, 24.684] - - - [18432, 128, 1, 128] - - [58, 27.171] - - - [12672, 2048, 1, 128] - - [29, 42.183] - - - [14208, 128, 1, 128] - - [40, 21.502] - - - [15360, 7553, 1, 128] - - [27, 45.889] - - - [26496, 1024, 1, 128] - - [35, 42.264] - - - [27136, 128, 1, 128] - - [76, 35.338] - - - [12032, 2048, 1, 128] - - [58, 42.76] - - - [11648, 1024, 1, 128] - - [37, 39.161] - - - [11776, 512, 1, 128] - - [34, 35.139] - - - [1024, 512, 1, 128] - - [132, 18.435] - - - [11264, 7681, 1, 128] - - [27, 45.447] - - - [19456, 11777, 1, 128] - - [37, 46.512] - - - [14080, 4096, 1, 128] - - [37, 43.534] - - - [7040, 3329, 1, 128] - - [35, 41.787] - - - [27392, 4096, 1, 128] - - [76, 44.661] - - - [14720, 7041, 1, 128] - - [29, 44.828] - - - [19584, 1024, 1, 128] - - [36, 41.301] - - - [21376, 13569, 1, 128] - - [29, 45.75] - - - [20480, 12801, 1, 128] - - [29, 47.065] - - - [21248, 128, 1, 128] - - [40, 29.72] - - - [9728, 1024, 1, 128] - - [34, 39.467] - - - [18688, 10881, 1, 128] - - [29, 45.714] - - - [21120, 13313, 1, 128] - - [51, 44.673] - - - [20096, 2048, 1, 128] - - [40, 43.167] - - - [16640, 4096, 1, 128] - - [25, 45.585] - - - [28160, 12161, 1, 128] - - [50, 45.981] - - - [640, 129, 1, 128] - - [166, 3.662] - - - [28672, 512, 1, 128] - - [37, 39.061] - - - [12416, 4096, 1, 128] - - [25, 44.607] - - - [25344, 9473, 1, 128] - - [57, 44.899] - - - [18304, 1024, 1, 128] - - [23, 40.763] - - - [25600, 4096, 1, 128] - - [37, 46.394] - - - [22272, 512, 1, 128] - - [60, 37.987] - - - [21504, 13825, 1, 128] - - [38, 46.611] - - - [4736, 128, 1, 128] - - [116, 20.684] - - - [26496, 10625, 1, 128] - - [64, 45.111] - - - [7040, 512, 1, 128] - - [35, 36.099] - - - [14336, 4096, 1, 128] - - [27, 46.045] - - - [9216, 512, 1, 128] - - [40, 32.251] - - - [1280, 641, 1, 128] - - [108, 24.901] - - - [16768, 8961, 1, 128] - - [29, 45.569] - - - [18944, 11137, 1, 128] - - [50, 46.171] - - - [21504, 2048, 1, 128] - - [61, 44.054] - - - [21888, 1024, 1, 128] - - [34, 37.4] - - - [11264, 512, 1, 128] - - [27, 34.059] - - - [27776, 8192, 1, 128] - - [25, 46.27] - - - [10368, 6785, 1, 128] - - [35, 44.491] - - - [18432, 10753, 1, 128] - - [37, 46.711] - - - [19968, 2048, 1, 128] - - [63, 43.687] - - - [16640, 512, 1, 128] - - [84, 35.332] - - - [24576, 8577, 1, 128] - - [38, 42.526] - - - [28672, 2048, 1, 128] - - [29, 43.883] - - - [11136, 128, 1, 128] - - [108, 29.674] - - - [12288, 4609, 1, 128] - - [25, 44.544] - - - [14848, 1024, 1, 128] - - [27, 40.215] - - - [14848, 128, 1, 128] - - [35, 22.334] - - - [7424, 1024, 1, 128] - - [25, 34.422] - - - [2560, 1024, 1, 128] - - [59, 29.199] - - - [6400, 128, 1, 128] - - [123, 25.674] - - - [15488, 7809, 1, 128] - - [29, 45.064] - - - [17920, 2048, 1, 128] - - [69, 43.624] - - - [5760, 512, 1, 128] - - [35, 31.739] - - - [16640, 1024, 1, 128] - - [27, 41.308] - - - [28160, 2048, 1, 128] - - [23, 44.07] - - - [5504, 3969, 1, 128] - - [68, 37.291] - - - [11776, 1024, 1, 128] - - [27, 39.659] - - - [18816, 128, 1, 128] - - [40, 27.194] - - - [27904, 12033, 1, 128] - - [50, 45.586] - - - [11520, 7937, 1, 128] - - [27, 44.746] - - - [18944, 11265, 1, 128] - - [64, 46.055] - - - [5376, 1024, 1, 128] - - [25, 32.116] - - - [12032, 4225, 1, 128] - - [25, 43.658] - - - [5376, 128, 1, 128] - - [108, 22.045] - - - [9856, 1024, 1, 128] - - [25, 39.055] - - - [26752, 10881, 1, 128] - - [27, 45.393] - - - [20352, 128, 1, 128] - - [59, 28.794] - - - [14464, 128, 1, 128] - - [57, 21.619] - - - [1024, 385, 1, 128] - - [122, 15.36] - - - [3840, 128, 1, 128] - - [115, 17.389] - - - [24192, 128, 1, 128] - - [60, 32.783] - - - [28544, 12673, 1, 128] - - [27, 45.789] - - - [1664, 128, 1, 128] - - [115, 9.305] - - - [26752, 8192, 1, 128] - - [22, 46.339] - - - [16896, 1024, 1, 128] - - [57, 41.586] - - - [9728, 128, 1, 128] - - [109, 30.297] - - - [11264, 2048, 1, 128] - - [29, 42.067] - - - [11392, 2048, 1, 128] - - [27, 41.914] - - - [20224, 2048, 1, 128] - - [40, 43.332] - - - [26880, 1024, 1, 128] - - [58, 42.478] - - - [15104, 512, 1, 128] - - [25, 34.75] - - - [26368, 2048, 1, 128] - - [25, 43.697] - - - [6784, 3073, 1, 128] - - [55, 40.31] - - - [23168, 128, 1, 128] - - [23, 31.797] - - - [8448, 1024, 1, 128] - - [67, 37.138] - - - [16896, 9089, 1, 128] - - [24, 46.11] - - - [17536, 128, 1, 128] - - [34, 25.85] - - - [22912, 512, 1, 128] - - [36, 38.232] - - - [28032, 128, 1, 128] - - [23, 36.13] - - - [19584, 512, 1, 128] - - [58, 39.162] - - - [27136, 11265, 1, 128] - - [24, 46.115] - - - [4992, 512, 1, 128] - - [33, 28.409] - - - [8448, 128, 1, 128] - - [171, 27.47] - - - [27648, 128, 1, 128] - - [23, 35.822] - - - [16640, 2048, 1, 128] - - [36, 42.998] - - - [26752, 10753, 1, 128] - - [38, 45.477] - - - [2944, 1281, 1, 128] - - [35, 26.563] - - - [5376, 3841, 1, 128] - - [27, 42.384] - - - [10496, 6913, 1, 128] - - [29, 44.723] - - - [17024, 512, 1, 128] - - [35, 37.559] - - - [11008, 7297, 1, 128] - - [100, 44.623] - - - [14080, 128, 1, 128] - - [60, 25.612] - - - [5888, 512, 1, 128] - - [34, 32.264] - - - [19200, 128, 1, 128] - - [34, 27.968] - - - [14208, 6529, 1, 128] - - [27, 44.534] - - - [22912, 4096, 1, 128] - - [34, 45.339] - - - [14336, 2048, 1, 128] - - [37, 43.096] - - - [17792, 128, 1, 128] - - [33, 26.123] - - - [22656, 14849, 1, 128] - - [22, 45.629] - - - [19712, 512, 1, 128] - - [27, 39.479] - - - [5248, 1024, 1, 128] - - [25, 35.108] - - - [3712, 2049, 1, 128] - - [27, 34.365] - - - [24448, 8449, 1, 128] - - [27, 45.255] - - - [8192, 512, 1, 128] - - [49, 28.879] - - - [25472, 4096, 1, 128] - - [25, 45.651] - - - [25088, 512, 1, 128] - - [27, 39.464] - - - [23168, 1024, 1, 128] - - [35, 42.044] - - - [24320, 8192, 1, 128] - - [57, 46.734] - - - [24192, 8192, 1, 128] - - [25, 46.474] - - - [2176, 512, 1, 128] - - [121, 28.055] - - - [4992, 3457, 1, 128] - - [35, 41.978] - - - [896, 257, 1, 128] - - [176, 9.692] - - - [28288, 1024, 1, 128] - - [37, 42.197] - - - [20864, 1024, 1, 128] - - [37, 42.728] - - - [18432, 2048, 1, 128] - - [25, 43.921] - - - [17280, 9601, 1, 128] - - [29, 45.382] - - - [18944, 4096, 1, 128] - - [31, 45.601] - - - [13440, 128, 1, 128] - - [172, 32.958] - - - [7424, 2048, 1, 128] - - [27, 40.427] - - - [768, 128, 1, 128] - - [116, 4.429] - - - [16128, 512, 1, 128] - - [57, 34.593] - - - [28288, 12289, 1, 128] - - [29, 45.397] - - - [23552, 128, 1, 128] - - [53, 31.909] - - - [24832, 8192, 1, 128] - - [59, 46.933] - - - [10240, 1024, 1, 128] - - [25, 40.434] - - - [8960, 2048, 1, 128] - - [49, 40.69] - - - [17664, 9985, 1, 128] - - [29, 45.395] - - - [25088, 4096, 1, 128] - - [67, 45.937] - - - [7552, 2048, 1, 128] - - [60, 40.591] - - - [15104, 7297, 1, 128] - - [25, 45.225] - - - [7168, 1024, 1, 128] - - [27, 33.982] - - - [26112, 8192, 1, 128] - - [57, 47.055] - - - [24192, 1024, 1, 128] - - [27, 42.824] - - - [22912, 2048, 1, 128] - - [35, 43.139] - - - [10368, 512, 1, 128] - - [36, 35.36] - - - [22528, 4096, 1, 128] - - [29, 46.425] - - - [6528, 128, 1, 128] - - [109, 26.046] - - - [26752, 4096, 1, 128] - - [38, 45.179] - - - [2816, 512, 1, 128] - - [108, 30.344] - - - [22016, 14209, 1, 128] - - [24, 46.54] - - - [8832, 1024, 1, 128] - - [80, 36.792] - - - [16384, 128, 1, 128] - - [60, 24.795] - - - [5120, 1024, 1, 128] - - [25, 34.644] - - - [24832, 8833, 1, 128] - - [50, 45.868] - - - [11520, 128, 1, 128] - - [121, 30.26] - - - [24960, 512, 1, 128] - - [25, 38.841] - - - [27520, 2048, 1, 128] - - [63, 43.992] - - - [22272, 14593, 1, 128] - - [50, 45.954] - - - [2048, 128, 1, 128] - - [116, 11.281] - - - [2176, 1537, 1, 128] - - [35, 25.126] - - - [10496, 1024, 1, 128] - - [25, 41.279] - - - [12160, 4353, 1, 128] - - [27, 43.944] - - - [6144, 1024, 1, 128] - - [25, 35.581] - - - [26752, 1024, 1, 128] - - [40, 42.239] - - - [17280, 4096, 1, 128] - - [34, 45.135] - - - [16896, 512, 1, 128] - - [72, 36.62] - - - [4480, 128, 1, 128] - - [116, 19.801] - - - [18944, 128, 1, 128] - - [56, 27.644] - - - [9600, 2048, 1, 128] - - [27, 41.598] - - - [19456, 1024, 1, 128] - - [37, 41.595] - - - [9984, 2048, 1, 128] - - [25, 42.255] - - - [25216, 9217, 1, 128] - - [22, 45.153] - - - [19968, 1024, 1, 128] - - [29, 42.029] - - - [13952, 2048, 1, 128] - - [63, 43.274] - - - [10496, 2048, 1, 128] - - [29, 42.812] - - - [12672, 1024, 1, 128] - - [27, 39.816] - - - [19072, 11393, 1, 128] - - [38, 45.652] - - - [11008, 2048, 1, 128] - - [75, 40.313] - - - [27520, 11649, 1, 128] - - [27, 45.664] - - - [10880, 512, 1, 128] - - [34, 32.199] - - - [14592, 6785, 1, 128] - - [59, 44.037] - - - [7424, 512, 1, 128] - - [27, 26.773] - - - [13056, 5249, 1, 128] - - [37, 44.423] - - - [23296, 15489, 1, 128] - - [25, 46.155] - - - [28416, 8192, 1, 128] - - [27, 46.394] - - - [11392, 7681, 1, 128] - - [27, 44.48] - - - [18048, 1024, 1, 128] - - [25, 40.637] - - - [15616, 7809, 1, 128] - - [29, 45.364] - - - [128, 128, 1, 128] - - [113, 0.75] - - - [24704, 512, 1, 128] - - [59, 36.765] - - - [7680, 4097, 1, 128] - - [27, 42.92] - - - [16640, 8961, 1, 128] - - [25, 45.941] - - - [18944, 1024, 1, 128] - - [49, 41.631] - - - [12928, 2048, 1, 128] - - [40, 42.458] - - - [22272, 2048, 1, 128] - - [53, 43.542] - - - [27904, 11905, 1, 128] - - [51, 45.645] - - - [26240, 2048, 1, 128] - - [29, 41.678] - - - [9728, 6017, 1, 128] - - [34, 44.661] - - - [20736, 1024, 1, 128] - - [37, 42.428] - - - [3456, 1921, 1, 128] - - [35, 37.335] - - - [8064, 512, 1, 128] - - [35, 27.566] - - - [4224, 1024, 1, 128] - - [27, 29.888] - - - [25984, 10113, 1, 128] - - [30, 45.563] - - - [13696, 6017, 1, 128] - - [85, 42.514] - - - [27520, 8192, 1, 128] - - [37, 46.441] - - - [18944, 512, 1, 128] - - [34, 38.516] - - - [6272, 128, 1, 128] - - [110, 25.437] - - - [27264, 4096, 1, 128] - - [49, 45.21] - - - [1792, 1153, 1, 128] - - [51, 23.607] - - - [17536, 9729, 1, 128] - - [37, 45.307] - - - [13184, 5505, 1, 128] - - [31, 43.612] - - - [2944, 128, 1, 128] - - [115, 15.411] - - - [25344, 512, 1, 128] - - [25, 39.459] - - - [23040, 15361, 1, 128] - - [37, 46.101] - - - [8704, 512, 1, 128] - - [40, 30.861] - - - [20864, 13057, 1, 128] - - [22, 45.465] - - - [19328, 4096, 1, 128] - - [37, 45.15] - - - [28288, 8192, 1, 128] - - [38, 46.469] - - - [10112, 1024, 1, 128] - - [27, 40.387] - - - [17536, 2048, 1, 128] - - [63, 43.799] - - - [7552, 128, 1, 128] - - [108, 24.997] - - - [15616, 7937, 1, 128] - - [29, 45.393] - - - [23040, 512, 1, 128] - - [69, 38.562] - - - [25984, 2048, 1, 128] - - [39, 43.611] - - - [14720, 128, 1, 128] - - [59, 22.186] - - - [23424, 1024, 1, 128] - - [25, 41.891] - - - [1920, 1281, 1, 128] - - [34, 27.133] - - - [27136, 2048, 1, 128] - - [44, 44.323] - - - [28800, 8192, 1, 128] - - [29, 46.396] - - - [15488, 128, 1, 128] - - [35, 23.15] - - - [28800, 12929, 1, 128] - - [37, 45.572] - - - [21888, 14081, 1, 128] - - [76, 41.962] - - - [25600, 1024, 1, 128] - - [34, 42.273] - - - [21632, 1024, 1, 128] - - [40, 41.233] - - - [24448, 1024, 1, 128] - - [49, 41.88] - - - [4352, 2689, 1, 128] - - [35, 38.62] - - - [20480, 512, 1, 128] - - [27, 39.934] - - - [7296, 128, 1, 128] - - [114, 24.703] - - - [4992, 1024, 1, 128] - - [25, 34.368] - - - [27264, 11393, 1, 128] - - [29, 45.662] - - - [26752, 128, 1, 128] - - [60, 35.086] - - - [24960, 1024, 1, 128] - - [40, 41.824] - - - [21504, 512, 1, 128] - - [23, 37.142] - - - [6272, 2561, 1, 128] - - [27, 40.26] - - - [25088, 9089, 1, 128] - - [47, 45.869] - - - [20864, 512, 1, 128] - - [35, 40.683] - - - [4224, 2561, 1, 128] - - [31, 36.881] - - - [15744, 8065, 1, 128] - - [25, 45.33] - - - [21632, 128, 1, 128] - - [57, 30.024] - - - [15104, 4096, 1, 128] - - [25, 45.379] - - - [20352, 512, 1, 128] - - [63, 40.397] - - - [25472, 9601, 1, 128] - - [29, 45.504] - - - [27904, 512, 1, 128] - - [27, 41.498] - - - [19968, 512, 1, 128] - - [34, 40.046] - - - [5760, 1024, 1, 128] - - [35, 34.613] - - - [28416, 12545, 1, 128] - - [37, 45.589] - - - [16512, 8833, 1, 128] - - [27, 44.373] - - - [6016, 128, 1, 128] - - [121, 24.399] - - - [13056, 4096, 1, 128] - - [49, 45.349] - - - [19968, 12289, 1, 128] - - [27, 45.801] - - - [7424, 3713, 1, 128] - - [34, 43.114] - - - [28800, 128, 1, 128] - - [40, 25.256] - - - [512, 512, 1, 128] - - [115, 10.955] - - - [24832, 2048, 1, 128] - - [39, 43.926] - - - [20736, 128, 1, 128] - - [58, 29.22] - - - [26368, 512, 1, 128] - - [76, 40.028] - - - [26496, 8192, 1, 128] - - [57, 45.86] - - - [13824, 4096, 1, 128] - - [34, 45.2] - - - [27264, 128, 1, 128] - - [40, 35.195] - - - [21760, 1024, 1, 128] - - [35, 41.232] - - - [2432, 1921, 1, 128] - - [35, 31.632] - - - [27136, 8192, 1, 128] - - [47, 47.092] - - - [6784, 2048, 1, 128] - - [37, 41.505] - - - [11264, 128, 1, 128] - - [177, 28.769] - - - [7552, 512, 1, 128] - - [58, 27.24] - - - [19328, 11649, 1, 128] - - [27, 45.711] - - - [17152, 2048, 1, 128] - - [60, 43.822] - - - [23808, 16129, 1, 128] - - [27, 46.207] - - - [20224, 12417, 1, 128] - - [22, 45.968] - - - [27904, 1024, 1, 128] - - [55, 42.724] - - - [3456, 512, 1, 128] - - [108, 33.676] - - - [13312, 512, 1, 128] - - [78, 37.959] - - - [26368, 4096, 1, 128] - - [25, 45.563] - - - [23296, 15617, 1, 128] - - [29, 46.156] - - - [26112, 10241, 1, 128] - - [24, 45.807] - - - [26240, 512, 1, 128] - - [34, 39.854] - - - [4352, 1024, 1, 128] - - [27, 30.751] - - - [10624, 2048, 1, 128] - - [49, 41.133] - - - [23808, 16001, 1, 128] - - [25, 46.117] - - - [17536, 9857, 1, 128] - - [25, 45.598] - - - [23936, 4096, 1, 128] - - [29, 45.442] - - - [1408, 128, 1, 128] - - [166, 7.933] - - - [14848, 512, 1, 128] - - [35, 34.702] - - - [8704, 4993, 1, 128] - - [27, 44.036] - - - [15104, 2048, 1, 128] - - [60, 43.131] - - - [2560, 512, 1, 128] - - [121, 30.724] - - - [27264, 8192, 1, 128] - - [29, 46.344] - - - [23808, 4096, 1, 128] - - [38, 45.544] - - - [14080, 6273, 1, 128] - - [35, 43.219] - - - [10112, 6529, 1, 128] - - [35, 43.982] - - - [27648, 512, 1, 128] - - [40, 41.305] - - - [20992, 128, 1, 128] - - [40, 29.53] - - - [15104, 128, 1, 128] - - [57, 22.807] - - - [7808, 128, 1, 128] - - [123, 25.729] - - - [3584, 1024, 1, 128] - - [55, 26.305] - - - [15232, 512, 1, 128] - - [25, 34.983] - - - [21376, 13697, 1, 128] - - [27, 45.838] - - - [11392, 7809, 1, 128] - - [54, 44.52] - - - [11904, 1024, 1, 128] - - [25, 39.736] - - - [28800, 2048, 1, 128] - - [36, 43.549] - - - [8960, 512, 1, 128] - - [55, 31.586] - - - [19456, 11649, 1, 128] - - [29, 46.451] - - - [11904, 128, 1, 128] - - [114, 31.381] - - - [18560, 512, 1, 128] - - [27, 37.97] - - - [6656, 128, 1, 128] - - [110, 26.414] - - - [17792, 2048, 1, 128] - - [53, 43.056] - - - [21632, 4096, 1, 128] - - [37, 45.217] - - - [25728, 4096, 1, 128] - - [49, 45.444] - - - [18048, 10241, 1, 128] - - [51, 44.043] - - - [1792, 1281, 1, 128] - - [57, 25.763] - - - [512, 385, 1, 128] - - [115, 8.547] - - - [26112, 512, 1, 128] - - [23, 40.26] - - - [16128, 1024, 1, 128] - - [23, 39.928] - - - [4480, 1024, 1, 128] - - [27, 31.324] - - - [14720, 4096, 1, 128] - - [29, 44.948] - - - [23552, 2048, 1, 128] - - [44, 44.203] - - - [22528, 512, 1, 128] - - [44, 38.262] - - - [22912, 128, 1, 128] - - [63, 31.394] - - - [25344, 1024, 1, 128] - - [34, 42.112] - - - [24064, 16257, 1, 128] - - [50, 46.571] - - - [9088, 5377, 1, 128] - - [29, 44.163] - - - [27776, 128, 1, 128] - - [60, 35.926] - - - [15616, 512, 1, 128] - - [76, 35.235] - - - [13568, 128, 1, 128] - - [172, 32.943] - - - [15488, 7681, 1, 128] - - [25, 44.725] - - - [20096, 512, 1, 128] - - [40, 39.969] - - - [24832, 4096, 1, 128] - - [27, 45.784] - - - [28800, 4096, 1, 128] - - [34, 45.168] - - - [11904, 4225, 1, 128] - - [37, 43.69] - - - [3968, 1024, 1, 128] - - [34, 28.48] - - - [6400, 2817, 1, 128] - - [27, 40.365] - - - [24576, 4096, 1, 128] - - [29, 42.846] - - - [9088, 128, 1, 128] - - [114, 29.292] - - - [17152, 4096, 1, 128] - - [35, 45.392] - - - [22528, 14721, 1, 128] - - [37, 46.947] - - - [27392, 2048, 1, 128] - - [86, 42.968] - - - [8832, 512, 1, 128] - - [36, 30.98] - - - [8960, 5249, 1, 128] - - [59, 43.23] - - - [3200, 1024, 1, 128] - - [80, 33.876] - - - [4736, 3073, 1, 128] - - [27, 39.057] - - - [28032, 2048, 1, 128] - - [40, 44.151] - - - [14592, 2048, 1, 128] - - [35, 41.79] - - - [13440, 1024, 1, 128] - - [27, 41.263] - - - [14464, 2048, 1, 128] - - [63, 42.686] - - - [6912, 2048, 1, 128] - - [27, 42.332] - - - [19584, 2048, 1, 128] - - [40, 43.263] - - - [17920, 128, 1, 128] - - [57, 26.677] - - - [19584, 11777, 1, 128] - - [37, 45.594] - - - [23936, 16129, 1, 128] - - [29, 46.1] - - - [10496, 6785, 1, 128] - - [27, 44.614] - - - [27648, 2048, 1, 128] - - [60, 44.374] - - - [23808, 128, 1, 128] - - [60, 32.621] - - - [20864, 2048, 1, 128] - - [29, 43.381] - - - [9088, 512, 1, 128] - - [80, 31.659] - - - [3584, 512, 1, 128] - - [39, 22.053] - - - [8576, 4993, 1, 128] - - [25, 43.512] - - - [3328, 1024, 1, 128] - - [79, 34.549] - - - [20608, 2048, 1, 128] - - [61, 43.039] - - - [23552, 15745, 1, 128] - - [37, 46.73] - - - [23424, 15617, 1, 128] - - [45, 43.609] - - - [21120, 512, 1, 128] - - [25, 41.201] - - - [6656, 512, 1, 128] - - [60, 35.556] - - - [12544, 128, 1, 128] - - [114, 31.709] - - - [24448, 8577, 1, 128] - - [37, 45.437] - - - [9984, 512, 1, 128] - - [36, 34.496] - - - [18304, 4096, 1, 128] - - [35, 45.075] - - - [17920, 512, 1, 128] - - [39, 37.359] - - - [12160, 4096, 1, 128] - - [35, 45.148] - - - [3968, 2433, 1, 128] - - [34, 38.283] - - - [27008, 4096, 1, 128] - - [64, 45.118] - - - [22272, 1024, 1, 128] - - [37, 41.907] - - - [14336, 512, 1, 128] - - [29, 33.551] - - - [18560, 10753, 1, 128] - - [37, 45.476] - - - [6272, 2048, 1, 128] - - [34, 39.859] - - - [12800, 1024, 1, 128] - - [29, 40.056] - - - [9600, 5889, 1, 128] - - [29, 43.887] - - - [13056, 128, 1, 128] - - [172, 32.667] - - - [7296, 2048, 1, 128] - - [35, 39.711] - - - [21376, 512, 1, 128] - - [40, 36.706] - - - [11904, 512, 1, 128] - - [34, 35.236] - - - [6400, 1024, 1, 128] - - [27, 37.431] - - - [27008, 1024, 1, 128] - - [35, 42.529] - - - [22400, 14721, 1, 128] - - [22, 45.543] - - - [6272, 1024, 1, 128] - - [80, 36.799] - - - [17408, 128, 1, 128] - - [40, 25.811] - - - [26624, 10625, 1, 128] - - [29, 46.721] - - - [22400, 1024, 1, 128] - - [49, 41.793] - - - [18304, 10625, 1, 128] - - [37, 45.509] - - - [15872, 1024, 1, 128] - - [57, 40.799] - - - [21120, 128, 1, 128] - - [61, 29.43] - - - [22784, 4096, 1, 128] - - [57, 45.163] - - - [25728, 9857, 1, 128] - - [29, 45.466] - - - [16256, 1024, 1, 128] - - [25, 39.961] - - - [18560, 4096, 1, 128] - - [37, 45.201] - - - [7936, 4225, 1, 128] - - [29, 43.039] - - - [7680, 3969, 1, 128] - - [35, 42.949] - - - [9472, 2048, 1, 128] - - [40, 41.446] - - - [28160, 128, 1, 128] - - [78, 35.981] - - - [18816, 512, 1, 128] - - [36, 38.09] - - - [9856, 512, 1, 128] - - [49, 33.895] - - - [17664, 9857, 1, 128] - - [38, 45.257] - - - [27392, 128, 1, 128] - - [58, 35.553] - - - [24448, 2048, 1, 128] - - [27, 43.225] - - - [7808, 512, 1, 128] - - [27, 27.757] - - - [13952, 512, 1, 128] - - [25, 39.585] - - - [24576, 512, 1, 128] - - [25, 39.435] - - - [27520, 128, 1, 128] - - [59, 35.463] - - - [26496, 512, 1, 128] - - [60, 40.326] - - - [8576, 512, 1, 128] - - [40, 30.226] - - - [11648, 512, 1, 128] - - [60, 34.583] - - - [17408, 2048, 1, 128] - - [60, 43.526] - - - [17920, 10241, 1, 128] - - [47, 45.751] - - - [16384, 1024, 1, 128] - - [27, 37.324] - - - [6016, 2048, 1, 128] - - [27, 40.14] - - - [9728, 512, 1, 128] - - [49, 33.763] - - - [19712, 128, 1, 128] - - [34, 28.104] - - - [26112, 1024, 1, 128] - - [60, 42.492] - - - [16768, 128, 1, 128] - - [39, 24.867] - - - [8960, 1024, 1, 128] - - [25, 37.168] - - - [6784, 128, 1, 128] - - [123, 26.778] - - - [12800, 4993, 1, 128] - - [59, 44.444] - - - [6144, 2561, 1, 128] - - [35, 39.951] - - - [26880, 10881, 1, 128] - - [38, 45.591] - - - [12928, 1024, 1, 128] - - [27, 40.278] - - - [7040, 3457, 1, 128] - - [49, 42.449] - - - [15744, 4096, 1, 128] - - [29, 45.477] - - - [20096, 4096, 1, 128] - - [55, 44.898] - - - [21760, 128, 1, 128] - - [57, 30.145] - - - [7936, 2048, 1, 128] - - [37, 40.79] - - - [24448, 8192, 1, 128] - - [29, 46.582] - - - [21120, 2048, 1, 128] - - [69, 42.844] - - - [12160, 1024, 1, 128] - - [29, 40.129] - - - [7168, 3457, 1, 128] - - [35, 41.781] - - - [15232, 7553, 1, 128] - - [38, 44.833] - - - [26624, 1024, 1, 128] - - [35, 42.803] - - - [25344, 2048, 1, 128] - - [27, 43.168] - - - [12544, 4865, 1, 128] - - [29, 44.396] - - - [21120, 4096, 1, 128] - - [59, 44.469] - - - [20224, 128, 1, 128] - - [36, 28.834] - - - [14592, 4096, 1, 128] - - [31, 43.855] - - - [16256, 8577, 1, 128] - - [29, 45.045] - - - [24192, 4096, 1, 128] - - [29, 45.486] - - - [21248, 1024, 1, 128] - - [27, 41.174] - - - [25216, 1024, 1, 128] - - [25, 42.089] - - - [5888, 2177, 1, 128] - - [29, 39.11] - - - [21504, 1024, 1, 128] - - [27, 41.566] - - - [17536, 1024, 1, 128] - - [29, 42.672] - - - [9728, 2048, 1, 128] - - [40, 41.96] - - - [13952, 6273, 1, 128] - - [29, 44.621] - - - [28800, 512, 1, 128] - - [72, 38.717] - - - [2304, 1793, 1, 128] - - [27, 28.752] - - - [12416, 128, 1, 128] - - [110, 31.601] - - - [20224, 1024, 1, 128] - - [35, 42.303] - - - [22144, 128, 1, 128] - - [34, 30.683] - - - [22784, 1024, 1, 128] - - [29, 42.032] - - - [27136, 4096, 1, 128] - - [49, 46.044] - - - [27264, 512, 1, 128] - - [27, 39.551] - - - [26240, 10241, 1, 128] - - [25, 41.87] - - - [27904, 4096, 1, 128] - - [64, 45.283] - - - [21504, 128, 1, 128] - - [34, 30.135] - - - [3712, 2177, 1, 128] - - [35, 35.942] - - - [18432, 1024, 1, 128] - - [37, 41.126] - - - [28672, 4096, 1, 128] - - [27, 46.523] - - - [25344, 4096, 1, 128] - - [67, 45.027] - - - [26880, 512, 1, 128] - - [63, 40.713] - - - [21888, 2048, 1, 128] - - [87, 36.567] - - - [1792, 128, 1, 128] - - [113, 10.02] - - - [6016, 1024, 1, 128] - - [49, 35.431] - - - [15104, 7425, 1, 128] - - [27, 45.481] - - - [22016, 2048, 1, 128] - - [72, 43.988] - - - [13952, 4096, 1, 128] - - [29, 44.891] - - - [20992, 4096, 1, 128] - - [49, 45.563] - - - [8064, 4481, 1, 128] - - [37, 42.758] - - - [12672, 4096, 1, 128] - - [27, 44.756] - - - [20096, 12289, 1, 128] - - [25, 45.138] - - - [14848, 2048, 1, 128] - - [25, 43.185] - - - [23168, 512, 1, 128] - - [40, 38.602] - - - [7680, 128, 1, 128] - - [108, 25.42] - - - [13312, 1024, 1, 128] - - [37, 40.722] - - - [10624, 1024, 1, 128] - - [37, 36.621] - - - [3840, 512, 1, 128] - - [27, 22.912] - - - [22144, 14337, 1, 128] - - [37, 45.549] - - - [3200, 128, 1, 128] - - [176, 16.517] - - - [25472, 9473, 1, 128] - - [37, 45.643] - - - [16768, 9089, 1, 128] - - [25, 45.453] - - - [12288, 2048, 1, 128] - - [27, 42.3] - - - [20608, 512, 1, 128] - - [76, 40.424] - - - [2816, 1024, 1, 128] - - [57, 31.092] - - - [7552, 1024, 1, 128] - - [27, 34.968] - - - [5120, 3457, 1, 128] - - [29, 40.119] - - - [25216, 2048, 1, 128] - - [23, 43.233] - - - [12672, 4865, 1, 128] - - [27, 43.916] - - - [10880, 2048, 1, 128] - - [49, 40.607] - - - [18176, 512, 1, 128] - - [29, 37.436] - - - [8320, 4609, 1, 128] - - [27, 42.483] - - - [16000, 4096, 1, 128] - - [35, 44.862] - - - [22144, 2048, 1, 128] - - [60, 43.669] - - - [22784, 512, 1, 128] - - [55, 38.3] - - - [4096, 2561, 1, 128] - - [59, 36.95] - - - [24576, 2048, 1, 128] - - [27, 41.044] - - - [26624, 4096, 1, 128] - - [25, 46.684] - - - [18560, 2048, 1, 128] - - [40, 43.454] - - - [19584, 128, 1, 128] - - [53, 28.304] - - - [23936, 2048, 1, 128] - - [40, 43.931] - - - [23552, 512, 1, 128] - - [34, 39.286] - - - [12032, 4096, 1, 128] - - [49, 44.644] - - - [3840, 2305, 1, 128] - - [34, 36.194] - - - [25088, 128, 1, 128] - - [23, 33.681] - - - [16640, 8833, 1, 128] - - [29, 46.213] - - - [896, 128, 1, 128] - - [113, 5.167] - - - [17280, 2048, 1, 128] - - [36, 43.519] - - - [16896, 2048, 1, 128] - - [23, 43.271] - - - [22656, 128, 1, 128] - - [23, 31.094] - - - [25728, 8192, 1, 128] - - [29, 46.455] - - - [16128, 128, 1, 128] - - [72, 24.353] - - - [3840, 1024, 1, 128] - - [27, 27.77] - - - [2944, 512, 1, 128] - - [123, 31.266] - - - [24064, 1024, 1, 128] - - [25, 42.816] - - - [896, 385, 1, 128] - - [178, 13.814] - - - [8064, 128, 1, 128] - - [114, 26.572] - - - [12416, 1024, 1, 128] - - [67, 39.02] - - - [20608, 128, 1, 128] - - [40, 28.874] - - - [2944, 1024, 1, 128] - - [59, 31.974] - - - [6656, 2048, 1, 128] - - [27, 40.68] - - - [24064, 128, 1, 128] - - [58, 32.609] - - - [15744, 7937, 1, 128] - - [25, 45.436] - - - [2688, 1024, 1, 128] - - [47, 30.021] - - - [24192, 8193, 1, 128] - - [29, 44.945] - - - [24320, 4096, 1, 128] - - [37, 45.703] - - - [24576, 8705, 1, 128] - - [22, 42.522] - - - [13824, 1024, 1, 128] - - [25, 41.992] - - - [27776, 512, 1, 128] - - [40, 41.391] - - - [10240, 128, 1, 128] - - [121, 30.977] - - - [26240, 10369, 1, 128] - - [37, 41.875] - - - [16512, 4096, 1, 128] - - [37, 42.737] - - - [9856, 6145, 1, 128] - - [25, 42.117] - - - [27392, 1024, 1, 128] - - [40, 40.737] - - - [14976, 1024, 1, 128] - - [27, 40.146] - - - [1280, 512, 1, 128] - - [166, 22.362] - - - [6528, 2817, 1, 128] - - [27, 40.738] - - - [12288, 512, 1, 128] - - [53, 35.624] - - - [5248, 512, 1, 128] - - [78, 29.412] - - - [28544, 4096, 1, 128] - - [55, 45.444] - - - [21248, 13569, 1, 128] - - [51, 45.898] - - - [26112, 2048, 1, 128] - - [60, 44.634] - - - [14208, 6401, 1, 128] - - [27, 44.543] - - - [13952, 128, 1, 128] - - [114, 33.545] - - - [2304, 1665, 1, 128] - - [34, 26.796] - - - [6912, 1024, 1, 128] - - [29, 39.297] - - - [28672, 1024, 1, 128] - - [49, 42.181] - - - [14592, 6913, 1, 128] - - [57, 43.881] - - - [24704, 1024, 1, 128] - - [63, 39.923] - - - [22400, 512, 1, 128] - - [27, 37.762] - - - [23424, 4096, 1, 128] - - [35, 43.442] - - - [24832, 128, 1, 128] - - [40, 33.52] - - - [23680, 2048, 1, 128] - - [37, 42.883] - - - [25984, 9985, 1, 128] - - [30, 45.49] - - - [15360, 512, 1, 128] - - [60, 35.422] - - - [21376, 2048, 1, 128] - - [36, 43.564] - - - [16128, 2048, 1, 128] - - [63, 43.058] - - - [15872, 512, 1, 128] - - [40, 36.255] - - - [3072, 128, 1, 128] - - [116, 16.313] - - - [27520, 4096, 1, 128] - - [27, 45.446] - - - [25216, 4096, 1, 128] - - [35, 45.281] - - - [28672, 12673, 1, 128] - - [37, 46.81] - - - [28288, 512, 1, 128] - - [34, 38.591] - - - [22400, 4096, 1, 128] - - [38, 45.103] - - - [25344, 9345, 1, 128] - - [47, 45.119] - - - [9984, 128, 1, 128] - - [114, 30.964] - - - [28416, 1024, 1, 128] - - [37, 42.268] - - - [27008, 8192, 1, 128] - - [50, 46.31] - - - [13184, 1024, 1, 128] - - [27, 40.381] - - - [10240, 512, 1, 128] - - [49, 34.804] - - - [3456, 128, 1, 128] - - [166, 17.354] - - - [16000, 8321, 1, 128] - - [29, 45.088] - - - [27520, 1024, 1, 128] - - [58, 42.634] - - - [25088, 1024, 1, 128] - - [25, 42.278] - - - [6784, 512, 1, 128] - - [60, 35.653] - - - [18432, 10625, 1, 128] - - [37, 46.727] - - - [16128, 4096, 1, 128] - - [27, 45.42] - - - [26880, 11009, 1, 128] - - [29, 45.732] - - - [28800, 12801, 1, 128] - - [29, 45.526] - - - [12288, 4096, 1, 128] - - [27, 45.625] - - - [20096, 12417, 1, 128] - - [29, 45.622] - - - [1920, 128, 1, 128] - - [116, 10.498] - - - [13056, 2048, 1, 128] - - [37, 42.641] - - - [384, 385, 1, 128] - - [179, 6.362] - - - [9088, 1024, 1, 128] - - [23, 37.203] - - - [6784, 1024, 1, 128] - - [27, 39.104] - - - [21760, 4096, 1, 128] - - [34, 45.548] - - - [27008, 11009, 1, 128] - - [50, 45.472] - - - [14208, 1024, 1, 128] - - [25, 39.018] - - - [25600, 512, 1, 128] - - [28, 39.615] - - - [23680, 1024, 1, 128] - - [25, 42.082] - - - [28160, 8192, 1, 128] - - [51, 46.763] - - - [22016, 4096, 1, 128] - - [30, 45.788] - - - [18688, 4096, 1, 128] - - [25, 45.475] - - - [10752, 1024, 1, 128] - - [27, 37.471] - - - [2432, 128, 1, 128] - - [180, 13.103] - - - [7296, 512, 1, 128] - - [36, 26.543] - - - [19200, 4096, 1, 128] - - [27, 45.06] - - - [4608, 2945, 1, 128] - - [34, 41.341] - - - [18816, 11009, 1, 128] - - [27, 45.335] - - - [9600, 1024, 1, 128] - - [80, 38.743] - - - [7168, 512, 1, 128] - - [60, 26.137] - - - [11904, 4097, 1, 128] - - [37, 43.378] - - - [17920, 1024, 1, 128] - - [25, 40.839] - - - [11520, 7809, 1, 128] - - [49, 44.699] - - - [22784, 14977, 1, 128] - - [47, 46.074] - - - [13696, 1024, 1, 128] - - [25, 41.476] - - - [15104, 1024, 1, 128] - - [37, 40.371] - - - [25216, 512, 1, 128] - - [35, 39.351] - - - [5376, 512, 1, 128] - - [57, 29.296] - - - [17408, 4096, 1, 128] - - [25, 45.895] - - - [25728, 512, 1, 128] - - [36, 39.372] - - - [896, 512, 1, 128] - - [118, 16.743] - - - [6912, 3329, 1, 128] - - [55, 41.631] - - - [22016, 512, 1, 128] - - [57, 37.306] - - - [22144, 4096, 1, 128] - - [55, 45.107] - - - [10368, 128, 1, 128] - - [108, 31.493] - - - [23296, 2048, 1, 128] - - [63, 43.486] - - - [17920, 10113, 1, 128] - - [64, 46.032] - - - [14848, 4096, 1, 128] - - [37, 45.374] - - - [26112, 128, 1, 128] - - [23, 34.81] - - - [28032, 8192, 1, 128] - - [37, 46.536] - - - [20096, 128, 1, 128] - - [33, 28.324] - - - [15360, 4096, 1, 128] - - [25, 45.83] - - - [3328, 128, 1, 128] - - [118, 17.059] - - - [25472, 512, 1, 128] - - [40, 39.729] - - - [18304, 128, 1, 128] - - [60, 26.658] - - - [20352, 12545, 1, 128] - - [29, 45.911] - - - [26624, 10753, 1, 128] - - [37, 46.719] - - - [20480, 2048, 1, 128] - - [25, 43.517] - - - [26496, 10497, 1, 128] - - [47, 45.036] - - - [22400, 128, 1, 128] - - [40, 30.98] - - - [9216, 5505, 1, 128] - - [22, 44.431] - - - [24064, 8193, 1, 128] - - [30, 45.758] - - - [4224, 128, 1, 128] - - [118, 18.782] - - - [6656, 3073, 1, 128] - - [37, 41.59] - - - [10880, 1024, 1, 128] - - [25, 36.012] - - - [23808, 512, 1, 128] - - [60, 39.24] - - - [15488, 1024, 1, 128] - - [49, 40.525] - - - [24704, 8705, 1, 128] - - [25, 44.615] - - - [12416, 4609, 1, 128] - - [57, 43.37] - - - [3712, 1024, 1, 128] - - [35, 26.643] - - - [25856, 8192, 1, 128] - - [38, 46.514] - - - [8320, 1024, 1, 128] - - [59, 35.307] - - - [16256, 512, 1, 128] - - [40, 34.925] - - - [18944, 2048, 1, 128] - - [58, 43.781] - - - [23168, 4096, 1, 128] - - [27, 45.243] - - - [15616, 2048, 1, 128] - - [37, 43.257] - - - [24320, 512, 1, 128] - - [57, 39.869] - - - [2688, 1025, 1, 128] - - [57, 29.542] - - - [12800, 5121, 1, 128] - - [47, 43.607] - - - [5120, 128, 1, 128] - - [116, 21.595] - - - [4352, 512, 1, 128] - - [34, 25.253] - - - [24576, 8192, 1, 128] - - [38, 43.71] - - - [8320, 512, 1, 128] - - [35, 27.956] - - - [12160, 4481, 1, 128] - - [29, 43.502] - - - [2560, 1025, 1, 128] - - [59, 28.887] - - - [19072, 1024, 1, 128] - - [35, 41.458] - - - [2816, 1153, 1, 128] - - [37, 33.809] - - - [6912, 128, 1, 128] - - [114, 27.43] - - - [9088, 2048, 1, 128] - - [37, 40.832] - - - [26368, 8192, 1, 128] - - [37, 46.613] - - - [17408, 9729, 1, 128] - - [37, 46.24] - - - [18816, 4096, 1, 128] - - [37, 45.154] - - - [4480, 512, 1, 128] - - [67, 25.95] - - - [11648, 128, 1, 128] - - [123, 30.816] - - - [1536, 897, 1, 128] - - [172, 27.683] - - - [11136, 1024, 1, 128] - - [60, 37.354] - - - [8704, 1024, 1, 128] - - [53, 38.492] - - - [19072, 2048, 1, 128] - - [23, 43.381] - - - [25856, 1024, 1, 128] - - [40, 42.188] - - - [7552, 3841, 1, 128] - - [29, 42.352] - - - [23296, 128, 1, 128] - - [60, 31.743] - - - [23424, 512, 1, 128] - - [53, 38.672] - - - [26368, 10497, 1, 128] - - [27, 45.865] - - - [18560, 1024, 1, 128] - - [27, 40.909] - - - [8192, 128, 1, 128] - - [109, 26.405] - - - [27776, 11905, 1, 128] - - [37, 45.554] - - - [18688, 1024, 1, 128] - - [60, 41.268] - - - [21248, 4096, 1, 128] - - [35, 45.353] - - - [16256, 8449, 1, 128] - - [29, 44.905] - - - [1920, 1409, 1, 128] - - [27, 28.905] - - - [24704, 4096, 1, 128] - - [61, 44.774] - - - [13824, 6145, 1, 128] - - [27, 44.647] - - - [6528, 512, 1, 128] - - [35, 34.74] - - - [21376, 128, 1, 128] - - [23, 29.786] - - - [11264, 1024, 1, 128] - - [25, 38.262] - - - [4352, 2817, 1, 128] - - [35, 38.359] - - - [22272, 4096, 1, 128] - - [22, 45.315] - - - [27264, 11265, 1, 128] - - [25, 45.335] - - - [28160, 1024, 1, 128] - - [31, 43.105] - - - [16256, 128, 1, 128] - - [72, 24.011] - - - [18688, 2048, 1, 128] - - [40, 43.547] - - - [9600, 6017, 1, 128] - - [55, 44.038] - - - [23552, 1024, 1, 128] - - [25, 42.519] - - - [8576, 128, 1, 128] - - [114, 27.764] - - - [20992, 13185, 1, 128] - - [29, 46.298] - - - [20992, 1024, 1, 128] - - [37, 42.981] - - - [14720, 512, 1, 128] - - [36, 34.106] - - - [28032, 1024, 1, 128] - - [25, 43.043] - - - [20352, 2048, 1, 128] - - [60, 43.425] - - - [15360, 128, 1, 128] - - [40, 23.052] - - - [8448, 2048, 1, 128] - - [29, 41.013] - - - [6272, 2689, 1, 128] - - [35, 41.19] - - - [7808, 4097, 1, 128] - - [25, 42.053] - - - [25472, 128, 1, 128] - - [60, 33.95] - - - [12288, 4481, 1, 128] - - [27, 44.377] - - - [28416, 4096, 1, 128] - - [34, 45.681] - - - [2176, 128, 1, 128] - - [116, 11.724] - - - [21760, 2048, 1, 128] - - [29, 43.441] - - - [21376, 1024, 1, 128] - - [27, 40.864] - - - [13696, 2048, 1, 128] - - [29, 41.689] - - - [28288, 12417, 1, 128] - - [25, 45.754] - - - [5632, 512, 1, 128] - - [59, 31.392] - - - [22016, 1024, 1, 128] - - [59, 41.391] - - - [25216, 128, 1, 128] - - [60, 33.676] - - - [25216, 8192, 1, 128] - - [22, 46.249] - - - [12032, 128, 1, 128] - - [172, 30.73] - - - [6144, 2048, 1, 128] - - [29, 39.869] - - - [23680, 128, 1, 128] - - [27, 32.26] - - - [15744, 128, 1, 128] - - [23, 23.537] - - - [3968, 512, 1, 128] - - [59, 23.346] - - - [16512, 1024, 1, 128] - - [75, 34.182] - - - [1536, 128, 1, 128] - - [119, 8.857] - - - [25984, 4096, 1, 128] - - [51, 45.164] - - - [19456, 512, 1, 128] - - [34, 39.313] - - - [9984, 1024, 1, 128] - - [27, 39.768] - - - [14080, 6401, 1, 128] - - [59, 43.172] - - - [20736, 2048, 1, 128] - - [63, 43.746] - - - [4224, 2689, 1, 128] - - [31, 37.664] - - - [13696, 512, 1, 128] - - [35, 38.896] - - - [17280, 1024, 1, 128] - - [35, 42.398] - - - [10752, 128, 1, 128] - - [114, 29.285] - - - [1536, 512, 1, 128] - - [123, 24.918] - - - [25728, 2048, 1, 128] - - [37, 43.634] - - - [9472, 128, 1, 128] - - [110, 29.878] - - - [7168, 3585, 1, 128] - - [27, 41.883] - - - [14720, 1024, 1, 128] - - [27, 39.831] - - - [25728, 128, 1, 128] - - [58, 34.113] - - - [14976, 128, 1, 128] - - [60, 22.567] - - - [24832, 1024, 1, 128] - - [40, 41.954] - - - [14080, 512, 1, 128] - - [34, 39.158] - - - [17152, 1024, 1, 128] - - [37, 42.312] - - - [19072, 512, 1, 128] - - [67, 38.216] - - - [21120, 1024, 1, 128] - - [27, 42.256] - - - [4864, 128, 1, 128] - - [113, 20.996] - - - [7936, 512, 1, 128] - - [27, 28.619] - - - [21248, 13441, 1, 128] - - [29, 45.979] - - - [12160, 2048, 1, 128] - - [63, 42.781] - - - [19712, 11905, 1, 128] - - [64, 44.366] - - - [23296, 1024, 1, 128] - - [35, 42.225] - - - [24832, 8961, 1, 128] - - [51, 45.757] - - - [13568, 2048, 1, 128] - - [63, 43.078] - - - [13696, 4096, 1, 128] - - [49, 43.197] - - - [5888, 128, 1, 128] - - [108, 24.145] - - - [10112, 2048, 1, 128] - - [37, 42.288] - - - [21632, 13953, 1, 128] - - [22, 45.908] - - - [19328, 512, 1, 128] - - [40, 38.887] - - - [6272, 512, 1, 128] - - [63, 33.934] - - - [4864, 3201, 1, 128] - - [49, 41.106] - - - [15232, 4096, 1, 128] - - [27, 44.986] - - - [23040, 4096, 1, 128] - - [35, 45.59] - - - [2816, 1281, 1, 128] - - [25, 26.313] - - - [8960, 128, 1, 128] - - [110, 28.755] - - - [9472, 1024, 1, 128] - - [49, 38.481] - - - [27648, 11777, 1, 128] - - [37, 46.371] - - - [28416, 2048, 1, 128] - - [29, 43.52] - - - [13952, 6145, 1, 128] - - [34, 44.191] - - - [13952, 1024, 1, 128] - - [27, 41.942] - - - [12544, 2048, 1, 128] - - [27, 42.423] - - - [10624, 7041, 1, 128] - - [29, 44.381] - - - [24704, 2048, 1, 128] - - [28, 43.324] - - - [17280, 9473, 1, 128] - - [38, 45.407] - - - [25088, 9217, 1, 128] - - [47, 45.765] - - - [10240, 6657, 1, 128] - - [29, 45.524] - - - [12800, 4096, 1, 128] - - [67, 44.802] - - - [17792, 1024, 1, 128] - - [29, 40.407] - - - [12160, 128, 1, 128] - - [172, 31.828] - - - [16512, 128, 1, 128] - - [63, 24.389] - - - [25856, 512, 1, 128] - - [35, 39.671] - - - [8576, 4865, 1, 128] - - [25, 43.84] - - - [25984, 1024, 1, 128] - - [33, 41.903] - - - [512, 128, 1, 128] - - [178, 2.93] - - - [10112, 128, 1, 128] - - [110, 31.099] - - - [28288, 2048, 1, 128] - - [23, 44.215] - - - [1152, 641, 1, 128] - - [109, 23.269] - - - [17920, 4096, 1, 128] - - [57, 45.539] - - - [2560, 1921, 1, 128] - - [35, 33.5] - - - [24704, 8833, 1, 128] - - [29, 44.724] - - - [3200, 512, 1, 128] - - [123, 32.245] - - - [6656, 2945, 1, 128] - - [57, 41.037] - - - [12672, 4993, 1, 128] - - [27, 43.969] - - - [4608, 1024, 1, 128] - - [27, 32.411] - - - [25856, 9985, 1, 128] - - [22, 45.586] - - - [23808, 2048, 1, 128] - - [60, 43.971] - - - [9728, 6145, 1, 128] - - [55, 44.553] - - - [28416, 12417, 1, 128] - - [29, 45.628] - - - [14464, 4096, 1, 128] - - [29, 44.862] - - - [21888, 128, 1, 128] - - [39, 30.215] - - - [23680, 15873, 1, 128] - - [50, 45.366] - - - [22144, 1024, 1, 128] - - [49, 41.574] - - - [17664, 512, 1, 256] - - [60, 56.371] - - - [25600, 1024, 1, 256] - - [53, 69.238] - - - [28928, 512, 1, 256] - - [63, 62.502] - - - [15104, 512, 1, 256] - - [23, 55.153] - - - [38912, 1024, 1, 256] - - [23, 70.324] - - - [34304, 8192, 1, 256] - - [42, 74.829] - - - [23552, 1024, 1, 256] - - [63, 68.047] - - - [39424, 23552, 1, 256] - - [41, 75.106] - - - [9472, 1024, 1, 256] - - [40, 59.523] - - - [28928, 13056, 1, 256] - - [51, 74.869] - - - [42496, 1024, 1, 256] - - [69, 70.355] - - - [18432, 1024, 1, 256] - - [40, 66.708] - - - [40192, 24320, 1, 256] - - [27, 74.909] - - - [33280, 17152, 1, 256] - - [24, 75.787] - - - [27904, 512, 1, 256] - - [27, 65.331] - - - [39680, 8192, 1, 256] - - [25, 74.356] - - - [28160, 8192, 1, 256] - - [29, 74.637] - - - [25088, 8192, 1, 256] - - [24, 74.701] - - - [23040, 15360, 1, 256] - - [25, 75.671] - - - [19712, 11776, 1, 256] - - [41, 74.182] - - - [43520, 27648, 1, 256] - - [41, 74.947] - - - [44544, 4096, 1, 256] - - [41, 73.451] - - - [20224, 4096, 1, 256] - - [27, 72.535] - - - [31744, 4096, 1, 256] - - [28, 73.158] - - - [33024, 16896, 1, 256] - - [42, 75.252] - - - [32768, 8192, 1, 256] - - [88, 58.919] - - - [42752, 4096, 1, 256] - - [27, 72.964] - - - [19968, 512, 1, 256] - - [58, 61.638] - - - [10496, 512, 1, 256] - - [35, 55.906] - - - [36864, 4096, 1, 256] - - [54, 72.634] - - - [12288, 1024, 1, 256] - - [37, 62.574] - - - [22784, 14848, 1, 256] - - [28, 75.154] - - - [17152, 9472, 1, 256] - - [31, 74.859] - - - [31488, 1024, 1, 256] - - [60, 69.495] - - - [25344, 1024, 1, 256] - - [69, 67.149] - - - [33536, 512, 1, 256] - - [40, 65.285] - - - [28672, 8192, 1, 256] - - [29, 74.718] - - - [15104, 7168, 1, 256] - - [27, 73.585] - - - [38144, 22272, 1, 256] - - [38, 74.833] - - - [25344, 4096, 1, 256] - - [43, 72.024] - - - [6400, 2560, 1, 256] - - [34, 64.98] - - - [21248, 13568, 1, 256] - - [37, 75.179] - - - [2304, 1536, 1, 256] - - [49, 53.512] - - - [20992, 512, 1, 256] - - [27, 63.931] - - - [3072, 1024, 1, 256] - - [35, 49.378] - - - [36864, 20736, 1, 256] - - [25, 75.147] - - - [39936, 24064, 1, 256] - - [37, 75.501] - - - [2816, 512, 1, 256] - - [181, 42.312] - - - [37888, 512, 1, 256] - - [23, 66.052] - - - [39680, 1024, 1, 256] - - [37, 69.234] - - - [35584, 19712, 1, 256] - - [22, 74.82] - - - [25600, 9728, 1, 256] - - [25, 75.44] - - - [2816, 1024, 1, 256] - - [59, 46.28] - - - [13056, 1024, 1, 256] - - [60, 62.574] - - - [39680, 4096, 1, 256] - - [37, 72.728] - - - [4864, 3072, 1, 256] - - [55, 64.205] - - - [27648, 11776, 1, 256] - - [29, 75.73] - - - [13056, 4096, 1, 256] - - [37, 72.06] - - - [4096, 2304, 1, 256] - - [34, 59.057] - - - [34048, 1024, 1, 256] - - [39, 68.468] - - - [6400, 512, 1, 256] - - [40, 51.306] - - - [15872, 4096, 1, 256] - - [37, 72.281] - - - [29440, 1024, 1, 256] - - [60, 68.859] - - - [7424, 512, 1, 256] - - [63, 41.766] - - - [19200, 4096, 1, 256] - - [37, 72.099] - - - [37376, 21504, 1, 256] - - [41, 75.215] - - - [37888, 1024, 1, 256] - - [58, 70.228] - - - [40704, 24832, 1, 256] - - [27, 74.84] - - - [26112, 1024, 1, 256] - - [69, 69.213] - - - [25088, 8960, 1, 256] - - [59, 75.437] - - - [27136, 512, 1, 256] - - [27, 64.388] - - - [4608, 512, 1, 256] - - [27, 38.941] - - - [31232, 8192, 1, 256] - - [41, 74.803] - - - [33024, 512, 1, 256] - - [35, 60.923] - - - [27648, 512, 1, 256] - - [84, 65.08] - - - [28928, 4096, 1, 256] - - [27, 72.464] - - - [44544, 2048, 1, 256] - - [41, 72.039] - - - [43776, 27648, 1, 256] - - [41, 74.437] - - - [19456, 4096, 1, 256] - - [25, 72.834] - - - [33536, 17664, 1, 256] - - [24, 75.192] - - - [35328, 4096, 1, 256] - - [43, 73.657] - - - [13312, 5376, 1, 256] - - [34, 74.004] - - - [32768, 1024, 1, 256] - - [38, 59.341] - - - [39168, 4096, 1, 256] - - [41, 73.18] - - - [15616, 7936, 1, 256] - - [37, 74.429] - - - [41472, 25600, 1, 256] - - [38, 74.75] - - - [14592, 4096, 1, 256] - - [31, 70.479] - - - [37632, 21760, 1, 256] - - [25, 74.813] - - - [37376, 21248, 1, 256] - - [37, 75.33] - - - [14336, 6656, 1, 256] - - [29, 74.692] - - - [36608, 20480, 1, 256] - - [41, 74.602] - - - [32256, 16384, 1, 256] - - [43, 75.223] - - - [44544, 28416, 1, 256] - - [38, 75.158] - - - [26112, 512, 1, 256] - - [35, 62.957] - - - [41216, 25344, 1, 256] - - [27, 74.919] - - - [16640, 512, 1, 256] - - [101, 56.008] - - - [30464, 14336, 1, 256] - - [41, 73.987] - - - [13312, 4096, 1, 256] - - [29, 72.101] - - - [22528, 1024, 1, 256] - - [40, 68.222] - - - [5632, 1024, 1, 256] - - [26, 50.714] - - - [27392, 1024, 1, 256] - - [51, 64.891] - - - [27648, 8192, 1, 256] - - [25, 75.078] - - - [26368, 1024, 1, 256] - - [23, 68.193] - - - [43776, 4096, 1, 256] - - [28, 73.189] - - - [23552, 15872, 1, 256] - - [25, 76.241] - - - [26624, 10496, 1, 256] - - [27, 75.778] - - - [27392, 8192, 1, 256] - - [28, 74.062] - - - [17408, 9728, 1, 256] - - [25, 75.456] - - - [16896, 9216, 1, 256] - - [27, 74.885] - - - [26880, 11008, 1, 256] - - [27, 74.818] - - - [31488, 512, 1, 256] - - [25, 64.98] - - - [14336, 6400, 1, 256] - - [49, 74.846] - - - [17152, 512, 1, 256] - - [63, 60.79] - - - [7168, 512, 1, 256] - - [49, 41.029] - - - [41984, 26112, 1, 256] - - [22, 75.545] - - - [11776, 512, 1, 256] - - [40, 52.738] - - - [16128, 8448, 1, 256] - - [34, 74.56] - - - [11520, 1024, 1, 256] - - [25, 61.695] - - - [27904, 1024, 1, 256] - - [39, 68.722] - - - [37888, 8192, 1, 256] - - [29, 74.991] - - - [20480, 12544, 1, 256] - - [22, 75.828] - - - [23552, 15616, 1, 256] - - [29, 76.268] - - - [21504, 13824, 1, 256] - - [27, 76.158] - - - [27136, 11008, 1, 256] - - [24, 75.587] - - - [32000, 512, 1, 256] - - [34, 63.572] - - - [26624, 1024, 1, 256] - - [40, 68.378] - - - [34816, 8192, 1, 256] - - [37, 75.041] - - - [23040, 512, 1, 256] - - [69, 61.25] - - - [36608, 1024, 1, 256] - - [40, 69.423] - - - [43264, 8192, 1, 256] - - [29, 74.358] - - - [30208, 14336, 1, 256] - - [42, 75.063] - - - [43520, 512, 1, 256] - - [59, 66.566] - - - [32256, 4096, 1, 256] - - [43, 73.355] - - - [33792, 17664, 1, 256] - - [27, 75.868] - - - [10752, 6912, 1, 256] - - [27, 74.308] - - - [29696, 8192, 1, 256] - - [29, 74.986] - - - [41472, 512, 1, 256] - - [25, 66.983] - - - [44544, 8192, 1, 256] - - [27, 74.517] - - - [41472, 8192, 1, 256] - - [43, 74.446] - - - [38656, 4096, 1, 256] - - [41, 73.31] - - - [44800, 512, 1, 256] - - [29, 66.471] - - - [37376, 4096, 1, 256] - - [41, 73.81] - - - [19200, 1024, 1, 256] - - [72, 66.532] - - - [39680, 23552, 1, 256] - - [37, 74.227] - - - [30976, 8192, 1, 256] - - [41, 73.668] - - - [25856, 1024, 1, 256] - - [40, 68.429] - - - [22016, 14336, 1, 256] - - [37, 75.309] - - - [17152, 9216, 1, 256] - - [25, 74.262] - - - [18432, 10752, 1, 256] - - [25, 75.858] - - - [5376, 1024, 1, 256] - - [40, 48.52] - - - [21760, 13824, 1, 256] - - [29, 75.487] - - - [15360, 512, 1, 256] - - [34, 55.596] - - - [2560, 512, 1, 256] - - [125, 43.314] - - - [36096, 8192, 1, 256] - - [28, 73.872] - - - [42752, 26624, 1, 256] - - [27, 74.446] - - - [35584, 19456, 1, 256] - - [41, 74.674] - - - [6144, 2304, 1, 256] - - [55, 66.961] - - - [42240, 1024, 1, 256] - - [44, 69.828] - - - [26880, 4096, 1, 256] - - [25, 72.673] - - - [28160, 12032, 1, 256] - - [27, 75.503] - - - [18688, 10752, 1, 256] - - [27, 75.22] - - - [43520, 8192, 1, 256] - - [29, 74.662] - - - [8192, 4352, 1, 256] - - [37, 70.832] - - - [6912, 3072, 1, 256] - - [34, 68.847] - - - [31744, 15616, 1, 256] - - [37, 75.832] - - - [36352, 20224, 1, 256] - - [41, 75.358] - - - [41216, 25088, 1, 256] - - [22, 74.747] - - - [37632, 1024, 1, 256] - - [23, 69.481] - - - [18944, 512, 1, 256] - - [25, 59.145] - - - [15616, 1024, 1, 256] - - [60, 65.832] - - - [44288, 512, 1, 256] - - [31, 66.064] - - - [24832, 8704, 1, 256] - - [51, 74.982] - - - [21504, 13568, 1, 256] - - [37, 76.086] - - - [18176, 10496, 1, 256] - - [27, 75.104] - - - [21248, 1024, 1, 256] - - [60, 66.339] - - - [16384, 1024, 1, 256] - - [54, 58.261] - - - [25600, 8192, 1, 256] - - [37, 75.08] - - - [28672, 12544, 1, 256] - - [22, 75.492] - - - [16128, 1024, 1, 256] - - [35, 62.905] - - - [22272, 14592, 1, 256] - - [47, 75.412] - - - [1280, 512, 1, 256] - - [121, 30.85] - - - [36864, 20992, 1, 256] - - [37, 75.008] - - - [3584, 1792, 1, 256] - - [55, 56.36] - - - [35072, 19200, 1, 256] - - [25, 74.991] - - - [32000, 4096, 1, 256] - - [54, 72.71] - - - [28416, 1024, 1, 256] - - [63, 67.534] - - - [20480, 12800, 1, 256] - - [25, 75.76] - - - [21760, 4096, 1, 256] - - [91, 72.584] - - - [44288, 8192, 1, 256] - - [41, 74.353] - - - [33280, 4096, 1, 256] - - [28, 73.587] - - - [32512, 1024, 1, 256] - - [33, 68.481] - - - [38400, 22528, 1, 256] - - [28, 75.187] - - - [40448, 1024, 1, 256] - - [60, 70.029] - - - [5120, 512, 1, 256] - - [35, 42.717] - - - [29952, 8192, 1, 256] - - [51, 74.456] - - - [40448, 24576, 1, 256] - - [43, 74.509] - - - [29696, 4096, 1, 256] - - [25, 73.234] - - - [21504, 1024, 1, 256] - - [58, 67.191] - - - [19968, 1024, 1, 256] - - [60, 66.898] - - - [16896, 512, 1, 256] - - [60, 59.514] - - - [33536, 17408, 1, 256] - - [28, 74.787] - - - [19712, 512, 1, 256] - - [27, 60.709] - - - [16384, 8704, 1, 256] - - [38, 63.818] - - - [29952, 13824, 1, 256] - - [30, 75.112] - - - [14592, 6656, 1, 256] - - [67, 72.415] - - - [36864, 1024, 1, 256] - - [37, 68.423] - - - [31744, 15872, 1, 256] - - [27, 75.897] - - - [24832, 8960, 1, 256] - - [31, 75.119] - - - [23808, 1024, 1, 256] - - [40, 68.233] - - - [19200, 11264, 1, 256] - - [28, 74.727] - - - [23296, 15360, 1, 256] - - [29, 75.187] - - - [34304, 18432, 1, 256] - - [42, 75.048] - - - [22016, 1024, 1, 256] - - [63, 67.571] - - - [40704, 4096, 1, 256] - - [28, 73.066] - - - [25600, 4096, 1, 256] - - [28, 73.152] - - - [3328, 1024, 1, 256] - - [35, 52.353] - - - [30464, 8192, 1, 256] - - [41, 73.421] - - - [39424, 8192, 1, 256] - - [37, 74.587] - - - [23808, 15872, 1, 256] - - [37, 75.524] - - - [8960, 1024, 1, 256] - - [53, 56.85] - - - [44032, 4096, 1, 256] - - [37, 73.194] - - - [35584, 8192, 1, 256] - - [42, 74.268] - - - [29184, 8192, 1, 256] - - [30, 74.597] - - - [13824, 1024, 1, 256] - - [44, 65.854] - - - [36608, 8192, 1, 256] - - [25, 74.225] - - - [30976, 512, 1, 256] - - [23, 64.406] - - - [33024, 4096, 1, 256] - - [42, 73.199] - - - [11776, 7936, 1, 256] - - [27, 74.374] - - - [23808, 16128, 1, 256] - - [27, 75.37] - - - [22272, 14336, 1, 256] - - [41, 74.944] - - - [27392, 11520, 1, 256] - - [42, 74.289] - - - [30464, 4096, 1, 256] - - [41, 71.791] - - - [20992, 13312, 1, 256] - - [28, 75.507] - - - [44800, 1024, 1, 256] - - [29, 69.404] - - - [32512, 4096, 1, 256] - - [89, 72.937] - - - [23296, 15616, 1, 256] - - [27, 75.408] - - - [9216, 1024, 1, 256] - - [36, 58.517] - - - [20224, 12544, 1, 256] - - [25, 75.301] - - - [32256, 1024, 1, 256] - - [60, 69.566] - - - [38400, 512, 1, 256] - - [72, 65.71] - - - [29952, 1024, 1, 256] - - [60, 68.339] - - - [36352, 512, 1, 256] - - [49, 65.269] - - - [41728, 25600, 1, 256] - - [41, 73.877] - - - [32000, 1024, 1, 256] - - [23, 68.824] - - - [38144, 22016, 1, 256] - - [22, 74.81] - - - [27136, 11264, 1, 256] - - [47, 75.389] - - - [34048, 18176, 1, 256] - - [65, 74.655] - - - [22016, 14080, 1, 256] - - [59, 76.113] - - - [19712, 12032, 1, 256] - - [41, 74.357] - - - [23552, 4096, 1, 256] - - [41, 72.963] - - - [15872, 1024, 1, 256] - - [40, 64.076] - - - [37120, 512, 1, 256] - - [49, 65.473] - - - [9984, 1024, 1, 256] - - [40, 61.703] - - - [32512, 8192, 1, 256] - - [24, 74.44] - - - [15360, 4096, 1, 256] - - [27, 72.39] - - - [13056, 512, 1, 256] - - [34, 57.597] - - - [44032, 8192, 1, 256] - - [37, 74.818] - - - [24576, 8192, 1, 256] - - [54, 70.1] - - - [36352, 8192, 1, 256] - - [28, 74.844] - - - [26368, 8192, 1, 256] - - [37, 74.449] - - - [20480, 1024, 1, 256] - - [27, 65.317] - - - [35072, 8192, 1, 256] - - [37, 74.34] - - - [32000, 15872, 1, 256] - - [27, 75.055] - - - [40704, 24576, 1, 256] - - [28, 74.2] - - - [15104, 7424, 1, 256] - - [29, 74.433] - - - [25856, 4096, 1, 256] - - [28, 72.553] - - - [14848, 512, 1, 256] - - [25, 54.42] - - - [39424, 4096, 1, 256] - - [65, 73.507] - - - [24832, 512, 1, 256] - - [23, 60.301] - - - [44288, 28416, 1, 256] - - [22, 74.674] - - - [12544, 4608, 1, 256] - - [35, 72.008] - - - [12800, 4864, 1, 256] - - [59, 73.386] - - - [29440, 512, 1, 256] - - [63, 63.112] - - - [40192, 24064, 1, 256] - - [37, 74.898] - - - [18176, 4096, 1, 256] - - [29, 72.486] - - - [40960, 8192, 1, 256] - - [54, 67.262] - - - [42240, 512, 1, 256] - - [36, 67.25] - - - [9728, 512, 1, 256] - - [49, 52.767] - - - [14848, 7168, 1, 256] - - [25, 73.661] - - - [44800, 28672, 1, 256] - - [54, 74.049] - - - [15616, 7680, 1, 256] - - [29, 74.594] - - - [33280, 17408, 1, 256] - - [42, 75.374] - - - [42752, 1024, 1, 256] - - [36, 69.863] - - - [35328, 8192, 1, 256] - - [47, 74.689] - - - [36352, 1024, 1, 256] - - [78, 69.823] - - - [35840, 1024, 1, 256] - - [74, 70.114] - - - [41472, 4096, 1, 256] - - [28, 73.656] - - - [3584, 1024, 1, 256] - - [60, 41.069] - - - [22528, 14592, 1, 256] - - [25, 76.245] - - - [44032, 512, 1, 256] - - [44, 66.675] - - - [30720, 1024, 1, 256] - - [58, 69.647] - - - [39680, 512, 1, 256] - - [58, 65.345] - - - [22272, 1024, 1, 256] - - [40, 67.222] - - - [42240, 26368, 1, 256] - - [29, 74.63] - - - [10240, 6400, 1, 256] - - [55, 74.368] - - - [30976, 14848, 1, 256] - - [28, 74.145] - - - [41728, 25856, 1, 256] - - [41, 74.256] - - - [28928, 12800, 1, 256] - - [47, 74.82] - - - [21760, 14080, 1, 256] - - [25, 75.486] - - - [5888, 1024, 1, 256] - - [60, 52.899] - - - [24576, 8704, 1, 256] - - [38, 70.872] - - - [38912, 4096, 1, 256] - - [25, 73.118] - - - [15360, 1024, 1, 256] - - [63, 65.031] - - - [18688, 512, 1, 256] - - [34, 58.254] - - - [27392, 512, 1, 256] - - [57, 62.874] - - - [22784, 512, 1, 256] - - [40, 61.037] - - - [40448, 4096, 1, 256] - - [61, 73.641] - - - [19200, 512, 1, 256] - - [35, 59.411] - - - [26368, 10496, 1, 256] - - [27, 74.947] - - - [25088, 9216, 1, 256] - - [42, 74.874] - - - [33536, 1024, 1, 256] - - [23, 68.643] - - - [25600, 9472, 1, 256] - - [27, 75.887] - - - [13824, 4096, 1, 256] - - [27, 71.984] - - - [5632, 3840, 1, 256] - - [34, 70.207] - - - [9216, 5376, 1, 256] - - [49, 72.915] - - - [8960, 5120, 1, 256] - - [31, 71.62] - - - [19456, 512, 1, 256] - - [27, 60.468] - - - [24576, 4096, 1, 256] - - [32, 68.369] - - - [27392, 11264, 1, 256] - - [41, 74.537] - - - [35072, 4096, 1, 256] - - [41, 72.993] - - - [44288, 4096, 1, 256] - - [41, 73.28] - - - [40448, 8192, 1, 256] - - [51, 74.465] - - - [33280, 512, 1, 256] - - [60, 64.893] - - - [22272, 4096, 1, 256] - - [41, 72.461] - - - [35584, 512, 1, 256] - - [23, 63.96] - - - [10752, 512, 1, 256] - - [35, 48.931] - - - [19968, 4096, 1, 256] - - [61, 72.937] - - - [34304, 1024, 1, 256] - - [39, 69.634] - - - [41216, 8192, 1, 256] - - [41, 74.477] - - - [35840, 19712, 1, 256] - - [27, 75.823] - - - [43520, 27392, 1, 256] - - [38, 75.44] - - - [30720, 14848, 1, 256] - - [27, 75.633] - - - [38400, 22272, 1, 256] - - [38, 75.441] - - - [1536, 1024, 1, 256] - - [182, 44.029] - - - [40192, 1024, 1, 256] - - [61, 69.757] - - - [44800, 256, 1, 256] - - [58, 58.147] - - - [1536, 512, 1, 256] - - [123, 35.992] - - - [34560, 18432, 1, 256] - - [25, 74.758] - - - [1792, 1024, 1, 256] - - [53, 30.733] - - - [5376, 3584, 1, 256] - - [35, 67.587] - - - [30208, 1024, 1, 256] - - [63, 68.896] - - - [31232, 512, 1, 256] - - [34, 65.399] - - - [23040, 4096, 1, 256] - - [89, 73.309] - - - [35840, 4096, 1, 256] - - [27, 73.139] - - - [38144, 512, 1, 256] - - [36, 65.417] - - - [31744, 512, 1, 256] - - [76, 63.607] - - - [14592, 6912, 1, 256] - - [68, 73.121] - - - [19456, 11520, 1, 256] - - [27, 76.138] - - - [7168, 1024, 1, 256] - - [40, 53.439] - - - [18944, 11264, 1, 256] - - [42, 75.342] - - - [19712, 1024, 1, 256] - - [27, 63.902] - - - [26112, 9984, 1, 256] - - [37, 75.563] - - - [38656, 22784, 1, 256] - - [47, 74.655] - - - [24320, 8192, 1, 256] - - [47, 74.704] - - - [4864, 1024, 1, 256] - - [49, 52.923] - - - [20480, 4096, 1, 256] - - [37, 72.506] - - - [10240, 1024, 1, 256] - - [60, 62.923] - - - [31232, 15360, 1, 256] - - [24, 75.415] - - - [24320, 4096, 1, 256] - - [47, 72.813] - - - [33792, 1024, 1, 256] - - [58, 69.904] - - - [12032, 1024, 1, 256] - - [60, 63.985] - - - [39168, 512, 1, 256] - - [35, 64.9] - - - [16896, 4096, 1, 256] - - [27, 72.514] - - - [36096, 1024, 1, 256] - - [76, 66.757] - - - [28416, 12544, 1, 256] - - [51, 74.959] - - - [30720, 4096, 1, 256] - - [25, 73.127] - - - [19712, 4096, 1, 256] - - [41, 71.466] - - - [37120, 21248, 1, 256] - - [37, 74.681] - - - [16384, 4096, 1, 256] - - [27, 61.887] - - - [18688, 11008, 1, 256] - - [29, 75.091] - - - [38400, 8192, 1, 256] - - [37, 74.678] - - - [11264, 7424, 1, 256] - - [27, 74.924] - - - [23296, 512, 1, 256] - - [35, 61.999] - - - [25344, 512, 1, 256] - - [35, 61.71] - - - [44544, 256, 1, 256] - - [36, 57.866] - - - [43264, 4096, 1, 256] - - [54, 72.828] - - - [32512, 16640, 1, 256] - - [30, 75.357] - - - [39936, 8192, 1, 256] - - [29, 74.974] - - - [43264, 512, 1, 256] - - [35, 65.863] - - - [16640, 8704, 1, 256] - - [37, 74.811] - - - [26624, 8192, 1, 256] - - [37, 75.099] - - - [35328, 19456, 1, 256] - - [42, 75.247] - - - [42752, 26880, 1, 256] - - [25, 74.838] - - - [25344, 9216, 1, 256] - - [43, 73.445] - - - [34048, 8192, 1, 256] - - [43, 73.799] - - - [18688, 4096, 1, 256] - - [29, 72.394] - - - [37632, 8192, 1, 256] - - [54, 74.221] - - - [19968, 12032, 1, 256] - - [27, 75.67] - - - [8448, 4608, 1, 256] - - [49, 71.474] - - - [2048, 1536, 1, 256] - - [57, 49.785] - - - [31488, 15616, 1, 256] - - [37, 74.939] - - - [35328, 512, 1, 256] - - [83, 64.709] - - - [37376, 8192, 1, 256] - - [41, 74.732] - - - [33792, 8192, 1, 256] - - [37, 74.959] - - - [36608, 4096, 1, 256] - - [41, 72.898] - - - [28416, 8192, 1, 256] - - [47, 74.104] - - - [5632, 512, 1, 256] - - [69, 46.409] - - - [13568, 4096, 1, 256] - - [27, 71.849] - - - [17664, 9728, 1, 256] - - [57, 74.224] - - - [13568, 1024, 1, 256] - - [23, 64.791] - - - [8448, 512, 1, 256] - - [35, 46.517] - - - [22528, 4096, 1, 256] - - [29, 72.967] - - - [33536, 8192, 1, 256] - - [47, 74.444] - - - [23296, 1024, 1, 256] - - [23, 67.242] - - - [43520, 4096, 1, 256] - - [41, 73.78] - - - [39936, 23808, 1, 256] - - [37, 75.482] - - - [12544, 4096, 1, 256] - - [34, 71.348] - - - [22016, 4096, 1, 256] - - [74, 73.077] - - - [14592, 512, 1, 256] - - [23, 53.515] - - - [39936, 4096, 1, 256] - - [27, 73.242] - - - [18176, 1024, 1, 256] - - [40, 65.761] - - - [44800, 2048, 1, 256] - - [43, 70.237] - - - [14848, 4096, 1, 256] - - [27, 72.5] - - - [20224, 12288, 1, 256] - - [41, 74.421] - - - [16896, 8960, 1, 256] - - [37, 75.55] - - - [43264, 27392, 1, 256] - - [37, 74.802] - - - [24064, 16128, 1, 256] - - [47, 75.907] - - - [1024, 512, 1, 256] - - [109, 24.986] - - - [24576, 8448, 1, 256] - - [22, 70.765] - - - [25344, 9472, 1, 256] - - [28, 74.117] - - - [3328, 1536, 1, 256] - - [35, 54.557] - - - [31488, 4096, 1, 256] - - [27, 72.728] - - - [43008, 8192, 1, 256] - - [37, 74.904] - - - [28672, 12800, 1, 256] - - [25, 75.357] - - - [20736, 13056, 1, 256] - - [27, 75.359] - - - [17664, 9984, 1, 256] - - [57, 74.504] - - - [17920, 1024, 1, 256] - - [39, 65.342] - - - [11008, 1024, 1, 256] - - [59, 57.863] - - - [44800, 4096, 1, 256] - - [32, 72.511] - - - [29952, 14080, 1, 256] - - [24, 75.218] - - - [39168, 23296, 1, 256] - - [41, 74.594] - - - [9472, 512, 1, 256] - - [23, 51.769] - - - [27904, 8192, 1, 256] - - [28, 74.273] - - - [5120, 1024, 1, 256] - - [40, 54.163] - - - [15872, 7936, 1, 256] - - [25, 75.157] - - - [13568, 5632, 1, 256] - - [34, 73.703] - - - [17920, 9984, 1, 256] - - [37, 75.469] - - - [16640, 8960, 1, 256] - - [51, 74.67] - - - [41984, 4096, 1, 256] - - [27, 73.1] - - - [6912, 512, 1, 256] - - [53, 54.305] - - - [28416, 4096, 1, 256] - - [32, 72.453] - - - [27648, 11520, 1, 256] - - [29, 75.978] - - - [7680, 3840, 1, 256] - - [49, 70.372] - - - [34048, 4096, 1, 256] - - [42, 72.436] - - - [11264, 512, 1, 256] - - [34, 50.753] - - - [26368, 4096, 1, 256] - - [37, 72.565] - - - [21248, 13312, 1, 256] - - [27, 75.106] - - - [15104, 1024, 1, 256] - - [40, 64.971] - - - [35072, 18944, 1, 256] - - [27, 74.936] - - - [6144, 1024, 1, 256] - - [27, 54.464] - - - [44800, 8192, 1, 256] - - [37, 74.033] - - - [25088, 512, 1, 256] - - [60, 61.479] - - - [27904, 12032, 1, 256] - - [43, 74.855] - - - [27648, 1024, 1, 256] - - [40, 68.968] - - - [28928, 8192, 1, 256] - - [47, 74.128] - - - [29440, 13312, 1, 256] - - [30, 75.126] - - - [43264, 27136, 1, 256] - - [38, 74.753] - - - [23552, 512, 1, 256] - - [63, 62.539] - - - [26880, 10752, 1, 256] - - [29, 75.05] - - - [44032, 28160, 1, 256] - - [37, 75.553] - - - [36096, 512, 1, 256] - - [59, 63.214] - - - [4352, 2560, 1, 256] - - [55, 60.33] - - - [38912, 8192, 1, 256] - - [25, 74.954] - - - [12032, 4096, 1, 256] - - [25, 71.167] - - - [37632, 512, 1, 256] - - [40, 65.67] - - - [30208, 512, 1, 256] - - [35, 64.176] - - - [2304, 512, 1, 256] - - [121, 41.352] - - - [24320, 8448, 1, 256] - - [57, 74.743] - - - [39424, 512, 1, 256] - - [57, 65.209] - - - [37632, 21504, 1, 256] - - [38, 74.577] - - - [17152, 1024, 1, 256] - - [40, 66.735] - - - [22784, 15104, 1, 256] - - [51, 75.335] - - - [27904, 11776, 1, 256] - - [42, 74.855] - - - [43008, 26880, 1, 256] - - [38, 75.604] - - - [41728, 4096, 1, 256] - - [41, 72.671] - - - [25344, 8192, 1, 256] - - [41, 73.596] - - - [44800, 28928, 1, 256] - - [25, 74.343] - - - [38912, 22784, 1, 256] - - [38, 75.639] - - - [44032, 1024, 1, 256] - - [44, 70.885] - - - [30976, 4096, 1, 256] - - [28, 71.622] - - - [15872, 8192, 1, 256] - - [25, 75.001] - - - [40960, 4096, 1, 256] - - [32, 66.076] - - - [35584, 1024, 1, 256] - - [63, 69.271] - - - [18944, 4096, 1, 256] - - [76, 72.836] - - - [36096, 20224, 1, 256] - - [28, 74.266] - - - [11008, 7168, 1, 256] - - [85, 71.456] - - - [7936, 1024, 1, 256] - - [27, 56.833] - - - [44288, 1024, 1, 256] - - [39, 69.726] - - - [38656, 8192, 1, 256] - - [41, 74.287] - - - [38144, 1024, 1, 256] - - [29, 69.01] - - - [41984, 1024, 1, 256] - - [58, 70.499] - - - [20736, 512, 1, 256] - - [60, 63.106] - - - [32768, 16640, 1, 256] - - [88, 59.277] - - - [40960, 1024, 1, 256] - - [25, 64.607] - - - [25856, 9984, 1, 256] - - [25, 74.885] - - - [29696, 13824, 1, 256] - - [29, 75.882] - - - [37120, 4096, 1, 256] - - [41, 73.15] - - - [37120, 20992, 1, 256] - - [25, 74.854] - - - [35072, 512, 1, 256] - - [36, 66.521] - - - [38656, 1024, 1, 256] - - [36, 69.426] - - - [37376, 512, 1, 256] - - [23, 65.878] - - - [32000, 16128, 1, 256] - - [29, 74.91] - - - [41984, 25856, 1, 256] - - [38, 75.626] - - - [23040, 15104, 1, 256] - - [30, 76.021] - - - [31232, 15104, 1, 256] - - [51, 75.761] - - - [25088, 4096, 1, 256] - - [61, 73.174] - - - [15360, 7424, 1, 256] - - [25, 75.274] - - - [16384, 8448, 1, 256] - - [22, 63.356] - - - [26624, 4096, 1, 256] - - [27, 73.175] - - - [14080, 6400, 1, 256] - - [85, 73.029] - - - [16128, 4096, 1, 256] - - [27, 72.085] - - - [43776, 27904, 1, 256] - - [41, 74.364] - - - [15872, 512, 1, 256] - - [60, 57.075] - - - [43776, 8192, 1, 256] - - [41, 74.087] - - - [10496, 6656, 1, 256] - - [35, 73.241] - - - [13312, 512, 1, 256] - - [27, 58.378] - - - [29184, 512, 1, 256] - - [36, 62.86] - - - [15360, 7680, 1, 256] - - [25, 75.288] - - - [40192, 8192, 1, 256] - - [25, 74.473] - - - [34560, 8192, 1, 256] - - [29, 74.307] - - - [25856, 8192, 1, 256] - - [25, 74.455] - - - [32512, 16384, 1, 256] - - [42, 75.174] - - - [12288, 4352, 1, 256] - - [25, 73.324] - - - [29440, 13568, 1, 256] - - [51, 75.133] - - - [28160, 1024, 1, 256] - - [23, 68.878] - - - [32768, 4096, 1, 256] - - [66, 58.306] - - - [24832, 4096, 1, 256] - - [41, 72.684] - - - [39680, 23808, 1, 256] - - [37, 74.405] - - - [22784, 4096, 1, 256] - - [61, 72.548] - - - [7936, 4096, 1, 256] - - [40, 69.283] - - - [8704, 4864, 1, 256] - - [55, 72.098] - - - [29696, 512, 1, 256] - - [63, 63.904] - - - [39424, 23296, 1, 256] - - [22, 75.413] - - - [17408, 9472, 1, 256] - - [25, 75.849] - - - [33792, 4096, 1, 256] - - [54, 72.911] - - - [17920, 512, 1, 256] - - [53, 56.82] - - - [25856, 512, 1, 256] - - [35, 62.098] - - - [44288, 28160, 1, 256] - - [22, 74.632] - - - [40192, 4096, 1, 256] - - [74, 73.012] - - - [21248, 512, 1, 256] - - [60, 58.27] - - - [3072, 512, 1, 256] - - [183, 44.68] - - - [29184, 13312, 1, 256] - - [25, 75.113] - - - [44544, 1024, 1, 256] - - [60, 70.345] - - - [37888, 21760, 1, 256] - - [27, 75.792] - - - [33792, 17920, 1, 256] - - [27, 75.738] - - - [6912, 1024, 1, 256] - - [60, 60.997] - - - [41216, 512, 1, 256] - - [39, 66.051] - - - [42240, 26112, 1, 256] - - [29, 74.753] - - - [30720, 8192, 1, 256] - - [37, 75.059] - - - [11776, 1024, 1, 256] - - [35, 62.652] - - - [43008, 4096, 1, 256] - - [38, 73.053] - - - [34560, 18688, 1, 256] - - [29, 74.923] - - - [41984, 512, 1, 256] - - [44, 67.355] - - - [41728, 512, 1, 256] - - [29, 67.03] - - - [2560, 1792, 1, 256] - - [35, 49.744] - - - [36864, 8192, 1, 256] - - [27, 74.544] - - - [40704, 8192, 1, 256] - - [27, 74.341] - - - [30720, 14592, 1, 256] - - [25, 75.998] - - - [32256, 512, 1, 256] - - [35, 64.09] - - - [40192, 512, 1, 256] - - [27, 65.943] - - - [8960, 512, 1, 256] - - [49, 49.372] - - - [16640, 4096, 1, 256] - - [48, 71.565] - - - [30976, 15104, 1, 256] - - [41, 74.383] - - - [27136, 8192, 1, 256] - - [30, 74.986] - - - [30208, 8192, 1, 256] - - [47, 74.616] - - - [21504, 512, 1, 256] - - [27, 58.901] - - - [9728, 5888, 1, 256] - - [34, 73.206] - - - [38912, 23040, 1, 256] - - [38, 75.609] - - - [7424, 1024, 1, 256] - - [34, 54.579] - - - [38656, 22528, 1, 256] - - [28, 74.572] - - - [26880, 512, 1, 256] - - [35, 63.666] - - - [29184, 13056, 1, 256] - - [47, 75.458] - - - [44032, 27904, 1, 256] - - [25, 75.577] - - - [38144, 8192, 1, 256] - - [41, 74.149] - - - [29952, 512, 1, 256] - - [35, 63.804] - - - [18432, 4096, 1, 256] - - [27, 72.431] - - - [28160, 12288, 1, 256] - - [28, 74.924] - - - [29696, 1024, 1, 256] - - [53, 69.307] - - - [39936, 1024, 1, 256] - - [58, 70.178] - - - [25600, 512, 1, 256] - - [37, 61.911] - - - [40448, 24320, 1, 256] - - [24, 75.043] - - - [40448, 512, 1, 256] - - [44, 66.253] - - - [7424, 3584, 1, 256] - - [55, 69.863] - - - [5376, 512, 1, 256] - - [33, 43.993] - - - [27136, 4096, 1, 256] - - [43, 73.399] - - - [35840, 19968, 1, 256] - - [27, 75.549] - - - [18944, 11008, 1, 256] - - [59, 75.428] - - - [34816, 18688, 1, 256] - - [27, 75.739] - - - [38400, 1024, 1, 256] - - [58, 70.053] - - - [36352, 20480, 1, 256] - - [28, 75.176] - - - [36608, 20736, 1, 256] - - [22, 74.849] - - - [28672, 1024, 1, 256] - - [29, 67.129] - - - [42496, 26624, 1, 256] - - [41, 75.087] - - - [31488, 15360, 1, 256] - - [27, 74.816] - - - [20992, 4096, 1, 256] - - [28, 73.044] - - - [12544, 512, 1, 256] - - [23, 55.681] - - - [24064, 8192, 1, 256] - - [47, 74.93] - - - [26880, 8192, 1, 256] - - [27, 74.524] - - - [4352, 512, 1, 256] - - [72, 36.778] - - - [7680, 1024, 1, 256] - - [60, 55.641] - - - [16128, 8192, 1, 256] - - [25, 74.553] - - - [39168, 8192, 1, 256] - - [28, 74.205] - - - [29440, 4096, 1, 256] - - [47, 72.927] - - - [33536, 4096, 1, 256] - - [41, 73.191] - - - [33024, 17152, 1, 256] - - [47, 75.207] - - - [34816, 18944, 1, 256] - - [25, 75.801] - - - [22016, 512, 1, 256] - - [34, 59.438] - - - [14848, 6912, 1, 256] - - [25, 74.685] - - - [20736, 12800, 1, 256] - - [25, 75.324] - - - [32256, 16128, 1, 256] - - [30, 75.679] - - - [7680, 512, 1, 256] - - [40, 43.371] - - - [19968, 12288, 1, 256] - - [41, 75.413] - - - [29184, 4096, 1, 256] - - [74, 73.343] - - - [15616, 4096, 1, 256] - - [29, 72.242] - - - [44544, 28672, 1, 256] - - [22, 74.533] - - - [26112, 4096, 1, 256] - - [65, 73.758] - - - [26624, 10752, 1, 256] - - [29, 75.829] - - - [15104, 4096, 1, 256] - - [27, 71.987] - - - [23296, 4096, 1, 256] - - [29, 72.489] - - - [37888, 22016, 1, 256] - - [29, 75.62] - - - [11520, 7680, 1, 256] - - [59, 73.86] - - - [41728, 1024, 1, 256] - - [27, 69.281] - - - [2304, 1792, 1, 256] - - [35, 45.323] - - - [34048, 17920, 1, 256] - - [24, 74.538] - - - [1536, 768, 1, 256] - - [121, 40.855] - - - [33280, 8192, 1, 256] - - [24, 74.883] - - - [11264, 1024, 1, 256] - - [40, 61.449] - - - [21760, 1024, 1, 256] - - [44, 66.154] - - - [18432, 10496, 1, 256] - - [27, 75.833] - - - [41216, 4096, 1, 256] - - [43, 73.138] - - - [41472, 25344, 1, 256] - - [37, 75.153] - - - [17408, 1024, 1, 256] - - [53, 66.969] - - - [19456, 1024, 1, 256] - - [63, 65.93] - - - [36096, 19968, 1, 256] - - [41, 74.202] - - - [8704, 512, 1, 256] - - [60, 48.105] - - - [30464, 1024, 1, 256] - - [59, 66.822] - - - [8192, 1024, 1, 256] - - [27, 57.825] - - - [11520, 512, 1, 256] - - [25, 51.67] - - - [44544, 512, 1, 256] - - [31, 66.525] - - - [20736, 4096, 1, 256] - - [25, 72.521] - - - [42752, 8192, 1, 256] - - [29, 74.408] - - - [39936, 512, 1, 256] - - [40, 65.84] - - - [42496, 26368, 1, 256] - - [27, 75.25] - - - [28672, 4096, 1, 256] - - [37, 72.526] - - - [35840, 8192, 1, 256] - - [29, 74.949] - - - [17664, 1024, 1, 256] - - [36, 63.994] - - - [21248, 4096, 1, 256] - - [37, 72.544] - - - [1280, 768, 1, 256] - - [121, 35.43] - - - [28160, 512, 1, 256] - - [58, 65.774] - - - [34304, 18176, 1, 256] - - [42, 75.541] - - - [19200, 11520, 1, 256] - - [27, 75.066] - - - [25856, 9728, 1, 256] - - [27, 74.614] - - - [35328, 19200, 1, 256] - - [51, 75.548] - - - [29440, 8192, 1, 256] - - [24, 74.577] - - - [20992, 13056, 1, 256] - - [25, 75.803] - - - [21760, 512, 1, 256] - - [35, 59.218] - - - [12800, 512, 1, 256] - - [40, 56.426] - - - [28416, 12288, 1, 256] - - [30, 74.006] - - - [29696, 13568, 1, 256] - - [27, 75.794] - - - [21504, 4096, 1, 256] - - [28, 72.907] - - - [30464, 14592, 1, 256] - - [47, 74.281] - - - [13056, 5120, 1, 256] - - [49, 73.164] - - - [34560, 4096, 1, 256] - - [29, 72.865] - - - [32768, 16896, 1, 256] - - [88, 59.383] - - - [13824, 5888, 1, 256] - - [22, 73.544] - - - [33024, 8192, 1, 256] - - [43, 74.541] - - - [14080, 4096, 1, 256] - - [47, 70.417] - - - [43008, 1024, 1, 256] - - [36, 70.968] - - - [31744, 1024, 1, 256] - - [40, 69.565] - - - [11008, 512, 1, 256] - - [34, 49.571] - - - [24832, 8192, 1, 256] - - [31, 74.582] - - - [43776, 512, 1, 256] - - [33, 63.657] - - - [24064, 1024, 1, 256] - - [72, 68.455] - - - [12800, 4096, 1, 256] - - [74, 71.929] - - - [19456, 11776, 1, 256] - - [25, 75.91] - - - [22528, 14848, 1, 256] - - [27, 75.994] - - - [30208, 14080, 1, 256] - - [24, 75.797] - - - [40704, 1024, 1, 256] - - [72, 69.656] - - - [35584, 4096, 1, 256] - - [43, 73.157] - - - [26112, 8192, 1, 256] - - [29, 74.852] - - - [9472, 5632, 1, 256] - - [34, 72.733] - - - [15616, 512, 1, 256] - - [34, 56.189] - - - [34816, 4096, 1, 256] - - [29, 73.248] - - - [31232, 4096, 1, 256] - - [41, 73.472] - - - [9728, 1024, 1, 256] - - [49, 60.648] - - - [13312, 1024, 1, 256] - - [63, 64.256] - - - [20224, 1024, 1, 256] - - [60, 66.946] - - - [4864, 512, 1, 256] - - [55, 40.82] - - - [34304, 4096, 1, 256] - - [42, 73.894] - - - [43776, 1024, 1, 256] - - [76, 67.204] - - - [37120, 8192, 1, 256] - - [25, 74.285] - - - [33792, 512, 1, 256] - - [34, 65.761] - - - [42496, 512, 1, 256] - - [25, 66.06] - - - [9216, 512, 1, 256] - - [25, 50.463] - - - [14336, 4096, 1, 256] - - [27, 72.202] - - - [43008, 27136, 1, 256] - - [22, 75.512] - - - [35840, 512, 1, 256] - - [49, 64.842] - - - [40960, 25088, 1, 256] - - [38, 67.979] - - - [17408, 512, 1, 256] - - [49, 60.954] - - - [12288, 4096, 1, 256] - - [27, 71.721] - - - [6656, 512, 1, 256] - - [25, 52.846] - - - [40960, 24832, 1, 256] - - [38, 68.288] - - - [39168, 23040, 1, 256] - - [28, 74.625] - - - [512, 1, 1, 128] - - [205, 0.025] - - - [384, 1, 1, 384] - - [418, 0.028] - - - [256, 1, 1, 256] - - [205, 0.017] - - - [128, 1, 1, 128] - - [205, 0.006] - - - [640, 1, 1, 128] - - [205, 0.032] - - - [1, 128, 1, 256] - - [205, 0.009] - - - [512, 128, 1, 256] - - [113, 4.27] - - - [2049, 128, 1, 256] - - [180, 14.336] - - - [49, 128, 1, 256] - - [158, 0.348] - - - [1537, 128, 1, 256] - - [184, 11.345] - - - [257, 128, 1, 256] - - [158, 1.888] - - - [9728, 128, 1, 256] - - [121, 42.487] - - - [3840, 128, 1, 256] - - [114, 23.522] - - - [1280, 128, 1, 256] - - [113, 10.616] - - - [7168, 128, 1, 256] - - [108, 32.965] - - - [6656, 128, 1, 256] - - [121, 37.792] - - - [2561, 128, 1, 256] - - [118, 17.834] - - - [6912, 128, 1, 256] - - [108, 38.946] - - - [2048, 128, 1, 256] - - [116, 15.996] - - - [2304, 128, 1, 256] - - [116, 17.807] - - - [1536, 128, 1, 256] - - [166, 12.739] - - - [4864, 128, 1, 256] - - [108, 29.428] - - - [8448, 128, 1, 256] - - [108, 38.138] - - - [3072, 128, 1, 256] - - [116, 23.02] - - - [3329, 128, 1, 256] - - [118, 22.543] - - - [3328, 128, 1, 256] - - [110, 24.565] - - - [8960, 128, 1, 256] - - [109, 39.721] - - - [9216, 128, 1, 256] - - [185, 38.761] - - - [2817, 128, 1, 256] - - [150, 19.524] - - - [6400, 128, 1, 256] - - [171, 37.641] - - - [561, 128, 1, 256] - - [118, 3.87] - - - [2816, 128, 1, 256] - - [114, 21.21] - - - [3073, 128, 1, 256] - - [160, 20.073] - - - [2097, 128, 1, 256] - - [118, 14.071] - - - [768, 128, 1, 256] - - [166, 6.405] - - - [9984, 128, 1, 256] - - [114, 42.97] - - - [3584, 128, 1, 256] - - [109, 22.61] - - - [817, 128, 1, 256] - - [128, 5.636] - - - [5632, 128, 1, 256] - - [109, 33.525] - - - [9472, 128, 1, 256] - - [114, 41.248] - - - [2305, 128, 1, 256] - - [128, 15.9] - - - [1329, 128, 1, 256] - - [118, 9.168] - - - [5888, 128, 1, 256] - - [110, 34.907] - - - [7680, 128, 1, 256] - - [177, 33.944] - - - [4608, 128, 1, 256] - - [114, 28.462] - - - [2353, 128, 1, 256] - - [165, 15.789] - - - [5120, 128, 1, 256] - - [108, 30.601] - - - [769, 128, 1, 256] - - [184, 5.733] - - - [1792, 128, 1, 256] - - [118, 14.072] - - - [1073, 128, 1, 256] - - [118, 7.437] - - - [513, 128, 1, 256] - - [132, 3.806] - - - [4096, 128, 1, 256] - - [110, 25.3] - - - [7424, 128, 1, 256] - - [123, 34.356] - - - [4352, 128, 1, 256] - - [110, 26.77] - - - [1793, 128, 1, 256] - - [118, 12.725] - - - [8192, 128, 1, 256] - - [186, 35.992] - - - [1281, 128, 1, 256] - - [130, 9.551] - - - [305, 128, 1, 256] - - [169, 2.104] - - - [2560, 128, 1, 256] - - [116, 19.481] - - - [2609, 128, 1, 256] - - [118, 17.831] - - - [1585, 128, 1, 256] - - [150, 10.783] - - - [8704, 128, 1, 256] - - [109, 37.244] - - - [10240, 128, 1, 256] - - [187, 41.644] - - - [256, 128, 1, 256] - - [115, 2.135] - - - [1025, 128, 1, 256] - - [184, 7.604] - - - [2865, 128, 1, 256] - - [143, 19.051] - - - [5376, 128, 1, 256] - - [110, 32.001] - - - [1841, 128, 1, 256] - - [165, 12.41] - - - [7936, 128, 1, 256] - - [108, 36.158] - - - [6144, 128, 1, 256] - - [110, 35.292] - - - [1024, 128, 1, 256] - - [115, 8.54] - - - [36096, 1281, 1, 256] - - [90, 63.047] - - - [38656, 2816, 1, 256] - - [74, 72.503] - - - [35072, 2048, 1, 256] - - [74, 71.133] - - - [39424, 2865, 1, 256] - - [54, 70.508] - - - [39168, 3328, 1, 256] - - [28, 72.947] - - - [36096, 2865, 1, 256] - - [51, 68.618] - - - [39216, 5632, 1, 256] - - [58, 62.951] - - - [38144, 6144, 1, 256] - - [27, 74.027] - - - [35328, 3072, 1, 256] - - [47, 73.434] - - - [39936, 256, 1, 256] - - [41, 59.512] - - - [36864, 3328, 1, 256] - - [29, 72.842] - - - [39168, 6144, 1, 256] - - [41, 73.921] - - - [36352, 4352, 1, 256] - - [29, 74.346] - - - [37680, 10240, 1, 256] - - [41, 62.962] - - - [38144, 256, 1, 256] - - [58, 57.666] - - - [37632, 1281, 1, 256] - - [74, 63.847] - - - [35632, 1792, 1, 256] - - [58, 62.655] - - - [36096, 4096, 1, 256] - - [41, 72.242] - - - [36144, 2816, 1, 256] - - [36, 62.834] - - - [36352, 256, 1, 256] - - [25, 55.862] - - - [35888, 2865, 1, 256] - - [53, 61.552] - - - [38912, 1280, 1, 256] - - [29, 71.695] - - - [37120, 3072, 1, 256] - - [25, 73.092] - - - [38448, 10240, 1, 256] - - [61, 62.59] - - - [39936, 3328, 1, 256] - - [27, 73.571] - - - [39168, 10240, 1, 256] - - [28, 74.409] - - - [39680, 3329, 1, 256] - - [54, 69.552] - - - [37168, 2865, 1, 256] - - [53, 61.603] - - - [38144, 5888, 1, 256] - - [37, 73.803] - - - [37120, 1281, 1, 256] - - [74, 64.097] - - - [37376, 10240, 1, 256] - - [25, 74.986] - - - [38704, 5120, 1, 256] - - [44, 62.977] - - - [39168, 5376, 1, 256] - - [28, 73.875] - - - [38656, 2865, 1, 256] - - [54, 69.906] - - - [37376, 3584, 1, 256] - - [25, 74.017] - - - [35072, 6144, 1, 256] - - [25, 74.189] - - - [39936, 6144, 1, 256] - - [37, 74.671] - - - [37632, 5376, 1, 256] - - [25, 73.89] - - - [36352, 2304, 1, 256] - - [25, 73.025] - - - [35840, 2048, 1, 256] - - [86, 71.487] - - - [36608, 1280, 1, 256] - - [25, 70.589] - - - [39936, 1792, 1, 256] - - [27, 73.01] - - - [36608, 3329, 1, 256] - - [54, 69.48] - - - [35072, 3329, 1, 256] - - [54, 69.553] - - - [37168, 3584, 1, 256] - - [44, 62.89] - - - [36096, 1792, 1, 256] - - [31, 70.106] - - - [39424, 3329, 1, 256] - - [41, 70.143] - - - [39424, 2048, 1, 256] - - [41, 71.915] - - - [39984, 2865, 1, 256] - - [36, 61.483] - - - [38448, 256, 1, 256] - - [102, 51.782] - - - [35584, 256, 1, 256] - - [44, 55.093] - - - [36608, 10240, 1, 256] - - [29, 74.583] - - - [38960, 5376, 1, 256] - - [28, 62.631] - - - [36352, 2048, 1, 256] - - [41, 71.757] - - - [39680, 1281, 1, 256] - - [27, 64.293] - - - [36608, 2304, 1, 256] - - [29, 72.296] - - - [39936, 1280, 1, 256] - - [27, 71.595] - - - [39680, 5376, 1, 256] - - [37, 74.025] - - - [35584, 10240, 1, 256] - - [24, 74.399] - - - [36864, 512, 1, 256] - - [49, 65.314] - - - [39424, 2816, 1, 256] - - [51, 73.142] - - - [35840, 2816, 1, 256] - - [25, 73.659] - - - [38192, 2816, 1, 256] - - [36, 62.921] - - - [35584, 2048, 1, 256] - - [41, 70.786] - - - [37936, 2865, 1, 256] - - [29, 61.41] - - - [39936, 2865, 1, 256] - - [29, 70.99] - - - [38656, 10240, 1, 256] - - [41, 74.41] - - - [36608, 2048, 1, 256] - - [61, 70.999] - - - [35120, 2816, 1, 256] - - [53, 63.338] - - - [39424, 5888, 1, 256] - - [25, 74.33] - - - [37680, 2816, 1, 256] - - [36, 63.798] - - - [36096, 6144, 1, 256] - - [41, 73.136] - - - [38144, 1281, 1, 256] - - [91, 64.27] - - - [37632, 2048, 1, 256] - - [61, 70.581] - - - [39680, 256, 1, 256] - - [58, 59.321] - - - [37680, 3840, 1, 256] - - [53, 63.791] - - - [39168, 2816, 1, 256] - - [41, 72.679] - - - [38192, 2865, 1, 256] - - [53, 60.856] - - - [38912, 4608, 1, 256] - - [27, 74.158] - - - [37120, 2048, 1, 256] - - [41, 71.342] - - - [35376, 1536, 1, 256] - - [36, 61.119] - - - [38448, 4864, 1, 256] - - [36, 63.358] - - - [38192, 10240, 1, 256] - - [74, 62.37] - - - [37632, 2816, 1, 256] - - [29, 72.628] - - - [39424, 1024, 1, 256] - - [76, 70.073] - - - [39168, 256, 1, 256] - - [58, 58.604] - - - [39984, 6144, 1, 256] - - [41, 63.37] - - - [38144, 4608, 1, 256] - - [37, 73.45] - - - [35840, 2865, 1, 256] - - [27, 71.032] - - - [36352, 6144, 1, 256] - - [41, 74.483] - - - [36864, 768, 1, 256] - - [36, 67.906] - - - [37888, 3328, 1, 256] - - [25, 73.568] - - - [36912, 3328, 1, 256] - - [41, 61.472] - - - [37120, 3584, 1, 256] - - [25, 73.471] - - - [38912, 1281, 1, 256] - - [27, 64.419] - - - [39472, 256, 1, 256] - - [29, 52.657] - - - [39936, 1281, 1, 256] - - [28, 64.808] - - - [37376, 5120, 1, 256] - - [37, 74.278] - - - [37888, 2048, 1, 256] - - [28, 71.546] - - - [37632, 1280, 1, 256] - - [29, 70.731] - - - [35376, 2816, 1, 256] - - [58, 63.641] - - - [38656, 3329, 1, 256] - - [30, 69.446] - - - [36912, 256, 1, 256] - - [29, 51.073] - - - [39168, 768, 1, 256] - - [69, 68.628] - - - [37424, 256, 1, 256] - - [37, 51.135] - - - [38448, 2816, 1, 256] - - [53, 63.43] - - - [35840, 3840, 1, 256] - - [25, 74.545] - - - [38912, 2865, 1, 256] - - [22, 71.049] - - - [36096, 1280, 1, 256] - - [57, 68.901] - - - [35328, 1024, 1, 256] - - [76, 69.48] - - - [39680, 3328, 1, 256] - - [29, 72.712] - - - [36352, 2816, 1, 256] - - [25, 73.066] - - - [38912, 256, 1, 256] - - [29, 58.888] - - - [39424, 3328, 1, 256] - - [28, 73.378] - - - [35888, 2816, 1, 256] - - [58, 63.218] - - - [36096, 2816, 1, 256] - - [64, 71.255] - - - [38960, 10240, 1, 256] - - [41, 62.954] - - - [35840, 3584, 1, 256] - - [27, 74.213] - - - [39424, 5120, 1, 256] - - [37, 74.236] - - - [37376, 1024, 1, 256] - - [72, 69.945] - - - [37632, 4096, 1, 256] - - [27, 72.873] - - - [36400, 2865, 1, 256] - - [53, 61.927] - - - [36144, 2560, 1, 256] - - [53, 63.37] - - - [36864, 1281, 1, 256] - - [27, 63.521] - - - [39424, 5376, 1, 256] - - [37, 74.471] - - - [36400, 2816, 1, 256] - - [58, 63.396] - - - [38656, 6144, 1, 256] - - [30, 73.826] - - - [37888, 5632, 1, 256] - - [27, 75.071] - - - [36912, 2865, 1, 256] - - [27, 59.852] - - - [38656, 4352, 1, 256] - - [30, 73.699] - - - [37632, 1536, 1, 256] - - [37, 70.658] - - - [35072, 2865, 1, 256] - - [29, 70.225] - - - [35888, 2304, 1, 256] - - [36, 63.831] - - - [38912, 3329, 1, 256] - - [54, 70.513] - - - [37680, 4096, 1, 256] - - [61, 62.597] - - - [38400, 6144, 1, 256] - - [28, 74.375] - - - [37888, 3840, 1, 256] - - [27, 74.613] - - - [36608, 3328, 1, 256] - - [28, 72.821] - - - [35328, 256, 1, 256] - - [84, 55.159] - - - [36096, 3329, 1, 256] - - [28, 68.384] - - - [37888, 5888, 1, 256] - - [37, 74.629] - - - [36864, 3329, 1, 256] - - [54, 69.763] - - - [35632, 256, 1, 256] - - [44, 50.1] - - - [38656, 4864, 1, 256] - - [47, 73.807] - - - [37888, 2816, 1, 256] - - [29, 73.767] - - - [37120, 3328, 1, 256] - - [28, 72.974] - - - [35328, 1536, 1, 256] - - [31, 70.794] - - - [35328, 1280, 1, 256] - - [57, 70.894] - - - [35888, 10240, 1, 256] - - [28, 63.793] - - - [36400, 10240, 1, 256] - - [41, 62.71] - - - [35072, 10240, 1, 256] - - [25, 74.629] - - - [39680, 2816, 1, 256] - - [37, 72.727] - - - [35584, 3329, 1, 256] - - [73, 69.36] - - - [36656, 256, 1, 256] - - [58, 50.805] - - - [38144, 4096, 1, 256] - - [74, 73.033] - - - [39936, 2816, 1, 256] - - [25, 73.724] - - - [36864, 3072, 1, 256] - - [25, 73.23] - - - [37936, 2816, 1, 256] - - [53, 63.534] - - - [37632, 3584, 1, 256] - - [25, 73.496] - - - [39984, 10240, 1, 256] - - [41, 63.649] - - - [38656, 512, 1, 256] - - [53, 66.314] - - - [35328, 10240, 1, 256] - - [51, 74.983] - - - [36096, 2048, 1, 256] - - [28, 69.762] - - - [37120, 4864, 1, 256] - - [25, 73.988] - - - [35840, 10240, 1, 256] - - [29, 75.284] - - - [39680, 5632, 1, 256] - - [37, 74.307] - - - [38144, 4352, 1, 256] - - [29, 73.886] - - - [36400, 2560, 1, 256] - - [36, 63.434] - - - [35840, 3329, 1, 256] - - [54, 70.438] - - - [37424, 10240, 1, 256] - - [41, 62.738] - - - [38912, 10240, 1, 256] - - [25, 75.232] - - - [35072, 768, 1, 256] - - [44, 68.0] - - - [36096, 3840, 1, 256] - - [54, 72.331] - - - [36656, 3072, 1, 256] - - [27, 62.236] - - - [39680, 1536, 1, 256] - - [32, 70.578] - - - [36656, 2865, 1, 256] - - [36, 61.222] - - - [38912, 512, 1, 256] - - [27, 65.257] - - - [38400, 256, 1, 256] - - [25, 57.601] - - - [38704, 10240, 1, 256] - - [28, 62.954] - - - [38912, 5376, 1, 256] - - [25, 74.867] - - - [35120, 256, 1, 256] - - [58, 53.951] - - - [38656, 3328, 1, 256] - - [28, 72.917] - - - [37888, 1536, 1, 256] - - [37, 71.305] - - - [39216, 5376, 1, 256] - - [36, 63.737] - - - [37376, 3329, 1, 256] - - [32, 70.049] - - - [37680, 256, 1, 256] - - [53, 51.618] - - - [39680, 6144, 1, 256] - - [29, 74.112] - - - [38400, 2865, 1, 256] - - [54, 70.515] - - - [36608, 2865, 1, 256] - - [54, 69.97] - - - [38912, 768, 1, 256] - - [40, 69.012] - - - [35584, 1792, 1, 256] - - [34, 71.502] - - - [39424, 256, 1, 256] - - [25, 58.69] - - - [36352, 1281, 1, 256] - - [74, 64.906] - - - [38400, 2048, 1, 256] - - [28, 71.748] - - - [38144, 3329, 1, 256] - - [54, 69.524] - - - [39680, 2048, 1, 256] - - [53, 69.64] - - - [38656, 256, 1, 256] - - [25, 57.937] - - - [39728, 2816, 1, 256] - - [58, 63.085] - - - [36352, 3329, 1, 256] - - [29, 70.011] - - - [38400, 10240, 1, 256] - - [25, 74.925] - - - [39984, 6400, 1, 256] - - [41, 63.356] - - - [37888, 4352, 1, 256] - - [37, 74.699] - - - [37888, 4096, 1, 256] - - [37, 73.073] - - - [35584, 1536, 1, 256] - - [35, 70.284] - - - [36096, 256, 1, 256] - - [36, 55.392] - - - [36864, 2048, 1, 256] - - [27, 68.862] - - - [36144, 2865, 1, 256] - - [53, 61.056] - - - [35584, 3584, 1, 256] - - [37, 73.116] - - - [35072, 1024, 1, 256] - - [40, 69.616] - - - [36352, 3328, 1, 256] - - [41, 73.516] - - - [39424, 1281, 1, 256] - - [91, 65.011] - - - [39728, 10240, 1, 256] - - [41, 62.94] - - - [37632, 2865, 1, 256] - - [54, 70.226] - - - [37168, 3328, 1, 256] - - [36, 63.032] - - - [37376, 5376, 1, 256] - - [37, 74.538] - - - [35328, 2865, 1, 256] - - [70, 70.435] - - - [35584, 6144, 1, 256] - - [25, 73.895] - - - [38704, 2816, 1, 256] - - [53, 63.386] - - - [36608, 3072, 1, 256] - - [37, 73.008] - - - [39680, 1280, 1, 256] - - [25, 70.822] - - - [35328, 1281, 1, 256] - - [74, 64.856] - - - [36608, 512, 1, 256] - - [40, 65.036] - - - [39936, 1536, 1, 256] - - [32, 71.413] - - - [39728, 5888, 1, 256] - - [58, 63.022] - - - [39168, 1281, 1, 256] - - [78, 64.281] - - - [37120, 256, 1, 256] - - [58, 56.339] - - - [38960, 2865, 1, 256] - - [27, 61.03] - - - [39168, 5120, 1, 256] - - [24, 73.651] - - - [36864, 256, 1, 256] - - [58, 56.308] - - - [36912, 2816, 1, 256] - - [28, 61.151] - - - [36096, 2304, 1, 256] - - [57, 70.752] - - - [35840, 3328, 1, 256] - - [27, 73.553] - - - [38704, 2865, 1, 256] - - [58, 61.61] - - - [38144, 1792, 1, 256] - - [37, 71.784] - - - [36608, 2560, 1, 256] - - [29, 72.736] - - - [35376, 10240, 1, 256] - - [41, 63.016] - - - [35840, 2304, 1, 256] - - [37, 73.298] - - - [35840, 1280, 1, 256] - - [37, 71.379] - - - [37376, 1280, 1, 256] - - [37, 71.316] - - - [35584, 3328, 1, 256] - - [42, 72.682] - - - [35584, 2865, 1, 256] - - [32, 69.888] - - - [39936, 10240, 1, 256] - - [27, 75.215] - - - [38912, 5120, 1, 256] - - [37, 74.698] - - - [37632, 3329, 1, 256] - - [32, 69.618] - - - [37888, 1792, 1, 256] - - [27, 72.793] - - - [36608, 1281, 1, 256] - - [28, 64.424] - - - [38192, 4352, 1, 256] - - [58, 62.726] - - - [39936, 2048, 1, 256] - - [41, 72.048] - - - [35072, 1281, 1, 256] - - [78, 64.274] - - - [39472, 2816, 1, 256] - - [53, 63.357] - - - [39728, 2865, 1, 256] - - [36, 61.192] - - - [38400, 2816, 1, 256] - - [25, 73.171] - - - [38400, 4608, 1, 256] - - [42, 73.988] - - - [39216, 10240, 1, 256] - - [61, 62.289] - - - [35072, 3072, 1, 256] - - [25, 72.885] - - - [38400, 4352, 1, 256] - - [29, 74.322] - - - [39216, 2816, 1, 256] - - [58, 63.182] - - - [35840, 1792, 1, 256] - - [29, 72.46] - - - [35632, 2048, 1, 256] - - [36, 62.969] - - - [38704, 256, 1, 256] - - [36, 52.174] - - - [37888, 3329, 1, 256] - - [29, 70.547] - - - [37888, 6144, 1, 256] - - [27, 74.7] - - - [37376, 6144, 1, 256] - - [41, 74.51] - - - [37376, 256, 1, 256] - - [31, 56.495] - - - [36400, 256, 1, 256] - - [36, 50.53] - - - [37936, 4096, 1, 256] - - [41, 62.768] - - - [38144, 10240, 1, 256] - - [37, 74.524] - - - [35376, 1792, 1, 256] - - [36, 62.315] - - - [37168, 10240, 1, 256] - - [61, 62.52] - - - [39984, 2816, 1, 256] - - [44, 63.532] - - - [37168, 2816, 1, 256] - - [53, 62.931] - - - [39424, 5632, 1, 256] - - [27, 74.685] - - - [36352, 1280, 1, 256] - - [38, 70.957] - - - [39680, 10240, 1, 256] - - [29, 74.481] - - - [38144, 3328, 1, 256] - - [65, 72.575] - - - [39168, 2048, 1, 256] - - [41, 71.015] - - - [35328, 6144, 1, 256] - - [43, 74.535] - - - [35632, 2865, 1, 256] - - [53, 61.238] - - - [36656, 10240, 1, 256] - - [41, 63.069] - - - [36608, 4352, 1, 256] - - [37, 73.899] - - - [35120, 2865, 1, 256] - - [44, 61.404] - - - [36608, 6144, 1, 256] - - [27, 74.021] - - - [37888, 2865, 1, 256] - - [37, 70.976] - - - [39168, 1024, 1, 256] - - [69, 69.668] - - - [38704, 4864, 1, 256] - - [44, 63.525] - - - [39168, 2865, 1, 256] - - [29, 69.967] - - - [38960, 5120, 1, 256] - - [28, 62.521] - - - [36864, 2816, 1, 256] - - [25, 73.037] - - - [38656, 1280, 1, 256] - - [31, 70.603] - - - [35584, 1281, 1, 256] - - [39, 64.143] - - - [39216, 2865, 1, 256] - - [36, 61.369] - - - [35120, 1280, 1, 256] - - [58, 62.736] - - - [36096, 3328, 1, 256] - - [42, 71.742] - - - [38912, 6144, 1, 256] - - [27, 74.651] - - - [37376, 3840, 1, 256] - - [29, 74.191] - - - [37424, 2816, 1, 256] - - [53, 63.705] - - - [36864, 10240, 1, 256] - - [27, 74.71] - - - [35328, 3328, 1, 256] - - [43, 73.449] - - - [37632, 5632, 1, 256] - - [37, 74.204] - - - [35072, 1536, 1, 256] - - [37, 70.539] - - - [36864, 2865, 1, 256] - - [29, 70.301] - - - [36864, 4608, 1, 256] - - [29, 73.415] - - - [37888, 1280, 1, 256] - - [34, 71.371] - - - [36864, 4864, 1, 256] - - [29, 74.336] - - - [37632, 256, 1, 256] - - [34, 57.126] - - - [38912, 2816, 1, 256] - - [27, 73.697] - - - [38656, 5120, 1, 256] - - [24, 73.766] - - - [35072, 1280, 1, 256] - - [49, 70.409] - - - [38400, 3329, 1, 256] - - [54, 70.087] - - - [35840, 1281, 1, 256] - - [84, 64.22] - - - [39680, 2865, 1, 256] - - [54, 70.163] - - - [38192, 256, 1, 256] - - [53, 51.414] - - - [37632, 10240, 1, 256] - - [37, 74.479] - - - [39984, 256, 1, 256] - - [52, 53.316] - - - [37424, 2865, 1, 256] - - [36, 61.628] - - - [37888, 256, 1, 256] - - [44, 57.593] - - - [36864, 6144, 1, 256] - - [37, 74.004] - - - [38656, 1281, 1, 256] - - [89, 64.467] - - - [37936, 256, 1, 256] - - [58, 52.269] - - - [39168, 4864, 1, 256] - - [30, 73.766] - - - [35840, 256, 1, 256] - - [78, 55.143] - - - [37888, 10240, 1, 256] - - [29, 75.274] - - - [39728, 6144, 1, 256] - - [28, 62.67] - - - [39680, 5888, 1, 256] - - [37, 73.797] - - - [38144, 2816, 1, 256] - - [27, 72.629] - - - [39728, 256, 1, 256] - - [27, 53.166] - - - [37376, 2816, 1, 256] - - [25, 73.31] - - - [36352, 2865, 1, 256] - - [27, 70.628] - - - [39216, 256, 1, 256] - - [58, 52.562] - - - [37888, 1281, 1, 256] - - [74, 64.407] - - - [39472, 10240, 1, 256] - - [28, 63.129] - - - [37376, 2048, 1, 256] - - [43, 71.636] - - - [36096, 10240, 1, 256] - - [54, 73.83] - - - [35584, 1280, 1, 256] - - [49, 70.281] - - - [39168, 5632, 1, 256] - - [47, 73.916] - - - [39936, 5632, 1, 256] - - [27, 74.951] - - - [35072, 256, 1, 256] - - [36, 59.305] - - - [35376, 2865, 1, 256] - - [53, 61.863] - - - [38400, 4864, 1, 256] - - [37, 74.544] - - - [35888, 256, 1, 256] - - [44, 50.379] - - - [35072, 3328, 1, 256] - - [41, 72.749] - - - [37936, 10240, 1, 256] - - [41, 63.616] - - - [36352, 10240, 1, 256] - - [41, 75.006] - - - [38656, 2048, 1, 256] - - [41, 71.141] - - - [35632, 2816, 1, 256] - - [36, 63.573] - - - [36912, 10240, 1, 256] - - [28, 62.221] - - - [39936, 5888, 1, 256] - - [29, 74.727] - - - [38448, 2865, 1, 256] - - [44, 61.645] - - - [38144, 3840, 1, 256] - - [27, 73.821] - - - [37632, 6144, 1, 256] - - [37, 74.0] - - - [37376, 3328, 1, 256] - - [41, 73.555] - - - [36608, 2816, 1, 256] - - [27, 72.815] - - - [36912, 3072, 1, 256] - - [28, 60.786] - - - [37120, 2816, 1, 256] - - [28, 72.663] - - - [38144, 2865, 1, 256] - - [54, 70.012] - - - [38912, 2048, 1, 256] - - [41, 70.594] - - - [38192, 4608, 1, 256] - - [58, 63.111] - - - [37120, 5120, 1, 256] - - [27, 74.028] - - - [38400, 3328, 1, 256] - - [28, 73.36] - - - [35632, 10240, 1, 256] - - [28, 62.964] - - - [38912, 4864, 1, 256] - - [27, 74.806] - - - [37120, 10240, 1, 256] - - [27, 74.59] - - - [37120, 3329, 1, 256] - - [32, 69.606] - - - [35840, 6144, 1, 256] - - [27, 74.653] - - - [38400, 1281, 1, 256] - - [42, 64.99] - - - [36144, 10240, 1, 256] - - [41, 62.723] - - - [38144, 1280, 1, 256] - - [35, 70.558] - - - [39424, 10240, 1, 256] - - [27, 74.889] - - - [39424, 6144, 1, 256] - - [41, 74.434] - - - [39424, 1280, 1, 256] - - [34, 70.931] - - - [35328, 3329, 1, 256] - - [47, 70.11] - - - [39472, 5888, 1, 256] - - [53, 63.366] - - - [36352, 4096, 1, 256] - - [28, 73.815] - - - [38656, 4608, 1, 256] - - [28, 73.336] - - - [37168, 256, 1, 256] - - [58, 51.069] - - - [38144, 2048, 1, 256] - - [41, 71.056] - - - [35840, 1536, 1, 256] - - [37, 71.264] - - - [37120, 1280, 1, 256] - - [35, 70.842] - - - [37424, 3840, 1, 256] - - [36, 63.367] - - - [37424, 3584, 1, 256] - - [53, 63.462] - - - [36864, 2560, 1, 256] - - [22, 72.877] - - - [39936, 6400, 1, 256] - - [29, 75.294] - - - [36096, 2560, 1, 256] - - [57, 71.482] - - - [37120, 768, 1, 256] - - [44, 68.579] - - - [35328, 2048, 1, 256] - - [41, 71.615] - - - [36608, 4608, 1, 256] - - [28, 73.42] - - - [38400, 4096, 1, 256] - - [41, 73.807] - - - [35328, 2816, 1, 256] - - [51, 73.251] - - - [36144, 256, 1, 256] - - [27, 50.175] - - - [36608, 256, 1, 256] - - [55, 55.964] - - - [39168, 3329, 1, 256] - - [24, 69.417] - - - [38448, 4608, 1, 256] - - [58, 62.818] - - - [37632, 3328, 1, 256] - - [37, 72.662] - - - [37680, 2865, 1, 256] - - [53, 61.765] - - - [35120, 10240, 1, 256] - - [44, 62.831] - - - [37120, 6144, 1, 256] - - [41, 73.953] - - - [36656, 2816, 1, 256] - - [58, 63.274] - - - [39936, 3329, 1, 256] - - [54, 70.508] - - - [35328, 1792, 1, 256] - - [31, 72.387] - - - [35120, 1536, 1, 256] - - [36, 61.566] - - - [39472, 2865, 1, 256] - - [36, 61.639] - - - [37936, 4352, 1, 256] - - [25, 63.579] - - - [35888, 2048, 1, 256] - - [53, 62.81] - - - [37888, 3584, 1, 256] - - [27, 74.221] - - - [37376, 2865, 1, 256] - - [37, 70.609] - - - [36864, 1280, 1, 256] - - [55, 70.756] - - - [39472, 5632, 1, 256] - - [36, 63.752] - - - [37120, 1024, 1, 256] - - [60, 69.346] - - - [37120, 2865, 1, 256] - - [54, 70.15] - - - [38400, 1280, 1, 256] - - [34, 71.196] - - - [35584, 2816, 1, 256] - - [35, 72.373] - - - [37376, 1281, 1, 256] - - [76, 64.736] - - - [36352, 2560, 1, 256] - - [27, 73.332] - - - [36144, 2304, 1, 256] - - [44, 62.853] - - - [37632, 3840, 1, 256] - - [27, 73.728] - - - [38960, 2816, 1, 256] - - [29, 62.02] - - - [37376, 3072, 1, 256] - - [38, 73.241] - - - [35072, 2816, 1, 256] - - [27, 72.753] - - - [38912, 3328, 1, 256] - - [37, 73.576] - - - [38960, 256, 1, 256] - - [58, 52.833] - - - [35376, 256, 1, 256] - - [58, 49.669] - - - [39168, 1280, 1, 256] - - [27, 70.66] - - - [44032, 5888, 1, 256] - - [37, 74.602] - - - [40192, 2865, 1, 256] - - [25, 70.132] - - - [43312, 256, 1, 256] - - [36, 50.632] - - - [43520, 1280, 1, 256] - - [25, 71.545] - - - [41216, 2816, 1, 256] - - [25, 72.943] - - - [41520, 7936, 1, 256] - - [36, 63.037] - - - [43008, 2048, 1, 256] - - [28, 70.834] - - - [42496, 2048, 1, 256] - - [41, 72.118] - - - [40704, 3328, 1, 256] - - [41, 72.847] - - - [41776, 7936, 1, 256] - - [44, 63.282] - - - [40192, 1792, 1, 256] - - [27, 72.033] - - - [43520, 6144, 1, 256] - - [37, 74.517] - - - [42032, 2865, 1, 256] - - [36, 61.445] - - - [41472, 3329, 1, 256] - - [54, 70.049] - - - [41008, 7424, 1, 256] - - [28, 61.948] - - - [40448, 2865, 1, 256] - - [27, 70.49] - - - [41264, 2865, 1, 256] - - [44, 62.119] - - - [43312, 9728, 1, 256] - - [58, 62.838] - - - [40704, 2816, 1, 256] - - [27, 72.846] - - - [42544, 8704, 1, 256] - - [53, 62.483] - - - [40960, 7168, 1, 256] - - [38, 66.483] - - - [41216, 3329, 1, 256] - - [73, 69.564] - - - [41984, 6144, 1, 256] - - [27, 74.663] - - - [42240, 10240, 1, 256] - - [27, 74.534] - - - [42752, 2865, 1, 256] - - [54, 70.315] - - - [41216, 1280, 1, 256] - - [31, 71.322] - - - [40704, 7168, 1, 256] - - [25, 73.285] - - - [41216, 10240, 1, 256] - - [28, 74.599] - - - [40960, 256, 1, 256] - - [36, 60.229] - - - [40704, 2560, 1, 256] - - [25, 73.113] - - - [42752, 3329, 1, 256] - - [32, 69.627] - - - [43264, 3329, 1, 256] - - [54, 69.794] - - - [40192, 6144, 1, 256] - - [25, 74.232] - - - [43008, 10240, 1, 256] - - [25, 75.153] - - - [43520, 1281, 1, 256] - - [41, 65.119] - - - [42496, 8960, 1, 256] - - [29, 75.285] - - - [43312, 10240, 1, 256] - - [44, 62.292] - - - [44032, 6144, 1, 256] - - [25, 74.572] - - - [40192, 256, 1, 256] - - [53, 59.814] - - - [41984, 1536, 1, 256] - - [27, 71.539] - - - [41216, 768, 1, 256] - - [36, 68.62] - - - [40752, 256, 1, 256] - - [25, 54.047] - - - [44288, 1280, 1, 256] - - [57, 70.78] - - - [43520, 9216, 1, 256] - - [41, 74.767] - - - [42032, 8192, 1, 256] - - [41, 63.425] - - - [41728, 3584, 1, 256] - - [50, 71.854] - - - [40448, 1280, 1, 256] - - [27, 71.25] - - - [41216, 7168, 1, 256] - - [27, 73.415] - - - [42496, 1280, 1, 256] - - [25, 71.711] - - - [40448, 6656, 1, 256] - - [24, 74.479] - - - [40240, 256, 1, 256] - - [25, 53.619] - - - [41264, 2816, 1, 256] - - [36, 63.512] - - - [43264, 3328, 1, 256] - - [27, 72.954] - - - [43008, 9216, 1, 256] - - [25, 74.511] - - - [42240, 1281, 1, 256] - - [41, 64.663] - - - [42288, 2865, 1, 256] - - [53, 61.45] - - - [43008, 3328, 1, 256] - - [25, 73.593] - - - [40496, 256, 1, 256] - - [25, 53.635] - - - [43264, 8960, 1, 256] - - [29, 74.747] - - - [43056, 9472, 1, 256] - - [41, 62.988] - - - [40448, 3328, 1, 256] - - [43, 73.139] - - - [41776, 8192, 1, 256] - - [41, 62.693] - - - [40704, 6400, 1, 256] - - [25, 74.452] - - - [41984, 7680, 1, 256] - - [37, 75.275] - - - [43312, 9472, 1, 256] - - [44, 62.742] - - - [40192, 1280, 1, 256] - - [37, 71.066] - - - [43776, 5632, 1, 256] - - [41, 73.586] - - - [41984, 2865, 1, 256] - - [32, 71.09] - - - [40448, 2816, 1, 256] - - [27, 73.248] - - - [42240, 3328, 1, 256] - - [41, 73.03] - - - [42752, 2048, 1, 256] - - [74, 70.609] - - - [42240, 256, 1, 256] - - [58, 61.761] - - - [43008, 3329, 1, 256] - - [32, 70.532] - - - [44032, 5632, 1, 256] - - [25, 75.031] - - - [40192, 2048, 1, 256] - - [74, 71.401] - - - [41216, 256, 1, 256] - - [26, 60.115] - - - [44288, 9984, 1, 256] - - [27, 74.699] - - - [43008, 1280, 1, 256] - - [29, 71.961] - - - [41984, 2816, 1, 256] - - [25, 73.74] - - - [42752, 6144, 1, 256] - - [37, 74.06] - - - [43776, 3329, 1, 256] - - [28, 68.689] - - - [43008, 2865, 1, 256] - - [22, 71.08] - - - [43776, 9728, 1, 256] - - [28, 74.27] - - - [42240, 7936, 1, 256] - - [27, 74.473] - - - [41472, 7424, 1, 256] - - [25, 74.806] - - - [43776, 5376, 1, 256] - - [28, 73.395] - - - [43008, 6144, 1, 256] - - [27, 74.574] - - - [41216, 3072, 1, 256] - - [37, 73.182] - - - [42496, 8192, 1, 256] - - [37, 74.717] - - - [40704, 6144, 1, 256] - - [25, 74.159] - - - [44032, 3329, 1, 256] - - [54, 70.575] - - - [43520, 2048, 1, 256] - - [61, 71.975] - - - [43264, 2048, 1, 256] - - [74, 71.216] - - - [40448, 1281, 1, 256] - - [61, 64.998] - - - [40496, 2865, 1, 256] - - [44, 61.745] - - - [40448, 6144, 1, 256] - - [25, 74.257] - - - [41008, 10240, 1, 256] - - [28, 62.391] - - - [43056, 2865, 1, 256] - - [27, 60.62] - - - [43264, 1280, 1, 256] - - [29, 71.169] - - - [40192, 10240, 1, 256] - - [27, 74.615] - - - [41216, 7680, 1, 256] - - [27, 74.657] - - - [41008, 7168, 1, 256] - - [41, 62.179] - - - [44288, 2048, 1, 256] - - [28, 71.786] - - - [41472, 6144, 1, 256] - - [43, 74.195] - - - [43264, 2865, 1, 256] - - [54, 70.362] - - - [40448, 6912, 1, 256] - - [30, 74.755] - - - [41216, 6912, 1, 256] - - [37, 74.568] - - - [41984, 1792, 1, 256] - - [27, 72.966] - - - [40192, 1281, 1, 256] - - [74, 64.585] - - - [40960, 3329, 1, 256] - - [54, 62.664] - - - [41520, 10240, 1, 256] - - [74, 62.859] - - - [44032, 10240, 1, 256] - - [25, 75.084] - - - [43264, 2816, 1, 256] - - [37, 72.884] - - - [43008, 4608, 1, 256] - - [37, 74.106] - - - [43776, 1281, 1, 256] - - [61, 63.973] - - - [40240, 6656, 1, 256] - - [44, 63.007] - - - [43264, 9216, 1, 256] - - [29, 74.059] - - - [40704, 3329, 1, 256] - - [32, 69.77] - - - [42752, 3328, 1, 256] - - [37, 72.653] - - - [41984, 2048, 1, 256] - - [41, 71.786] - - - [44288, 3329, 1, 256] - - [28, 69.201] - - - [40192, 3328, 1, 256] - - [41, 72.906] - - - [40960, 10240, 1, 256] - - [38, 67.284] - - - [42496, 256, 1, 256] - - [63, 56.259] - - - [40496, 10240, 1, 256] - - [74, 62.619] - - - [40496, 6912, 1, 256] - - [53, 63.78] - - - [43776, 6144, 1, 256] - - [28, 73.627] - - - [40960, 1280, 1, 256] - - [25, 64.486] - - - [42288, 8704, 1, 256] - - [28, 62.698] - - - [42496, 3328, 1, 256] - - [28, 73.671] - - - [41216, 2865, 1, 256] - - [54, 70.173] - - - [42496, 3329, 1, 256] - - [32, 70.164] - - - [41984, 7936, 1, 256] - - [27, 75.243] - - - [41472, 1281, 1, 256] - - [61, 65.032] - - - [41776, 256, 1, 256] - - [53, 55.02] - - - [42752, 8960, 1, 256] - - [22, 74.694] - - - [41472, 7168, 1, 256] - - [25, 73.532] - - - [40240, 10240, 1, 256] - - [41, 62.723] - - - [41728, 1280, 1, 256] - - [59, 69.927] - - - [40752, 2865, 1, 256] - - [36, 61.872] - - - [40960, 2048, 1, 256] - - [27, 62.991] - - - [41472, 7680, 1, 256] - - [37, 74.81] - - - [41472, 10240, 1, 256] - - [25, 74.705] - - - [41264, 7680, 1, 256] - - [53, 63.108] - - - [42800, 8960, 1, 256] - - [53, 63.579] - - - [41728, 10240, 1, 256] - - [41, 74.303] - - - [44032, 3328, 1, 256] - - [29, 73.632] - - - [40704, 6912, 1, 256] - - [27, 74.49] - - - [41472, 2048, 1, 256] - - [48, 71.859] - - - [40960, 6144, 1, 256] - - [38, 67.029] - - - [43776, 3328, 1, 256] - - [28, 72.192] - - - [42496, 2865, 1, 256] - - [27, 70.925] - - - [40960, 3328, 1, 256] - - [38, 66.416] - - - [41728, 7936, 1, 256] - - [41, 73.453] - - - [41984, 3329, 1, 256] - - [25, 70.588] - - - [43008, 256, 1, 256] - - [27, 57.193] - - - [42240, 1280, 1, 256] - - [22, 70.817] - - - [43776, 10240, 1, 256] - - [28, 74.271] - - - [42752, 8448, 1, 256] - - [27, 74.38] - - - [42496, 1281, 1, 256] - - [41, 65.226] - - - [44032, 1536, 1, 256] - - [25, 71.765] - - - [40960, 2816, 1, 256] - - [22, 66.747] - - - [44288, 1792, 1, 256] - - [25, 71.649] - - - [43264, 1281, 1, 256] - - [74, 64.133] - - - [43008, 8704, 1, 256] - - [25, 75.246] - - - [41728, 1536, 1, 256] - - [57, 69.755] - - - [41728, 2048, 1, 256] - - [89, 69.934] - - - [43520, 9728, 1, 256] - - [27, 74.909] - - - [42032, 256, 1, 256] - - [36, 55.382] - - - [43776, 256, 1, 256] - - [58, 56.609] - - - [43008, 9472, 1, 256] - - [25, 75.471] - - - [44032, 1792, 1, 256] - - [37, 72.939] - - - [40704, 2865, 1, 256] - - [37, 70.238] - - - [42240, 1792, 1, 256] - - [27, 71.959] - - - [40704, 2304, 1, 256] - - [27, 72.693] - - - [42800, 9216, 1, 256] - - [28, 62.893] - - - [42240, 8704, 1, 256] - - [37, 74.597] - - - [42496, 6144, 1, 256] - - [25, 74.501] - - - [43568, 9728, 1, 256] - - [41, 63.01] - - - [40704, 2048, 1, 256] - - [28, 71.158] - - - [41472, 7936, 1, 256] - - [25, 74.77] - - - [42752, 2816, 1, 256] - - [27, 72.973] - - - [41008, 2865, 1, 256] - - [28, 58.776] - - - [40960, 6912, 1, 256] - - [38, 68.041] - - - [44032, 256, 1, 256] - - [36, 57.631] - - - [42496, 4352, 1, 256] - - [25, 74.491] - - - [42032, 8448, 1, 256] - - [28, 63.624] - - - [42752, 4608, 1, 256] - - [29, 73.474] - - - [44032, 1280, 1, 256] - - [27, 72.047] - - - [44288, 6144, 1, 256] - - [41, 74.003] - - - [42800, 2865, 1, 256] - - [36, 61.011] - - - [41008, 2816, 1, 256] - - [41, 61.755] - - - [41984, 8192, 1, 256] - - [25, 74.9] - - - [43264, 256, 1, 256] - - [60, 56.868] - - - [41728, 2865, 1, 256] - - [51, 68.836] - - - [43520, 5120, 1, 256] - - [25, 74.521] - - - [41984, 3584, 1, 256] - - [37, 74.444] - - - [41216, 3328, 1, 256] - - [28, 73.012] - - - [43520, 9472, 1, 256] - - [27, 75.262] - - - [43264, 9728, 1, 256] - - [25, 74.48] - - - [41728, 1281, 1, 256] - - [91, 64.039] - - - [40704, 1281, 1, 256] - - [41, 64.284] - - - [42288, 256, 1, 256] - - [25, 49.73] - - - [40960, 512, 1, 256] - - [27, 64.162] - - - [42752, 4352, 1, 256] - - [25, 74.11] - - - [40752, 10240, 1, 256] - - [28, 62.91] - - - [41728, 3328, 1, 256] - - [41, 72.265] - - - [43568, 2816, 1, 256] - - [36, 62.54] - - - [43008, 512, 1, 256] - - [44, 66.535] - - - [41216, 2048, 1, 256] - - [41, 71.246] - - - [42800, 256, 1, 256] - - [53, 49.994] - - - [43312, 2816, 1, 256] - - [36, 63.107] - - - [40192, 6400, 1, 256] - - [27, 74.557] - - - [41264, 7424, 1, 256] - - [36, 63.342] - - - [42544, 8960, 1, 256] - - [44, 63.039] - - - [41472, 256, 1, 256] - - [37, 61.092] - - - [42288, 10240, 1, 256] - - [41, 62.646] - - - [43520, 1024, 1, 256] - - [53, 70.335] - - - [42288, 8448, 1, 256] - - [44, 62.928] - - - [43776, 9472, 1, 256] - - [74, 74.186] - - - [43008, 1281, 1, 256] - - [25, 64.667] - - - [43008, 8960, 1, 256] - - [38, 75.419] - - - [41728, 256, 1, 256] - - [53, 61.757] - - - [41520, 7680, 1, 256] - - [28, 62.77] - - - [42240, 3329, 1, 256] - - [32, 69.539] - - - [41472, 2816, 1, 256] - - [47, 72.969] - - - [41216, 6144, 1, 256] - - [27, 74.096] - - - [40752, 2816, 1, 256] - - [44, 62.707] - - - [42496, 8704, 1, 256] - - [29, 75.149] - - - [40448, 6400, 1, 256] - - [37, 74.798] - - - [44032, 1281, 1, 256] - - [25, 64.951] - - - [41472, 1024, 1, 256] - - [63, 70.004] - - - [41216, 7424, 1, 256] - - [25, 74.529] - - - [43312, 2865, 1, 256] - - [36, 61.386] - - - [40960, 768, 1, 256] - - [29, 61.525] - - - [40240, 2865, 1, 256] - - [44, 61.619] - - - [43264, 768, 1, 256] - - [58, 68.7] - - - [40192, 3329, 1, 256] - - [37, 69.591] - - - [42800, 10240, 1, 256] - - [41, 62.897] - - - [42752, 512, 1, 256] - - [55, 65.629] - - - [40752, 6912, 1, 256] - - [36, 63.427] - - - [42240, 8192, 1, 256] - - [28, 74.303] - - - [42288, 2816, 1, 256] - - [36, 62.835] - - - [40960, 2865, 1, 256] - - [54, 63.581] - - - [42800, 2816, 1, 256] - - [58, 62.968] - - - [42496, 2816, 1, 256] - - [29, 73.572] - - - [41728, 7680, 1, 256] - - [28, 73.673] - - - [42240, 8448, 1, 256] - - [41, 74.494] - - - [41984, 1281, 1, 256] - - [74, 64.815] - - - [41984, 3328, 1, 256] - - [27, 73.564] - - - [40240, 6400, 1, 256] - - [44, 63.269] - - - [44288, 256, 1, 256] - - [23, 57.649] - - - [42496, 4096, 1, 256] - - [41, 73.815] - - - [43520, 3329, 1, 256] - - [54, 70.231] - - - [44288, 5888, 1, 256] - - [74, 73.788] - - - [42752, 1281, 1, 256] - - [37, 64.042] - - - [43776, 9984, 1, 256] - - [41, 73.921] - - - [41008, 256, 1, 256] - - [25, 53.831] - - - [40960, 1281, 1, 256] - - [38, 57.818] - - - [40704, 6656, 1, 256] - - [29, 74.314] - - - [40192, 2816, 1, 256] - - [27, 72.942] - - - [43264, 10240, 1, 256] - - [27, 74.641] - - - [44032, 9984, 1, 256] - - [27, 75.513] - - - [43520, 2865, 1, 256] - - [27, 70.933] - - - [42240, 3840, 1, 256] - - [25, 73.925] - - - [43056, 9216, 1, 256] - - [28, 62.923] - - - [43520, 10240, 1, 256] - - [25, 74.931] - - - [42544, 10240, 1, 256] - - [41, 62.521] - - - [40448, 2304, 1, 256] - - [29, 72.902] - - - [40704, 1280, 1, 256] - - [57, 70.86] - - - [43520, 2816, 1, 256] - - [25, 73.456] - - - [43520, 5376, 1, 256] - - [27, 74.646] - - - [41984, 256, 1, 256] - - [27, 61.95] - - - [43776, 1280, 1, 256] - - [61, 68.773] - - - [43568, 2865, 1, 256] - - [53, 61.386] - - - [41520, 256, 1, 256] - - [44, 55.016] - - - [41472, 3328, 1, 256] - - [43, 73.35] - - - [40192, 6656, 1, 256] - - [25, 74.325] - - - [40448, 2048, 1, 256] - - [28, 71.758] - - - [41520, 2816, 1, 256] - - [58, 63.286] - - - [43520, 9984, 1, 256] - - [25, 75.254] - - - [42544, 2865, 1, 256] - - [58, 61.581] - - - [42240, 2048, 1, 256] - - [41, 71.349] - - - [41472, 1280, 1, 256] - - [67, 71.256] - - - [40192, 5888, 1, 256] - - [27, 73.973] - - - [42240, 2865, 1, 256] - - [27, 70.024] - - - [41984, 10240, 1, 256] - - [25, 75.156] - - - [41264, 10240, 1, 256] - - [27, 62.503] - - - [42752, 10240, 1, 256] - - [27, 74.606] - - - [41216, 1024, 1, 256] - - [72, 69.891] - - - [41776, 10240, 1, 256] - - [41, 62.884] - - - [40960, 7424, 1, 256] - - [38, 67.903] - - - [40960, 2560, 1, 256] - - [22, 66.523] - - - [41216, 1281, 1, 256] - - [89, 64.548] - - - [41984, 1280, 1, 256] - - [27, 71.781] - - - [40448, 3329, 1, 256] - - [32, 69.924] - - - [41776, 2816, 1, 256] - - [53, 63.494] - - - [40704, 256, 1, 256] - - [36, 60.412] - - - [43264, 4864, 1, 256] - - [29, 74.253] - - - [42240, 6144, 1, 256] - - [27, 74.187] - - - [43520, 3328, 1, 256] - - [41, 73.577] - - - [42752, 256, 1, 256] - - [44, 56.12] - - - [40752, 7168, 1, 256] - - [28, 62.197] - - - [43776, 1536, 1, 256] - - [91, 67.546] - - - [42032, 10240, 1, 256] - - [28, 63.661] - - - [43008, 4864, 1, 256] - - [25, 74.853] - - - [40704, 10240, 1, 256] - - [27, 74.63] - - - [44288, 1281, 1, 256] - - [74, 64.844] - - - [41520, 2865, 1, 256] - - [44, 61.695] - - - [41264, 256, 1, 256] - - [103, 53.76] - - - [40496, 6656, 1, 256] - - [44, 63.334] - - - [42240, 4096, 1, 256] - - [41, 73.206] - - - [43568, 9984, 1, 256] - - [53, 63.212] - - - [43264, 9472, 1, 256] - - [27, 74.68] - - - [43008, 768, 1, 256] - - [36, 69.577] - - - [43776, 2816, 1, 256] - - [41, 72.397] - - - [43008, 2816, 1, 256] - - [22, 73.792] - - - [41984, 8448, 1, 256] - - [29, 75.147] - - - [43520, 256, 1, 256] - - [53, 56.961] - - - [43776, 2865, 1, 256] - - [28, 68.682] - - - [41984, 3840, 1, 256] - - [27, 74.662] - - - [42544, 256, 1, 256] - - [37, 49.843] - - - [43056, 2816, 1, 256] - - [41, 61.699] - - - [41472, 3072, 1, 256] - - [37, 73.12] - - - [41776, 2865, 1, 256] - - [53, 61.177] - - - [43056, 256, 1, 256] - - [37, 50.543] - - - [41728, 6144, 1, 256] - - [28, 73.428] - - - [42496, 8448, 1, 256] - - [28, 74.955] - - - [43568, 256, 1, 256] - - [58, 50.65] - - - [42752, 8704, 1, 256] - - [27, 74.639] - - - [42544, 2816, 1, 256] - - [53, 63.276] - - - [40448, 10240, 1, 256] - - [47, 74.672] - - - [41728, 2816, 1, 256] - - [91, 71.895] - - - [43568, 10240, 1, 256] - - [41, 62.816] - - - [44032, 2048, 1, 256] - - [41, 71.938] - - - [41472, 2865, 1, 256] - - [25, 70.643] - - - [40448, 256, 1, 256] - - [37, 60.316] - - - [41728, 3329, 1, 256] - - [28, 68.274] - - - [43264, 6144, 1, 256] - - [25, 74.082] - - - [40960, 6656, 1, 256] - - [22, 67.684] - - - [42752, 9216, 1, 256] - - [29, 74.194] - - - [40496, 2816, 1, 256] - - [44, 63.636] - - - [40704, 512, 1, 256] - - [53, 66.176] - - - [43056, 10240, 1, 256] - - [41, 63.048] - - - [44032, 9728, 1, 256] - - [27, 75.072] - - - [41728, 8192, 1, 256] - - [41, 74.078] - - - [43264, 1024, 1, 256] - - [53, 69.973] - - - [43776, 2048, 1, 256] - - [28, 70.652] - - - [40240, 2816, 1, 256] - - [58, 62.853] - - - [42752, 1280, 1, 256] - - [27, 71.138] - - - [44288, 10240, 1, 256] - - [41, 74.537] - - - [42240, 2816, 1, 256] - - [27, 72.834] - - - [41728, 7424, 1, 256] - - [28, 73.837] - - - [44288, 3328, 1, 256] - - [28, 72.991] - - - [43264, 5120, 1, 256] - - [27, 74.085] - - - [42032, 2816, 1, 256] - - [53, 63.145] - - - [11776, 6144, 1, 256] - - [37, 73.095] - - - [11264, 1792, 1, 256] - - [25, 67.19] - - - [4352, 2865, 1, 256] - - [25, 59.825] - - - [14640, 1536, 1, 256] - - [27, 59.341] - - - [4096, 2865, 1, 256] - - [59, 61.805] - - - [5168, 256, 1, 256] - - [121, 41.239] - - - [19968, 3328, 1, 256] - - [27, 72.493] - - - [12544, 3328, 1, 256] - - [35, 70.717] - - - [15408, 2816, 1, 256] - - [25, 62.707] - - - [16640, 3329, 1, 256] - - [25, 67.757] - - - [768, 768, 1, 256] - - [108, 28.462] - - - [3840, 512, 1, 256] - - [59, 32.591] - - - [7424, 5888, 1, 256] - - [27, 71.721] - - - [48, 49, 1, 256] - - [166, 0.141] - - - [16384, 768, 1, 256] - - [25, 54.969] - - - [15664, 2865, 1, 256] - - [25, 61.915] - - - [12544, 2048, 1, 256] - - [23, 67.949] - - - [7680, 4096, 1, 256] - - [27, 70.334] - - - [8240, 5376, 1, 256] - - [37, 65.489] - - - [11520, 256, 1, 256] - - [25, 47.463] - - - [12800, 256, 1, 256] - - [55, 51.019] - - - [10544, 2865, 1, 256] - - [27, 59.435] - - - [10032, 6912, 1, 256] - - [36, 63.949] - - - [3072, 3072, 1, 256] - - [55, 59.056] - - - [5888, 2865, 1, 256] - - [49, 64.896] - - - [8448, 3328, 1, 256] - - [34, 69.62] - - - [17920, 4096, 1, 256] - - [74, 72.746] - - - [19200, 5376, 1, 256] - - [25, 73.642] - - - [16432, 2865, 1, 256] - - [38, 64.164] - - - [12032, 3329, 1, 256] - - [35, 67.426] - - - [11776, 8704, 1, 256] - - [67, 74.604] - - - [11520, 1281, 1, 256] - - [60, 60.407] - - - [19760, 10240, 1, 256] - - [44, 62.731] - - - [15360, 1281, 1, 256] - - [25, 61.715] - - - [19712, 2865, 1, 256] - - [30, 67.478] - - - [9216, 6400, 1, 256] - - [25, 73.86] - - - [18944, 3329, 1, 256] - - [30, 69.341] - - - [5632, 2816, 1, 256] - - [55, 67.037] - - - [13872, 256, 1, 256] - - [35, 51.278] - - - [9984, 1280, 1, 256] - - [34, 61.779] - - - [19248, 10240, 1, 256] - - [36, 63.303] - - - [14128, 256, 1, 256] - - [35, 38.254] - - - [12080, 9216, 1, 256] - - [53, 63.728] - - - [18224, 5120, 1, 256] - - [58, 63.714] - - - [2352, 256, 1, 256] - - [123, 28.229] - - - [17712, 4608, 1, 256] - - [58, 63.487] - - - [8192, 5376, 1, 256] - - [35, 71.577] - - - [8752, 5888, 1, 256] - - [36, 63.141] - - - [11264, 3584, 1, 256] - - [49, 72.09] - - - [816, 256, 1, 256] - - [115, 12.953] - - - [5376, 3328, 1, 256] - - [25, 68.034] - - - [6144, 2560, 1, 256] - - [35, 67.197] - - - [9264, 256, 1, 256] - - [27, 38.214] - - - [8960, 5376, 1, 256] - - [57, 72.299] - - - [2608, 2353, 1, 256] - - [35, 50.213] - - - [2096, 256, 1, 256] - - [123, 25.259] - - - [9984, 7168, 1, 256] - - [34, 72.618] - - - [7424, 3329, 1, 256] - - [27, 65.354] - - - [2352, 2304, 1, 256] - - [36, 46.217] - - - [9984, 512, 1, 256] - - [34, 53.859] - - - [6656, 3840, 1, 256] - - [49, 69.201] - - - [17408, 3329, 1, 256] - - [37, 69.421] - - - [8496, 5376, 1, 256] - - [29, 64.353] - - - [11264, 3840, 1, 256] - - [34, 72.724] - - - [13312, 2865, 1, 256] - - [25, 69.024] - - - [3584, 768, 1, 256] - - [23, 44.664] - - - [11520, 6144, 1, 256] - - [59, 72.709] - - - [15360, 2048, 1, 256] - - [61, 69.024] - - - [7936, 3328, 1, 256] - - [63, 69.124] - - - [6144, 1281, 1, 256] - - [35, 55.503] - - - [19968, 6656, 1, 256] - - [37, 74.609] - - - [15152, 256, 1, 256] - - [49, 41.36] - - - [18432, 4608, 1, 256] - - [35, 73.713] - - - [1072, 256, 1, 256] - - [116, 16.399] - - - [6400, 4864, 1, 256] - - [55, 70.408] - - - [19712, 1281, 1, 256] - - [86, 59.967] - - - [1792, 1280, 1, 256] - - [59, 37.644] - - - [8192, 2865, 1, 256] - - [35, 65.783] - - - [3376, 256, 1, 256] - - [108, 37.471] - - - [10544, 2816, 1, 256] - - [58, 61.779] - - - [14336, 2816, 1, 256] - - [27, 71.657] - - - [16384, 1280, 1, 256] - - [22, 58.499] - - - [1280, 256, 1, 256] - - [116, 19.581] - - - [12544, 8960, 1, 256] - - [27, 74.392] - - - [13824, 1281, 1, 256] - - [53, 61.987] - - - [3072, 256, 1, 256] - - [109, 35.708] - - - [19760, 2816, 1, 256] - - [58, 63.795] - - - [8448, 5376, 1, 256] - - [49, 71.936] - - - [11824, 2865, 1, 256] - - [27, 60.194] - - - [6656, 3584, 1, 256] - - [35, 68.057] - - - [12288, 8704, 1, 256] - - [37, 75.086] - - - [11312, 256, 1, 256] - - [36, 44.077] - - - [15920, 2816, 1, 256] - - [53, 64.786] - - - [12032, 8448, 1, 256] - - [34, 73.765] - - - [14080, 2048, 1, 256] - - [92, 67.032] - - - [6400, 5120, 1, 256] - - [35, 70.545] - - - [7216, 2865, 1, 256] - - [27, 60.212] - - - [4400, 1280, 1, 256] - - [25, 47.188] - - - [5376, 3840, 1, 256] - - [34, 67.618] - - - [7168, 2816, 1, 256] - - [34, 67.248] - - - [19200, 5632, 1, 256] - - [27, 73.869] - - - [4144, 1024, 1, 256] - - [27, 44.266] - - - [12800, 3329, 1, 256] - - [67, 68.406] - - - [6400, 2865, 1, 256] - - [37, 64.555] - - - [12800, 5376, 1, 256] - - [57, 73.266] - - - [7168, 1536, 1, 256] - - [49, 59.413] - - - [19968, 1281, 1, 256] - - [58, 63.204] - - - [17664, 1281, 1, 256] - - [39, 62.694] - - - [11264, 3329, 1, 256] - - [29, 68.433] - - - [17712, 256, 1, 256] - - [23, 46.132] - - - [6656, 5376, 1, 256] - - [59, 71.81] - - - [13056, 5376, 1, 256] - - [34, 73.45] - - - [11568, 2865, 1, 256] - - [25, 58.964] - - - [3328, 1281, 1, 256] - - [49, 46.053] - - - [19968, 2048, 1, 256] - - [58, 70.791] - - - [2304, 2048, 1, 256] - - [35, 50.783] - - - [7728, 256, 1, 256] - - [53, 32.336] - - - [7424, 4352, 1, 256] - - [55, 71.451] - - - [5376, 2048, 1, 256] - - [35, 58.901] - - - [19456, 2816, 1, 256] - - [25, 72.622] - - - [7216, 2816, 1, 256] - - [25, 60.371] - - - [18688, 5376, 1, 256] - - [25, 73.843] - - - [4656, 1792, 1, 256] - - [27, 54.518] - - - [10240, 768, 1, 256] - - [44, 56.426] - - - [19456, 1280, 1, 256] - - [55, 69.213] - - - [18432, 3329, 1, 256] - - [29, 69.63] - - - [17920, 6144, 1, 256] - - [29, 74.033] - - - [1536, 1280, 1, 256] - - [47, 32.732] - - - [19456, 6400, 1, 256] - - [37, 75.291] - - - [15360, 6144, 1, 256] - - [25, 74.341] - - - [15664, 10240, 1, 256] - - [25, 63.248] - - - [3840, 256, 1, 256] - - [188, 34.149] - - - [4864, 3328, 1, 256] - - [55, 67.47] - - - [18224, 2865, 1, 256] - - [37, 61.253] - - - [13056, 9984, 1, 256] - - [25, 75.033] - - - [12288, 256, 1, 256] - - [49, 49.522] - - - [7168, 3840, 1, 256] - - [35, 70.01] - - - [17712, 4352, 1, 256] - - [58, 63.205] - - - [14592, 10240, 1, 256] - - [68, 73.199] - - - [8704, 5376, 1, 256] - - [49, 72.78] - - - [16128, 2816, 1, 256] - - [27, 71.212] - - - [4352, 3329, 1, 256] - - [49, 61.736] - - - [13568, 512, 1, 256] - - [35, 59.368] - - - [15872, 2865, 1, 256] - - [25, 69.974] - - - [12032, 1281, 1, 256] - - [34, 59.533] - - - [11520, 2048, 1, 256] - - [63, 66.369] - - - [12032, 2048, 1, 256] - - [58, 67.9] - - - [5632, 1281, 1, 256] - - [34, 51.805] - - - [13312, 9984, 1, 256] - - [29, 75.719] - - - [4912, 2865, 1, 256] - - [37, 54.733] - - - [15408, 2304, 1, 256] - - [27, 63.5] - - - [7472, 2816, 1, 256] - - [53, 61.407] - - - [18688, 10240, 1, 256] - - [29, 74.915] - - - [10752, 7936, 1, 256] - - [25, 74.542] - - - [2048, 1793, 1, 256] - - [35, 40.502] - - - [11776, 1280, 1, 256] - - [67, 64.065] - - - [10032, 256, 1, 256] - - [34, 39.665] - - - [17408, 1536, 1, 256] - - [49, 68.826] - - - [14080, 2865, 1, 256] - - [67, 68.381] - - - [16688, 3328, 1, 256] - - [44, 65.797] - - - [18944, 1024, 1, 256] - - [40, 67.129] - - - [2352, 2097, 1, 256] - - [37, 49.148] - - - [11008, 2048, 1, 256] - - [104, 64.317] - - - [10240, 6912, 1, 256] - - [49, 74.534] - - - [8448, 768, 1, 256] - - [72, 55.654] - - - [16640, 1024, 1, 256] - - [60, 63.0] - - - [11824, 8960, 1, 256] - - [53, 64.102] - - - [7936, 1280, 1, 256] - - [34, 62.118] - - - [6960, 3840, 1, 256] - - [29, 62.457] - - - [3328, 2048, 1, 256] - - [35, 58.869] - - - [16944, 2865, 1, 256] - - [37, 63.601] - - - [1024, 256, 1, 256] - - [116, 15.829] - - - [16944, 3840, 1, 256] - - [53, 64.63] - - - [3376, 2816, 1, 256] - - [29, 53.634] - - - [12288, 768, 1, 256] - - [36, 57.894] - - - [17152, 3329, 1, 256] - - [55, 69.147] - - - [6192, 2865, 1, 256] - - [25, 56.627] - - - [5888, 1281, 1, 256] - - [35, 53.849] - - - [11824, 256, 1, 256] - - [27, 45.599] - - - [18688, 1280, 1, 256] - - [34, 67.925] - - - [11520, 7936, 1, 256] - - [59, 73.686] - - - [15616, 1281, 1, 256] - - [36, 61.056] - - - [16944, 10240, 1, 256] - - [25, 64.198] - - - [12032, 4352, 1, 256] - - [35, 71.886] - - - [9984, 6656, 1, 256] - - [35, 73.374] - - - [17408, 1281, 1, 256] - - [61, 62.295] - - - [6912, 3329, 1, 256] - - [35, 65.222] - - - [16176, 2865, 1, 256] - - [29, 63.843] - - - [7936, 4864, 1, 256] - - [31, 71.543] - - - [7168, 256, 1, 256] - - [51, 30.817] - - - [9728, 6144, 1, 256] - - [27, 72.968] - - - [10752, 7680, 1, 256] - - [25, 74.394] - - - [13056, 5632, 1, 256] - - [37, 73.854] - - - [17152, 2865, 1, 256] - - [27, 68.888] - - - [4096, 512, 1, 256] - - [59, 34.82] - - - [3584, 2304, 1, 256] - - [34, 59.247] - - - [11264, 2048, 1, 256] - - [58, 67.528] - - - [18944, 5376, 1, 256] - - [67, 74.399] - - - [8960, 3329, 1, 256] - - [55, 66.204] - - - [7936, 1281, 1, 256] - - [40, 54.833] - - - [12848, 2816, 1, 256] - - [36, 61.496] - - - [9472, 3328, 1, 256] - - [40, 70.075] - - - [2816, 2816, 1, 256] - - [55, 56.977] - - - [15616, 10240, 1, 256] - - [29, 74.822] - - - [2816, 256, 1, 256] - - [109, 33.257] - - - [48, 256, 1, 256] - - [113, 0.779] - - - [17408, 1792, 1, 256] - - [55, 70.156] - - - [10032, 2865, 1, 256] - - [27, 59.522] - - - [3584, 2865, 1, 256] - - [35, 62.625] - - - [9472, 2816, 1, 256] - - [35, 69.497] - - - [2096, 2048, 1, 256] - - [35, 44.187] - - - [9216, 1536, 1, 256] - - [49, 65.881] - - - [5936, 256, 1, 256] - - [182, 41.044] - - - [11520, 1280, 1, 256] - - [49, 63.231] - - - [16896, 3328, 1, 256] - - [37, 72.404] - - - [7984, 4864, 1, 256] - - [25, 64.026] - - - [11008, 1280, 1, 256] - - [31, 63.427] - - - [18432, 6144, 1, 256] - - [25, 74.435] - - - [2096, 1841, 1, 256] - - [34, 40.144] - - - [8448, 1024, 1, 256] - - [23, 59.313] - - - [17968, 10240, 1, 256] - - [36, 63.221] - - - [1536, 1536, 1, 256] - - [34, 38.782] - - - [7728, 4864, 1, 256] - - [25, 61.971] - - - [18944, 3328, 1, 256] - - [31, 72.475] - - - [4608, 1792, 1, 256] - - [49, 58.808] - - - [8960, 6144, 1, 256] - - [35, 72.367] - - - [18736, 2816, 1, 256] - - [44, 62.4] - - - [8704, 5120, 1, 256] - - [35, 72.429] - - - [19456, 6144, 1, 256] - - [25, 74.371] - - - [19456, 1281, 1, 256] - - [61, 62.089] - - - [17200, 3840, 1, 256] - - [27, 64.105] - - - [2352, 2353, 1, 256] - - [36, 46.546] - - - [17408, 2816, 1, 256] - - [55, 72.23] - - - [13312, 2816, 1, 256] - - [34, 71.629] - - - [8960, 2816, 1, 256] - - [49, 70.027] - - - [2048, 1792, 1, 256] - - [49, 41.156] - - - [17152, 10240, 1, 256] - - [57, 74.66] - - - [16176, 10240, 1, 256] - - [27, 65.472] - - - [10288, 2865, 1, 256] - - [29, 60.769] - - - [8704, 2816, 1, 256] - - [34, 69.527] - - - [7424, 4096, 1, 256] - - [27, 70.152] - - - [6656, 1024, 1, 256] - - [27, 58.781] - - - [2304, 256, 1, 256] - - [108, 28.11] - - - [16384, 2865, 1, 256] - - [22, 59.821] - - - [7680, 2816, 1, 256] - - [25, 69.69] - - - [11520, 3329, 1, 256] - - [31, 66.836] - - - [10752, 1280, 1, 256] - - [34, 64.9] - - - [3120, 2816, 1, 256] - - [34, 56.127] - - - [15872, 1281, 1, 256] - - [23, 62.2] - - - [13824, 6144, 1, 256] - - [29, 73.534] - - - [6912, 3584, 1, 256] - - [34, 69.944] - - - [12032, 3328, 1, 256] - - [55, 70.445] - - - [11264, 1281, 1, 256] - - [53, 60.24] - - - [19456, 5632, 1, 256] - - [25, 74.96] - - - [17200, 2816, 1, 256] - - [58, 63.832] - - - [11520, 3840, 1, 256] - - [67, 71.601] - - - [11520, 2865, 1, 256] - - [55, 68.192] - - - [14848, 1280, 1, 256] - - [35, 66.953] - - - [16176, 256, 1, 256] - - [60, 43.839] - - - [16384, 256, 1, 256] - - [40, 45.395] - - - [4096, 768, 1, 256] - - [57, 49.726] - - - [4864, 2816, 1, 256] - - [34, 65.058] - - - [13568, 256, 1, 256] - - [40, 53.158] - - - [4608, 2048, 1, 256] - - [49, 58.79] - - - [9984, 6144, 1, 256] - - [35, 72.711] - - - [3632, 768, 1, 256] - - [44, 42.625] - - - [19200, 5888, 1, 256] - - [34, 73.564] - - - [5632, 2865, 1, 256] - - [35, 63.299] - - - [15360, 1280, 1, 256] - - [49, 67.381] - - - [12800, 1280, 1, 256] - - [34, 64.295] - - - [7168, 3328, 1, 256] - - [49, 68.647] - - - [11264, 8448, 1, 256] - - [37, 74.919] - - - [18176, 3328, 1, 256] - - [49, 72.027] - - - [4096, 2560, 1, 256] - - [55, 63.739] - - - [12544, 768, 1, 256] - - [53, 59.132] - - - [11568, 8448, 1, 256] - - [44, 63.978] - - - [8704, 1280, 1, 256] - - [34, 60.386] - - - [13056, 1536, 1, 256] - - [25, 65.618] - - - [2304, 1024, 1, 256] - - [57, 38.726] - - - [3072, 1281, 1, 256] - - [35, 43.281] - - - [6912, 1280, 1, 256] - - [35, 61.687] - - - [9216, 2816, 1, 256] - - [35, 69.713] - - - [17152, 6144, 1, 256] - - [27, 73.807] - - - [18992, 2865, 1, 256] - - [25, 60.345] - - - [10240, 2560, 1, 256] - - [25, 70.12] - - - [560, 256, 1, 256] - - [115, 8.889] - - - [2304, 1280, 1, 256] - - [35, 46.866] - - - [7680, 6144, 1, 256] - - [34, 72.08] - - - [15920, 2560, 1, 256] - - [37, 64.2] - - - [17456, 10240, 1, 256] - - [53, 63.79] - - - [14080, 3328, 1, 256] - - [57, 70.137] - - - [13360, 10240, 1, 256] - - [27, 64.032] - - - [8448, 5632, 1, 256] - - [49, 72.533] - - - [17408, 3584, 1, 256] - - [37, 73.531] - - - [6704, 2865, 1, 256] - - [25, 57.23] - - - [12592, 9472, 1, 256] - - [36, 64.017] - - - [18992, 10240, 1, 256] - - [44, 63.555] - - - [5376, 2865, 1, 256] - - [55, 63.726] - - - [18480, 5120, 1, 256] - - [37, 62.77] - - - [14336, 2048, 1, 256] - - [61, 68.004] - - - [7424, 3328, 1, 256] - - [40, 69.576] - - - [256, 49, 1, 256] - - [115, 0.778] - - - [12288, 1280, 1, 256] - - [27, 65.855] - - - [13568, 3329, 1, 256] - - [25, 67.986] - - - [15360, 1792, 1, 256] - - [49, 69.304] - - - [7168, 3584, 1, 256] - - [27, 69.608] - - - [10240, 3328, 1, 256] - - [25, 71.06] - - - [6400, 1281, 1, 256] - - [25, 57.395] - - - [11008, 6144, 1, 256] - - [75, 72.223] - - - [512, 513, 1, 256] - - [116, 15.777] - - - [19248, 256, 1, 256] - - [63, 48.424] - - - [2608, 256, 1, 256] - - [114, 30.8] - - - [16688, 3584, 1, 256] - - [29, 65.642] - - - [17920, 4864, 1, 256] - - [31, 74.051] - - - [18688, 1281, 1, 256] - - [40, 62.826] - - - [18224, 4864, 1, 256] - - [36, 63.328] - - - [10496, 2816, 1, 256] - - [63, 69.463] - - - [12288, 4864, 1, 256] - - [37, 73.366] - - - [9216, 2865, 1, 256] - - [27, 66.685] - - - [17664, 3329, 1, 256] - - [27, 68.974] - - - [3632, 512, 1, 256] - - [35, 29.845] - - - [11776, 3329, 1, 256] - - [57, 68.167] - - - [19456, 1792, 1, 256] - - [55, 70.657] - - - [12592, 256, 1, 256] - - [105, 47.498] - - - [10752, 3072, 1, 256] - - [27, 70.504] - - - [10800, 2816, 1, 256] - - [60, 63.108] - - - [6192, 3072, 1, 256] - - [25, 60.601] - - - [17152, 1536, 1, 256] - - [55, 68.551] - - - [2096, 2097, 1, 256] - - [25, 44.726] - - - [8192, 4608, 1, 256] - - [25, 70.244] - - - [13056, 6144, 1, 256] - - [27, 73.792] - - - [16640, 10240, 1, 256] - - [25, 74.759] - - - [12592, 9728, 1, 256] - - [58, 63.963] - - - [18176, 2816, 1, 256] - - [27, 71.857] - - - [18176, 4864, 1, 256] - - [25, 73.781] - - - [18944, 6144, 1, 256] - - [24, 74.1] - - - [12544, 256, 1, 256] - - [27, 50.485] - - - [13056, 1281, 1, 256] - - [63, 59.592] - - - [304, 49, 1, 256] - - [166, 0.914] - - - [17920, 2816, 1, 256] - - [64, 71.961] - - - [4656, 256, 1, 256] - - [114, 39.388] - - - [7728, 2865, 1, 256] - - [27, 59.378] - - - [15872, 1280, 1, 256] - - [57, 66.943] - - - [17456, 256, 1, 256] - - [34, 45.992] - - - [18176, 4608, 1, 256] - - [49, 73.101] - - - [7168, 5632, 1, 256] - - [49, 72.334] - - - [13616, 256, 1, 256] - - [36, 50.658] - - - [15104, 3329, 1, 256] - - [37, 68.888] - - - [19712, 3328, 1, 256] - - [67, 70.314] - - - [10032, 7168, 1, 256] - - [53, 63.028] - - - [11008, 3328, 1, 256] - - [75, 69.176] - - - [10496, 6144, 1, 256] - - [35, 72.906] - - - [6656, 2865, 1, 256] - - [59, 65.529] - - - [17664, 3840, 1, 256] - - [59, 72.723] - - - [6960, 4096, 1, 256] - - [44, 62.828] - - - [4608, 256, 1, 256] - - [175, 39.897] - - - [10496, 6912, 1, 256] - - [34, 73.557] - - - [16128, 2560, 1, 256] - - [59, 71.226] - - - [15872, 256, 1, 256] - - [34, 44.267] - - - [6656, 1281, 1, 256] - - [92, 53.011] - - - [3584, 512, 1, 256] - - [59, 31.134] - - - [11520, 1792, 1, 256] - - [34, 67.255] - - - [11264, 8192, 1, 256] - - [37, 74.546] - - - [10752, 2048, 1, 256] - - [40, 67.186] - - - [18688, 3329, 1, 256] - - [27, 69.295] - - - [4352, 768, 1, 256] - - [23, 51.77] - - - [18432, 512, 1, 256] - - [27, 58.069] - - - [18992, 256, 1, 256] - - [40, 47.829] - - - [13568, 2865, 1, 256] - - [25, 68.418] - - - [14640, 256, 1, 256] - - [49, 39.97] - - - [11264, 256, 1, 256] - - [34, 46.464] - - - [16896, 3329, 1, 256] - - [25, 69.544] - - - [18944, 5120, 1, 256] - - [47, 74.02] - - - [768, 513, 1, 256] - - [108, 22.833] - - - [14080, 10240, 1, 256] - - [24, 73.542] - - - [15872, 2560, 1, 256] - - [49, 71.856] - - - [6912, 5632, 1, 256] - - [49, 71.919] - - - [13360, 2865, 1, 256] - - [25, 61.513] - - - [6400, 3072, 1, 256] - - [35, 68.625] - - - [15616, 3329, 1, 256] - - [27, 69.058] - - - [9264, 2816, 1, 256] - - [37, 61.26] - - - [18176, 512, 1, 256] - - [40, 57.477] - - - [11264, 1280, 1, 256] - - [35, 66.843] - - - [1328, 1329, 1, 256] - - [49, 29.006] - - - [18736, 5376, 1, 256] - - [58, 62.844] - - - [5376, 1792, 1, 256] - - [35, 59.533] - - - [6144, 4608, 1, 256] - - [25, 70.913] - - - [6400, 3328, 1, 256] - - [34, 69.436] - - - [12032, 2865, 1, 256] - - [34, 68.265] - - - [12288, 4608, 1, 256] - - [27, 72.894] - - - [16128, 3072, 1, 256] - - [59, 71.509] - - - [2048, 256, 1, 256] - - [123, 25.3] - - - [4096, 256, 1, 256] - - [186, 36.316] - - - [5888, 2304, 1, 256] - - [35, 64.782] - - - [2816, 2561, 1, 256] - - [34, 52.938] - - - [3072, 1536, 1, 256] - - [35, 50.641] - - - [2304, 1281, 1, 256] - - [27, 46.337] - - - [15616, 2048, 1, 256] - - [63, 69.79] - - - [12800, 1024, 1, 256] - - [23, 62.114] - - - [8960, 3328, 1, 256] - - [35, 69.542] - - - [18432, 1280, 1, 256] - - [27, 67.656] - - - [8448, 2048, 1, 256] - - [40, 65.944] - - - [19712, 6400, 1, 256] - - [24, 73.172] - - - [14384, 1280, 1, 256] - - [29, 58.663] - - - [6448, 2816, 1, 256] - - [36, 58.008] - - - [18176, 2048, 1, 256] - - [23, 70.358] - - - [3072, 1792, 1, 256] - - [27, 49.282] - - - [12080, 8960, 1, 256] - - [36, 63.832] - - - [13312, 1281, 1, 256] - - [25, 60.158] - - - [16688, 2816, 1, 256] - - [27, 64.939] - - - [6400, 256, 1, 256] - - [183, 46.542] - - - [2048, 2048, 1, 256] - - [35, 46.271] - - - [14336, 256, 1, 256] - - [78, 41.029] - - - [11520, 2816, 1, 256] - - [67, 70.272] - - - [14384, 10240, 1, 256] - - [27, 63.156] - - - [7472, 256, 1, 256] - - [44, 31.265] - - - [1280, 1280, 1, 256] - - [121, 47.358] - - - [8704, 1024, 1, 256] - - [23, 61.245] - - - [9520, 2865, 1, 256] - - [44, 60.006] - - - [18480, 256, 1, 256] - - [60, 47.093] - - - [18176, 256, 1, 256] - - [40, 49.441] - - - [15872, 6144, 1, 256] - - [27, 74.514] - - - [304, 256, 1, 256] - - [116, 4.8] - - - [13568, 5888, 1, 256] - - [49, 73.35] - - - [3328, 3328, 1, 256] - - [34, 59.627] - - - [6656, 5120, 1, 256] - - [49, 71.447] - - - [9520, 2816, 1, 256] - - [58, 61.723] - - - [1536, 1537, 1, 256] - - [34, 38.752] - - - [3072, 2865, 1, 256] - - [80, 55.246] - - - [10032, 2816, 1, 256] - - [53, 62.235] - - - [12032, 9216, 1, 256] - - [54, 73.508] - - - [13872, 10240, 1, 256] - - [44, 63.139] - - - [13824, 2048, 1, 256] - - [44, 69.586] - - - [12544, 9728, 1, 256] - - [37, 74.151] - - - [17664, 4352, 1, 256] - - [57, 73.069] - - - [4096, 1281, 1, 256] - - [49, 47.228] - - - [17408, 1280, 1, 256] - - [49, 68.046] - - - [18432, 2816, 1, 256] - - [55, 72.467] - - - [5120, 256, 1, 256] - - [189, 41.991] - - - [18736, 2865, 1, 256] - - [58, 60.889] - - - [19200, 256, 1, 256] - - [34, 51.609] - - - [2048, 512, 1, 256] - - [174, 36.208] - - - [11008, 7680, 1, 256] - - [71, 73.369] - - - [5888, 3072, 1, 256] - - [35, 65.822] - - - [11776, 8192, 1, 256] - - [27, 74.076] - - - [5888, 512, 1, 256] - - [27, 48.316] - - - [7936, 2816, 1, 256] - - [55, 67.886] - - - [5120, 2865, 1, 256] - - [49, 62.345] - - - [12032, 2816, 1, 256] - - [49, 70.599] - - - [256, 257, 1, 256] - - [179, 4.192] - - - [13104, 2865, 1, 256] - - [44, 60.974] - - - [5680, 2865, 1, 256] - - [37, 57.715] - - - [15408, 10240, 1, 256] - - [29, 63.988] - - - [18432, 4864, 1, 256] - - [27, 74.451] - - - [17712, 2865, 1, 256] - - [27, 60.28] - - - [768, 256, 1, 256] - - [115, 12.668] - - - [9728, 3328, 1, 256] - - [29, 71.105] - - - [12848, 9728, 1, 256] - - [53, 64.542] - - - [2304, 2305, 1, 256] - - [57, 46.736] - - - [10240, 6144, 1, 256] - - [29, 73.496] - - - [13312, 1280, 1, 256] - - [49, 66.37] - - - [9008, 5888, 1, 256] - - [36, 63.331] - - - [7424, 3840, 1, 256] - - [49, 70.87] - - - [12032, 1280, 1, 256] - - [49, 65.217] - - - [18480, 2816, 1, 256] - - [29, 62.353] - - - [18432, 5120, 1, 256] - - [25, 74.354] - - - [7424, 4608, 1, 256] - - [23, 70.809] - - - [9776, 2865, 1, 256] - - [36, 58.455] - - - [5632, 2560, 1, 256] - - [34, 67.482] - - - [7680, 2048, 1, 256] - - [40, 65.594] - - - [6704, 2816, 1, 256] - - [58, 59.919] - - - [13872, 2816, 1, 256] - - [36, 63.15] - - - [17968, 2816, 1, 256] - - [44, 62.628] - - - [4144, 2865, 1, 256] - - [29, 58.002] - - - [14640, 1280, 1, 256] - - [27, 58.845] - - - [16432, 2816, 1, 256] - - [36, 67.138] - - - [16128, 1280, 1, 256] - - [34, 66.94] - - - [8240, 5120, 1, 256] - - [27, 65.285] - - - [4352, 2816, 1, 256] - - [35, 64.474] - - - [12544, 2865, 1, 256] - - [27, 68.603] - - - [6144, 2048, 1, 256] - - [63, 63.394] - - - [13616, 512, 1, 256] - - [40, 55.1] - - - [5632, 2048, 1, 256] - - [27, 60.737] - - - [13312, 2048, 1, 256] - - [23, 67.86] - - - [9728, 1281, 1, 256] - - [27, 59.385] - - - [7424, 1281, 1, 256] - - [35, 58.342] - - - [10800, 256, 1, 256] - - [36, 41.971] - - - [2048, 1281, 1, 256] - - [57, 42.578] - - - [5376, 1280, 1, 256] - - [34, 59.027] - - - [15664, 2816, 1, 256] - - [36, 63.151] - - - [256, 256, 1, 256] - - [166, 4.319] - - - [2048, 1280, 1, 256] - - [57, 42.717] - - - [9776, 256, 1, 256] - - [53, 39.723] - - - [4096, 3329, 1, 256] - - [59, 63.485] - - - [9728, 2304, 1, 256] - - [55, 68.567] - - - [19968, 2865, 1, 256] - - [25, 70.059] - - - [13568, 6144, 1, 256] - - [29, 73.621] - - - [15360, 2304, 1, 256] - - [55, 71.231] - - - [9264, 6400, 1, 256] - - [27, 63.431] - - - [19200, 2048, 1, 256] - - [23, 69.891] - - - [11520, 4096, 1, 256] - - [40, 70.668] - - - [18688, 5632, 1, 256] - - [25, 74.26] - - - [11776, 256, 1, 256] - - [36, 48.248] - - - [17152, 256, 1, 256] - - [40, 47.478] - - - [5120, 1280, 1, 256] - - [34, 57.075] - - - [14896, 1792, 1, 256] - - [44, 61.582] - - - [10288, 2816, 1, 256] - - [29, 60.366] - - - [7984, 2865, 1, 256] - - [37, 58.999] - - - [4864, 1281, 1, 256] - - [35, 54.304] - - - [7216, 256, 1, 256] - - [40, 30.145] - - - [5888, 3328, 1, 256] - - [35, 68.189] - - - [7424, 2816, 1, 256] - - [34, 68.231] - - - [15360, 3328, 1, 256] - - [35, 72.12] - - - [10544, 256, 1, 256] - - [34, 41.412] - - - [9776, 2816, 1, 256] - - [53, 61.887] - - - [8240, 2816, 1, 256] - - [27, 63.984] - - - [6656, 3072, 1, 256] - - [34, 66.992] - - - [18224, 10240, 1, 256] - - [53, 63.963] - - - [13824, 2865, 1, 256] - - [35, 69.183] - - - [5376, 1281, 1, 256] - - [35, 50.065] - - - [13568, 9984, 1, 256] - - [37, 74.874] - - - [18176, 4352, 1, 256] - - [37, 73.61] - - - [11776, 1281, 1, 256] - - [44, 58.768] - - - [15616, 6144, 1, 256] - - [37, 73.791] - - - [4400, 256, 1, 256] - - [123, 37.764] - - - [18992, 2816, 1, 256] - - [36, 63.728] - - - [14640, 10240, 1, 256] - - [58, 63.516] - - - [5120, 2048, 1, 256] - - [27, 61.987] - - - [19968, 10240, 1, 256] - - [27, 75.102] - - - [19200, 2865, 1, 256] - - [59, 69.04] - - - [15152, 2816, 1, 256] - - [53, 63.055] - - - [2560, 2560, 1, 256] - - [34, 57.421] - - - [8448, 2816, 1, 256] - - [55, 67.887] - - - [8704, 5632, 1, 256] - - [55, 72.624] - - - [1024, 769, 1, 256] - - [123, 35.338] - - - [17200, 4096, 1, 256] - - [44, 63.61] - - - [5376, 256, 1, 256] - - [183, 40.594] - - - [6656, 256, 1, 256] - - [182, 47.239] - - - [18688, 3328, 1, 256] - - [35, 71.918] - - - [13056, 256, 1, 256] - - [27, 51.76] - - - [13104, 2816, 1, 256] - - [36, 63.036] - - - [7424, 1792, 1, 256] - - [34, 63.557] - - - [14592, 2816, 1, 256] - - [59, 70.771] - - - [12336, 2865, 1, 256] - - [37, 59.727] - - - [17920, 256, 1, 256] - - [63, 49.106] - - - [12800, 2048, 1, 256] - - [36, 68.659] - - - [3632, 256, 1, 256] - - [121, 32.495] - - - [18688, 768, 1, 256] - - [53, 66.3] - - - [16384, 2816, 1, 256] - - [22, 62.061] - - - [14896, 10240, 1, 256] - - [25, 63.546] - - - [816, 817, 1, 256] - - [121, 30.034] - - - [9008, 2865, 1, 256] - - [27, 59.73] - - - [14848, 1024, 1, 256] - - [60, 64.645] - - - [16640, 256, 1, 256] - - [39, 45.255] - - - [7424, 256, 1, 256] - - [31, 31.555] - - - [10240, 3329, 1, 256] - - [29, 68.28] - - - [18176, 5120, 1, 256] - - [27, 73.841] - - - [6912, 2865, 1, 256] - - [25, 64.594] - - - [1024, 1025, 1, 256] - - [185, 35.087] - - - [5632, 4096, 1, 256] - - [40, 68.981] - - - [12544, 1024, 1, 256] - - [63, 61.314] - - - [2864, 2609, 1, 256] - - [27, 49.44] - - - [16896, 2048, 1, 256] - - [60, 69.794] - - - [3840, 1280, 1, 256] - - [49, 52.859] - - - [11008, 1281, 1, 256] - - [31, 56.944] - - - [15104, 1281, 1, 256] - - [23, 62.75] - - - [7168, 3329, 1, 256] - - [29, 66.789] - - - [12800, 5120, 1, 256] - - [59, 73.081] - - - [512, 257, 1, 256] - - [166, 8.338] - - - [12288, 8960, 1, 256] - - [27, 75.243] - - - [9728, 6912, 1, 256] - - [49, 74.008] - - - [9728, 6656, 1, 256] - - [34, 73.905] - - - [2560, 2304, 1, 256] - - [55, 52.115] - - - [10544, 7424, 1, 256] - - [27, 63.368] - - - [5888, 3329, 1, 256] - - [25, 64.312] - - - [3888, 2816, 1, 256] - - [27, 53.57] - - - [18944, 10240, 1, 256] - - [47, 75.238] - - - [17200, 10240, 1, 256] - - [29, 63.374] - - - [4144, 1280, 1, 256] - - [35, 53.132] - - - [9728, 1280, 1, 256] - - [34, 65.329] - - - [14896, 1536, 1, 256] - - [25, 60.265] - - - [5888, 4352, 1, 256] - - [34, 69.203] - - - [1024, 1024, 1, 256] - - [172, 36.317] - - - [4912, 2816, 1, 256] - - [37, 59.825] - - - [19456, 3329, 1, 256] - - [27, 69.935] - - - [7680, 4608, 1, 256] - - [55, 71.244] - - - [8496, 2865, 1, 256] - - [25, 59.372] - - - [3584, 2048, 1, 256] - - [34, 53.777] - - - [9984, 3329, 1, 256] - - [55, 66.875] - - - [10800, 7680, 1, 256] - - [44, 64.193] - - - [13616, 2816, 1, 256] - - [58, 62.517] - - - [15104, 10240, 1, 256] - - [27, 74.662] - - - [10240, 6656, 1, 256] - - [29, 74.233] - - - [16128, 1281, 1, 256] - - [69, 60.91] - - - [16896, 1280, 1, 256] - - [55, 68.869] - - - [12544, 9472, 1, 256] - - [25, 74.527] - - - [11008, 7424, 1, 256] - - [85, 73.158] - - - [9472, 3329, 1, 256] - - [37, 66.568] - - - [6912, 2816, 1, 256] - - [35, 67.987] - - - [2048, 1841, 1, 256] - - [40, 41.594] - - - [17152, 4096, 1, 256] - - [54, 72.113] - - - [12544, 5120, 1, 256] - - [49, 72.708] - - - [13824, 3328, 1, 256] - - [49, 71.534] - - - [6912, 2048, 1, 256] - - [40, 65.866] - - - [9472, 256, 1, 256] - - [57, 39.795] - - - [9216, 1281, 1, 256] - - [36, 56.597] - - - [7168, 1281, 1, 256] - - [49, 55.992] - - - [10752, 7424, 1, 256] - - [29, 74.342] - - - [16176, 3072, 1, 256] - - [37, 65.657] - - - [12288, 9216, 1, 256] - - [29, 74.52] - - - [14336, 512, 1, 256] - - [27, 52.927] - - - [14336, 3328, 1, 256] - - [25, 72.042] - - - [4864, 1280, 1, 256] - - [34, 54.748] - - - [19760, 2865, 1, 256] - - [36, 60.832] - - - [8240, 256, 1, 256] - - [34, 34.619] - - - [18688, 1024, 1, 256] - - [40, 66.573] - - - [16128, 10240, 1, 256] - - [27, 74.965] - - - [5632, 256, 1, 256] - - [183, 42.857] - - - [5680, 2560, 1, 256] - - [37, 57.404] - - - [7680, 1281, 1, 256] - - [27, 59.333] - - - [17408, 2048, 1, 256] - - [58, 69.586] - - - [10752, 2865, 1, 256] - - [25, 68.123] - - - [14848, 1281, 1, 256] - - [53, 62.3] - - - [560, 512, 1, 256] - - [109, 17.046] - - - [19968, 1280, 1, 256] - - [35, 68.465] - - - [16384, 10240, 1, 256] - - [22, 63.0] - - - [512, 305, 1, 256] - - [170, 9.788] - - - [19200, 6144, 1, 256] - - [29, 73.718] - - - [8448, 5120, 1, 256] - - [25, 72.211] - - - [13824, 3329, 1, 256] - - [29, 68.838] - - - [7984, 2816, 1, 256] - - [27, 61.507] - - - [17920, 3329, 1, 256] - - [51, 69.263] - - - [16688, 2865, 1, 256] - - [49, 63.796] - - - [12032, 256, 1, 256] - - [25, 49.22] - - - [7424, 2865, 1, 256] - - [27, 65.069] - - - [14336, 10240, 1, 256] - - [25, 75.425] - - - [17152, 2048, 1, 256] - - [63, 69.707] - - - [14896, 2816, 1, 256] - - [44, 62.421] - - - [16384, 2048, 1, 256] - - [38, 59.349] - - - [8192, 2816, 1, 256] - - [35, 68.313] - - - [6192, 256, 1, 256] - - [182, 42.027] - - - [2304, 768, 1, 256] - - [124, 48.821] - - - [18688, 256, 1, 256] - - [35, 50.462] - - - [8960, 1281, 1, 256] - - [49, 59.542] - - - [19968, 6400, 1, 256] - - [34, 74.803] - - - [8752, 2816, 1, 256] - - [29, 62.219] - - - [19456, 3328, 1, 256] - - [49, 72.717] - - - [2560, 2561, 1, 256] - - [34, 57.087] - - - [15920, 2865, 1, 256] - - [27, 61.767] - - - [12544, 6144, 1, 256] - - [25, 73.28] - - - [19200, 3328, 1, 256] - - [53, 71.858] - - - [3328, 2865, 1, 256] - - [34, 58.513] - - - [7936, 3329, 1, 256] - - [57, 65.631] - - - [11264, 2865, 1, 256] - - [29, 68.379] - - - [6144, 3329, 1, 256] - - [49, 66.13] - - - [16128, 3329, 1, 256] - - [25, 68.765] - - - [12800, 9728, 1, 256] - - [74, 74.84] - - - [512, 256, 1, 256] - - [166, 8.445] - - - [11264, 2816, 1, 256] - - [55, 70.812] - - - [12544, 3329, 1, 256] - - [27, 67.735] - - - [14848, 3329, 1, 256] - - [27, 69.124] - - - [1328, 256, 1, 256] - - [108, 19.803] - - - [3120, 256, 1, 256] - - [123, 35.025] - - - [1024, 768, 1, 256] - - [121, 35.569] - - - [7728, 2816, 1, 256] - - [23, 59.059] - - - [1024, 817, 1, 256] - - [121, 36.274] - - - [10288, 7424, 1, 256] - - [25, 63.45] - - - [19968, 6144, 1, 256] - - [41, 74.294] - - - [13616, 10240, 1, 256] - - [36, 63.511] - - - [1536, 1329, 1, 256] - - [69, 33.699] - - - [9984, 3328, 1, 256] - - [55, 70.3] - - - [9472, 5888, 1, 256] - - [29, 72.702] - - - [11264, 7936, 1, 256] - - [55, 74.907] - - - [8496, 256, 1, 256] - - [58, 35.102] - - - [17664, 1792, 1, 256] - - [34, 69.801] - - - [11824, 2816, 1, 256] - - [44, 62.525] - - - [16944, 2816, 1, 256] - - [29, 64.078] - - - [19968, 6912, 1, 256] - - [27, 74.983] - - - [3376, 2865, 1, 256] - - [27, 54.171] - - - [3840, 2560, 1, 256] - - [49, 60.629] - - - [11776, 8448, 1, 256] - - [27, 74.414] - - - [19248, 6144, 1, 256] - - [58, 63.548] - - - [14080, 512, 1, 256] - - [27, 61.021] - - - [16128, 3328, 1, 256] - - [49, 71.602] - - - [6656, 2048, 1, 256] - - [23, 63.698] - - - [15664, 256, 1, 256] - - [55, 42.258] - - - [17664, 1280, 1, 256] - - [35, 67.773] - - - [16384, 6144, 1, 256] - - [22, 63.174] - - - [9984, 256, 1, 256] - - [51, 41.767] - - - [14592, 1281, 1, 256] - - [34, 60.083] - - - [4608, 3329, 1, 256] - - [49, 64.498] - - - [8960, 2048, 1, 256] - - [34, 64.636] - - - [2864, 2865, 1, 256] - - [27, 53.819] - - - [2816, 2609, 1, 256] - - [34, 53.455] - - - [14080, 1281, 1, 256] - - [25, 60.65] - - - [1792, 1536, 1, 256] - - [31, 44.237] - - - [10240, 7424, 1, 256] - - [37, 74.721] - - - [5936, 2816, 1, 256] - - [37, 58.611] - - - [19712, 256, 1, 256] - - [40, 52.927] - - - [18944, 5888, 1, 256] - - [31, 74.241] - - - [9728, 3329, 1, 256] - - [27, 68.007] - - - [19248, 2816, 1, 256] - - [53, 62.64] - - - [13568, 1792, 1, 256] - - [49, 68.552] - - - [1584, 1585, 1, 256] - - [29, 38.322] - - - [8704, 2048, 1, 256] - - [40, 67.585] - - - [13056, 9728, 1, 256] - - [25, 74.687] - - - [12800, 2865, 1, 256] - - [31, 68.74] - - - [14336, 6144, 1, 256] - - [27, 74.178] - - - [5120, 1536, 1, 256] - - [35, 56.426] - - - [18432, 1281, 1, 256] - - [25, 61.726] - - - [10240, 256, 1, 256] - - [34, 42.778] - - - [12544, 9216, 1, 256] - - [37, 73.968] - - - [12800, 1281, 1, 256] - - [33, 61.721] - - - [8704, 5888, 1, 256] - - [25, 72.741] - - - [15360, 3329, 1, 256] - - [27, 69.087] - - - [11312, 8448, 1, 256] - - [25, 63.968] - - - [17152, 3328, 1, 256] - - [25, 71.81] - - - [16384, 3328, 1, 256] - - [38, 61.752] - - - [13824, 2816, 1, 256] - - [49, 71.392] - - - [560, 305, 1, 256] - - [113, 10.423] - - - [16432, 256, 1, 256] - - [23, 44.904] - - - [3632, 2865, 1, 256] - - [37, 51.217] - - - [3584, 3328, 1, 256] - - [35, 64.076] - - - [3840, 768, 1, 256] - - [58, 46.673] - - - [19504, 256, 1, 256] - - [60, 48.745] - - - [1280, 1073, 1, 256] - - [190, 40.001] - - - [17712, 10240, 1, 256] - - [53, 63.117] - - - [2816, 1536, 1, 256] - - [35, 47.445] - - - [12800, 6144, 1, 256] - - [59, 73.468] - - - [4656, 2816, 1, 256] - - [36, 57.374] - - - [17920, 10240, 1, 256] - - [37, 75.114] - - - [9984, 2816, 1, 256] - - [35, 69.679] - - - [4352, 256, 1, 256] - - [123, 38.47] - - - [11312, 2865, 1, 256] - - [25, 60.631] - - - [18432, 3328, 1, 256] - - [29, 72.629] - - - [4096, 1280, 1, 256] - - [106, 54.347] - - - [4864, 3329, 1, 256] - - [35, 63.146] - - - [14640, 2865, 1, 256] - - [37, 59.354] - - - [17152, 2816, 1, 256] - - [34, 71.888] - - - [7680, 1280, 1, 256] - - [35, 60.456] - - - [1584, 1536, 1, 256] - - [34, 37.94] - - - [14080, 1280, 1, 256] - - [35, 67.641] - - - [13824, 512, 1, 256] - - [27, 60.187] - - - [7936, 256, 1, 256] - - [26, 33.726] - - - [12592, 2865, 1, 256] - - [37, 60.455] - - - [2816, 2560, 1, 256] - - [49, 62.208] - - - [6912, 1536, 1, 256] - - [55, 64.344] - - - [12800, 9984, 1, 256] - - [25, 75.144] - - - [10496, 256, 1, 256] - - [40, 43.732] - - - [18176, 2865, 1, 256] - - [27, 69.649] - - - [4608, 1536, 1, 256] - - [35, 60.86] - - - [3328, 2816, 1, 256] - - [49, 58.508] - - - [3840, 1024, 1, 256] - - [35, 43.255] - - - [13824, 1280, 1, 256] - - [34, 67.666] - - - [3840, 1281, 1, 256] - - [49, 52.315] - - - [17152, 1281, 1, 256] - - [40, 62.238] - - - [13568, 1281, 1, 256] - - [40, 61.146] - - - [14848, 1792, 1, 256] - - [49, 70.07] - - - [13056, 9472, 1, 256] - - [25, 74.981] - - - [18176, 1281, 1, 256] - - [44, 62.024] - - - [5680, 256, 1, 256] - - [182, 39.74] - - - [13056, 2816, 1, 256] - - [35, 71.111] - - - [11824, 8704, 1, 256] - - [58, 64.185] - - - [7936, 4352, 1, 256] - - [57, 70.426] - - - [8704, 256, 1, 256] - - [59, 37.042] - - - [5424, 2304, 1, 256] - - [53, 56.662] - - - [14128, 768, 1, 256] - - [58, 52.887] - - - [10752, 1024, 1, 256] - - [40, 58.928] - - - [9264, 6144, 1, 256] - - [27, 63.829] - - - [4352, 3328, 1, 256] - - [35, 63.523] - - - [18944, 5632, 1, 256] - - [67, 74.526] - - - [12032, 8704, 1, 256] - - [49, 74.067] - - - [2048, 2049, 1, 256] - - [35, 45.631] - - - [6400, 3329, 1, 256] - - [35, 64.596] - - - [15616, 2560, 1, 256] - - [35, 71.001] - - - [7472, 2865, 1, 256] - - [29, 57.811] - - - [14848, 1536, 1, 256] - - [35, 68.167] - - - [18736, 10240, 1, 256] - - [58, 63.377] - - - [6400, 1024, 1, 256] - - [35, 57.281] - - - [7936, 5120, 1, 256] - - [37, 71.824] - - - [4656, 1536, 1, 256] - - [35, 47.999] - - - [3328, 256, 1, 256] - - [114, 37.36] - - - [3072, 1280, 1, 256] - - [35, 43.671] - - - [2864, 2816, 1, 256] - - [35, 53.201] - - - [9472, 6144, 1, 256] - - [55, 72.458] - - - [3840, 2304, 1, 256] - - [25, 62.014] - - - [17408, 2865, 1, 256] - - [25, 69.641] - - - [16384, 2560, 1, 256] - - [38, 60.625] - - - [16384, 3329, 1, 256] - - [29, 58.92] - - - [16688, 10240, 1, 256] - - [25, 65.624] - - - [18688, 2048, 1, 256] - - [63, 69.803] - - - [7936, 4608, 1, 256] - - [25, 70.813] - - - [9472, 6400, 1, 256] - - [34, 73.356] - - - [14336, 3329, 1, 256] - - [29, 69.103] - - - [4608, 1024, 1, 256] - - [55, 50.887] - - - [16896, 6144, 1, 256] - - [27, 74.464] - - - [10752, 3329, 1, 256] - - [27, 68.157] - - - [6704, 256, 1, 256] - - [181, 44.38] - - - [17408, 6144, 1, 256] - - [25, 74.582] - - - [9984, 2048, 1, 256] - - [63, 66.727] - - - [17968, 4864, 1, 256] - - [36, 63.918] - - - [5120, 3584, 1, 256] - - [34, 66.629] - - - [14336, 2865, 1, 256] - - [37, 69.431] - - - [18736, 256, 1, 256] - - [53, 47.713] - - - [13568, 2048, 1, 256] - - [40, 68.59] - - - [17456, 4352, 1, 256] - - [25, 64.048] - - - [5424, 2816, 1, 256] - - [53, 59.939] - - - [17664, 4608, 1, 256] - - [59, 72.479] - - - [7984, 256, 1, 256] - - [44, 33.496] - - - [6400, 3584, 1, 256] - - [49, 69.343] - - - [19712, 3329, 1, 256] - - [47, 67.597] - - - [10752, 2816, 1, 256] - - [59, 70.312] - - - [17152, 1280, 1, 256] - - [35, 67.098] - - - [560, 561, 1, 256] - - [114, 18.392] - - - [8192, 1281, 1, 256] - - [27, 55.243] - - - [4864, 1792, 1, 256] - - [34, 61.43] - - - [3632, 2816, 1, 256] - - [44, 57.042] - - - [11520, 8448, 1, 256] - - [29, 73.729] - - - [5168, 2865, 1, 256] - - [27, 57.066] - - - [13568, 10240, 1, 256] - - [29, 74.74] - - - [12544, 2816, 1, 256] - - [27, 70.656] - - - [19504, 6144, 1, 256] - - [29, 63.873] - - - [11776, 2048, 1, 256] - - [39, 67.352] - - - [18688, 2865, 1, 256] - - [34, 69.555] - - - [14336, 768, 1, 256] - - [58, 59.929] - - - [18688, 6144, 1, 256] - - [29, 74.125] - - - [11776, 2816, 1, 256] - - [31, 70.084] - - - [12288, 9472, 1, 256] - - [37, 75.314] - - - [5120, 1792, 1, 256] - - [27, 57.697] - - - [16128, 512, 1, 256] - - [49, 56.775] - - - [5376, 3329, 1, 256] - - [35, 63.193] - - - [9216, 3329, 1, 256] - - [34, 67.666] - - - [9008, 2816, 1, 256] - - [25, 59.908] - - - [6448, 3328, 1, 256] - - [27, 58.765] - - - [19968, 3329, 1, 256] - - [27, 69.474] - - - [11520, 8704, 1, 256] - - [67, 74.005] - - - [13824, 256, 1, 256] - - [35, 54.449] - - - [1584, 256, 1, 256] - - [116, 23.152] - - - [10496, 7168, 1, 256] - - [37, 72.807] - - - [5376, 2304, 1, 256] - - [34, 64.441] - - - [10752, 7168, 1, 256] - - [27, 73.268] - - - [18432, 2048, 1, 256] - - [78, 68.684] - - - [12080, 256, 1, 256] - - [36, 46.156] - - - [8704, 3328, 1, 256] - - [34, 69.305] - - - [4608, 1280, 1, 256] - - [34, 51.827] - - - [6192, 3328, 1, 256] - - [27, 60.553] - - - [8704, 3329, 1, 256] - - [55, 67.535] - - - [5424, 2560, 1, 256] - - [23, 61.133] - - - [11008, 2816, 1, 256] - - [85, 67.657] - - - [11776, 4352, 1, 256] - - [59, 72.393] - - - [11008, 1536, 1, 256] - - [57, 62.887] - - - [13312, 3328, 1, 256] - - [27, 71.901] - - - [7168, 4096, 1, 256] - - [27, 69.656] - - - [9216, 256, 1, 256] - - [40, 38.837] - - - [19504, 2865, 1, 256] - - [37, 61.604] - - - [5936, 2865, 1, 256] - - [25, 59.11] - - - [1840, 1792, 1, 256] - - [25, 47.791] - - - [19968, 2816, 1, 256] - - [49, 72.336] - - - [9008, 256, 1, 256] - - [58, 36.85] - - - [9728, 256, 1, 256] - - [25, 40.936] - - - [11056, 7936, 1, 256] - - [27, 64.522] - - - [7680, 3329, 1, 256] - - [37, 67.112] - - - [1792, 256, 1, 256] - - [114, 22.324] - - - [17664, 10240, 1, 256] - - [51, 74.207] - - - [11776, 2865, 1, 256] - - [31, 67.578] - - - [512, 512, 1, 256] - - [176, 15.996] - - - [16640, 768, 1, 256] - - [23, 59.572] - - - [4352, 2048, 1, 256] - - [25, 61.808] - - - [19504, 2816, 1, 256] - - [36, 63.411] - - - [12080, 2865, 1, 256] - - [27, 60.842] - - - [14080, 768, 1, 256] - - [60, 64.176] - - - [7936, 512, 1, 256] - - [40, 44.612] - - - [5376, 2560, 1, 256] - - [55, 64.677] - - - [5632, 3329, 1, 256] - - [27, 65.402] - - - [5120, 3840, 1, 256] - - [49, 68.426] - - - [6192, 2816, 1, 256] - - [27, 60.678] - - - [4608, 3072, 1, 256] - - [34, 67.182] - - - [19712, 6656, 1, 256] - - [59, 73.027] - - - [14896, 256, 1, 256] - - [34, 40.437] - - - [6400, 1280, 1, 256] - - [34, 58.678] - - - [12800, 9216, 1, 256] - - [74, 74.395] - - - [15616, 256, 1, 256] - - [23, 43.934] - - - [17920, 4608, 1, 256] - - [61, 73.278] - - - [7936, 2865, 1, 256] - - [31, 66.845] - - - [13312, 3329, 1, 256] - - [27, 68.915] - - - [5168, 2304, 1, 256] - - [25, 57.61] - - - [14128, 10240, 1, 256] - - [58, 63.75] - - - [3840, 2816, 1, 256] - - [34, 65.639] - - - [8960, 1536, 1, 256] - - [34, 64.968] - - - [3328, 3073, 1, 256] - - [27, 61.561] - - - [4096, 3328, 1, 256] - - [55, 64.747] - - - [14592, 2048, 1, 256] - - [39, 67.195] - - - [9728, 2048, 1, 256] - - [23, 65.56] - - - [13312, 5888, 1, 256] - - [55, 74.216] - - - [15616, 2304, 1, 256] - - [27, 71.067] - - - [19712, 2816, 1, 256] - - [51, 70.242] - - - [9216, 5888, 1, 256] - - [35, 73.129] - - - [7168, 4352, 1, 256] - - [27, 71.0] - - - [9520, 6400, 1, 256] - - [36, 62.601] - - - [13568, 3328, 1, 256] - - [34, 70.67] - - - [17408, 4352, 1, 256] - - [34, 74.236] - - - [11056, 2865, 1, 256] - - [29, 59.852] - - - [18480, 2865, 1, 256] - - [29, 60.402] - - - [13824, 768, 1, 256] - - [40, 63.874] - - - [17664, 6144, 1, 256] - - [67, 73.429] - - - [7216, 4352, 1, 256] - - [27, 62.523] - - - [14128, 2865, 1, 256] - - [37, 60.047] - - - [11520, 3328, 1, 256] - - [36, 69.856] - - - [18992, 5888, 1, 256] - - [44, 63.717] - - - [17408, 10240, 1, 256] - - [37, 75.574] - - - [15104, 6144, 1, 256] - - [27, 73.579] - - - [16640, 1280, 1, 256] - - [60, 65.558] - - - [13056, 2865, 1, 256] - - [27, 68.468] - - - [11776, 8960, 1, 256] - - [49, 74.797] - - - [11312, 2816, 1, 256] - - [27, 62.941] - - - [11264, 3328, 1, 256] - - [55, 71.228] - - - [8192, 512, 1, 256] - - [78, 44.765] - - - [14848, 6144, 1, 256] - - [37, 73.826] - - - [10496, 7680, 1, 256] - - [37, 73.763] - - - [2816, 2817, 1, 256] - - [35, 56.435] - - - [15104, 2865, 1, 256] - - [27, 69.417] - - - [18176, 3329, 1, 256] - - [25, 69.35] - - - [3328, 1792, 1, 256] - - [27, 52.483] - - - [6144, 3328, 1, 256] - - [35, 67.918] - - - [12288, 6144, 1, 256] - - [37, 73.86] - - - [8960, 5888, 1, 256] - - [57, 71.764] - - - [3584, 1280, 1, 256] - - [34, 50.073] - - - [7728, 4608, 1, 256] - - [53, 63.336] - - - [18176, 6144, 1, 256] - - [29, 74.114] - - - [16944, 256, 1, 256] - - [34, 45.256] - - - [3888, 768, 1, 256] - - [40, 45.092] - - - [8448, 3329, 1, 256] - - [29, 66.474] - - - [3072, 3073, 1, 256] - - [55, 57.596] - - - [4912, 256, 1, 256] - - [109, 40.173] - - - [5936, 3072, 1, 256] - - [27, 58.598] - - - [7168, 2865, 1, 256] - - [27, 66.573] - - - [19456, 10240, 1, 256] - - [25, 75.579] - - - [1840, 1585, 1, 256] - - [60, 43.249] - - - [18992, 5632, 1, 256] - - [44, 64.215] - - - [4912, 1792, 1, 256] - - [49, 56.232] - - - [8704, 6144, 1, 256] - - [34, 72.779] - - - [816, 768, 1, 256] - - [123, 29.143] - - - [18432, 2865, 1, 256] - - [25, 69.917] - - - [3120, 2865, 1, 256] - - [27, 51.19] - - - [6448, 2865, 1, 256] - - [35, 57.933] - - - [12080, 2816, 1, 256] - - [53, 62.716] - - - [10496, 3328, 1, 256] - - [60, 70.271] - - - [15920, 10240, 1, 256] - - [53, 64.072] - - - [15872, 2048, 1, 256] - - [58, 68.977] - - - [11568, 2816, 1, 256] - - [36, 61.211] - - - [19200, 10240, 1, 256] - - [37, 74.775] - - - [13312, 5632, 1, 256] - - [37, 74.331] - - - [15360, 2816, 1, 256] - - [29, 71.927] - - - [12288, 2865, 1, 256] - - [32, 68.287] - - - [19760, 6400, 1, 256] - - [44, 64.027] - - - [19968, 256, 1, 256] - - [44, 53.43] - - - [7680, 4352, 1, 256] - - [55, 71.265] - - - [11008, 3584, 1, 256] - - [75, 70.462] - - - [3072, 2817, 1, 256] - - [83, 54.528] - - - [11264, 6144, 1, 256] - - [29, 73.603] - - - [5424, 256, 1, 256] - - [110, 38.221] - - - [13568, 1280, 1, 256] - - [25, 66.879] - - - [3840, 2048, 1, 256] - - [27, 56.612] - - - [6144, 3072, 1, 256] - - [25, 67.753] - - - [19200, 1536, 1, 256] - - [57, 68.062] - - - [10240, 1280, 1, 256] - - [49, 62.93] - - - [3376, 512, 1, 256] - - [183, 44.698] - - - [12544, 1281, 1, 256] - - [40, 61.48] - - - [9776, 6656, 1, 256] - - [44, 63.826] - - - [7680, 2865, 1, 256] - - [25, 66.362] - - - [10544, 7680, 1, 256] - - [53, 63.447] - - - [15616, 3328, 1, 256] - - [25, 71.606] - - - [3328, 1280, 1, 256] - - [34, 46.673] - - - [2560, 1024, 1, 256] - - [34, 42.786] - - - [17456, 4096, 1, 256] - - [58, 64.105] - - - [6912, 3328, 1, 256] - - [55, 68.767] - - - [3584, 2816, 1, 256] - - [34, 62.134] - - - [17408, 3328, 1, 256] - - [25, 72.343] - - - [19200, 2816, 1, 256] - - [55, 71.241] - - - [15104, 1792, 1, 256] - - [35, 68.538] - - - [6144, 256, 1, 256] - - [182, 44.901] - - - [8192, 6144, 1, 256] - - [27, 72.019] - - - [12032, 4608, 1, 256] - - [35, 72.088] - - - [1840, 256, 1, 256] - - [108, 22.542] - - - [13312, 256, 1, 256] - - [34, 52.493] - - - [9216, 1792, 1, 256] - - [27, 65.175] - - - [14592, 3329, 1, 256] - - [57, 68.572] - - - [8448, 1280, 1, 256] - - [55, 64.286] - - - [11520, 8192, 1, 256] - - [57, 73.806] - - - [2608, 2560, 1, 256] - - [27, 53.501] - - - [5120, 2304, 1, 256] - - [49, 62.892] - - - [13056, 3328, 1, 256] - - [35, 70.928] - - - [11008, 8192, 1, 256] - - [75, 72.75] - - - [14896, 2865, 1, 256] - - [27, 60.701] - - - [6704, 3840, 1, 256] - - [27, 61.112] - - - [15872, 3329, 1, 256] - - [25, 69.435] - - - [7168, 1792, 1, 256] - - [55, 62.425] - - - [4656, 2865, 1, 256] - - [37, 57.651] - - - [18736, 5632, 1, 256] - - [58, 63.345] - - - [768, 512, 1, 256] - - [108, 23.02] - - - [16432, 3072, 1, 256] - - [38, 65.493] - - - [14848, 2865, 1, 256] - - [27, 68.873] - - - [4864, 1536, 1, 256] - - [35, 53.95] - - - [9472, 2865, 1, 256] - - [34, 66.928] - - - [10496, 2048, 1, 256] - - [60, 68.578] - - - [14336, 1024, 1, 256] - - [40, 63.079] - - - [18432, 256, 1, 256] - - [40, 50.137] - - - [16896, 3840, 1, 256] - - [67, 73.437] - - - [10240, 512, 1, 256] - - [40, 54.602] - - - [15664, 2304, 1, 256] - - [58, 62.137] - - - [10496, 3329, 1, 256] - - [37, 66.755] - - - [19456, 1536, 1, 256] - - [35, 69.248] - - - [17920, 1281, 1, 256] - - [33, 63.538] - - - [8960, 256, 1, 256] - - [34, 37.859] - - - [10496, 768, 1, 256] - - [53, 56.926] - - - [5120, 2816, 1, 256] - - [49, 66.912] - - - [12288, 2048, 1, 256] - - [37, 66.331] - - - [11568, 8704, 1, 256] - - [36, 63.361] - - - [10496, 1024, 1, 256] - - [40, 64.161] - - - [10288, 256, 1, 256] - - [34, 41.229] - - - [5168, 2048, 1, 256] - - [25, 59.001] - - - [11776, 3328, 1, 256] - - [57, 71.261] - - - [15152, 10240, 1, 256] - - [53, 63.573] - - - [14384, 2865, 1, 256] - - [27, 60.698] - - - [12288, 512, 1, 256] - - [84, 54.504] - - - [16688, 256, 1, 256] - - [34, 44.404] - - - [6912, 4096, 1, 256] - - [25, 69.673] - - - [4864, 2048, 1, 256] - - [35, 60.744] - - - [4096, 1024, 1, 256] - - [49, 45.746] - - - [12848, 9984, 1, 256] - - [53, 64.505] - - - [16896, 1281, 1, 256] - - [53, 61.455] - - - [768, 561, 1, 256] - - [114, 24.478] - - - [16896, 3584, 1, 256] - - [27, 73.171] - - - [14592, 6144, 1, 256] - - [67, 72.431] - - - [17664, 4096, 1, 256] - - [47, 71.584] - - - [8448, 2865, 1, 256] - - [35, 67.332] - - - [18432, 768, 1, 256] - - [53, 65.406] - - - [12032, 512, 1, 256] - - [23, 53.772] - - - [11008, 256, 1, 256] - - [79, 44.969] - - - [15360, 1536, 1, 256] - - [35, 67.238] - - - [5888, 2048, 1, 256] - - [40, 63.123] - - - [13104, 256, 1, 256] - - [63, 49.375] - - - [11264, 7680, 1, 256] - - [29, 74.867] - - - [19248, 2865, 1, 256] - - [58, 60.917] - - - [17200, 2865, 1, 256] - - [27, 61.583] - - - [8192, 2048, 1, 256] - - [37, 64.148] - - - [7472, 4608, 1, 256] - - [53, 61.625] - - - [7168, 2048, 1, 256] - - [23, 63.091] - - - [13360, 2816, 1, 256] - - [58, 62.643] - - - [17920, 4352, 1, 256] - - [57, 73.831] - - - [15408, 256, 1, 256] - - [40, 41.789] - - - [19200, 1281, 1, 256] - - [60, 63.126] - - - [15360, 256, 1, 256] - - [55, 43.165] - - - [9984, 6400, 1, 256] - - [35, 73.374] - - - [18944, 2865, 1, 256] - - [37, 69.777] - - - [3840, 2865, 1, 256] - - [35, 59.129] - - - [8192, 3328, 1, 256] - - [27, 68.008] - - - [5888, 256, 1, 256] - - [191, 43.899] - - - [15616, 2816, 1, 256] - - [27, 71.289] - - - [17664, 2865, 1, 256] - - [59, 69.269] - - - [14592, 768, 1, 256] - - [40, 59.758] - - - [18944, 1281, 1, 256] - - [72, 63.675] - - - [11264, 1536, 1, 256] - - [49, 66.394] - - - [8496, 5632, 1, 256] - - [34, 64.342] - - - [17664, 3328, 1, 256] - - [31, 71.465] - - - [14848, 2048, 1, 256] - - [40, 69.541] - - - [15408, 2865, 1, 256] - - [25, 61.565] - - - [4096, 2048, 1, 256] - - [59, 58.762] - - - [14128, 1024, 1, 256] - - [44, 55.462] - - - [1072, 817, 1, 256] - - [123, 30.424] - - - [17152, 3840, 1, 256] - - [25, 72.897] - - - [17664, 2048, 1, 256] - - [53, 68.715] - - - [16896, 256, 1, 256] - - [34, 46.769] - - - [2304, 2097, 1, 256] - - [49, 51.622] - - - [5888, 2560, 1, 256] - - [49, 64.841] - - - [9472, 1792, 1, 256] - - [25, 65.661] - - - [1328, 1280, 1, 256] - - [34, 36.117] - - - [19200, 1280, 1, 256] - - [57, 68.368] - - - [12544, 1280, 1, 256] - - [25, 66.492] - - - [16432, 3328, 1, 256] - - [28, 66.565] - - - [17920, 1280, 1, 256] - - [57, 68.79] - - - [8752, 5632, 1, 256] - - [29, 63.332] - - - [7936, 2048, 1, 256] - - [44, 63.392] - - - [9472, 1280, 1, 256] - - [27, 63.42] - - - [16896, 1024, 1, 256] - - [40, 66.25] - - - [6656, 3329, 1, 256] - - [59, 65.83] - - - [17456, 2865, 1, 256] - - [29, 62.344] - - - [5632, 2304, 1, 256] - - [49, 62.835] - - - [14080, 1024, 1, 256] - - [33, 65.233] - - - [15872, 3328, 1, 256] - - [34, 72.279] - - - [5168, 2816, 1, 256] - - [25, 56.571] - - - [13312, 9728, 1, 256] - - [25, 75.464] - - - [1584, 1329, 1, 256] - - [27, 33.18] - - - [15664, 2560, 1, 256] - - [37, 62.896] - - - [2048, 768, 1, 256] - - [117, 44.901] - - - [17712, 2816, 1, 256] - - [53, 63.326] - - - [16128, 2865, 1, 256] - - [27, 68.888] - - - [15872, 2816, 1, 256] - - [35, 72.252] - - - [18224, 2816, 1, 256] - - [44, 62.778] - - - [5632, 4352, 1, 256] - - [35, 69.961] - - - [1792, 1281, 1, 256] - - [49, 37.52] - - - [6656, 2816, 1, 256] - - [55, 66.455] - - - [16640, 1281, 1, 256] - - [65, 60.184] - - - [13056, 10240, 1, 256] - - [29, 74.986] - - - [17968, 256, 1, 256] - - [25, 46.195] - - - [5376, 4096, 1, 256] - - [27, 67.138] - - - [15152, 2048, 1, 256] - - [58, 60.877] - - - [13568, 2816, 1, 256] - - [34, 70.507] - - - [12800, 2816, 1, 256] - - [31, 70.881] - - - [6960, 2816, 1, 256] - - [29, 61.413] - - - [17968, 4608, 1, 256] - - [44, 63.285] - - - [15104, 3328, 1, 256] - - [23, 71.526] - - - [7472, 4352, 1, 256] - - [25, 61.464] - - - [15872, 2304, 1, 256] - - [35, 71.22] - - - [4400, 2816, 1, 256] - - [27, 59.033] - - - [16128, 6144, 1, 256] - - [27, 73.909] - - - [18944, 2816, 1, 256] - - [67, 72.175] - - - [5424, 2865, 1, 256] - - [44, 60.264] - - - [8192, 768, 1, 256] - - [34, 52.953] - - - [12848, 256, 1, 256] - - [40, 48.34] - - - [12288, 1281, 1, 256] - - [27, 60.225] - - - [13872, 512, 1, 256] - - [25, 55.706] - - - [5888, 1280, 1, 256] - - [49, 54.731] - - - [2816, 1281, 1, 256] - - [35, 41.044] - - - [19200, 3329, 1, 256] - - [67, 68.61] - - - [12800, 3328, 1, 256] - - [57, 71.351] - - - [15360, 2865, 1, 256] - - [27, 69.192] - - - [17152, 3584, 1, 256] - - [55, 72.518] - - - [17456, 2816, 1, 256] - - [29, 63.271] - - - [18176, 10240, 1, 256] - - [25, 74.987] - - - [6144, 2816, 1, 256] - - [49, 67.471] - - - [18176, 1280, 1, 256] - - [34, 68.702] - - - [16384, 1281, 1, 256] - - [27, 53.321] - - - [9216, 3328, 1, 256] - - [25, 69.932] - - - [14080, 2816, 1, 256] - - [67, 70.205] - - - [18688, 2816, 1, 256] - - [35, 71.742] - - - [15872, 10240, 1, 256] - - [25, 75.281] - - - [10800, 7936, 1, 256] - - [44, 64.944] - - - [9984, 1281, 1, 256] - - [40, 60.139] - - - [4144, 256, 1, 256] - - [110, 36.091] - - - [16640, 6144, 1, 256] - - [27, 73.468] - - - [11776, 4096, 1, 256] - - [74, 71.643] - - - [11056, 8192, 1, 256] - - [29, 64.782] - - - [5376, 2816, 1, 256] - - [55, 64.272] - - - [19712, 1280, 1, 256] - - [25, 67.903] - - - [4608, 2816, 1, 256] - - [55, 63.153] - - - [19456, 2865, 1, 256] - - [27, 70.001] - - - [14080, 256, 1, 256] - - [53, 54.728] - - - [7216, 4096, 1, 256] - - [29, 61.81] - - - [2816, 1280, 1, 256] - - [49, 54.728] - - - [10496, 3072, 1, 256] - - [34, 70.522] - - - [12544, 4864, 1, 256] - - [55, 72.899] - - - [9984, 6912, 1, 256] - - [34, 73.374] - - - [4912, 2048, 1, 256] - - [44, 56.406] - - - [9984, 2304, 1, 256] - - [35, 68.987] - - - [19248, 5888, 1, 256] - - [36, 63.654] - - - [19712, 2048, 1, 256] - - [84, 67.452] - - - [9728, 2816, 1, 256] - - [34, 69.216] - - - [19504, 6400, 1, 256] - - [44, 64.123] - - - [16896, 3072, 1, 256] - - [34, 72.467] - - - [15104, 1536, 1, 256] - - [35, 67.917] - - - [2608, 2609, 1, 256] - - [25, 46.653] - - - [14384, 256, 1, 256] - - [40, 39.048] - - - [17664, 2816, 1, 256] - - [31, 71.206] - - - [9776, 6912, 1, 256] - - [44, 64.424] - - - [1792, 512, 1, 256] - - [108, 33.381] - - - [13312, 1536, 1, 256] - - [49, 66.916] - - - [1072, 1073, 1, 256] - - [108, 36.546] - - - [9472, 2048, 1, 256] - - [40, 67.028] - - - [4608, 3328, 1, 256] - - [55, 65.809] - - - [9984, 2560, 1, 256] - - [35, 68.769] - - - [6912, 5376, 1, 256] - - [49, 71.656] - - - [16640, 2865, 1, 256] - - [34, 68.013] - - - [4352, 3072, 1, 256] - - [34, 63.961] - - - [5632, 3328, 1, 256] - - [35, 67.157] - - - [9216, 5632, 1, 256] - - [29, 73.616] - - - [3328, 3329, 1, 256] - - [34, 58.687] - - - [13824, 10240, 1, 256] - - [51, 74.807] - - - [12288, 3329, 1, 256] - - [27, 68.787] - - - [2864, 256, 1, 256] - - [110, 33.16] - - - [19712, 1792, 1, 256] - - [57, 68.68] - - - [6656, 1280, 1, 256] - - [49, 59.572] - - - [13056, 2048, 1, 256] - - [60, 68.872] - - - [6912, 1281, 1, 256] - - [49, 55.428] - - - [16176, 2816, 1, 256] - - [25, 65.223] - - - [14592, 3328, 1, 256] - - [57, 70.537] - - - [10496, 2865, 1, 256] - - [25, 66.963] - - - [9728, 6400, 1, 256] - - [34, 73.801] - - - [5888, 4608, 1, 256] - - [55, 68.523] - - - [16432, 10240, 1, 256] - - [54, 68.845] - - - [19456, 5888, 1, 256] - - [27, 74.612] - - - [3888, 256, 1, 256] - - [110, 34.576] - - - [12336, 2816, 1, 256] - - [27, 61.954] - - - [19456, 256, 1, 256] - - [40, 52.392] - - - [14384, 2816, 1, 256] - - [29, 62.004] - - - [14384, 1024, 1, 256] - - [25, 56.205] - - - [16640, 2816, 1, 256] - - [34, 70.493] - - - [3840, 3329, 1, 256] - - [35, 61.31] - - - [2304, 2304, 1, 256] - - [35, 56.085] - - - [10240, 2816, 1, 256] - - [27, 71.023] - - - [13104, 10240, 1, 256] - - [36, 64.67] - - - [1536, 256, 1, 256] - - [116, 22.905] - - - [11008, 2865, 1, 256] - - [85, 67.12] - - - [13104, 9984, 1, 256] - - [44, 64.266] - - - [10240, 7168, 1, 256] - - [25, 73.574] - - - [3888, 2865, 1, 256] - - [25, 53.93] - - - [8192, 4864, 1, 256] - - [25, 70.814] - - - [15920, 256, 1, 256] - - [23, 43.417] - - - [6448, 3584, 1, 256] - - [53, 61.758] - - - [16128, 2304, 1, 256] - - [35, 70.651] - - - [9728, 2865, 1, 256] - - [27, 68.808] - - - [6144, 4864, 1, 256] - - [37, 71.369] - - - [14848, 256, 1, 256] - - [60, 42.175] - - - [4352, 1024, 1, 256] - - [34, 48.505] - - - [15360, 10240, 1, 256] - - [29, 75.504] - - - [19504, 10240, 1, 256] - - [28, 63.61] - - - [3328, 3072, 1, 256] - - [34, 62.752] - - - [1536, 1281, 1, 256] - - [31, 32.663] - - - [19760, 6656, 1, 256] - - [58, 63.827] - - - [3584, 3329, 1, 256] - - [34, 62.683] - - - [14848, 2816, 1, 256] - - [34, 71.299] - - - [4400, 2865, 1, 256] - - [27, 54.738] - - - [3888, 1024, 1, 256] - - [44, 41.614] - - - [16640, 2048, 1, 256] - - [40, 68.188] - - - [4096, 2816, 1, 256] - - [59, 61.935] - - - [14640, 2816, 1, 256] - - [44, 62.79] - - - [9472, 1281, 1, 256] - - [34, 57.761] - - - [8192, 1280, 1, 256] - - [35, 62.359] - - - [8960, 2865, 1, 256] - - [49, 67.185] - - - [4144, 2816, 1, 256] - - [25, 57.389] - - - [10288, 7168, 1, 256] - - [25, 62.857] - - - [14592, 256, 1, 256] - - [40, 41.321] - - - [10240, 2048, 1, 256] - - [36, 66.868] - - - [17920, 2865, 1, 256] - - [47, 69.514] - - - [12592, 2816, 1, 256] - - [58, 62.814] - - - [14592, 1536, 1, 256] - - [35, 66.733] - - - [11568, 256, 1, 256] - - [29, 44.265] - - - [6704, 3584, 1, 256] - - [29, 61.656] - - - [5120, 3328, 1, 256] - - [49, 66.584] - - - [4400, 1536, 1, 256] - - [35, 53.849] - - - [18944, 256, 1, 256] - - [35, 51.106] - - - [19712, 5888, 1, 256] - - [41, 72.495] - - - [7984, 5120, 1, 256] - - [34, 64.029] - - - [8240, 2865, 1, 256] - - [35, 61.677] - - - [6144, 1280, 1, 256] - - [49, 56.718] - - - [8496, 2816, 1, 256] - - [37, 61.931] - - - [14592, 1024, 1, 256] - - [27, 62.882] - - - [14592, 2865, 1, 256] - - [31, 68.747] - - - [13360, 256, 1, 256] - - [53, 50.34] - - - [8448, 256, 1, 256] - - [59, 36.005] - - - [16896, 2816, 1, 256] - - [55, 72.165] - - - [15152, 2865, 1, 256] - - [37, 61.182] - - - [11056, 2816, 1, 256] - - [44, 62.578] - - - [15616, 1280, 1, 256] - - [49, 66.097] - - - [8192, 5120, 1, 256] - - [27, 70.894] - - - [17408, 256, 1, 256] - - [40, 48.142] - - - [18432, 10240, 1, 256] - - [37, 75.394] - - - [14592, 1280, 1, 256] - - [35, 65.738] - - - [3328, 512, 1, 256] - - [181, 48.048] - - - [14336, 1280, 1, 256] - - [49, 66.254] - - - [13616, 2865, 1, 256] - - [36, 60.21] - - - [8192, 256, 1, 256] - - [35, 35.021] - - - [10240, 1281, 1, 256] - - [49, 60.343] - - - [1840, 1841, 1, 256] - - [34, 35.183] - - - [12800, 9472, 1, 256] - - [29, 75.03] - - - [17664, 256, 1, 256] - - [40, 48.45] - - - [768, 769, 1, 256] - - [109, 28.381] - - - [19456, 2048, 1, 256] - - [36, 69.833] - - - [13056, 3329, 1, 256] - - [37, 67.803] - - - [11056, 256, 1, 256] - - [27, 43.423] - - - [7424, 6144, 1, 256] - - [35, 71.699] - - - [14848, 3328, 1, 256] - - [29, 71.856] - - - [6656, 3328, 1, 256] - - [53, 67.747] - - - [10752, 1281, 1, 256] - - [40, 59.249] - - - [9984, 2865, 1, 256] - - [27, 67.115] - - - [14080, 3329, 1, 256] - - [57, 68.263] - - - [17920, 3328, 1, 256] - - [76, 72.159] - - - [13312, 10240, 1, 256] - - [27, 75.296] - - - [16640, 3584, 1, 256] - - [34, 71.857] - - - [17408, 3840, 1, 256] - - [37, 73.865] - - - [12032, 8960, 1, 256] - - [34, 74.177] - - - [10800, 2865, 1, 256] - - [53, 61.201] - - - [3072, 2816, 1, 256] - - [34, 60.572] - - - [14128, 2816, 1, 256] - - [58, 61.885] - - - [11312, 8192, 1, 256] - - [29, 64.421] - - - [2560, 2305, 1, 256] - - [35, 51.575] - - - [16640, 3072, 1, 256] - - [27, 70.804] - - - [16128, 2048, 1, 256] - - [40, 68.171] - - - [6144, 512, 1, 256] - - [60, 49.726] - - - [18688, 4864, 1, 256] - - [35, 73.833] - - - [17200, 256, 1, 256] - - [40, 45.57] - - - [8752, 2865, 1, 256] - - [29, 58.991] - - - [18944, 1280, 1, 256] - - [49, 68.497] - - - [16640, 3328, 1, 256] - - [48, 70.911] - - - [304, 305, 1, 256] - - [119, 5.569] - - - [15104, 256, 1, 256] - - [40, 42.985] - - - [7680, 3328, 1, 256] - - [35, 68.466] - - - [12336, 9216, 1, 256] - - [25, 63.141] - - - [14080, 6144, 1, 256] - - [67, 72.26] - - - [7168, 5888, 1, 256] - - [27, 72.398] - - - [7424, 1280, 1, 256] - - [34, 59.134] - - - [4864, 3584, 1, 256] - - [55, 67.73] - - - [1280, 1025, 1, 256] - - [183, 38.019] - - - [10240, 2865, 1, 256] - - [27, 68.642] - - - [18480, 10240, 1, 256] - - [25, 62.84] - - - [7680, 256, 1, 256] - - [47, 32.732] - - - [9472, 6656, 1, 256] - - [34, 73.381] - - - [12032, 6144, 1, 256] - - [29, 72.815] - - - [5120, 3329, 1, 256] - - [34, 65.105] - - - [10752, 256, 1, 256] - - [27, 44.79] - - - [6960, 256, 1, 256] - - [34, 39.312] - - - [9008, 6144, 1, 256] - - [58, 63.58] - - - [7424, 2048, 1, 256] - - [60, 64.503] - - - [5632, 1280, 1, 256] - - [35, 61.021] - - - [19712, 10240, 1, 256] - - [41, 73.761] - - - [6400, 768, 1, 256] - - [63, 52.653] - - - [10752, 3328, 1, 256] - - [27, 71.314] - - - [18432, 5376, 1, 256] - - [27, 74.701] - - - [9520, 256, 1, 256] - - [40, 38.314] - - - [5680, 2816, 1, 256] - - [25, 60.823] - - - [11008, 3329, 1, 256] - - [75, 65.949] - - - [4608, 2865, 1, 256] - - [35, 62.836] - - - [6448, 256, 1, 256] - - [192, 43.664] - - - [3584, 256, 1, 256] - - [114, 33.276] - - - [12336, 9472, 1, 256] - - [37, 63.293] - - - [1280, 1281, 1, 256] - - [63, 36.52] - - - [7936, 6144, 1, 256] - - [59, 71.94] - - - [15152, 1792, 1, 256] - - [37, 60.278] - - - [4352, 1281, 1, 256] - - [27, 49.182] - - - [12848, 2865, 1, 256] - - [37, 59.68] - - - [16944, 3584, 1, 256] - - [36, 64.55] - - - [8752, 256, 1, 256] - - [34, 36.102] - - - [6912, 256, 1, 256] - - [124, 48.589] - - - [14336, 1281, 1, 256] - - [36, 59.694] - - - [2304, 2049, 1, 256] - - [35, 50.153] - - - [9216, 6144, 1, 256] - - [29, 73.135] - - - [1072, 1024, 1, 256] - - [110, 36.172] - - - [10752, 6144, 1, 256] - - [27, 73.602] - - - [1792, 1537, 1, 256] - - [35, 43.657] - - - [17968, 2865, 1, 256] - - [25, 61.071] - - - [8448, 4864, 1, 256] - - [34, 72.259] - - - [15408, 2048, 1, 256] - - [53, 62.18] - - - [15104, 1280, 1, 256] - - [34, 67.135] - - - [9264, 2865, 1, 256] - - [25, 59.195] - - - [15616, 2865, 1, 256] - - [35, 69.029] - - - [16896, 10240, 1, 256] - - [25, 75.376] - - - [15104, 2816, 1, 256] - - [60, 71.358] - - - [13872, 2865, 1, 256] - - [25, 60.894] - - - [13056, 1280, 1, 256] - - [34, 65.37] - - - [12288, 3328, 1, 256] - - [27, 71.562] - - - [3840, 3328, 1, 256] - - [34, 62.248] - - - [9216, 1280, 1, 256] - - [34, 62.921] - - - [8448, 1281, 1, 256] - - [34, 56.762] - - - [2560, 256, 1, 256] - - [123, 31.104] - - - [4608, 1281, 1, 256] - - [35, 51.593] - - - [6144, 2865, 1, 256] - - [25, 63.396] - - - [5888, 2816, 1, 256] - - [35, 65.221] - - - [3584, 1281, 1, 256] - - [34, 49.145] - - - [18688, 5120, 1, 256] - - [25, 73.865] - - - [12288, 2816, 1, 256] - - [25, 70.398] - - - [4864, 2865, 1, 256] - - [34, 65.382] - - - [9216, 2048, 1, 256] - - [58, 66.26] - - - [13872, 768, 1, 256] - - [36, 58.587] - - - [10496, 7424, 1, 256] - - [25, 73.594] - - - [16384, 512, 1, 256] - - [34, 54.525] - - - [14848, 10240, 1, 256] - - [27, 74.949] - - - [17920, 2048, 1, 256] - - [39, 69.882] - - - [11008, 7936, 1, 256] - - [71, 73.234] - - - [1792, 1792, 1, 256] - - [27, 49.797] - - - [7680, 4864, 1, 256] - - [34, 72.155] - - - [19760, 256, 1, 256] - - [58, 49.25] - - - [15616, 1792, 1, 256] - - [35, 69.504] - - - [1792, 1793, 1, 256] - - [25, 49.099] - - - [8192, 3329, 1, 256] - - [27, 66.799] - - - [2560, 1280, 1, 256] - - [34, 51.366] - - - [1328, 1073, 1, 256] - - [117, 38.13] - - - [16896, 2865, 1, 256] - - [37, 69.675] - - - [8960, 1280, 1, 256] - - [49, 61.258] - - - [6960, 2865, 1, 256] - - [27, 58.909] - - - [1280, 1024, 1, 256] - - [125, 42.945] - - - [6400, 2048, 1, 256] - - [55, 62.695] - - - [18480, 5376, 1, 256] - - [29, 62.704] - - - [18944, 2048, 1, 256] - - [40, 70.368] - - - [9520, 6656, 1, 256] - - [36, 63.438] - - - [4352, 1536, 1, 256] - - [35, 57.987] - - - [19712, 6144, 1, 256] - - [51, 72.546] - - - [6400, 2816, 1, 256] - - [29, 68.737] - - - [1792, 1585, 1, 256] - - [25, 44.837] - - - [13312, 6144, 1, 256] - - [25, 74.102] - - - [17408, 4096, 1, 256] - - [27, 72.71] - - - [16128, 256, 1, 256] - - [69, 44.519] - - - [15104, 2048, 1, 256] - - [60, 68.882] - - - [8704, 2865, 1, 256] - - [29, 66.521] - - - [6144, 768, 1, 256] - - [60, 50.604] - - - [10496, 1280, 1, 256] - - [35, 63.682] - - - [816, 561, 1, 256] - - [109, 21.462] - - - [6912, 3840, 1, 256] - - [55, 70.36] - - - [8704, 1281, 1, 256] - - [23, 58.499] - - - [13312, 1792, 1, 256] - - [35, 68.51] - - - [5120, 1281, 1, 256] - - [55, 55.721] - - - [10496, 1281, 1, 256] - - [60, 58.196] - - - [8448, 6144, 1, 256] - - [34, 72.701] - - - [2560, 2353, 1, 256] - - [34, 52.819] - - - [4352, 1280, 1, 256] - - [34, 49.744] - - - [12336, 256, 1, 256] - - [36, 47.326] - - - [21504, 10240, 1, 256] - - [29, 75.583] - - - [31744, 6144, 1, 256] - - [27, 74.524] - - - [27648, 1280, 1, 256] - - [34, 70.518] - - - [22272, 512, 1, 256] - - [23, 60.583] - - - [29184, 256, 1, 256] - - [34, 52.208] - - - [23808, 4096, 1, 256] - - [74, 72.707] - - - [30720, 7168, 1, 256] - - [37, 73.869] - - - [29440, 2865, 1, 256] - - [27, 70.006] - - - [25600, 5632, 1, 256] - - [27, 75.075] - - - [24832, 10240, 1, 256] - - [27, 74.967] - - - [22784, 2865, 1, 256] - - [29, 69.758] - - - [24368, 768, 1, 256] - - [58, 59.639] - - - [21760, 8192, 1, 256] - - [37, 74.528] - - - [29184, 10240, 1, 256] - - [24, 74.869] - - - [26368, 6144, 1, 256] - - [27, 74.073] - - - [23088, 10240, 1, 256] - - [53, 63.352] - - - [29952, 4096, 1, 256] - - [29, 72.803] - - - [24320, 6144, 1, 256] - - [30, 74.167] - - - [32256, 2048, 1, 256] - - [65, 71.243] - - - [29488, 2865, 1, 256] - - [37, 60.926] - - - [21808, 2816, 1, 256] - - [58, 63.226] - - - [32000, 7936, 1, 256] - - [27, 74.425] - - - [23040, 10240, 1, 256] - - [25, 75.226] - - - [31792, 10240, 1, 256] - - [28, 63.633] - - - [24320, 1281, 1, 256] - - [69, 64.001] - - - [27136, 3072, 1, 256] - - [51, 73.24] - - - [31488, 6144, 1, 256] - - [27, 74.077] - - - [34096, 2865, 1, 256] - - [44, 62.013] - - - [33024, 8960, 1, 256] - - [30, 74.681] - - - [28928, 1280, 1, 256] - - [34, 69.965] - - - [31488, 5632, 1, 256] - - [27, 74.077] - - - [27696, 4096, 1, 256] - - [25, 62.899] - - - [31488, 10240, 1, 256] - - [37, 74.538] - - - [28928, 10240, 1, 256] - - [29, 74.419] - - - [26160, 10240, 1, 256] - - [36, 62.983] - - - [26112, 3328, 1, 256] - - [89, 73.184] - - - [28928, 4864, 1, 256] - - [29, 73.672] - - - [27904, 3328, 1, 256] - - [81, 72.375] - - - [29184, 5376, 1, 256] - - [30, 74.353] - - - [29952, 1281, 1, 256] - - [33, 64.138] - - - [24832, 6144, 1, 256] - - [24, 74.183] - - - [28160, 4096, 1, 256] - - [42, 73.284] - - - [24320, 1280, 1, 256] - - [57, 69.334] - - - [34816, 768, 1, 256] - - [53, 68.672] - - - [34816, 1281, 1, 256] - - [27, 63.991] - - - [27136, 2816, 1, 256] - - [57, 73.176] - - - [32256, 8192, 1, 256] - - [30, 74.855] - - - [26624, 2865, 1, 256] - - [27, 70.664] - - - [23808, 3840, 1, 256] - - [27, 73.539] - - - [29440, 5376, 1, 256] - - [57, 74.171] - - - [30464, 10240, 1, 256] - - [24, 73.683] - - - [29232, 10240, 1, 256] - - [36, 62.969] - - - [27136, 1280, 1, 256] - - [57, 70.23] - - - [27904, 6144, 1, 256] - - [28, 73.746] - - - [33024, 2816, 1, 256] - - [42, 72.269] - - - [34816, 3329, 1, 256] - - [25, 70.685] - - - [34048, 1792, 1, 256] - - [31, 71.4] - - - [21248, 7424, 1, 256] - - [37, 74.819] - - - [29952, 256, 1, 256] - - [23, 53.356] - - - [34560, 256, 1, 256] - - [58, 59.115] - - - [26368, 3072, 1, 256] - - [37, 72.767] - - - [23600, 2865, 1, 256] - - [37, 61.367] - - - [30720, 512, 1, 256] - - [35, 64.809] - - - [30768, 10240, 1, 256] - - [41, 62.81] - - - [28928, 1024, 1, 256] - - [63, 68.461] - - - [26624, 256, 1, 256] - - [25, 57.116] - - - [26928, 10240, 1, 256] - - [53, 63.381] - - - [21248, 7936, 1, 256] - - [27, 74.601] - - - [34304, 2816, 1, 256] - - [59, 73.087] - - - [29696, 3840, 1, 256] - - [27, 74.502] - - - [27696, 10240, 1, 256] - - [27, 63.529] - - - [24064, 2048, 1, 256] - - [72, 70.68] - - - [33536, 6144, 1, 256] - - [30, 74.074] - - - [32512, 8704, 1, 256] - - [47, 74.767] - - - [21552, 2816, 1, 256] - - [53, 63.568] - - - [27648, 10240, 1, 256] - - [37, 75.468] - - - [22272, 2048, 1, 256] - - [33, 70.036] - - - [28976, 5632, 1, 256] - - [53, 63.367] - - - [30720, 10240, 1, 256] - - [25, 75.334] - - - [26112, 2816, 1, 256] - - [27, 73.015] - - - [20528, 10240, 1, 256] - - [28, 62.326] - - - [29696, 1536, 1, 256] - - [55, 70.945] - - - [31536, 2865, 1, 256] - - [53, 61.567] - - - [32000, 3328, 1, 256] - - [49, 72.394] - - - [20784, 2865, 1, 256] - - [25, 60.364] - - - [33280, 9984, 1, 256] - - [51, 75.53] - - - [25600, 3329, 1, 256] - - [32, 70.116] - - - [27904, 4096, 1, 256] - - [65, 72.935] - - - [29488, 256, 1, 256] - - [58, 48.225] - - - [32048, 10240, 1, 256] - - [27, 63.133] - - - [31280, 2865, 1, 256] - - [53, 61.474] - - - [32816, 2816, 1, 256] - - [28, 65.244] - - - [34096, 2816, 1, 256] - - [44, 63.755] - - - [20992, 3328, 1, 256] - - [25, 72.666] - - - [32768, 1281, 1, 256] - - [95, 51.657] - - - [24576, 4864, 1, 256] - - [38, 70.216] - - - [30464, 3328, 1, 256] - - [59, 71.563] - - - [28208, 256, 1, 256] - - [44, 46.063] - - - [23552, 1280, 1, 256] - - [49, 70.066] - - - [20528, 7168, 1, 256] - - [25, 61.991] - - - [34560, 2865, 1, 256] - - [37, 70.251] - - - [20736, 2816, 1, 256] - - [34, 72.019] - - - [26880, 3328, 1, 256] - - [25, 72.345] - - - [31536, 8192, 1, 256] - - [58, 62.985] - - - [31744, 8448, 1, 256] - - [25, 75.166] - - - [20224, 2865, 1, 256] - - [27, 69.451] - - - [22528, 2048, 1, 256] - - [61, 69.864] - - - [24320, 2048, 1, 256] - - [69, 70.443] - - - [32512, 8960, 1, 256] - - [47, 74.928] - - - [33072, 10240, 1, 256] - - [29, 64.214] - - - [24880, 10240, 1, 256] - - [29, 63.243] - - - [21040, 7680, 1, 256] - - [36, 64.019] - - - [26368, 10240, 1, 256] - - [25, 74.681] - - - [32304, 8704, 1, 256] - - [58, 63.345] - - - [33536, 1281, 1, 256] - - [74, 63.723] - - - [27136, 1024, 1, 256] - - [23, 68.778] - - - [33792, 1281, 1, 256] - - [41, 64.358] - - - [33584, 256, 1, 256] - - [27, 54.154] - - - [20528, 7424, 1, 256] - - [27, 62.655] - - - [28928, 2865, 1, 256] - - [29, 69.869] - - - [22016, 2048, 1, 256] - - [48, 70.212] - - - [29440, 3328, 1, 256] - - [61, 72.706] - - - [30208, 2048, 1, 256] - - [44, 71.22] - - - [20480, 2816, 1, 256] - - [27, 72.352] - - - [25904, 256, 1, 256] - - [34, 51.108] - - - [20736, 10240, 1, 256] - - [27, 74.9] - - - [32816, 256, 1, 256] - - [58, 53.638] - - - [33792, 3328, 1, 256] - - [25, 73.479] - - - [22272, 1281, 1, 256] - - [69, 63.01] - - - [25600, 1280, 1, 256] - - [29, 69.893] - - - [33280, 3329, 1, 256] - - [24, 70.393] - - - [22784, 1281, 1, 256] - - [39, 63.786] - - - [25392, 2816, 1, 256] - - [53, 63.703] - - - [33280, 3328, 1, 256] - - [29, 73.324] - - - [21760, 1280, 1, 256] - - [35, 68.663] - - - [33024, 768, 1, 256] - - [23, 65.517] - - - [25088, 1792, 1, 256] - - [55, 71.376] - - - [26368, 3329, 1, 256] - - [54, 69.349] - - - [34560, 3328, 1, 256] - - [61, 72.635] - - - [23040, 6144, 1, 256] - - [47, 74.371] - - - [30464, 2048, 1, 256] - - [89, 68.048] - - - [28672, 3328, 1, 256] - - [29, 72.77] - - - [30464, 6912, 1, 256] - - [30, 73.595] - - - [32048, 2816, 1, 256] - - [36, 63.364] - - - [33792, 9728, 1, 256] - - [25, 75.271] - - - [27392, 1536, 1, 256] - - [31, 66.59] - - - [24112, 512, 1, 256] - - [29, 57.899] - - - [28160, 256, 1, 256] - - [34, 59.165] - - - [34816, 2048, 1, 256] - - [28, 70.698] - - - [25648, 10240, 1, 256] - - [41, 63.438] - - - [20992, 10240, 1, 256] - - [25, 75.383] - - - [22528, 1281, 1, 256] - - [25, 63.117] - - - [25904, 2304, 1, 256] - - [53, 62.878] - - - [27952, 2865, 1, 256] - - [58, 61.685] - - - [30976, 768, 1, 256] - - [39, 66.489] - - - [20480, 3329, 1, 256] - - [32, 69.603] - - - [33072, 256, 1, 256] - - [25, 52.096] - - - [26624, 2560, 1, 256] - - [25, 73.237] - - - [28208, 2865, 1, 256] - - [58, 61.173] - - - [26672, 3328, 1, 256] - - [25, 61.969] - - - [26880, 2865, 1, 256] - - [54, 69.839] - - - [26112, 2304, 1, 256] - - [37, 72.648] - - - [29184, 5120, 1, 256] - - [37, 74.128] - - - [29744, 6144, 1, 256] - - [41, 62.89] - - - [30464, 3329, 1, 256] - - [70, 68.208] - - - [22272, 2560, 1, 256] - - [31, 72.039] - - - [25344, 2048, 1, 256] - - [39, 69.446] - - - [31792, 256, 1, 256] - - [44, 51.858] - - - [21248, 2816, 1, 256] - - [35, 71.886] - - - [32816, 10240, 1, 256] - - [74, 65.69] - - - [27136, 3840, 1, 256] - - [59, 74.215] - - - [34096, 10240, 1, 256] - - [41, 62.84] - - - [24576, 4608, 1, 256] - - [22, 69.32] - - - [32256, 1281, 1, 256] - - [39, 64.32] - - - [26928, 2865, 1, 256] - - [53, 60.727] - - - [20784, 7424, 1, 256] - - [36, 63.384] - - - [24112, 2816, 1, 256] - - [58, 64.168] - - - [22272, 256, 1, 256] - - [36, 49.692] - - - [30208, 1281, 1, 256] - - [69, 64.377] - - - [28720, 2816, 1, 256] - - [25, 61.361] - - - [20992, 1280, 1, 256] - - [34, 69.476] - - - [31488, 1536, 1, 256] - - [29, 70.299] - - - [21296, 8192, 1, 256] - - [53, 63.224] - - - [30512, 7168, 1, 256] - - [58, 62.559] - - - [27136, 2865, 1, 256] - - [25, 70.677] - - - [25088, 3329, 1, 256] - - [51, 69.78] - - - [29696, 3329, 1, 256] - - [37, 70.496] - - - [23040, 1280, 1, 256] - - [60, 69.267] - - - [30000, 256, 1, 256] - - [53, 48.765] - - - [20224, 3329, 1, 256] - - [37, 69.242] - - - [29232, 2816, 1, 256] - - [44, 63.791] - - - [31232, 7424, 1, 256] - - [47, 75.246] - - - [29488, 2816, 1, 256] - - [58, 63.082] - - - [25904, 2865, 1, 256] - - [36, 61.333] - - - [30512, 2816, 1, 256] - - [53, 63.905] - - - [20736, 768, 1, 256] - - [23, 65.491] - - - [20480, 256, 1, 256] - - [23, 53.911] - - - [28672, 6144, 1, 256] - - [27, 74.135] - - - [26624, 2816, 1, 256] - - [37, 73.242] - - - [28928, 768, 1, 256] - - [40, 66.936] - - - [27648, 256, 1, 256] - - [78, 58.412] - - - [32256, 6144, 1, 256] - - [47, 74.469] - - - [30720, 6144, 1, 256] - - [37, 74.708] - - - [32560, 2865, 1, 256] - - [53, 63.06] - - - [23088, 9728, 1, 256] - - [44, 63.913] - - - [22784, 9728, 1, 256] - - [28, 74.598] - - - [33024, 6144, 1, 256] - - [43, 73.922] - - - [27392, 2865, 1, 256] - - [41, 67.581] - - - [21504, 1280, 1, 256] - - [55, 68.976] - - - [30720, 6656, 1, 256] - - [27, 75.049] - - - [24880, 2865, 1, 256] - - [36, 62.105] - - - [25392, 1792, 1, 256] - - [44, 61.999] - - - [20224, 2816, 1, 256] - - [55, 71.778] - - - [20224, 256, 1, 256] - - [40, 53.677] - - - [25856, 3329, 1, 256] - - [25, 69.116] - - - [30976, 256, 1, 256] - - [40, 54.826] - - - [26880, 6144, 1, 256] - - [37, 74.125] - - - [26672, 2816, 1, 256] - - [37, 62.709] - - - [25600, 256, 1, 256] - - [80, 55.352] - - - [28160, 1281, 1, 256] - - [74, 64.386] - - - [20480, 10240, 1, 256] - - [27, 75.272] - - - [21504, 7936, 1, 256] - - [25, 75.534] - - - [20272, 7168, 1, 256] - - [44, 62.663] - - - [24880, 2816, 1, 256] - - [44, 64.271] - - - [23296, 9728, 1, 256] - - [29, 74.626] - - - [34816, 2865, 1, 256] - - [27, 71.003] - - - [31792, 2865, 1, 256] - - [29, 62.045] - - - [29488, 6144, 1, 256] - - [44, 62.997] - - - [23856, 2865, 1, 256] - - [36, 60.587] - - - [25088, 256, 1, 256] - - [53, 54.654] - - - [22016, 8960, 1, 256] - - [31, 75.62] - - - [23040, 3072, 1, 256] - - [27, 72.891] - - - [23856, 512, 1, 256] - - [25, 57.075] - - - [33792, 3329, 1, 256] - - [27, 70.407] - - - [22784, 9216, 1, 256] - - [28, 74.474] - - - [30720, 4864, 1, 256] - - [37, 74.739] - - - [32000, 8192, 1, 256] - - [27, 74.349] - - - [28160, 3329, 1, 256] - - [32, 69.566] - - - [28672, 256, 1, 256] - - [44, 51.858] - - - [27648, 1281, 1, 256] - - [28, 63.629] - - - [23808, 6144, 1, 256] - - [37, 74.136] - - - [23344, 10240, 1, 256] - - [36, 62.842] - - - [20736, 7680, 1, 256] - - [25, 74.828] - - - [33024, 9216, 1, 256] - - [42, 74.582] - - - [26160, 2816, 1, 256] - - [58, 63.826] - - - [24064, 10240, 1, 256] - - [47, 75.204] - - - [24320, 768, 1, 256] - - [72, 65.439] - - - [28208, 10240, 1, 256] - - [74, 62.743] - - - [34560, 1024, 1, 256] - - [40, 69.938] - - - [33792, 1792, 1, 256] - - [37, 72.448] - - - [30720, 2816, 1, 256] - - [29, 73.516] - - - [24624, 2816, 1, 256] - - [41, 62.429] - - - [20736, 3329, 1, 256] - - [29, 68.948] - - - [21760, 1792, 1, 256] - - [27, 70.593] - - - [21760, 8704, 1, 256] - - [27, 75.019] - - - [34608, 10240, 1, 256] - - [44, 63.342] - - - [22784, 9472, 1, 256] - - [67, 74.935] - - - [31536, 2816, 1, 256] - - [58, 63.566] - - - [27904, 4352, 1, 256] - - [59, 73.693] - - - [23552, 2865, 1, 256] - - [29, 70.547] - - - [24064, 256, 1, 256] - - [40, 52.765] - - - [34304, 2048, 1, 256] - - [41, 71.565] - - - [30464, 1280, 1, 256] - - [67, 68.821] - - - [29440, 5632, 1, 256] - - [31, 74.412] - - - [21808, 8704, 1, 256] - - [36, 63.111] - - - [30464, 6656, 1, 256] - - [24, 73.443] - - - [20736, 1024, 1, 256] - - [60, 68.087] - - - [24832, 1024, 1, 256] - - [69, 67.376] - - - [24576, 1024, 1, 256] - - [27, 64.881] - - - [29184, 2048, 1, 256] - - [74, 71.095] - - - [30976, 4864, 1, 256] - - [67, 72.422] - - - [25344, 1536, 1, 256] - - [57, 69.677] - - - [22016, 1280, 1, 256] - - [34, 69.22] - - - [32560, 8960, 1, 256] - - [53, 64.641] - - - [31536, 7936, 1, 256] - - [53, 64.014] - - - [26880, 3072, 1, 256] - - [37, 72.76] - - - [28464, 2865, 1, 256] - - [53, 61.458] - - - [20224, 6400, 1, 256] - - [27, 74.537] - - - [26624, 3328, 1, 256] - - [55, 73.411] - - - [24320, 512, 1, 256] - - [57, 63.719] - - - [34352, 768, 1, 256] - - [36, 61.293] - - - [30720, 768, 1, 256] - - [40, 67.188] - - - [34560, 10240, 1, 256] - - [37, 74.662] - - - [22016, 3328, 1, 256] - - [34, 72.823] - - - [20480, 1281, 1, 256] - - [27, 62.783] - - - [31232, 2816, 1, 256] - - [37, 72.918] - - - [31232, 6144, 1, 256] - - [30, 74.376] - - - [27136, 256, 1, 256] - - [27, 57.836] - - - [23344, 256, 1, 256] - - [27, 47.781] - - - [30208, 4352, 1, 256] - - [51, 74.258] - - - [32000, 6144, 1, 256] - - [29, 74.002] - - - [29184, 6144, 1, 256] - - [27, 74.248] - - - [29232, 5632, 1, 256] - - [58, 63.1] - - - [22576, 2816, 1, 256] - - [37, 62.212] - - - [31488, 1280, 1, 256] - - [34, 70.391] - - - [23856, 2816, 1, 256] - - [53, 63.075] - - - [29184, 2865, 1, 256] - - [57, 70.156] - - - [21248, 6144, 1, 256] - - [27, 74.121] - - - [30720, 4608, 1, 256] - - [37, 74.138] - - - [27952, 256, 1, 256] - - [29, 53.969] - - - [32512, 10240, 1, 256] - - [30, 74.736] - - - [31744, 3328, 1, 256] - - [25, 73.375] - - - [22528, 3328, 1, 256] - - [29, 73.174] - - - [34048, 3329, 1, 256] - - [73, 69.195] - - - [31744, 2816, 1, 256] - - [34, 73.414] - - - [27904, 256, 1, 256] - - [35, 58.921] - - - [21552, 256, 1, 256] - - [36, 45.306] - - - [29952, 6144, 1, 256] - - [24, 74.05] - - - [22784, 3328, 1, 256] - - [42, 72.078] - - - [20784, 256, 1, 256] - - [58, 51.079] - - - [30208, 2816, 1, 256] - - [91, 72.979] - - - [31232, 5376, 1, 256] - - [48, 74.579] - - - [30256, 256, 1, 256] - - [25, 48.965] - - - [21248, 1280, 1, 256] - - [55, 68.285] - - - [28160, 1280, 1, 256] - - [31, 70.64] - - - [30720, 3329, 1, 256] - - [37, 70.407] - - - [34560, 3329, 1, 256] - - [32, 69.597] - - - [31024, 2816, 1, 256] - - [58, 63.748] - - - [32000, 256, 1, 256] - - [58, 56.124] - - - [20528, 256, 1, 256] - - [34, 50.618] - - - [24624, 10240, 1, 256] - - [41, 63.948] - - - [21504, 7680, 1, 256] - - [25, 75.495] - - - [33536, 9728, 1, 256] - - [24, 74.661] - - - [33280, 6144, 1, 256] - - [51, 74.485] - - - [20480, 2865, 1, 256] - - [29, 70.005] - - - [30720, 1281, 1, 256] - - [27, 63.715] - - - [21760, 6144, 1, 256] - - [37, 74.339] - - - [30976, 6912, 1, 256] - - [41, 73.297] - - - [27648, 2816, 1, 256] - - [25, 73.351] - - - [20992, 3329, 1, 256] - - [29, 69.637] - - - [26672, 3072, 1, 256] - - [25, 62.488] - - - [24832, 2816, 1, 256] - - [57, 72.466] - - - [23552, 9728, 1, 256] - - [25, 75.537] - - - [26880, 1280, 1, 256] - - [49, 69.337] - - - [25088, 1280, 1, 256] - - [25, 70.249] - - - [33280, 9472, 1, 256] - - [24, 75.514] - - - [27136, 3328, 1, 256] - - [31, 73.302] - - - [28416, 2816, 1, 256] - - [59, 72.194] - - - [20480, 3328, 1, 256] - - [35, 72.255] - - - [31232, 256, 1, 256] - - [55, 55.016] - - - [33328, 9728, 1, 256] - - [58, 63.695] - - - [26416, 256, 1, 256] - - [27, 51.781] - - - [31744, 2865, 1, 256] - - [25, 70.877] - - - [22784, 6144, 1, 256] - - [24, 73.854] - - - [32000, 5888, 1, 256] - - [27, 73.811] - - - [28160, 4864, 1, 256] - - [25, 74.305] - - - [34352, 2865, 1, 256] - - [44, 61.36] - - - [29696, 256, 1, 256] - - [27, 52.782] - - - [26112, 2048, 1, 256] - - [56, 70.991] - - - [25088, 5376, 1, 256] - - [67, 74.491] - - - [29952, 3329, 1, 256] - - [37, 69.674] - - - [21296, 10240, 1, 256] - - [58, 63.537] - - - [31744, 1280, 1, 256] - - [29, 71.007] - - - [21760, 256, 1, 256] - - [37, 48.404] - - - [31488, 2048, 1, 256] - - [61, 70.757] - - - [30976, 1281, 1, 256] - - [89, 62.501] - - - [23040, 256, 1, 256] - - [84, 50.859] - - - [34304, 6144, 1, 256] - - [43, 74.528] - - - [31744, 3329, 1, 256] - - [54, 70.267] - - - [31744, 5888, 1, 256] - - [37, 74.707] - - - [29184, 1281, 1, 256] - - [74, 64.472] - - - [23856, 10240, 1, 256] - - [53, 63.188] - - - [23808, 1792, 1, 256] - - [49, 71.018] - - - [32000, 1792, 1, 256] - - [27, 71.451] - - - [26880, 2816, 1, 256] - - [35, 72.335] - - - [28416, 3328, 1, 256] - - [59, 72.264] - - - [27136, 6144, 1, 256] - - [27, 74.561] - - - [28416, 4608, 1, 256] - - [51, 72.828] - - - [33536, 1280, 1, 256] - - [34, 70.538] - - - [27440, 2865, 1, 256] - - [53, 61.881] - - - [25088, 2865, 1, 256] - - [24, 70.235] - - - [30976, 2816, 1, 256] - - [67, 70.984] - - - [26672, 10240, 1, 256] - - [41, 62.731] - - - [34048, 10240, 1, 256] - - [47, 74.096] - - - [34352, 2816, 1, 256] - - [58, 63.786] - - - [22064, 2865, 1, 256] - - [36, 61.128] - - - [28208, 4864, 1, 256] - - [44, 63.705] - - - [22528, 1280, 1, 256] - - [27, 70.17] - - - [26624, 3072, 1, 256] - - [37, 73.536] - - - [33072, 2865, 1, 256] - - [29, 62.888] - - - [22576, 256, 1, 256] - - [29, 46.957] - - - [34560, 2048, 1, 256] - - [78, 71.126] - - - [29440, 5888, 1, 256] - - [31, 73.964] - - - [34560, 1280, 1, 256] - - [25, 70.838] - - - [32000, 10240, 1, 256] - - [29, 74.514] - - - [32304, 2816, 1, 256] - - [58, 64.167] - - - [30976, 2865, 1, 256] - - [57, 68.745] - - - [30208, 6400, 1, 256] - - [47, 74.958] - - - [29232, 2865, 1, 256] - - [53, 61.737] - - - [33072, 2816, 1, 256] - - [53, 64.978] - - - [30512, 2865, 1, 256] - - [53, 61.791] - - - [20016, 2816, 1, 256] - - [58, 63.492] - - - [28416, 4352, 1, 256] - - [67, 73.602] - - - [25648, 2816, 1, 256] - - [58, 63.004] - - - [25344, 1280, 1, 256] - - [57, 70.141] - - - [24576, 10240, 1, 256] - - [38, 70.414] - - - [33024, 1281, 1, 256] - - [91, 63.336] - - - [33584, 10240, 1, 256] - - [25, 63.044] - - - [28416, 4864, 1, 256] - - [31, 73.651] - - - [23296, 3329, 1, 256] - - [32, 69.072] - - - [30464, 4352, 1, 256] - - [30, 72.6] - - - [29696, 5632, 1, 256] - - [37, 75.118] - - - [25136, 256, 1, 256] - - [27, 49.859] - - - [20528, 2865, 1, 256] - - [27, 60.85] - - - [27440, 2816, 1, 256] - - [44, 64.396] - - - [28160, 2048, 1, 256] - - [74, 70.872] - - - [24320, 2816, 1, 256] - - [59, 72.422] - - - [20736, 6144, 1, 256] - - [37, 73.931] - - - [28416, 5120, 1, 256] - - [31, 73.518] - - - [21552, 8448, 1, 256] - - [36, 64.011] - - - [20736, 1281, 1, 256] - - [40, 62.761] - - - [28464, 4864, 1, 256] - - [58, 63.713] - - - [30512, 10240, 1, 256] - - [41, 62.92] - - - [34304, 512, 1, 256] - - [57, 66.06] - - - [22784, 10240, 1, 256] - - [59, 74.611] - - - [25648, 2048, 1, 256] - - [44, 62.668] - - - [25856, 10240, 1, 256] - - [37, 74.853] - - - [32256, 8960, 1, 256] - - [24, 75.392] - - - [20736, 2865, 1, 256] - - [37, 69.727] - - - [20992, 7680, 1, 256] - - [37, 75.084] - - - [31024, 10240, 1, 256] - - [61, 62.465] - - - [26112, 256, 1, 256] - - [27, 56.263] - - - [30000, 2865, 1, 256] - - [53, 61.012] - - - [25904, 2560, 1, 256] - - [58, 63.309] - - - [24832, 768, 1, 256] - - [39, 65.797] - - - [25088, 6144, 1, 256] - - [51, 74.353] - - - [24624, 1280, 1, 256] - - [22, 61.013] - - - [22016, 8192, 1, 256] - - [37, 74.983] - - - [29952, 3328, 1, 256] - - [57, 72.644] - - - [31232, 2048, 1, 256] - - [43, 71.486] - - - [30256, 6656, 1, 256] - - [36, 63.547] - - - [20992, 2816, 1, 256] - - [37, 72.698] - - - [33792, 1536, 1, 256] - - [29, 71.307] - - - [20224, 1280, 1, 256] - - [34, 68.324] - - - [25600, 5888, 1, 256] - - [27, 74.704] - - - [26624, 768, 1, 256] - - [23, 66.916] - - - [32256, 2816, 1, 256] - - [67, 73.279] - - - [21760, 1281, 1, 256] - - [69, 63.106] - - - [25392, 10240, 1, 256] - - [58, 63.224] - - - [32768, 256, 1, 256] - - [34, 57.515] - - - [22528, 3329, 1, 256] - - [27, 70.282] - - - [23552, 3329, 1, 256] - - [27, 70.102] - - - [33024, 2865, 1, 256] - - [54, 69.197] - - - [29696, 2816, 1, 256] - - [25, 73.527] - - - [27392, 10240, 1, 256] - - [41, 74.262] - - - [23040, 2048, 1, 256] - - [69, 70.576] - - - [27648, 6144, 1, 256] - - [27, 74.648] - - - [22016, 2304, 1, 256] - - [34, 71.872] - - - [34560, 1281, 1, 256] - - [78, 64.272] - - - [27136, 1281, 1, 256] - - [33, 63.812] - - - [32000, 1281, 1, 256] - - [37, 63.609] - - - [27184, 3840, 1, 256] - - [58, 63.75] - - - [24880, 1536, 1, 256] - - [29, 61.665] - - - [28672, 768, 1, 256] - - [23, 65.658] - - - [34816, 2816, 1, 256] - - [22, 73.569] - - - [26160, 256, 1, 256] - - [44, 51.484] - - - [30464, 7168, 1, 256] - - [24, 72.343] - - - [30208, 3328, 1, 256] - - [43, 73.144] - - - [32304, 10240, 1, 256] - - [37, 63.119] - - - [26624, 1280, 1, 256] - - [35, 70.172] - - - [29696, 10240, 1, 256] - - [27, 75.448] - - - [32000, 8704, 1, 256] - - [37, 74.667] - - - [27392, 1281, 1, 256] - - [76, 62.129] - - - [26416, 2865, 1, 256] - - [36, 60.973] - - - [26160, 2560, 1, 256] - - [36, 63.321] - - - [28672, 3329, 1, 256] - - [27, 69.719] - - - [23808, 256, 1, 256] - - [59, 52.087] - - - [27184, 10240, 1, 256] - - [53, 63.078] - - - [33280, 2048, 1, 256] - - [43, 71.202] - - - [33280, 2816, 1, 256] - - [51, 73.287] - - - [23040, 9984, 1, 256] - - [27, 75.6] - - - [26112, 1280, 1, 256] - - [35, 70.457] - - - [33328, 9984, 1, 256] - - [36, 64.001] - - - [32560, 9216, 1, 256] - - [37, 63.849] - - - [22832, 9728, 1, 256] - - [58, 63.859] - - - [27904, 1280, 1, 256] - - [31, 70.022] - - - [33280, 1281, 1, 256] - - [47, 64.269] - - - [33280, 1280, 1, 256] - - [27, 70.983] - - - [32048, 256, 1, 256] - - [27, 52.152] - - - [27184, 2865, 1, 256] - - [36, 61.678] - - - [26880, 3329, 1, 256] - - [32, 69.209] - - - [20784, 7680, 1, 256] - - [53, 63.953] - - - [24832, 3329, 1, 256] - - [30, 69.453] - - - [25856, 1280, 1, 256] - - [35, 69.588] - - - [34560, 2816, 1, 256] - - [27, 72.636] - - - [20016, 256, 1, 256] - - [40, 49.813] - - - [23600, 256, 1, 256] - - [29, 48.338] - - - [22576, 9216, 1, 256] - - [41, 62.894] - - - [25344, 5632, 1, 256] - - [65, 73.279] - - - [28928, 5632, 1, 256] - - [27, 74.019] - - - [31024, 256, 1, 256] - - [35, 50.513] - - - [21552, 2865, 1, 256] - - [27, 61.492] - - - [29184, 3072, 1, 256] - - [25, 72.996] - - - [24320, 2865, 1, 256] - - [57, 70.104] - - - [20480, 6656, 1, 256] - - [27, 74.673] - - - [33536, 10240, 1, 256] - - [51, 74.79] - - - [20736, 1280, 1, 256] - - [55, 69.078] - - - [24832, 1280, 1, 256] - - [25, 69.452] - - - [29488, 10240, 1, 256] - - [37, 63.086] - - - [27392, 6144, 1, 256] - - [28, 73.273] - - - [29440, 3329, 1, 256] - - [47, 69.625] - - - [25856, 1281, 1, 256] - - [36, 62.913] - - - [34560, 768, 1, 256] - - [60, 68.277] - - - [31488, 7680, 1, 256] - - [25, 74.434] - - - [29184, 5632, 1, 256] - - [27, 74.567] - - - [32512, 512, 1, 256] - - [57, 62.875] - - - [26112, 2865, 1, 256] - - [37, 70.579] - - - [32512, 1280, 1, 256] - - [55, 70.012] - - - [20992, 1024, 1, 256] - - [40, 68.585] - - - [27904, 10240, 1, 256] - - [29, 74.585] - - - [29952, 6656, 1, 256] - - [51, 74.395] - - - [21248, 2048, 1, 256] - - [40, 70.141] - - - [34352, 256, 1, 256] - - [36, 53.882] - - - [24064, 512, 1, 256] - - [55, 63.612] - - - [32816, 2865, 1, 256] - - [28, 62.415] - - - [33840, 256, 1, 256] - - [35, 53.89] - - - [33792, 1280, 1, 256] - - [35, 71.087] - - - [21296, 7936, 1, 256] - - [44, 63.676] - - - [34096, 256, 1, 256] - - [36, 53.881] - - - [32256, 8704, 1, 256] - - [47, 75.201] - - - [30464, 1281, 1, 256] - - [59, 62.873] - - - [28464, 2816, 1, 256] - - [58, 63.429] - - - [25136, 2865, 1, 256] - - [36, 61.245] - - - [31792, 8448, 1, 256] - - [63, 64.158] - - - [24320, 4608, 1, 256] - - [57, 73.571] - - - [25088, 5120, 1, 256] - - [31, 74.212] - - - [31744, 2048, 1, 256] - - [61, 71.217] - - - [30720, 1280, 1, 256] - - [37, 70.942] - - - [34048, 256, 1, 256] - - [34, 58.331] - - - [28416, 512, 1, 256] - - [34, 61.95] - - - [22272, 10240, 1, 256] - - [47, 74.77] - - - [32512, 3328, 1, 256] - - [65, 72.772] - - - [29744, 10240, 1, 256] - - [41, 63.34] - - - [22784, 2048, 1, 256] - - [39, 70.375] - - - [23552, 2048, 1, 256] - - [41, 70.837] - - - [25344, 2816, 1, 256] - - [57, 71.619] - - - [27440, 3840, 1, 256] - - [58, 64.087] - - - [21552, 10240, 1, 256] - - [27, 63.385] - - - [21808, 256, 1, 256] - - [58, 45.844] - - - [24576, 6144, 1, 256] - - [32, 69.698] - - - [29744, 256, 1, 256] - - [44, 48.643] - - - [31488, 3328, 1, 256] - - [41, 72.631] - - - [33536, 3329, 1, 256] - - [25, 69.443] - - - [21040, 256, 1, 256] - - [23, 51.366] - - - [22272, 9216, 1, 256] - - [41, 74.35] - - - [27648, 4096, 1, 256] - - [41, 73.226] - - - [29440, 1280, 1, 256] - - [57, 70.202] - - - [31744, 7936, 1, 256] - - [27, 75.389] - - - [26624, 1281, 1, 256] - - [27, 63.119] - - - [28672, 2048, 1, 256] - - [37, 68.271] - - - [24064, 3328, 1, 256] - - [57, 72.942] - - - [25344, 3329, 1, 256] - - [47, 68.833] - - - [33280, 9728, 1, 256] - - [42, 75.157] - - - [22320, 8960, 1, 256] - - [53, 64.407] - - - [30464, 6144, 1, 256] - - [51, 72.887] - - - [34304, 2304, 1, 256] - - [31, 72.816] - - - [28928, 256, 1, 256] - - [25, 51.625] - - - [27392, 1280, 1, 256] - - [23, 67.051] - - - [26672, 2865, 1, 256] - - [29, 61.236] - - - [28720, 10240, 1, 256] - - [41, 62.448] - - - [25088, 2816, 1, 256] - - [34, 72.671] - - - [31280, 256, 1, 256] - - [53, 50.995] - - - [29488, 5888, 1, 256] - - [58, 63.407] - - - [30720, 2048, 1, 256] - - [74, 70.334] - - - [21808, 10240, 1, 256] - - [53, 63.114] - - - [24576, 2865, 1, 256] - - [54, 65.98] - - - [23808, 1280, 1, 256] - - [23, 69.635] - - - [33280, 1024, 1, 256] - - [72, 69.675] - - - [25856, 256, 1, 256] - - [90, 55.701] - - - [25648, 2304, 1, 256] - - [53, 63.621] - - - [29952, 2865, 1, 256] - - [25, 70.007] - - - [23040, 1024, 1, 256] - - [60, 66.825] - - - [34304, 3328, 1, 256] - - [42, 73.509] - - - [31792, 8192, 1, 256] - - [25, 63.352] - - - [24576, 2816, 1, 256] - - [38, 68.802] - - - [27648, 1536, 1, 256] - - [29, 70.642] - - - [23296, 9472, 1, 256] - - [27, 75.084] - - - [24624, 256, 1, 256] - - [53, 49.754] - - - [20736, 2048, 1, 256] - - [63, 70.538] - - - [28720, 5376, 1, 256] - - [41, 62.018] - - - [20480, 512, 1, 256] - - [60, 62.552] - - - [33840, 2865, 1, 256] - - [25, 62.278] - - - [24064, 2865, 1, 256] - - [47, 70.099] - - - [24064, 2816, 1, 256] - - [31, 72.665] - - - [20992, 256, 1, 256] - - [34, 55.358] - - - [33328, 256, 1, 256] - - [36, 53.917] - - - [28928, 5120, 1, 256] - - [25, 73.722] - - - [34304, 256, 1, 256] - - [34, 58.916] - - - [34304, 1281, 1, 256] - - [61, 64.906] - - - [31744, 1281, 1, 256] - - [28, 64.217] - - - [33584, 2816, 1, 256] - - [53, 63.774] - - - [24064, 4352, 1, 256] - - [59, 74.349] - - - [20224, 6912, 1, 256] - - [25, 74.537] - - - [21504, 1281, 1, 256] - - [76, 62.721] - - - [33536, 3328, 1, 256] - - [41, 72.67] - - - [34816, 3328, 1, 256] - - [37, 73.475] - - - [31024, 7680, 1, 256] - - [58, 63.222] - - - [22016, 3329, 1, 256] - - [29, 69.596] - - - [25344, 1281, 1, 256] - - [39, 63.325] - - - [31744, 7680, 1, 256] - - [27, 75.385] - - - [27952, 10240, 1, 256] - - [44, 63.351] - - - [23808, 2048, 1, 256] - - [60, 70.427] - - - [32768, 2816, 1, 256] - - [38, 59.065] - - - [34816, 256, 1, 256] - - [25, 59.726] - - - [27904, 2865, 1, 256] - - [31, 69.642] - - - [31232, 1280, 1, 256] - - [55, 70.796] - - - [22016, 1281, 1, 256] - - [72, 63.18] - - - [22528, 8704, 1, 256] - - [25, 75.669] - - - [22528, 9216, 1, 256] - - [27, 74.957] - - - [34816, 1280, 1, 256] - - [25, 71.284] - - - [23808, 10240, 1, 256] - - [27, 75.037] - - - [32512, 2048, 1, 256] - - [42, 70.782] - - - [34816, 1024, 1, 256] - - [23, 70.048] - - - [34048, 2048, 1, 256] - - [91, 70.237] - - - [30768, 2816, 1, 256] - - [25, 61.946] - - - [22272, 3329, 1, 256] - - [51, 68.958] - - - [25600, 3328, 1, 256] - - [34, 73.271] - - - [34048, 2816, 1, 256] - - [57, 72.312] - - - [22064, 8704, 1, 256] - - [36, 63.169] - - - [25648, 256, 1, 256] - - [58, 50.986] - - - [22784, 768, 1, 256] - - [40, 66.495] - - - [27904, 2048, 1, 256] - - [23, 70.333] - - - [22528, 9472, 1, 256] - - [27, 75.892] - - - [21504, 2865, 1, 256] - - [27, 70.242] - - - [28672, 5376, 1, 256] - - [27, 74.327] - - - [22576, 9472, 1, 256] - - [25, 62.863] - - - [24576, 256, 1, 256] - - [25, 53.728] - - - [28672, 5120, 1, 256] - - [27, 74.274] - - - [24576, 3328, 1, 256] - - [22, 68.462] - - - [32816, 9472, 1, 256] - - [28, 66.519] - - - [27440, 256, 1, 256] - - [29, 53.249] - - - [22272, 8704, 1, 256] - - [51, 74.842] - - - [30000, 2816, 1, 256] - - [53, 63.72] - - - [26928, 2816, 1, 256] - - [58, 63.237] - - - [22064, 2816, 1, 256] - - [58, 63.866] - - - [23552, 3328, 1, 256] - - [27, 73.147] - - - [28416, 256, 1, 256] - - [60, 51.395] - - - [28928, 6144, 1, 256] - - [25, 73.774] - - - [32768, 512, 1, 256] - - [29, 57.901] - - - [22272, 2865, 1, 256] - - [37, 69.352] - - - [26928, 256, 1, 256] - - [25, 52.393] - - - [21760, 10240, 1, 256] - - [27, 74.893] - - - [26368, 512, 1, 256] - - [55, 62.805] - - - [26672, 256, 1, 256] - - [53, 52.318] - - - [33328, 2865, 1, 256] - - [44, 62.95] - - - [30720, 3328, 1, 256] - - [34, 73.371] - - - [25856, 2865, 1, 256] - - [29, 69.722] - - - [25088, 3328, 1, 256] - - [67, 72.895] - - - [28416, 2560, 1, 256] - - [57, 72.27] - - - [33536, 9472, 1, 256] - - [24, 74.945] - - - [20480, 1280, 1, 256] - - [55, 68.503] - - - [30208, 6144, 1, 256] - - [30, 74.374] - - - [34864, 1024, 1, 256] - - [25, 61.22] - - - [33280, 256, 1, 256] - - [35, 56.664] - - - [23296, 3328, 1, 256] - - [37, 72.158] - - - [32560, 256, 1, 256] - - [63, 52.359] - - - [32560, 2816, 1, 256] - - [44, 64.998] - - - [33536, 256, 1, 256] - - [40, 58.138] - - - [34608, 768, 1, 256] - - [53, 60.915] - - - [24832, 5120, 1, 256] - - [25, 73.988] - - - [25856, 2048, 1, 256] - - [78, 70.029] - - - [30768, 256, 1, 256] - - [36, 50.495] - - - [30000, 6656, 1, 256] - - [44, 63.377] - - - [24320, 1024, 1, 256] - - [69, 68.601] - - - [33280, 9216, 1, 256] - - [42, 74.887] - - - [31488, 5376, 1, 256] - - [27, 73.859] - - - [28416, 1281, 1, 256] - - [72, 62.992] - - - [27392, 3584, 1, 256] - - [41, 71.904] - - - [26368, 2048, 1, 256] - - [74, 70.201] - - - [22528, 256, 1, 256] - - [35, 50.074] - - - [32768, 2048, 1, 256] - - [22, 56.064] - - - [30256, 6912, 1, 256] - - [36, 63.907] - - - [28672, 512, 1, 256] - - [34, 62.915] - - - [21760, 8448, 1, 256] - - [25, 74.631] - - - [34560, 6144, 1, 256] - - [29, 74.161] - - - [27696, 2816, 1, 256] - - [44, 63.638] - - - [29952, 2048, 1, 256] - - [23, 70.524] - - - [22576, 10240, 1, 256] - - [41, 62.743] - - - [25600, 1792, 1, 256] - - [34, 71.951] - - - [28976, 10240, 1, 256] - - [53, 62.845] - - - [29952, 1280, 1, 256] - - [55, 70.288] - - - [26368, 2816, 1, 256] - - [34, 72.464] - - - [26416, 3072, 1, 256] - - [58, 62.558] - - - [27648, 3329, 1, 256] - - [37, 70.276] - - - [34560, 2560, 1, 256] - - [29, 72.91] - - - [32048, 8448, 1, 256] - - [53, 63.167] - - - [30464, 2865, 1, 256] - - [73, 68.347] - - - [34048, 3328, 1, 256] - - [43, 72.364] - - - [23808, 2865, 1, 256] - - [25, 69.84] - - - [25600, 2816, 1, 256] - - [25, 73.099] - - - [20736, 6912, 1, 256] - - [25, 74.702] - - - [24576, 512, 1, 256] - - [58, 62.802] - - - [33792, 256, 1, 256] - - [53, 58.332] - - - [22576, 2865, 1, 256] - - [37, 60.853] - - - [30464, 256, 1, 256] - - [34, 54.235] - - - [24368, 2816, 1, 256] - - [58, 63.888] - - - [20224, 512, 1, 256] - - [23, 62.059] - - - [30512, 6912, 1, 256] - - [58, 64.4] - - - [20272, 2816, 1, 256] - - [36, 63.653] - - - [23296, 256, 1, 256] - - [29, 51.395] - - - [27904, 2816, 1, 256] - - [57, 72.231] - - - [29184, 1280, 1, 256] - - [57, 70.572] - - - [24112, 10240, 1, 256] - - [36, 63.496] - - - [31280, 7680, 1, 256] - - [41, 63.157] - - - [24064, 6144, 1, 256] - - [30, 74.387] - - - [26624, 6144, 1, 256] - - [29, 74.59] - - - [30768, 2865, 1, 256] - - [29, 61.038] - - - [20528, 2816, 1, 256] - - [25, 61.315] - - - [25392, 2865, 1, 256] - - [29, 61.179] - - - [22272, 6144, 1, 256] - - [30, 73.834] - - - [25088, 10240, 1, 256] - - [30, 75.251] - - - [25344, 2865, 1, 256] - - [31, 69.458] - - - [23552, 1792, 1, 256] - - [55, 71.439] - - - [23296, 3584, 1, 256] - - [34, 73.272] - - - [28160, 2816, 1, 256] - - [67, 72.65] - - - [20272, 2865, 1, 256] - - [29, 60.656] - - - [22832, 9472, 1, 256] - - [44, 64.024] - - - [21760, 7936, 1, 256] - - [29, 74.725] - - - [26928, 3328, 1, 256] - - [53, 63.435] - - - [33072, 9472, 1, 256] - - [44, 64.768] - - - [33024, 1280, 1, 256] - - [34, 69.175] - - - [34352, 512, 1, 256] - - [36, 60.082] - - - [26368, 2865, 1, 256] - - [27, 70.011] - - - [27952, 4352, 1, 256] - - [53, 63.569] - - - [21504, 8192, 1, 256] - - [27, 75.141] - - - [22320, 9216, 1, 256] - - [44, 63.672] - - - [31232, 2865, 1, 256] - - [27, 70.381] - - - [21248, 7680, 1, 256] - - [37, 74.765] - - - [24368, 256, 1, 256] - - [27, 49.499] - - - [25648, 2865, 1, 256] - - [27, 60.967] - - - [21248, 2865, 1, 256] - - [27, 69.655] - - - [28416, 2865, 1, 256] - - [25, 69.557] - - - [24320, 3329, 1, 256] - - [51, 69.541] - - - [27648, 2048, 1, 256] - - [86, 71.036] - - - [27648, 2865, 1, 256] - - [29, 70.739] - - - [26880, 2048, 1, 256] - - [44, 70.716] - - - [28672, 2560, 1, 256] - - [37, 72.702] - - - [24064, 1280, 1, 256] - - [55, 69.177] - - - [30256, 2865, 1, 256] - - [36, 61.639] - - - [22064, 10240, 1, 256] - - [27, 63.217] - - - [30464, 4608, 1, 256] - - [47, 72.193] - - - [22016, 6144, 1, 256] - - [29, 74.548] - - - [29440, 2816, 1, 256] - - [31, 72.647] - - - [25392, 2048, 1, 256] - - [44, 62.586] - - - [20992, 2048, 1, 256] - - [58, 70.296] - - - [33024, 3329, 1, 256] - - [32, 69.227] - - - [20224, 3328, 1, 256] - - [34, 71.953] - - - [28208, 4608, 1, 256] - - [36, 63.809] - - - [25344, 6144, 1, 256] - - [28, 72.892] - - - [30464, 512, 1, 256] - - [72, 63.734] - - - [21248, 3329, 1, 256] - - [27, 69.165] - - - [29696, 6144, 1, 256] - - [29, 74.778] - - - [20992, 7936, 1, 256] - - [25, 75.164] - - - [33024, 9472, 1, 256] - - [65, 74.823] - - - [32000, 3329, 1, 256] - - [54, 69.733] - - - [21248, 1281, 1, 256] - - [58, 63.451] - - - [24624, 1024, 1, 256] - - [28, 59.008] - - - [22272, 2816, 1, 256] - - [59, 71.99] - - - [29440, 1281, 1, 256] - - [39, 64.108] - - - [30464, 6400, 1, 256] - - [64, 73.497] - - - [25136, 10240, 1, 256] - - [53, 63.177] - - - [23040, 9472, 1, 256] - - [51, 75.599] - - - [33840, 2816, 1, 256] - - [63, 63.323] - - - [30976, 1024, 1, 256] - - [72, 68.274] - - - [34048, 6144, 1, 256] - - [51, 73.573] - - - [32000, 2048, 1, 256] - - [58, 69.97] - - - [32048, 2865, 1, 256] - - [58, 61.728] - - - [33328, 10240, 1, 256] - - [29, 63.577] - - - [25088, 1536, 1, 256] - - [34, 69.981] - - - [30512, 256, 1, 256] - - [44, 49.716] - - - [20480, 6912, 1, 256] - - [25, 74.992] - - - [34608, 2816, 1, 256] - - [44, 63.396] - - - [22064, 256, 1, 256] - - [27, 46.283] - - - [25600, 2865, 1, 256] - - [29, 70.481] - - - [26880, 1024, 1, 256] - - [23, 68.001] - - - [27392, 2048, 1, 256] - - [78, 69.256] - - - [30208, 10240, 1, 256] - - [47, 75.228] - - - [20016, 10240, 1, 256] - - [58, 63.271] - - - [26880, 10240, 1, 256] - - [37, 74.707] - - - [28160, 3328, 1, 256] - - [41, 72.965] - - - [33536, 2048, 1, 256] - - [84, 70.838] - - - [31232, 7936, 1, 256] - - [47, 75.186] - - - [31536, 10240, 1, 256] - - [44, 62.995] - - - [24832, 1536, 1, 256] - - [55, 69.465] - - - [32768, 768, 1, 256] - - [22, 56.214] - - - [29440, 6144, 1, 256] - - [51, 74.249] - - - [26112, 2560, 1, 256] - - [37, 73.095] - - - [33792, 6144, 1, 256] - - [37, 74.818] - - - [22528, 10240, 1, 256] - - [25, 75.441] - - - [20480, 768, 1, 256] - - [60, 64.292] - - - [22320, 256, 1, 256] - - [44, 46.987] - - - [23808, 3328, 1, 256] - - [27, 72.471] - - - [28464, 256, 1, 256] - - [55, 46.705] - - - [27136, 2048, 1, 256] - - [48, 70.97] - - - [29744, 6400, 1, 256] - - [53, 63.787] - - - [20480, 7168, 1, 256] - - [29, 73.728] - - - [22832, 256, 1, 256] - - [53, 47.124] - - - [21552, 8192, 1, 256] - - [27, 63.331] - - - [25856, 2560, 1, 256] - - [37, 72.187] - - - [28160, 6144, 1, 256] - - [41, 74.237] - - - [31280, 2816, 1, 256] - - [36, 63.623] - - - [23600, 10240, 1, 256] - - [28, 63.677] - - - [26368, 1281, 1, 256] - - [60, 63.562] - - - [24576, 1280, 1, 256] - - [25, 65.76] - - - [33536, 1536, 1, 256] - - [35, 70.649] - - - [23088, 2816, 1, 256] - - [53, 63.162] - - - [26624, 2048, 1, 256] - - [61, 70.108] - - - [29952, 2816, 1, 256] - - [25, 72.497] - - - [21760, 2048, 1, 256] - - [63, 70.422] - - - [30976, 6144, 1, 256] - - [28, 72.357] - - - [29696, 1280, 1, 256] - - [34, 70.63] - - - [30208, 4096, 1, 256] - - [41, 73.684] - - - [24832, 2865, 1, 256] - - [25, 69.873] - - - [31488, 1281, 1, 256] - - [40, 64.289] - - - [34304, 2865, 1, 256] - - [51, 70.546] - - - [32512, 256, 1, 256] - - [58, 56.034] - - - [25136, 1536, 1, 256] - - [27, 61.602] - - - [26112, 3329, 1, 256] - - [54, 70.013] - - - [24880, 1280, 1, 256] - - [53, 63.141] - - - [28208, 2816, 1, 256] - - [36, 63.499] - - - [29184, 5888, 1, 256] - - [47, 74.176] - - - [28160, 4352, 1, 256] - - [30, 74.11] - - - [34352, 10240, 1, 256] - - [28, 62.689] - - - [23856, 256, 1, 256] - - [44, 48.368] - - - [25344, 10240, 1, 256] - - [28, 73.908] - - - [20992, 1281, 1, 256] - - [33, 63.172] - - - [26624, 512, 1, 256] - - [34, 63.853] - - - [21040, 10240, 1, 256] - - [37, 62.806] - - - [23040, 3328, 1, 256] - - [78, 72.72] - - - [30976, 7168, 1, 256] - - [41, 72.418] - - - [25856, 2304, 1, 256] - - [25, 71.728] - - - [24368, 1024, 1, 256] - - [44, 63.2] - - - [33280, 2865, 1, 256] - - [51, 70.866] - - - [23296, 1536, 1, 256] - - [34, 69.712] - - - [21504, 6144, 1, 256] - - [27, 74.607] - - - [23552, 2816, 1, 256] - - [27, 73.152] - - - [30464, 2816, 1, 256] - - [31, 71.467] - - - [22832, 2865, 1, 256] - - [36, 61.219] - - - [24576, 2048, 1, 256] - - [27, 64.874] - - - [22272, 8448, 1, 256] - - [51, 74.457] - - - [32256, 1280, 1, 256] - - [67, 71.001] - - - [25856, 5888, 1, 256] - - [27, 73.834] - - - [30976, 5120, 1, 256] - - [41, 72.377] - - - [29184, 3329, 1, 256] - - [73, 69.841] - - - [24112, 2865, 1, 256] - - [58, 61.649] - - - [29744, 2816, 1, 256] - - [53, 63.337] - - - [21760, 2816, 1, 256] - - [35, 72.102] - - - [25600, 2048, 1, 256] - - [41, 71.033] - - - [32000, 1280, 1, 256] - - [34, 70.416] - - - [25856, 3328, 1, 256] - - [41, 72.256] - - - [20016, 6656, 1, 256] - - [53, 64.154] - - - [32256, 2865, 1, 256] - - [32, 70.537] - - - [22272, 3328, 1, 256] - - [67, 72.025] - - - [21504, 3328, 1, 256] - - [25, 72.919] - - - [31232, 5120, 1, 256] - - [24, 74.185] - - - [24112, 256, 1, 256] - - [27, 48.82] - - - [30208, 1280, 1, 256] - - [55, 70.586] - - - [22064, 8960, 1, 256] - - [58, 64.15] - - - [28160, 10240, 1, 256] - - [29, 74.999] - - - [21504, 1536, 1, 256] - - [49, 69.413] - - - [31744, 5632, 1, 256] - - [22, 74.894] - - - [20272, 6912, 1, 256] - - [58, 63.868] - - - [29952, 1792, 1, 256] - - [25, 71.606] - - - [25904, 10240, 1, 256] - - [28, 62.643] - - - [25344, 1792, 1, 256] - - [67, 70.475] - - - [32512, 8448, 1, 256] - - [65, 74.783] - - - [25088, 2048, 1, 256] - - [58, 70.772] - - - [23808, 9984, 1, 256] - - [29, 75.123] - - - [32768, 3329, 1, 256] - - [103, 56.152] - - - [34816, 6144, 1, 256] - - [37, 74.652] - - - [32256, 256, 1, 256] - - [31, 56.472] - - - [26368, 3328, 1, 256] - - [27, 72.308] - - - [23296, 1280, 1, 256] - - [49, 69.319] - - - [34608, 1024, 1, 256] - - [44, 62.214] - - - [30976, 1280, 1, 256] - - [59, 69.531] - - - [22528, 6144, 1, 256] - - [27, 74.61] - - - [21248, 10240, 1, 256] - - [27, 74.773] - - - [22528, 2865, 1, 256] - - [37, 70.368] - - - [22528, 768, 1, 256] - - [40, 66.36] - - - [22016, 8704, 1, 256] - - [47, 75.464] - - - [30720, 6912, 1, 256] - - [27, 75.396] - - - [33024, 2048, 1, 256] - - [42, 70.653] - - - [31232, 3329, 1, 256] - - [37, 69.855] - - - [33024, 3328, 1, 256] - - [43, 72.644] - - - [30976, 7424, 1, 256] - - [41, 73.418] - - - [27136, 3584, 1, 256] - - [25, 73.896] - - - [34048, 1280, 1, 256] - - [59, 70.127] - - - [34864, 1280, 1, 256] - - [25, 62.169] - - - [25600, 2304, 1, 256] - - [29, 72.842] - - - [21760, 3329, 1, 256] - - [29, 69.145] - - - [26928, 3584, 1, 256] - - [53, 63.394] - - - [28976, 2816, 1, 256] - - [53, 63.623] - - - [24832, 4864, 1, 256] - - [57, 74.103] - - - [21248, 1536, 1, 256] - - [34, 68.786] - - - [23808, 2816, 1, 256] - - [34, 72.377] - - - [32768, 9472, 1, 256] - - [22, 59.284] - - - [27392, 3328, 1, 256] - - [43, 71.508] - - - [26880, 3584, 1, 256] - - [25, 73.397] - - - [23552, 1281, 1, 256] - - [74, 63.035] - - - [27648, 3840, 1, 256] - - [37, 74.357] - - - [22016, 10240, 1, 256] - - [47, 75.337] - - - [34816, 2560, 1, 256] - - [25, 73.621] - - - [31536, 256, 1, 256] - - [53, 50.925] - - - [34816, 10240, 1, 256] - - [27, 75.289] - - - [27904, 1792, 1, 256] - - [57, 71.214] - - - [33792, 10240, 1, 256] - - [27, 75.354] - - - [23296, 2816, 1, 256] - - [27, 72.116] - - - [31024, 7424, 1, 256] - - [36, 63.587] - - - [22784, 1280, 1, 256] - - [34, 68.723] - - - [30976, 2048, 1, 256] - - [93, 68.691] - - - [27392, 4096, 1, 256] - - [43, 72.243] - - - [33792, 2816, 1, 256] - - [37, 73.562] - - - [32560, 10240, 1, 256] - - [37, 64.059] - - - [20736, 7424, 1, 256] - - [29, 74.802] - - - [28672, 2865, 1, 256] - - [25, 70.166] - - - [31488, 256, 1, 256] - - [23, 55.599] - - - [20992, 7424, 1, 256] - - [27, 75.243] - - - [21504, 1792, 1, 256] - - [55, 71.135] - - - [27696, 2865, 1, 256] - - [37, 61.694] - - - [33024, 1024, 1, 256] - - [23, 67.629] - - - [22016, 256, 1, 256] - - [58, 49.001] - - - [23088, 256, 1, 256] - - [58, 47.102] - - - [28976, 256, 1, 256] - - [55, 47.864] - - - [27392, 256, 1, 256] - - [36, 57.922] - - - [34304, 3329, 1, 256] - - [47, 69.933] - - - [32512, 9216, 1, 256] - - [43, 74.559] - - - [31488, 3329, 1, 256] - - [54, 69.544] - - - [20016, 2865, 1, 256] - - [36, 61.493] - - - [22016, 8448, 1, 256] - - [51, 75.105] - - - [31024, 2865, 1, 256] - - [53, 61.593] - - - [29440, 256, 1, 256] - - [60, 52.381] - - - [34608, 2865, 1, 256] - - [36, 61.135] - - - [20480, 2048, 1, 256] - - [27, 67.869] - - - [28160, 2865, 1, 256] - - [54, 69.887] - - - [28416, 2304, 1, 256] - - [55, 71.87] - - - [23552, 6144, 1, 256] - - [25, 74.767] - - - [21296, 256, 1, 256] - - [44, 44.768] - - - [28672, 4864, 1, 256] - - [25, 74.407] - - - [27648, 1792, 1, 256] - - [37, 72.137] - - - [31488, 7424, 1, 256] - - [37, 74.459] - - - [23040, 2865, 1, 256] - - [30, 69.926] - - - [30976, 3328, 1, 256] - - [31, 71.027] - - - [25856, 1792, 1, 256] - - [34, 71.058] - - - [33536, 9984, 1, 256] - - [51, 74.907] - - - [24832, 1281, 1, 256] - - [69, 63.319] - - - [29184, 3328, 1, 256] - - [61, 72.913] - - - [32000, 2816, 1, 256] - - [27, 72.454] - - - [34304, 768, 1, 256] - - [39, 68.64] - - - [24576, 1281, 1, 256] - - [27, 58.687] - - - [25088, 1281, 1, 256] - - [39, 63.796] - - - [29744, 2865, 1, 256] - - [53, 61.455] - - - [25136, 2816, 1, 256] - - [58, 63.662] - - - [29696, 1281, 1, 256] - - [28, 64.089] - - - [27392, 3329, 1, 256] - - [84, 67.992] - - - [31488, 2816, 1, 256] - - [37, 72.59] - - - [30976, 10240, 1, 256] - - [28, 73.744] - - - [26624, 3329, 1, 256] - - [25, 70.492] - - - [34304, 1280, 1, 256] - - [59, 71.143] - - - [25392, 256, 1, 256] - - [53, 50.845] - - - [26624, 10240, 1, 256] - - [29, 75.441] - - - [26112, 6144, 1, 256] - - [25, 74.593] - - - [29696, 3328, 1, 256] - - [29, 73.404] - - - [32304, 2865, 1, 256] - - [58, 62.87] - - - [24368, 2865, 1, 256] - - [44, 62.435] - - - [31488, 8192, 1, 256] - - [37, 74.293] - - - [20224, 6656, 1, 256] - - [25, 74.229] - - - [31232, 1281, 1, 256] - - [33, 64.494] - - - [21296, 2865, 1, 256] - - [53, 61.548] - - - [24112, 768, 1, 256] - - [44, 58.71] - - - [32000, 8448, 1, 256] - - [27, 74.291] - - - [23552, 1536, 1, 256] - - [29, 69.852] - - - [30976, 7680, 1, 256] - - [75, 72.959] - - - [31280, 10240, 1, 256] - - [41, 63.142] - - - [23344, 9984, 1, 256] - - [44, 63.465] - - - [21248, 8192, 1, 256] - - [29, 74.558] - - - [29696, 6400, 1, 256] - - [27, 75.374] - - - [32304, 8960, 1, 256] - - [58, 64.13] - - - [27184, 256, 1, 256] - - [58, 52.727] - - - [28464, 10240, 1, 256] - - [41, 62.755] - - - [20736, 256, 1, 256] - - [27, 54.594] - - - [31232, 10240, 1, 256] - - [51, 75.057] - - - [25856, 6144, 1, 256] - - [37, 74.009] - - - [27440, 10240, 1, 256] - - [37, 62.863] - - - [23088, 2865, 1, 256] - - [36, 61.571] - - - [29696, 3584, 1, 256] - - [25, 74.288] - - - [23040, 9728, 1, 256] - - [65, 75.267] - - - [31744, 10240, 1, 256] - - [27, 75.225] - - - [31744, 1792, 1, 256] - - [35, 72.252] - - - [24320, 256, 1, 256] - - [44, 53.207] - - - [27696, 256, 1, 256] - - [58, 53.65] - - - [29696, 2865, 1, 256] - - [27, 70.8] - - - [22784, 3072, 1, 256] - - [27, 72.397] - - - [29952, 5888, 1, 256] - - [59, 73.889] - - - [28928, 2816, 1, 256] - - [27, 72.431] - - - [30768, 7424, 1, 256] - - [37, 62.446] - - - [27440, 4096, 1, 256] - - [53, 63.269] - - - [24064, 4096, 1, 256] - - [91, 73.037] - - - [32256, 3329, 1, 256] - - [30, 70.183] - - - [30976, 3329, 1, 256] - - [59, 68.061] - - - [25600, 10240, 1, 256] - - [25, 75.488] - - - [20224, 6144, 1, 256] - - [25, 73.983] - - - [21040, 7936, 1, 256] - - [40, 64.061] - - - [26368, 2560, 1, 256] - - [49, 72.493] - - - [32512, 1281, 1, 256] - - [91, 63.536] - - - [28928, 3072, 1, 256] - - [27, 72.55] - - - [34864, 2865, 1, 256] - - [27, 61.273] - - - [23552, 9984, 1, 256] - - [27, 75.93] - - - [21040, 2865, 1, 256] - - [53, 61.177] - - - [34048, 1281, 1, 256] - - [89, 63.624] - - - [23296, 10240, 1, 256] - - [29, 74.732] - - - [32768, 6144, 1, 256] - - [22, 58.987] - - - [25904, 2816, 1, 256] - - [44, 63.446] - - - [31232, 1024, 1, 256] - - [60, 69.645] - - - [27648, 3328, 1, 256] - - [25, 73.377] - - - [34864, 256, 1, 256] - - [44, 54.082] - - - [21248, 256, 1, 256] - - [27, 47.768] - - - [26416, 10240, 1, 256] - - [53, 62.855] - - - [27184, 3584, 1, 256] - - [58, 63.863] - - - [23296, 2048, 1, 256] - - [53, 69.765] - - - [34048, 512, 1, 256] - - [35, 65.439] - - - [21760, 2865, 1, 256] - - [25, 69.561] - - - [28672, 2816, 1, 256] - - [25, 72.765] - - - [28672, 4608, 1, 256] - - [25, 73.631] - - - [34560, 512, 1, 256] - - [35, 66.38] - - - [32768, 2865, 1, 256] - - [103, 56.916] - - - [30208, 6912, 1, 256] - - [47, 75.246] - - - [32512, 6144, 1, 256] - - [30, 74.065] - - - [24832, 3328, 1, 256] - - [57, 72.518] - - - [27392, 2816, 1, 256] - - [76, 70.787] - - - [32768, 8704, 1, 256] - - [38, 59.402] - - - [23552, 10240, 1, 256] - - [27, 75.547] - - - [32816, 9216, 1, 256] - - [28, 66.618] - - - [33024, 10240, 1, 256] - - [43, 74.697] - - - [34608, 256, 1, 256] - - [25, 54.121] - - - [20736, 3328, 1, 256] - - [29, 72.065] - - - [31232, 7680, 1, 256] - - [47, 75.166] - - - [22528, 512, 1, 256] - - [34, 60.959] - - - [30208, 2865, 1, 256] - - [37, 70.349] - - - [22272, 2304, 1, 256] - - [34, 71.492] - - - [32512, 2816, 1, 256] - - [24, 72.51] - - - [31488, 7936, 1, 256] - - [37, 74.414] - - - [28416, 2048, 1, 256] - - [48, 69.489] - - - [22784, 3329, 1, 256] - - [29, 69.258] - - - [23040, 2816, 1, 256] - - [35, 72.336] - - - [24320, 3328, 1, 256] - - [59, 72.697] - - - [24064, 1281, 1, 256] - - [39, 64.011] - - - [33072, 9728, 1, 256] - - [53, 64.079] - - - [29440, 10240, 1, 256] - - [30, 74.925] - - - [30208, 6656, 1, 256] - - [30, 74.77] - - - [32768, 3328, 1, 256] - - [22, 58.272] - - - [28416, 6144, 1, 256] - - [27, 73.683] - - - [27904, 4608, 1, 256] - - [28, 73.253] - - - [27184, 2816, 1, 256] - - [58, 63.994] - - - [29184, 1024, 1, 256] - - [39, 69.176] - - - [31744, 1536, 1, 256] - - [55, 71.1] - - - [28416, 10240, 1, 256] - - [47, 74.509] - - - [24368, 10240, 1, 256] - - [29, 63.847] - - - [27904, 3329, 1, 256] - - [30, 69.138] - - - [25344, 3328, 1, 256] - - [59, 71.605] - - - [29952, 6400, 1, 256] - - [24, 74.528] - - - [29440, 2048, 1, 256] - - [74, 70.786] - - - [28928, 1281, 1, 256] - - [33, 63.378] - - - [30208, 3329, 1, 256] - - [37, 69.876] - - - [23088, 9984, 1, 256] - - [58, 64.288] - - - [29184, 2816, 1, 256] - - [24, 72.862] - - - [22528, 2560, 1, 256] - - [27, 72.954] - - - [33328, 2816, 1, 256] - - [58, 64.013] - - - [26368, 256, 1, 256] - - [44, 56.773] - - - [22832, 10240, 1, 256] - - [58, 63.212] - - - [31792, 2816, 1, 256] - - [40, 63.597] - - - [24832, 2048, 1, 256] - - [72, 70.487] - - - [24880, 256, 1, 256] - - [29, 49.526] - - - [33840, 10240, 1, 256] - - [41, 63.711] - - - [33584, 9984, 1, 256] - - [44, 63.613] - - - [28672, 10240, 1, 256] - - [27, 74.999] - - - [24832, 256, 1, 256] - - [57, 53.701] - - - [31488, 2865, 1, 256] - - [37, 70.05] - - - [30720, 7424, 1, 256] - - [29, 75.416] - - - [33536, 2816, 1, 256] - - [57, 72.7] - - - [30000, 6400, 1, 256] - - [58, 63.97] - - - [20224, 1281, 1, 256] - - [60, 63.716] - - - [22832, 2816, 1, 256] - - [53, 63.64] - - - [25600, 6144, 1, 256] - - [27, 74.742] - - - [24320, 4352, 1, 256] - - [57, 73.984] - - - [32768, 10240, 1, 256] - - [66, 58.859] - - - [26880, 768, 1, 256] - - [40, 66.503] - - - [24576, 3329, 1, 256] - - [32, 65.155] - - - [27904, 3840, 1, 256] - - [57, 73.388] - - - [30256, 2816, 1, 256] - - [53, 63.427] - - - [23296, 1281, 1, 256] - - [40, 62.497] - - - [26880, 256, 1, 256] - - [36, 57.446] - - - [23344, 2816, 1, 256] - - [36, 63.552] - - - [33792, 2048, 1, 256] - - [86, 71.385] - - - [21504, 3329, 1, 256] - - [29, 70.177] - - - [20272, 256, 1, 256] - - [60, 50.146] - - - [32768, 1280, 1, 256] - - [38, 56.265] - - - [32256, 10240, 1, 256] - - [51, 75.088] - - - [27952, 2816, 1, 256] - - [58, 63.766] - - - [28928, 5376, 1, 256] - - [47, 73.7] - - - [20992, 6144, 1, 256] - - [29, 74.373] - - - [20224, 2048, 1, 256] - - [23, 69.842] - - - [33280, 10240, 1, 256] - - [27, 75.116] - - - [24064, 3329, 1, 256] - - [51, 69.727] - - - [32768, 9216, 1, 256] - - [66, 58.928] - - - [20016, 6912, 1, 256] - - [53, 64.096] - - - [22320, 10240, 1, 256] - - [44, 63.577] - - - [22784, 256, 1, 256] - - [34, 50.256] - - - [34816, 512, 1, 256] - - [29, 66.676] - - - [32048, 8704, 1, 256] - - [53, 63.329] - - - [29232, 5888, 1, 256] - - [36, 63.527] - - - [24064, 768, 1, 256] - - [63, 65.425] - - - [33792, 9984, 1, 256] - - [29, 75.596] - - - [32512, 3329, 1, 256] - - [73, 69.597] - - - [21504, 2048, 1, 256] - - [41, 70.357] - - - [28160, 2304, 1, 256] - - [27, 72.19] - - - [20784, 10240, 1, 256] - - [37, 63.029] - - - [20224, 7168, 1, 256] - - [29, 73.51] - - - [28976, 2865, 1, 256] - - [44, 61.422] - - - [21296, 2816, 1, 256] - - [44, 63.29] - - - [23552, 256, 1, 256] - - [34, 51.689] - - - [26160, 2865, 1, 256] - - [36, 61.461] - - - [23600, 2816, 1, 256] - - [36, 63.383] - - - [20480, 7424, 1, 256] - - [25, 75.186] - - - [28928, 3329, 1, 256] - - [54, 69.267] - - - [20784, 2816, 1, 256] - - [53, 63.356] - - - [25344, 256, 1, 256] - - [53, 55.05] - - - [20224, 10240, 1, 256] - - [37, 74.867] - - - [28672, 1280, 1, 256] - - [49, 69.999] - - - [29232, 256, 1, 256] - - [44, 47.674] - - - [28720, 2865, 1, 256] - - [25, 59.965] - - - [22016, 2816, 1, 256] - - [27, 72.506] - - - [25600, 1536, 1, 256] - - [34, 70.369] - - - [26112, 10240, 1, 256] - - [27, 75.339] - - - [27136, 10240, 1, 256] - - [51, 75.178] - - - [31744, 8192, 1, 256] - - [27, 74.947] - - - [24320, 10240, 1, 256] - - [51, 74.908] - - - [29952, 10240, 1, 256] - - [30, 74.674] - - - [23296, 9984, 1, 256] - - [25, 75.04] - - - [34560, 2304, 1, 256] - - [35, 72.547] - - - [32000, 2865, 1, 256] - - [54, 70.214] - - - [25088, 1024, 1, 256] - - [63, 68.105] - - - [20272, 10240, 1, 256] - - [44, 63.816] - - - [25344, 5376, 1, 256] - - [42, 73.025] - - - [21760, 3328, 1, 256] - - [25, 72.312] - - - [32768, 8960, 1, 256] - - [38, 59.008] - - - [29952, 3840, 1, 256] - - [31, 73.676] - - - [32512, 2865, 1, 256] - - [73, 69.874] - - - [23344, 2865, 1, 256] - - [36, 60.816] - - - [24576, 768, 1, 256] - - [34, 60.724] - - - [27648, 3584, 1, 256] - - [25, 74.08] - - - [27952, 4608, 1, 256] - - [53, 63.343] - - - [29440, 3584, 1, 256] - - [27, 73.461] - - - [34096, 512, 1, 256] - - [53, 59.465] - - - [32304, 256, 1, 256] - - [53, 52.548] - - - [21040, 2816, 1, 256] - - [36, 64.179] - - - [22784, 1024, 1, 256] - - [33, 67.558] - - - [22784, 2816, 1, 256] - - [35, 71.985] - - - [25856, 2816, 1, 256] - - [49, 72.336] - - - [23296, 6144, 1, 256] - - [27, 74.154] - - - [28160, 4608, 1, 256] - - [91, 73.703] - - - [25136, 1792, 1, 256] - - [29, 62.163] - - - [30208, 256, 1, 256] - - [23, 53.789] - - - [23808, 1281, 1, 256] - - [60, 63.94] - - - [26368, 2304, 1, 256] - - [35, 71.856] - - - [27648, 4352, 1, 256] - - [27, 74.703] - - - [31280, 7936, 1, 256] - - [44, 63.951] - - - [22320, 2865, 1, 256] - - [25, 61.292] - - - [22320, 2816, 1, 256] - - [58, 63.845] - - - [28720, 5120, 1, 256] - - [28, 61.738] - - - [22272, 1280, 1, 256] - - [34, 69.431] - - - [31232, 3328, 1, 256] - - [41, 73.148] - - - [29696, 2048, 1, 256] - - [41, 71.313] - - - [34048, 9984, 1, 256] - - [51, 74.355] - - - [28416, 1280, 1, 256] - - [55, 69.731] - - - [21504, 2816, 1, 256] - - [49, 72.798] - - - [33536, 2865, 1, 256] - - [37, 69.879] - - - [23552, 3840, 1, 256] - - [37, 74.249] - - - [31744, 256, 1, 256] - - [27, 55.586] - - - [25600, 1281, 1, 256] - - [41, 63.258] - - - [30768, 7168, 1, 256] - - [28, 62.311] - - - [23808, 3329, 1, 256] - - [27, 69.505] - - - [32256, 3328, 1, 256] - - [24, 73.232] - - - [23040, 9216, 1, 256] - - [41, 74.942] - - - [33024, 256, 1, 256] - - [56, 54.622] - - - [33584, 2865, 1, 256] - - [58, 62.267] - - - [21504, 8448, 1, 256] - - [27, 75.477] - - - [27904, 1281, 1, 256] - - [40, 63.938] - - - [34304, 10240, 1, 256] - - [43, 75.096] - - - [20992, 2865, 1, 256] - - [37, 70.053] - - - [22528, 8960, 1, 256] - - [29, 75.783] - - - [28928, 3328, 1, 256] - - [25, 72.199] - - - [21808, 2865, 1, 256] - - [53, 61.674] - - - [26416, 2816, 1, 256] - - [53, 63.431] - - - [27392, 3840, 1, 256] - - [41, 72.145] - - - [26112, 1281, 1, 256] - - [33, 64.084] - - - [34864, 10240, 1, 256] - - [28, 62.895] - - - [29440, 1536, 1, 256] - - [35, 70.113] - - - [30256, 10240, 1, 256] - - [44, 62.69] - - - [22528, 2816, 1, 256] - - [37, 73.081] - - - [28928, 2048, 1, 256] - - [39, 69.845] - - - [28976, 5376, 1, 256] - - [36, 63.383] - - - [20736, 7168, 1, 256] - - [25, 73.681] - - - [22016, 2865, 1, 256] - - [27, 70.061] - - - [26368, 1280, 1, 256] - - [34, 69.846] - - - [24624, 2865, 1, 256] - - [55, 59.672] - - - [23040, 3329, 1, 256] - - [73, 69.455] - - - [23296, 2865, 1, 256] - - [27, 69.699] - - - [28416, 3329, 1, 256] - - [54, 69.052] - - - [23040, 1281, 1, 256] - - [56, 63.922] - - - [21808, 8448, 1, 256] - - [36, 64.072] - - - [30720, 2865, 1, 256] - - [27, 70.964] - - - [22272, 8960, 1, 256] - - [31, 74.925] - - - [34864, 2816, 1, 256] - - [29, 62.188] - - - [31232, 7168, 1, 256] - - [51, 73.728] - - - [27696, 4352, 1, 256] - - [25, 63.284] - - - [21504, 256, 1, 256] - - [49, 48.344] - - - [28672, 1281, 1, 256] - - [29, 63.048] - - - [29696, 1792, 1, 256] - - [27, 72.277] - - - [28464, 5120, 1, 256] - - [44, 63.253] - - - [27136, 3329, 1, 256] - - [47, 70.101] - - - [21248, 3328, 1, 256] - - [55, 72.083] - - - [26880, 1281, 1, 256] - - [58, 63.574] - - - [32256, 8448, 1, 256] - - [47, 75.036] - - - [20480, 6144, 1, 256] - - [27, 74.369] - - - [34048, 2865, 1, 256] - - [73, 69.561] - - - [29696, 5888, 1, 256] - - [37, 74.749] - - - [28720, 256, 1, 256] - - [29, 46.969] - - - [33792, 2865, 1, 256] - - [37, 71.066] - - - [22784, 8960, 1, 256] - - [27, 74.796] - - - [30720, 256, 1, 256] - - [60, 54.635] - - - [23808, 512, 1, 256] - - [40, 63.037] - - - [33024, 9728, 1, 256] - - [43, 74.736] - - - [42624, 13824, 1, 384] - - [28, 88.278] - - - [33024, 3840, 1, 384] - - [36, 89.14] - - - [33408, 15360, 1, 384] - - [29, 90.765] - - - [44160, 8832, 1, 384] - - [53, 90.868] - - - [31488, 2688, 1, 384] - - [36, 89.239] - - - [39168, 3072, 1, 384] - - [25, 89.312] - - - [31872, 5760, 1, 384] - - [37, 90.116] - - - [36096, 13440, 1, 384] - - [28, 89.824] - - - [41856, 1152, 1, 384] - - [53, 87.469] - - - [32256, 1153, 1, 384] - - [53, 78.124] - - - [44160, 1153, 1, 384] - - [58, 78.2] - - - [31488, 7296, 1, 384] - - [29, 90.342] - - - [43008, 9216, 1, 384] - - [28, 88.337] - - - [31872, 6144, 1, 384] - - [37, 89.989] - - - [32640, 7297, 1, 384] - - [54, 85.715] - - - [33792, 1152, 1, 384] - - [27, 85.923] - - - [43776, 13441, 1, 384] - - [54, 87.285] - - - [36480, 1153, 1, 384] - - [36, 78.52] - - - [37632, 1152, 1, 384] - - [36, 85.178] - - - [37248, 8448, 1, 384] - - [27, 90.158] - - - [31872, 7297, 1, 384] - - [29, 88.03] - - - [41856, 7296, 1, 384] - - [58, 90.634] - - - [39936, 7297, 1, 384] - - [27, 87.892] - - - [35712, 1153, 1, 384] - - [36, 77.05] - - - [35712, 3072, 1, 384] - - [37, 89.547] - - - [31488, 1153, 1, 384] - - [25, 76.43] - - - [36480, 1152, 1, 384] - - [53, 85.411] - - - [36864, 9216, 1, 384] - - [54, 86.953] - - - [42624, 15360, 1, 384] - - [28, 87.319] - - - [37632, 8832, 1, 384] - - [58, 90.755] - - - [32640, 1153, 1, 384] - - [78, 73.418] - - - [36864, 3072, 1, 384] - - [29, 87.687] - - - [32640, 6912, 1, 384] - - [37, 88.737] - - - [31872, 13440, 1, 384] - - [27, 90.709] - - - [39168, 3840, 1, 384] - - [29, 89.866] - - - [39168, 10368, 1, 384] - - [53, 90.644] - - - [33792, 3072, 1, 384] - - [37, 88.025] - - - [39552, 1536, 1, 384] - - [27, 87.432] - - - [38784, 7296, 1, 384] - - [36, 90.329] - - - [40320, 1153, 1, 384] - - [36, 78.245] - - - [42240, 1152, 1, 384] - - [44, 87.93] - - - [43776, 14976, 1, 384] - - [41, 89.905] - - - [38784, 9216, 1, 384] - - [58, 90.782] - - - [33024, 4224, 1, 384] - - [53, 89.263] - - - [43776, 7297, 1, 384] - - [54, 86.204] - - - [34560, 9216, 1, 384] - - [56, 90.365] - - - [43392, 8064, 1, 384] - - [27, 90.729] - - - [34944, 7296, 1, 384] - - [27, 90.377] - - - [38400, 7296, 1, 384] - - [37, 90.343] - - - [41856, 6912, 1, 384] - - [44, 90.762] - - - [40704, 3072, 1, 384] - - [29, 89.188] - - - [41472, 12672, 1, 384] - - [25, 91.04] - - - [36864, 1920, 1, 384] - - [37, 87.498] - - - [43008, 1920, 1, 384] - - [27, 88.654] - - - [43008, 13824, 1, 384] - - [29, 90.147] - - - [31104, 13441, 1, 384] - - [29, 89.23] - - - [41472, 12288, 1, 384] - - [61, 89.607] - - - [31488, 7297, 1, 384] - - [25, 87.888] - - - [35712, 6912, 1, 384] - - [25, 90.541] - - - [40704, 5376, 1, 384] - - [58, 90.373] - - - [36480, 9216, 1, 384] - - [44, 90.768] - - - [38784, 13440, 1, 384] - - [37, 90.948] - - - [36096, 15360, 1, 384] - - [30, 89.555] - - - [41856, 15360, 1, 384] - - [36, 91.147] - - - [37632, 2688, 1, 384] - - [44, 88.713] - - - [33792, 4608, 1, 384] - - [27, 88.828] - - - [38400, 13440, 1, 384] - - [37, 90.905] - - - [31104, 3072, 1, 384] - - [25, 88.396] - - - [33792, 13440, 1, 384] - - [25, 90.724] - - - [34176, 5376, 1, 384] - - [36, 89.995] - - - [31872, 3072, 1, 384] - - [25, 88.853] - - - [33792, 1920, 1, 384] - - [29, 88.614] - - - [34560, 1153, 1, 384] - - [36, 77.466] - - - [43392, 15360, 1, 384] - - [27, 90.778] - - - [39168, 4224, 1, 384] - - [25, 90.179] - - - [43776, 1153, 1, 384] - - [51, 76.234] - - - [41472, 6528, 1, 384] - - [58, 90.214] - - - [42240, 1153, 1, 384] - - [25, 78.784] - - - [36480, 13441, 1, 384] - - [27, 88.921] - - - [31488, 5760, 1, 384] - - [29, 90.215] - - - [34560, 13440, 1, 384] - - [25, 90.784] - - - [32256, 3072, 1, 384] - - [25, 88.739] - - - [37632, 15360, 1, 384] - - [25, 90.795] - - - [43776, 8448, 1, 384] - - [41, 89.663] - - - [37248, 13440, 1, 384] - - [37, 90.849] - - - [34944, 13440, 1, 384] - - [29, 90.894] - - - [41088, 3072, 1, 384] - - [32, 86.816] - - - [43008, 14208, 1, 384] - - [25, 90.513] - - - [33792, 7296, 1, 384] - - [25, 90.356] - - - [43392, 8448, 1, 384] - - [53, 90.947] - - - [31104, 7297, 1, 384] - - [37, 88.048] - - - [31104, 2304, 1, 384] - - [53, 88.538] - - - [35712, 1152, 1, 384] - - [36, 86.709] - - - [39552, 13440, 1, 384] - - [25, 90.967] - - - [37632, 2304, 1, 384] - - [53, 88.3] - - - [31872, 1153, 1, 384] - - [44, 77.391] - - - [39552, 3072, 1, 384] - - [25, 89.481] - - - [36864, 15360, 1, 384] - - [32, 88.677] - - - [33408, 4608, 1, 384] - - [58, 89.822] - - - [43392, 7297, 1, 384] - - [27, 88.191] - - - [32256, 7296, 1, 384] - - [27, 90.393] - - - [41472, 7296, 1, 384] - - [25, 90.726] - - - [38016, 9216, 1, 384] - - [36, 90.741] - - - [38784, 1153, 1, 384] - - [25, 77.728] - - - [34944, 2688, 1, 384] - - [44, 88.484] - - - [36864, 1152, 1, 384] - - [25, 85.952] - - - [39168, 7297, 1, 384] - - [37, 88.429] - - - [33024, 768, 1, 384] - - [86, 79.528] - - - [34560, 13441, 1, 384] - - [27, 89.15] - - - [33792, 7680, 1, 384] - - [29, 90.317] - - - [36864, 1153, 1, 384] - - [25, 76.423] - - - [40320, 4992, 1, 384] - - [53, 90.42] - - - [31488, 13440, 1, 384] - - [25, 90.896] - - - [39552, 10752, 1, 384] - - [36, 91.09] - - - [36096, 1152, 1, 384] - - [54, 83.95] - - - [44160, 1152, 1, 384] - - [37, 86.378] - - - [37632, 9216, 1, 384] - - [44, 90.767] - - - [37248, 15360, 1, 384] - - [37, 90.911] - - - [34944, 5760, 1, 384] - - [27, 90.259] - - - [41088, 15360, 1, 384] - - [61, 89.764] - - - [41088, 11904, 1, 384] - - [74, 89.957] - - - [35328, 6528, 1, 384] - - [29, 90.045] - - - [32640, 15360, 1, 384] - - [54, 89.211] - - - [33024, 7297, 1, 384] - - [32, 86.942] - - - [31104, 1153, 1, 384] - - [25, 75.891] - - - [40704, 1153, 1, 384] - - [36, 78.6] - - - [42240, 13440, 1, 384] - - [58, 91.201] - - - [41472, 7297, 1, 384] - - [37, 88.199] - - - [33408, 3072, 1, 384] - - [25, 89.593] - - - [40704, 13440, 1, 384] - - [25, 91.016] - - - [39168, 7296, 1, 384] - - [25, 90.527] - - - [34176, 9216, 1, 384] - - [56, 90.031] - - - [35328, 15360, 1, 384] - - [27, 90.572] - - - [38400, 1152, 1, 384] - - [53, 86.421] - - - [37248, 3072, 1, 384] - - [27, 89.373] - - - [31488, 2304, 1, 384] - - [53, 87.961] - - - [40704, 1152, 1, 384] - - [29, 85.429] - - - [39168, 768, 1, 384] - - [36, 85.104] - - - [34944, 1153, 1, 384] - - [53, 78.011] - - - [39936, 13440, 1, 384] - - [25, 90.764] - - - [43008, 7297, 1, 384] - - [25, 87.793] - - - [33024, 15360, 1, 384] - - [27, 90.317] - - - [34176, 1920, 1, 384] - - [25, 87.757] - - - [40320, 15360, 1, 384] - - [29, 90.659] - - - [37632, 3072, 1, 384] - - [25, 88.886] - - - [40320, 11136, 1, 384] - - [44, 90.855] - - - [34944, 1152, 1, 384] - - [58, 85.364] - - - [44160, 14976, 1, 384] - - [44, 90.918] - - - [33792, 1536, 1, 384] - - [29, 86.629] - - - [38016, 13441, 1, 384] - - [29, 88.959] - - - [37632, 7296, 1, 384] - - [37, 90.485] - - - [41856, 6528, 1, 384] - - [36, 90.77] - - - [36096, 6912, 1, 384] - - [51, 89.39] - - - [39936, 15360, 1, 384] - - [32, 89.281] - - - [43776, 9216, 1, 384] - - [74, 89.701] - - - [38400, 9600, 1, 384] - - [29, 90.744] - - - [39552, 15360, 1, 384] - - [37, 90.804] - - - [37248, 2304, 1, 384] - - [29, 88.9] - - - [33792, 1153, 1, 384] - - [27, 78.249] - - - [42624, 1152, 1, 384] - - [29, 85.746] - - - [35328, 3072, 1, 384] - - [25, 88.672] - - - [37632, 13440, 1, 384] - - [27, 90.935] - - - [38400, 3072, 1, 384] - - [25, 88.81] - - - [32640, 1152, 1, 384] - - [37, 82.416] - - - [31872, 1152, 1, 384] - - [44, 84.902] - - - [40320, 3072, 1, 384] - - [25, 89.67] - - - [38016, 15360, 1, 384] - - [29, 91.029] - - - [35712, 9216, 1, 384] - - [58, 90.628] - - - [33024, 13441, 1, 384] - - [32, 88.236] - - - [36096, 3072, 1, 384] - - [30, 87.825] - - - [36864, 13440, 1, 384] - - [25, 90.3] - - - [33408, 13441, 1, 384] - - [27, 89.408] - - - [37248, 9216, 1, 384] - - [29, 90.012] - - - [31488, 1152, 1, 384] - - [29, 84.3] - - - [31488, 3072, 1, 384] - - [27, 88.897] - - - [35328, 1152, 1, 384] - - [37, 86.19] - - - [37248, 7297, 1, 384] - - [24, 87.954] - - - [34944, 6144, 1, 384] - - [27, 89.899] - - - [36480, 1536, 1, 384] - - [29, 85.898] - - - [39168, 15360, 1, 384] - - [37, 90.611] - - - [43392, 13441, 1, 384] - - [25, 89.041] - - - [42624, 1536, 1, 384] - - [32, 86.032] - - - [36480, 7296, 1, 384] - - [58, 90.429] - - - [33792, 9216, 1, 384] - - [61, 88.885] - - - [36096, 768, 1, 384] - - [62, 81.715] - - - [33408, 1536, 1, 384] - - [27, 86.603] - - - [31872, 13441, 1, 384] - - [25, 88.734] - - - [43008, 13440, 1, 384] - - [37, 90.679] - - - [33024, 1152, 1, 384] - - [53, 83.603] - - - [34560, 5376, 1, 384] - - [36, 89.95] - - - [32640, 3840, 1, 384] - - [29, 87.392] - - - [33408, 1153, 1, 384] - - [27, 77.608] - - - [32256, 1152, 1, 384] - - [27, 85.845] - - - [41856, 13440, 1, 384] - - [44, 91.068] - - - [43776, 2688, 1, 384] - - [74, 87.926] - - - [34560, 8832, 1, 384] - - [29, 90.698] - - - [32256, 6528, 1, 384] - - [25, 89.983] - - - [33408, 13440, 1, 384] - - [25, 90.855] - - - [36096, 7296, 1, 384] - - [30, 89.202] - - - [43776, 3072, 1, 384] - - [30, 86.725] - - - [38784, 7297, 1, 384] - - [25, 87.841] - - - [39936, 7296, 1, 384] - - [27, 90.369] - - - [37632, 8448, 1, 384] - - [53, 90.624] - - - [43392, 9216, 1, 384] - - [58, 90.379] - - - [41856, 13056, 1, 384] - - [53, 91.0] - - - [30720, 13441, 1, 384] - - [29, 88.74] - - - [36864, 7680, 1, 384] - - [25, 89.447] - - - [41472, 1152, 1, 384] - - [44, 86.836] - - - [39168, 13440, 1, 384] - - [37, 90.989] - - - [43776, 2304, 1, 384] - - [74, 87.775] - - - [34176, 15360, 1, 384] - - [25, 90.734] - - - [36096, 7297, 1, 384] - - [30, 86.378] - - - [33792, 4992, 1, 384] - - [27, 90.089] - - - [35712, 15360, 1, 384] - - [29, 90.899] - - - [39168, 9984, 1, 384] - - [25, 90.689] - - - [36096, 9216, 1, 384] - - [74, 89.158] - - - [43008, 1536, 1, 384] - - [29, 86.668] - - - [33408, 9216, 1, 384] - - [62, 90.217] - - - [40704, 7296, 1, 384] - - [58, 90.655] - - - [38016, 2688, 1, 384] - - [36, 89.371] - - - [39168, 13441, 1, 384] - - [29, 89.187] - - - [39168, 9216, 1, 384] - - [37, 89.958] - - - [38400, 15360, 1, 384] - - [37, 90.558] - - - [43392, 2304, 1, 384] - - [44, 89.144] - - - [38400, 13441, 1, 384] - - [25, 88.937] - - - [43008, 1152, 1, 384] - - [25, 86.658] - - - [39936, 4608, 1, 384] - - [29, 88.688] - - - [43392, 14592, 1, 384] - - [27, 90.978] - - - [34176, 13441, 1, 384] - - [30, 88.89] - - - [38784, 9984, 1, 384] - - [29, 90.687] - - - [44160, 13441, 1, 384] - - [37, 88.866] - - - [31488, 5376, 1, 384] - - [44, 90.263] - - - [39936, 13441, 1, 384] - - [37, 88.989] - - - [34176, 1152, 1, 384] - - [37, 86.48] - - - [32640, 3072, 1, 384] - - [25, 86.224] - - - [34560, 15360, 1, 384] - - [29, 90.733] - - - [34944, 15360, 1, 384] - - [37, 90.744] - - - [37632, 13441, 1, 384] - - [27, 89.069] - - - [40320, 5376, 1, 384] - - [36, 90.308] - - - [41856, 12672, 1, 384] - - [44, 91.102] - - - [34176, 4992, 1, 384] - - [29, 89.892] - - - [42624, 7297, 1, 384] - - [51, 84.477] - - - [41856, 1153, 1, 384] - - [25, 78.384] - - - [41472, 9216, 1, 384] - - [74, 89.586] - - - [40704, 2304, 1, 384] - - [53, 88.393] - - - [36864, 8064, 1, 384] - - [25, 90.015] - - - [40704, 5760, 1, 384] - - [37, 90.322] - - - [41088, 7297, 1, 384] - - [54, 86.365] - - - [38784, 1152, 1, 384] - - [25, 87.044] - - - [38784, 3072, 1, 384] - - [29, 89.002] - - - [34560, 2304, 1, 384] - - [37, 88.169] - - - [36096, 1153, 1, 384] - - [62, 76.971] - - - [35712, 13440, 1, 384] - - [29, 90.934] - - - [39936, 1152, 1, 384] - - [27, 86.321] - - - [43392, 14208, 1, 384] - - [25, 90.644] - - - [39552, 1153, 1, 384] - - [36, 77.084] - - - [35712, 6528, 1, 384] - - [44, 90.358] - - - [31104, 5376, 1, 384] - - [58, 90.199] - - - [31104, 9216, 1, 384] - - [44, 90.311] - - - [33024, 9216, 1, 384] - - [61, 89.636] - - - [39936, 11136, 1, 384] - - [25, 90.513] - - - [43008, 3072, 1, 384] - - [25, 87.535] - - - [41856, 768, 1, 384] - - [53, 85.598] - - - [43776, 1152, 1, 384] - - [74, 85.56] - - - [34176, 7297, 1, 384] - - [24, 87.794] - - - [38016, 7297, 1, 384] - - [53, 88.208] - - - [36480, 7680, 1, 384] - - [58, 90.618] - - - [38400, 7297, 1, 384] - - [37, 88.136] - - - [44160, 2688, 1, 384] - - [44, 89.787] - - - [33792, 15360, 1, 384] - - [54, 89.277] - - - [40704, 2688, 1, 384] - - [36, 89.517] - - - [38784, 3840, 1, 384] - - [53, 90.014] - - - [44160, 7296, 1, 384] - - [58, 90.668] - - - [41088, 2688, 1, 384] - - [36, 88.083] - - - [38016, 3072, 1, 384] - - [29, 89.602] - - - [42240, 7296, 1, 384] - - [36, 90.815] - - - [41856, 9216, 1, 384] - - [58, 90.762] - - - [32640, 13440, 1, 384] - - [27, 89.719] - - - [40320, 13441, 1, 384] - - [27, 88.933] - - - [36480, 13440, 1, 384] - - [27, 90.863] - - - [41856, 7297, 1, 384] - - [58, 88.115] - - - [41088, 7296, 1, 384] - - [38, 89.193] - - - [33408, 1152, 1, 384] - - [36, 85.116] - - - [43392, 1920, 1, 384] - - [27, 88.388] - - - [31104, 1920, 1, 384] - - [37, 86.98] - - - [31488, 15360, 1, 384] - - [395, 91.493] - - - [31872, 7296, 1, 384] - - [25, 90.056] - - - [43008, 7680, 1, 384] - - [29, 90.127] - - - [35328, 13440, 1, 384] - - [25, 90.941] - - - [43776, 15360, 1, 384] - - [61, 89.613] - - - [34944, 3072, 1, 384] - - [25, 89.225] - - - [37248, 1153, 1, 384] - - [36, 77.474] - - - [31104, 1152, 1, 384] - - [58, 86.317] - - - [34560, 7297, 1, 384] - - [37, 88.375] - - - [43776, 14592, 1, 384] - - [41, 89.832] - - - [33408, 7296, 1, 384] - - [25, 90.328] - - - [33024, 7296, 1, 384] - - [29, 89.596] - - - [33024, 13440, 1, 384] - - [27, 90.417] - - - [31104, 7296, 1, 384] - - [25, 90.339] - - - [42240, 9216, 1, 384] - - [53, 91.084] - - - [34944, 13441, 1, 384] - - [37, 89.08] - - - [33792, 7297, 1, 384] - - [37, 88.085] - - - [35328, 13441, 1, 384] - - [25, 89.221] - - - [34176, 7296, 1, 384] - - [37, 90.146] - - - [40320, 1920, 1, 384] - - [53, 89.136] - - - [31872, 15360, 1, 384] - - [25, 90.709] - - - [39168, 1153, 1, 384] - - [53, 78.445] - - - [31104, 4992, 1, 384] - - [37, 89.792] - - - [41088, 1152, 1, 384] - - [74, 84.162] - - - [39552, 10368, 1, 384] - - [44, 90.976] - - - [40704, 11520, 1, 384] - - [25, 90.92] - - - [36864, 7297, 1, 384] - - [27, 87.1] - - - [42240, 15360, 1, 384] - - [36, 91.13] - - - [34560, 1152, 1, 384] - - [27, 84.888] - - - [31104, 13440, 1, 384] - - [29, 90.806] - - - [31488, 9216, 1, 384] - - [29, 89.809] - - - [34176, 3072, 1, 384] - - [29, 88.904] - - - [41088, 1153, 1, 384] - - [58, 76.435] - - - [43392, 1153, 1, 384] - - [53, 78.69] - - - [42240, 6912, 1, 384] - - [36, 90.914] - - - [43008, 15360, 1, 384] - - [54, 89.299] - - - [42240, 7297, 1, 384] - - [27, 88.344] - - - [43776, 7296, 1, 384] - - [41, 89.302] - - - [35712, 7296, 1, 384] - - [37, 90.409] - - - [38400, 9216, 1, 384] - - [56, 89.62] - - - [39936, 9216, 1, 384] - - [61, 89.072] - - - [32256, 6144, 1, 384] - - [27, 89.891] - - - [42624, 7680, 1, 384] - - [25, 88.093] - - - [33408, 4224, 1, 384] - - [36, 89.662] - - - [38784, 768, 1, 384] - - [44, 84.311] - - - [38016, 7296, 1, 384] - - [29, 90.545] - - - [34560, 5760, 1, 384] - - [29, 90.265] - - - [34944, 7297, 1, 384] - - [27, 87.974] - - - [38016, 8832, 1, 384] - - [37, 90.763] - - - [39936, 1920, 1, 384] - - [27, 88.364] - - - [40320, 11520, 1, 384] - - [27, 91.004] - - - [32256, 7297, 1, 384] - - [27, 88.391] - - - [33792, 13441, 1, 384] - - [25, 88.944] - - - [41472, 3072, 1, 384] - - [37, 89.299] - - - [33024, 1153, 1, 384] - - [58, 76.254] - - - [36864, 7296, 1, 384] - - [37, 89.855] - - - [38016, 1153, 1, 384] - - [27, 78.612] - - - [40320, 7297, 1, 384] - - [29, 88.106] - - - [42624, 13441, 1, 384] - - [30, 84.193] - - - [43008, 13441, 1, 384] - - [25, 88.6] - - - [39552, 9216, 1, 384] - - [44, 90.808] - - - [35328, 9216, 1, 384] - - [74, 89.367] - - - [42624, 3072, 1, 384] - - [25, 88.928] - - - [40320, 13440, 1, 384] - - [27, 91.065] - - - [42240, 13441, 1, 384] - - [29, 89.158] - - - [39936, 10752, 1, 384] - - [37, 90.046] - - - [41472, 6144, 1, 384] - - [29, 89.786] - - - [36864, 1536, 1, 384] - - [25, 85.867] - - - [33408, 7297, 1, 384] - - [27, 88.0] - - - [31872, 2688, 1, 384] - - [53, 88.748] - - - [41472, 1153, 1, 384] - - [27, 78.0] - - - [38400, 1153, 1, 384] - - [36, 77.123] - - - [38400, 3456, 1, 384] - - [36, 89.537] - - - [41856, 13441, 1, 384] - - [29, 88.889] - - - [43392, 1152, 1, 384] - - [36, 87.397] - - - [39552, 4608, 1, 384] - - [44, 89.957] - - - [40704, 15360, 1, 384] - - [29, 90.792] - - - [42240, 3072, 1, 384] - - [29, 89.467] - - - [32640, 3456, 1, 384] - - [27, 86.766] - - - [35712, 768, 1, 384] - - [44, 82.475] - - - [31104, 15360, 1, 384] - - [407, 91.195] - - - [40704, 13441, 1, 384] - - [27, 89.252] - - - [32640, 7296, 1, 384] - - [29, 88.652] - - - [34176, 8448, 1, 384] - - [93, 90.045] - - - [32640, 13441, 1, 384] - - [54, 87.325] - - - [36864, 13441, 1, 384] - - [25, 87.818] - - - [34176, 13440, 1, 384] - - [37, 90.629] - - - [37248, 1152, 1, 384] - - [53, 87.015] - - - [44160, 7297, 1, 384] - - [44, 88.025] - - - [41088, 6144, 1, 384] - - [61, 88.988] - - - [39936, 1536, 1, 384] - - [25, 86.0] - - - [44160, 15360, 1, 384] - - [44, 90.988] - - - [35712, 7297, 1, 384] - - [25, 88.087] - - - [35328, 6144, 1, 384] - - [29, 89.727] - - - [42624, 7296, 1, 384] - - [50, 88.397] - - - [33408, 7680, 1, 384] - - [29, 90.515] - - - [41472, 13441, 1, 384] - - [37, 89.163] - - - [43776, 8832, 1, 384] - - [74, 89.563] - - - [32256, 15360, 1, 384] - - [25, 90.612] - - - [32256, 9216, 1, 384] - - [27, 89.77] - - - [31872, 9216, 1, 384] - - [26, 90.157] - - - [37248, 7296, 1, 384] - - [37, 90.367] - - - [40320, 1152, 1, 384] - - [44, 87.104] - - - [34560, 8448, 1, 384] - - [62, 90.284] - - - [38784, 3456, 1, 384] - - [53, 89.687] - - - [41472, 15360, 1, 384] - - [25, 90.467] - - - [41856, 3072, 1, 384] - - [29, 89.414] - - - [41088, 13441, 1, 384] - - [54, 87.66] - - - [39936, 1153, 1, 384] - - [27, 77.349] - - - [37248, 1920, 1, 384] - - [27, 88.403] - - - [39552, 7296, 1, 384] - - [36, 90.452] - - - [40320, 2304, 1, 384] - - [44, 89.112] - - - [34560, 2688, 1, 384] - - [53, 89.033] - - - [42240, 13056, 1, 384] - - [37, 90.791] - - - [40320, 9216, 1, 384] - - [36, 90.595] - - - [40704, 7297, 1, 384] - - [53, 88.281] - - - [43776, 13440, 1, 384] - - [28, 90.065] - - - [39936, 4992, 1, 384] - - [27, 90.119] - - - [42624, 13440, 1, 384] - - [22, 87.56] - - - [37632, 1153, 1, 384] - - [44, 78.129] - - - [33024, 3072, 1, 384] - - [29, 88.041] - - - [40704, 9216, 1, 384] - - [53, 90.812] - - - [42624, 1153, 1, 384] - - [25, 77.163] - - - [43392, 13440, 1, 384] - - [25, 90.96] - - - [36480, 3072, 1, 384] - - [37, 88.881] - - - [41088, 12288, 1, 384] - - [74, 89.769] - - - [39168, 1152, 1, 384] - - [25, 85.477] - - - [39936, 3072, 1, 384] - - [54, 87.229] - - - [35712, 13441, 1, 384] - - [25, 89.034] - - - [41088, 13440, 1, 384] - - [37, 90.307] - - - [43392, 3072, 1, 384] - - [27, 89.924] - - - [33792, 8064, 1, 384] - - [37, 90.4] - - - [32256, 13440, 1, 384] - - [27, 90.922] - - - [35328, 7297, 1, 384] - - [25, 88.144] - - - [40704, 11904, 1, 384] - - [25, 90.964] - - - [33024, 6912, 1, 384] - - [37, 89.944] - - - [38784, 15360, 1, 384] - - [25, 90.615] - - - [42240, 768, 1, 384] - - [53, 86.132] - - - [44160, 13440, 1, 384] - - [36, 91.059] - - - [39552, 7297, 1, 384] - - [25, 88.337] - - - [32640, 768, 1, 384] - - [76, 78.277] - - - [44160, 9216, 1, 384] - - [58, 90.934] - - - [32640, 6528, 1, 384] - - [29, 88.489] - - - [39552, 13441, 1, 384] - - [25, 88.971] - - - [31488, 13441, 1, 384] - - [29, 89.238] - - - [43008, 7296, 1, 384] - - [29, 90.267] - - - [41088, 5760, 1, 384] - - [54, 88.944] - - - [41472, 13440, 1, 384] - - [25, 90.964] - - - [43392, 7296, 1, 384] - - [25, 90.595] - - - [34944, 9216, 1, 384] - - [36, 90.258] - - - [43008, 1153, 1, 384] - - [25, 78.043] - - - [32640, 9216, 1, 384] - - [28, 88.653] - - - [36096, 13441, 1, 384] - - [32, 87.233] - - - [39552, 1152, 1, 384] - - [36, 86.13] - - - [37632, 7297, 1, 384] - - [27, 88.148] - - - [42624, 9216, 1, 384] - - [47, 86.819] - - - [43008, 8064, 1, 384] - - [27, 90.492] - - - [38784, 9600, 1, 384] - - [44, 90.879] - - - [37248, 8064, 1, 384] - - [37, 90.624] - - - [30720, 15360, 1, 384] - - [395, 91.732] - - - [38016, 13440, 1, 384] - - [29, 90.967] - - - [34944, 8832, 1, 384] - - [25, 90.609] - - - [37248, 13441, 1, 384] - - [29, 88.869] - - - [34560, 7296, 1, 384] - - [37, 90.444] - - - [44160, 3072, 1, 384] - - [29, 89.316] - - - [40320, 7296, 1, 384] - - [37, 90.462] - - - [34176, 2304, 1, 384] - - [29, 88.599] - - - [41088, 9216, 1, 384] - - [74, 89.557] - - - [34176, 1153, 1, 384] - - [25, 76.738] - - - [39552, 4224, 1, 384] - - [44, 90.185] - - - [38784, 13441, 1, 384] - - [25, 88.737] - - - [36480, 7297, 1, 384] - - [25, 87.916] - - - [32256, 3456, 1, 384] - - [27, 89.567] - - - [34176, 8064, 1, 384] - - [25, 90.295] - - - [36480, 15360, 1, 384] - - [58, 91.186] - - - [34560, 3072, 1, 384] - - [27, 89.269] - - - [35328, 7296, 1, 384] - - [29, 90.268] - - - [32256, 13441, 1, 384] - - [25, 89.398] - - - [38016, 1152, 1, 384] - - [58, 85.896] - - - [35328, 1153, 1, 384] - - [25, 76.639] - - - [23040, 7296, 1, 384] - - [25, 89.767] - - - [12672, 7296, 1, 384] - - [25, 88.855] - - - [4224, 4225, 1, 384] - - [27, 76.404] - - - [19968, 13440, 1, 384] - - [29, 90.89] - - - [16128, 3072, 1, 384] - - [25, 86.309] - - - [19968, 9216, 1, 384] - - [53, 89.746] - - - [24576, 13440, 1, 384] - - [38, 87.615] - - - [17280, 3072, 1, 384] - - [27, 87.102] - - - [16512, 9216, 1, 384] - - [61, 87.679] - - - [21120, 1536, 1, 384] - - [37, 85.882] - - - [18432, 13441, 1, 384] - - [29, 88.684] - - - [21120, 9216, 1, 384] - - [56, 90.106] - - - [27264, 3072, 1, 384] - - [25, 88.364] - - - [12288, 4608, 1, 384] - - [27, 86.933] - - - [22272, 5376, 1, 384] - - [44, 89.631] - - - [7296, 6912, 1, 384] - - [27, 88.378] - - - [26880, 9216, 1, 384] - - [44, 90.633] - - - [3072, 2688, 1, 384] - - [35, 72.028] - - - [16512, 2688, 1, 384] - - [84, 82.82] - - - [8064, 7680, 1, 384] - - [27, 87.628] - - - [22656, 1153, 1, 384] - - [44, 73.989] - - - [24960, 8064, 1, 384] - - [29, 90.562] - - - [23808, 9216, 1, 384] - - [53, 90.726] - - - [29568, 15360, 1, 384] - - [51, 90.385] - - - [1920, 1152, 1, 384] - - [462, 56.041] - - - [11136, 10752, 1, 384] - - [27, 89.659] - - - [25728, 1152, 1, 384] - - [27, 84.086] - - - [19584, 3072, 1, 384] - - [37, 87.287] - - - [3840, 1153, 1, 384] - - [55, 58.679] - - - [15360, 7296, 1, 384] - - [395, 91.611] - - - [13056, 12673, 1, 384] - - [29, 89.234] - - - [5376, 5377, 1, 384] - - [29, 82.103] - - - [28416, 13440, 1, 384] - - [36, 91.161] - - - [11904, 4224, 1, 384] - - [44, 88.137] - - - [24576, 10752, 1, 384] - - [54, 86.561] - - - [20352, 7297, 1, 384] - - [25, 88.385] - - - [16512, 7296, 1, 384] - - [41, 87.442] - - - [17280, 13441, 1, 384] - - [29, 89.218] - - - [24192, 10368, 1, 384] - - [44, 90.498] - - - [20352, 6528, 1, 384] - - [44, 90.017] - - - [1920, 1536, 1, 384] - - [35, 55.125] - - - [15744, 8064, 1, 384] - - [400, 92.375] - - - [13056, 3072, 1, 384] - - [27, 85.097] - - - [20352, 7296, 1, 384] - - [29, 89.816] - - - [10368, 1152, 1, 384] - - [60, 76.853] - - - [16128, 1152, 1, 384] - - [27, 79.04] - - - [13440, 7297, 1, 384] - - [27, 87.271] - - - [19200, 13441, 1, 384] - - [27, 89.246] - - - [13440, 13441, 1, 384] - - [58, 88.895] - - - [7680, 7297, 1, 384] - - [25, 86.162] - - - [27648, 14208, 1, 384] - - [25, 90.692] - - - [23424, 9216, 1, 384] - - [27, 89.747] - - - [24960, 1153, 1, 384] - - [36, 76.623] - - - [28032, 2304, 1, 384] - - [36, 88.371] - - - [30720, 3072, 1, 384] - - [29, 88.071] - - - [11904, 1152, 1, 384] - - [25, 77.734] - - - [24576, 3072, 1, 384] - - [29, 84.927] - - - [26112, 1153, 1, 384] - - [58, 75.83] - - - [10368, 10369, 1, 384] - - [25, 87.889] - - - [14976, 1536, 1, 384] - - [25, 82.728] - - - [11520, 7296, 1, 384] - - [36, 89.153] - - - [5376, 5376, 1, 384] - - [25, 82.183] - - - [28800, 7296, 1, 384] - - [44, 90.21] - - - [22656, 3072, 1, 384] - - [27, 87.879] - - - [11904, 7296, 1, 384] - - [27, 88.549] - - - [13824, 3072, 1, 384] - - [27, 85.697] - - - [21504, 13440, 1, 384] - - [37, 90.62] - - - [28800, 13440, 1, 384] - - [44, 90.974] - - - [13824, 7296, 1, 384] - - [25, 89.863] - - - [28416, 13441, 1, 384] - - [27, 89.206] - - - [20736, 7296, 1, 384] - - [37, 90.329] - - - [4992, 4608, 1, 384] - - [25, 84.18] - - - [21888, 1153, 1, 384] - - [57, 73.218] - - - [6912, 3072, 1, 384] - - [400, 84.104] - - - [7680, 7680, 1, 384] - - [25, 88.391] - - - [11904, 11905, 1, 384] - - [25, 88.706] - - - [9600, 1920, 1, 384] - - [27, 79.204] - - - [25728, 2688, 1, 384] - - [36, 87.846] - - - [29568, 3840, 1, 384] - - [24, 89.362] - - - [9984, 7297, 1, 384] - - [25, 86.231] - - - [13056, 2688, 1, 384] - - [53, 84.981] - - - [3456, 1920, 1, 384] - - [376, 75.644] - - - [19200, 1152, 1, 384] - - [37, 80.905] - - - [15744, 2304, 1, 384] - - [25, 84.376] - - - [17664, 7296, 1, 384] - - [53, 89.781] - - - [3072, 3072, 1, 384] - - [229, 71.128] - - - [21888, 7296, 1, 384] - - [64, 87.508] - - - [16128, 13440, 1, 384] - - [27, 90.31] - - - [23040, 1153, 1, 384] - - [58, 74.924] - - - [21504, 9216, 1, 384] - - [61, 88.698] - - - [21120, 4608, 1, 384] - - [56, 89.236] - - - [10368, 1153, 1, 384] - - [36, 68.413] - - - [29184, 13441, 1, 384] - - [47, 89.21] - - - [8832, 1536, 1, 384] - - [25, 77.661] - - - [30336, 3072, 1, 384] - - [27, 89.232] - - - [24192, 1153, 1, 384] - - [36, 74.563] - - - [16128, 2304, 1, 384] - - [25, 85.472] - - - [20736, 13440, 1, 384] - - [27, 90.841] - - - [24960, 7297, 1, 384] - - [25, 88.598] - - - [18048, 1536, 1, 384] - - [25, 82.772] - - - [19200, 5760, 1, 384] - - [25, 89.282] - - - [13440, 13056, 1, 384] - - [25, 90.019] - - - [6144, 1152, 1, 384] - - [34, 73.359] - - - [1920, 1920, 1, 384] - - [228, 59.067] - - - [18816, 5376, 1, 384] - - [36, 88.84] - - - [28800, 2688, 1, 384] - - [53, 88.798] - - - [20352, 3840, 1, 384] - - [27, 88.528] - - - [3840, 3841, 1, 384] - - [29, 76.356] - - - [17280, 768, 1, 384] - - [53, 76.137] - - - [21888, 2304, 1, 384] - - [54, 85.589] - - - [28416, 14592, 1, 384] - - [58, 90.909] - - - [18816, 3072, 1, 384] - - [25, 86.239] - - - [25344, 13440, 1, 384] - - [29, 91.119] - - - [20736, 6912, 1, 384] - - [53, 89.894] - - - [26880, 1152, 1, 384] - - [37, 83.551] - - - [29952, 3072, 1, 384] - - [37, 88.052] - - - [24960, 8448, 1, 384] - - [53, 90.766] - - - [15360, 8064, 1, 384] - - [29, 89.763] - - - [27648, 1920, 1, 384] - - [25, 87.035] - - - [3456, 2304, 1, 384] - - [55, 69.753] - - - [23040, 6528, 1, 384] - - [36, 89.69] - - - [14208, 1153, 1, 384] - - [36, 69.98] - - - [27648, 1153, 1, 384] - - [37, 76.656] - - - [1920, 1921, 1, 384] - - [40, 49.571] - - - [19584, 13441, 1, 384] - - [27, 89.274] - - - [8448, 3072, 1, 384] - - [29, 82.728] - - - [16512, 13441, 1, 384] - - [32, 86.662] - - - [4992, 768, 1, 384] - - [23, 52.095] - - - [28416, 14976, 1, 384] - - [58, 90.971] - - - [8448, 1152, 1, 384] - - [35, 72.41] - - - [20352, 9216, 1, 384] - - [58, 90.003] - - - [19584, 1153, 1, 384] - - [27, 76.048] - - - [20736, 768, 1, 384] - - [53, 81.601] - - - [28416, 2688, 1, 384] - - [53, 88.293] - - - [27264, 13440, 1, 384] - - [395, 91.225] - - - [16128, 7296, 1, 384] - - [29, 89.025] - - - [27648, 13440, 1, 384] - - [395, 91.636] - - - [26880, 13056, 1, 384] - - [394, 90.855] - - - [6528, 1920, 1, 384] - - [49, 80.083] - - - [20352, 13441, 1, 384] - - [25, 89.041] - - - [12288, 7297, 1, 384] - - [25, 87.098] - - - [21120, 7680, 1, 384] - - [25, 90.502] - - - [13824, 13441, 1, 384] - - [27, 88.762] - - - [26112, 13440, 1, 384] - - [37, 90.947] - - - [16512, 7297, 1, 384] - - [32, 84.44] - - - [6144, 5761, 1, 384] - - [25, 82.18] - - - [24960, 1152, 1, 384] - - [25, 85.839] - - - [9600, 9216, 1, 384] - - [36, 88.337] - - - [22272, 1153, 1, 384] - - [58, 76.45] - - - [24960, 2304, 1, 384] - - [37, 88.378] - - - [11136, 7296, 1, 384] - - [27, 88.208] - - - [28800, 3072, 1, 384] - - [27, 87.63] - - - [6912, 2688, 1, 384] - - [27, 79.957] - - - [25728, 3072, 1, 384] - - [29, 88.902] - - - [15744, 13441, 1, 384] - - [27, 89.503] - - - [18816, 7296, 1, 384] - - [36, 89.387] - - - [18816, 7297, 1, 384] - - [25, 87.989] - - - [13440, 13440, 1, 384] - - [25, 90.071] - - - [29184, 3456, 1, 384] - - [25, 89.49] - - - [8064, 768, 1, 384] - - [35, 63.254] - - - [4992, 4609, 1, 384] - - [25, 78.626] - - - [26496, 13056, 1, 384] - - [395, 91.047] - - - [21504, 4608, 1, 384] - - [29, 88.464] - - - [18048, 9216, 1, 384] - - [407, 91.513] - - - [14592, 13441, 1, 384] - - [37, 89.353] - - - [22656, 1152, 1, 384] - - [44, 83.362] - - - [14976, 3072, 1, 384] - - [29, 86.849] - - - [24960, 13441, 1, 384] - - [36, 89.752] - - - [768, 768, 1, 384] - - [108, 32.393] - - - [12672, 4992, 1, 384] - - [25, 87.364] - - - [11136, 3072, 1, 384] - - [27, 86.907] - - - [19584, 1152, 1, 384] - - [27, 82.418] - - - [16896, 3456, 1, 384] - - [29, 87.275] - - - [23040, 1152, 1, 384] - - [27, 84.389] - - - [6528, 6528, 1, 384] - - [27, 86.8] - - - [25344, 3072, 1, 384] - - [37, 87.695] - - - [2688, 1536, 1, 384] - - [467, 66.899] - - - [5760, 1536, 1, 384] - - [35, 76.188] - - - [6144, 5760, 1, 384] - - [27, 85.977] - - - [21504, 8064, 1, 384] - - [37, 89.897] - - - [12288, 12288, 1, 384] - - [29, 88.86] - - - [16128, 13441, 1, 384] - - [27, 89.135] - - - [25344, 8448, 1, 384] - - [29, 90.218] - - - [23808, 7297, 1, 384] - - [58, 88.195] - - - [15744, 7296, 1, 384] - - [25, 89.763] - - - [16896, 13441, 1, 384] - - [25, 89.152] - - - [15360, 1920, 1, 384] - - [25, 83.634] - - - [21504, 1152, 1, 384] - - [29, 84.059] - - - [6912, 1152, 1, 384] - - [34, 69.651] - - - [16512, 3072, 1, 384] - - [32, 82.448] - - - [28800, 1153, 1, 384] - - [53, 76.181] - - - [21888, 8064, 1, 384] - - [50, 87.855] - - - [20736, 7297, 1, 384] - - [27, 88.128] - - - [10752, 10753, 1, 384] - - [29, 88.712] - - - [8832, 7297, 1, 384] - - [36, 86.031] - - - [28032, 7297, 1, 384] - - [29, 88.513] - - - [23424, 9600, 1, 384] - - [25, 90.657] - - - [23040, 13440, 1, 384] - - [37, 90.83] - - - [26880, 13441, 1, 384] - - [441, 89.574] - - - [4224, 4224, 1, 384] - - [35, 82.929] - - - [9600, 9600, 1, 384] - - [25, 88.649] - - - [26112, 1152, 1, 384] - - [58, 85.034] - - - [29568, 3456, 1, 384] - - [53, 88.517] - - - [28032, 9216, 1, 384] - - [53, 90.625] - - - [27648, 9216, 1, 384] - - [61, 88.785] - - - [17664, 1153, 1, 384] - - [53, 73.888] - - - [12672, 12289, 1, 384] - - [53, 87.114] - - - [21888, 1152, 1, 384] - - [29, 82.958] - - - [21888, 9216, 1, 384] - - [28, 87.296] - - - [10752, 10369, 1, 384] - - [37, 88.454] - - - [22656, 7296, 1, 384] - - [29, 90.219] - - - [13440, 13057, 1, 384] - - [44, 88.973] - - - [10752, 1153, 1, 384] - - [34, 70.09] - - - [12672, 3072, 1, 384] - - [25, 86.363] - - - [23424, 13440, 1, 384] - - [37, 90.874] - - - [29952, 3840, 1, 384] - - [53, 89.78] - - - [18432, 1920, 1, 384] - - [29, 85.746] - - - [26112, 7297, 1, 384] - - [29, 88.499] - - - [18816, 1153, 1, 384] - - [53, 73.484] - - - [17664, 4224, 1, 384] - - [27, 88.407] - - - [11520, 11521, 1, 384] - - [29, 88.63] - - - [30720, 1920, 1, 384] - - [25, 87.964] - - - [15360, 13441, 1, 384] - - [25, 89.017] - - - [17664, 13441, 1, 384] - - [53, 89.74] - - - [26496, 3072, 1, 384] - - [27, 87.645] - - - [20736, 4224, 1, 384] - - [58, 89.19] - - - [18816, 13441, 1, 384] - - [44, 89.24] - - - [18048, 13441, 1, 384] - - [36, 89.388] - - - [20352, 3072, 1, 384] - - [27, 88.042] - - - [1152, 768, 1, 384] - - [331, 43.744] - - - [16896, 7296, 1, 384] - - [25, 89.528] - - - [28800, 9216, 1, 384] - - [58, 90.373] - - - [9600, 1152, 1, 384] - - [53, 72.126] - - - [29952, 1153, 1, 384] - - [37, 76.04] - - - [20736, 1153, 1, 384] - - [36, 75.779] - - - [19584, 5760, 1, 384] - - [37, 89.61] - - - [29568, 7296, 1, 384] - - [53, 89.833] - - - [7296, 3072, 1, 384] - - [27, 82.095] - - - [27264, 1152, 1, 384] - - [25, 83.817] - - - [12288, 4992, 1, 384] - - [37, 86.87] - - - [5760, 5376, 1, 384] - - [37, 83.708] - - - [30720, 1152, 1, 384] - - [29, 85.402] - - - [14208, 13441, 1, 384] - - [44, 89.455] - - - [21504, 7296, 1, 384] - - [27, 89.468] - - - [7296, 6913, 1, 384] - - [27, 85.463] - - - [23808, 6912, 1, 384] - - [53, 90.007] - - - [20352, 768, 1, 384] - - [53, 79.924] - - - [2688, 2688, 1, 384] - - [228, 65.669] - - - [13056, 12672, 1, 384] - - [29, 90.482] - - - [29568, 13440, 1, 384] - - [36, 90.69] - - - [11904, 1153, 1, 384] - - [53, 70.203] - - - [2688, 2689, 1, 384] - - [49, 63.266] - - - [9984, 9985, 1, 384] - - [27, 87.492] - - - [22272, 13440, 1, 384] - - [27, 90.599] - - - [30336, 15360, 1, 384] - - [395, 91.719] - - - [21504, 7680, 1, 384] - - [29, 90.021] - - - [24192, 13441, 1, 384] - - [37, 89.274] - - - [15360, 1536, 1, 384] - - [27, 80.28] - - - [24576, 7297, 1, 384] - - [37, 83.448] - - - [11136, 3456, 1, 384] - - [27, 85.406] - - - [9600, 1153, 1, 384] - - [40, 70.514] - - - [18048, 7297, 1, 384] - - [29, 87.754] - - - [6144, 1153, 1, 384] - - [55, 61.339] - - - [23040, 9600, 1, 384] - - [25, 90.44] - - - [26880, 1153, 1, 384] - - [58, 75.08] - - - [10752, 7297, 1, 384] - - [25, 86.643] - - - [6912, 6529, 1, 384] - - [25, 85.275] - - - [29184, 9216, 1, 384] - - [56, 89.646] - - - [20736, 9216, 1, 384] - - [58, 90.051] - - - [23808, 1152, 1, 384] - - [44, 82.731] - - - [11136, 1153, 1, 384] - - [40, 72.462] - - - [25344, 1152, 1, 384] - - [25, 83.112] - - - [25344, 13441, 1, 384] - - [24, 89.36] - - - [14976, 7296, 1, 384] - - [395, 91.797] - - - [14592, 13440, 1, 384] - - [25, 90.678] - - - [7680, 7681, 1, 384] - - [25, 85.928] - - - [29568, 768, 1, 384] - - [72, 81.546] - - - [5760, 1152, 1, 384] - - [25, 68.56] - - - [21888, 13441, 1, 384] - - [28, 84.654] - - - [17664, 768, 1, 384] - - [53, 77.465] - - - [25728, 11904, 1, 384] - - [44, 90.922] - - - [9984, 2688, 1, 384] - - [36, 85.359] - - - [28416, 1153, 1, 384] - - [25, 75.623] - - - [17664, 3072, 1, 384] - - [25, 86.289] - - - [23040, 7297, 1, 384] - - [27, 88.179] - - - [8448, 8448, 1, 384] - - [53, 88.552] - - - [4608, 4225, 1, 384] - - [25, 76.866] - - - [4224, 2688, 1, 384] - - [49, 74.08] - - - [3072, 1152, 1, 384] - - [467, 68.888] - - - [29184, 1152, 1, 384] - - [25, 85.452] - - - [13440, 3072, 1, 384] - - [37, 87.428] - - - [6912, 6913, 1, 384] - - [37, 86.541] - - - [18432, 13440, 1, 384] - - [37, 90.518] - - - [14208, 7296, 1, 384] - - [395, 91.627] - - - [5376, 768, 1, 384] - - [35, 55.528] - - - [29184, 7296, 1, 384] - - [25, 90.039] - - - [20352, 1152, 1, 384] - - [27, 80.551] - - - [2304, 1153, 1, 384] - - [492, 54.02] - - - [23808, 9984, 1, 384] - - [25, 90.845] - - - [8448, 8065, 1, 384] - - [27, 86.48] - - - [24576, 1152, 1, 384] - - [25, 82.609] - - - [1536, 1537, 1, 384] - - [59, 44.725] - - - [4224, 3072, 1, 384] - - [27, 74.892] - - - [19968, 7296, 1, 384] - - [25, 90.127] - - - [19200, 5376, 1, 384] - - [36, 88.93] - - - [4608, 1152, 1, 384] - - [35, 69.296] - - - [18432, 4992, 1, 384] - - [25, 88.376] - - - [26880, 7297, 1, 384] - - [58, 88.499] - - - [15744, 3072, 1, 384] - - [27, 87.77] - - - [22272, 7296, 1, 384] - - [25, 89.771] - - - [20352, 6912, 1, 384] - - [27, 89.537] - - - [26880, 13440, 1, 384] - - [395, 90.962] - - - [4224, 3840, 1, 384] - - [25, 82.671] - - - [23424, 13441, 1, 384] - - [47, 88.888] - - - [16512, 13440, 1, 384] - - [32, 89.118] - - - [21120, 1152, 1, 384] - - [27, 82.711] - - - [10368, 3072, 1, 384] - - [27, 85.551] - - - [28032, 13440, 1, 384] - - [58, 90.785] - - - [14208, 6528, 1, 384] - - [36, 89.104] - - - [768, 769, 1, 384] - - [335, 31.74] - - - [3456, 1152, 1, 384] - - [34, 54.395] - - - [12672, 1152, 1, 384] - - [35, 75.651] - - - [7680, 3072, 1, 384] - - [25, 80.713] - - - [19200, 2304, 1, 384] - - [53, 86.548] - - - [13056, 1153, 1, 384] - - [36, 69.844] - - - [27264, 1153, 1, 384] - - [27, 75.153] - - - [29568, 1153, 1, 384] - - [58, 77.447] - - - [11520, 11136, 1, 384] - - [53, 89.532] - - - [9216, 9216, 1, 384] - - [25, 87.436] - - - [18048, 1153, 1, 384] - - [58, 75.368] - - - [8064, 1152, 1, 384] - - [25, 68.874] - - - [22272, 7297, 1, 384] - - [44, 88.518] - - - [22272, 13441, 1, 384] - - [58, 89.718] - - - [22656, 2688, 1, 384] - - [53, 88.542] - - - [19584, 6144, 1, 384] - - [58, 89.72] - - - [8064, 7297, 1, 384] - - [27, 85.223] - - - [8064, 7681, 1, 384] - - [25, 86.659] - - - [23808, 7296, 1, 384] - - [29, 90.104] - - - [24960, 7296, 1, 384] - - [27, 89.816] - - - [14208, 6912, 1, 384] - - [399, 91.718] - - - [19968, 6528, 1, 384] - - [44, 89.457] - - - [28416, 7296, 1, 384] - - [44, 90.064] - - - [29952, 13440, 1, 384] - - [37, 90.917] - - - [17280, 7297, 1, 384] - - [29, 87.711] - - - [1536, 1152, 1, 384] - - [182, 59.095] - - - [8832, 1153, 1, 384] - - [39, 65.553] - - - [28032, 1153, 1, 384] - - [27, 77.624] - - - [2688, 2305, 1, 384] - - [34, 64.521] - - - [8064, 3072, 1, 384] - - [27, 84.009] - - - [28032, 3072, 1, 384] - - [37, 89.145] - - - [3840, 3456, 1, 384] - - [35, 77.031] - - - [21888, 1920, 1, 384] - - [38, 83.085] - - - [11904, 11520, 1, 384] - - [25, 89.564] - - - [9600, 9601, 1, 384] - - [58, 88.003] - - - [21120, 13440, 1, 384] - - [29, 90.747] - - - [19584, 2688, 1, 384] - - [36, 86.563] - - - [6912, 6528, 1, 384] - - [36, 85.626] - - - [29568, 1152, 1, 384] - - [54, 85.277] - - - [23808, 3072, 1, 384] - - [37, 88.181] - - - [18816, 4992, 1, 384] - - [53, 88.521] - - - [29952, 9216, 1, 384] - - [58, 90.581] - - - [22656, 13440, 1, 384] - - [27, 90.994] - - - [20352, 3456, 1, 384] - - [53, 87.447] - - - [3456, 1153, 1, 384] - - [35, 53.265] - - - [3840, 3457, 1, 384] - - [55, 76.26] - - - [15744, 8448, 1, 384] - - [53, 90.086] - - - [26112, 3072, 1, 384] - - [29, 88.29] - - - [28032, 14208, 1, 384] - - [441, 91.378] - - - [21504, 1536, 1, 384] - - [37, 83.608] - - - [11520, 768, 1, 384] - - [53, 75.835] - - - [6528, 6144, 1, 384] - - [27, 85.453] - - - [18432, 1153, 1, 384] - - [27, 72.221] - - - [3072, 1920, 1, 384] - - [55, 62.213] - - - [25344, 9216, 1, 384] - - [58, 90.665] - - - [30336, 7297, 1, 384] - - [37, 87.864] - - - [8832, 1152, 1, 384] - - [49, 75.114] - - - [26112, 9216, 1, 384] - - [58, 90.379] - - - [29952, 7296, 1, 384] - - [27, 90.178] - - - [11520, 11137, 1, 384] - - [29, 89.105] - - - [16896, 13440, 1, 384] - - [27, 90.773] - - - [29568, 13441, 1, 384] - - [51, 88.47] - - - [30336, 9216, 1, 384] - - [53, 90.28] - - - [2688, 1152, 1, 384] - - [467, 62.165] - - - [10368, 10368, 1, 384] - - [58, 89.575] - - - [25344, 11520, 1, 384] - - [27, 91.069] - - - [24576, 1920, 1, 384] - - [29, 83.311] - - - [11904, 4608, 1, 384] - - [53, 87.354] - - - [12672, 5376, 1, 384] - - [53, 88.794] - - - [11520, 3072, 1, 384] - - [25, 85.887] - - - [3072, 3073, 1, 384] - - [27, 69.368] - - - [24960, 11136, 1, 384] - - [53, 90.638] - - - [9984, 9600, 1, 384] - - [25, 88.848] - - - [19200, 2688, 1, 384] - - [53, 87.33] - - - [26496, 7296, 1, 384] - - [25, 90.121] - - - [23040, 3072, 1, 384] - - [37, 87.388] - - - [5760, 5761, 1, 384] - - [27, 84.571] - - - [5760, 5377, 1, 384] - - [29, 83.087] - - - [26880, 768, 1, 384] - - [53, 81.248] - - - [13824, 7297, 1, 384] - - [29, 88.168] - - - [13440, 7296, 1, 384] - - [25, 89.037] - - - [16128, 8448, 1, 384] - - [36, 89.625] - - - [24960, 3072, 1, 384] - - [27, 88.551] - - - [6144, 6144, 1, 384] - - [27, 86.096] - - - [27648, 13441, 1, 384] - - [395, 90.086] - - - [10368, 7297, 1, 384] - - [36, 87.132] - - - [22272, 2304, 1, 384] - - [25, 87.15] - - - [30720, 1153, 1, 384] - - [37, 77.37] - - - [24192, 13440, 1, 384] - - [27, 90.816] - - - [9984, 9984, 1, 384] - - [27, 89.105] - - - [29952, 1152, 1, 384] - - [25, 84.039] - - - [26112, 12672, 1, 384] - - [395, 91.15] - - - [8448, 7296, 1, 384] - - [27, 87.405] - - - [19584, 13440, 1, 384] - - [37, 90.462] - - - [21120, 1153, 1, 384] - - [37, 76.613] - - - [8832, 8449, 1, 384] - - [53, 86.262] - - - [28032, 13441, 1, 384] - - [27, 89.389] - - - [7680, 1153, 1, 384] - - [23, 65.5] - - - [19584, 9216, 1, 384] - - [53, 89.89] - - - [28800, 1152, 1, 384] - - [58, 84.104] - - - [29952, 768, 1, 384] - - [53, 83.522] - - - [12288, 1152, 1, 384] - - [27, 80.098] - - - [9600, 9217, 1, 384] - - [62, 86.678] - - - [14976, 13441, 1, 384] - - [29, 89.511] - - - [25344, 8832, 1, 384] - - [27, 90.401] - - - [18432, 4608, 1, 384] - - [25, 87.45] - - - [2304, 1920, 1, 384] - - [49, 59.968] - - - [11520, 4224, 1, 384] - - [27, 88.351] - - - [26496, 1153, 1, 384] - - [53, 77.269] - - - [28416, 2304, 1, 384] - - [53, 87.819] - - - [19200, 3072, 1, 384] - - [37, 87.757] - - - [26112, 7296, 1, 384] - - [29, 90.453] - - - [21504, 7297, 1, 384] - - [25, 87.803] - - - [4224, 1152, 1, 384] - - [60, 64.294] - - - [17664, 3840, 1, 384] - - [58, 88.524] - - - [6144, 1536, 1, 384] - - [49, 70.378] - - - [28032, 14592, 1, 384] - - [36, 91.167] - - - [8064, 8064, 1, 384] - - [27, 87.259] - - - [11136, 1152, 1, 384] - - [27, 74.068] - - - [13056, 7297, 1, 384] - - [25, 87.919] - - - [19968, 3456, 1, 384] - - [53, 87.779] - - - [25344, 7297, 1, 384] - - [29, 88.445] - - - [17280, 3840, 1, 384] - - [27, 88.77] - - - [28416, 1152, 1, 384] - - [27, 83.942] - - - [21120, 3072, 1, 384] - - [25, 88.775] - - - [28416, 7297, 1, 384] - - [37, 88.08] - - - [6528, 6529, 1, 384] - - [27, 83.475] - - - [26496, 9216, 1, 384] - - [36, 90.68] - - - [14592, 7296, 1, 384] - - [394, 91.305] - - - [14208, 1152, 1, 384] - - [27, 76.797] - - - [24576, 1536, 1, 384] - - [37, 83.575] - - - [18048, 7296, 1, 384] - - [37, 89.317] - - - [4608, 3072, 1, 384] - - [35, 80.893] - - - [28800, 14976, 1, 384] - - [44, 90.941] - - - [17664, 1152, 1, 384] - - [36, 80.529] - - - [24576, 7680, 1, 384] - - [37, 86.674] - - - [16896, 9216, 1, 384] - - [44, 89.757] - - - [20736, 3840, 1, 384] - - [44, 88.309] - - - [27264, 9216, 1, 384] - - [25, 89.624] - - - [21888, 3072, 1, 384] - - [30, 84.822] - - - [24576, 11136, 1, 384] - - [22, 87.405] - - - [14592, 1153, 1, 384] - - [36, 71.604] - - - [23424, 7296, 1, 384] - - [25, 90.291] - - - [22272, 3072, 1, 384] - - [27, 88.543] - - - [8832, 8832, 1, 384] - - [37, 88.346] - - - [8064, 7296, 1, 384] - - [27, 88.017] - - - [22656, 8832, 1, 384] - - [58, 90.025] - - - [22272, 2688, 1, 384] - - [44, 87.261] - - - [6528, 1152, 1, 384] - - [49, 65.966] - - - [8832, 8833, 1, 384] - - [36, 87.997] - - - [28800, 15360, 1, 384] - - [36, 90.849] - - - [23424, 1153, 1, 384] - - [27, 75.606] - - - [13440, 1152, 1, 384] - - [53, 79.284] - - - [10752, 10368, 1, 384] - - [36, 89.934] - - - [3456, 3456, 1, 384] - - [55, 77.322] - - - [4608, 4608, 1, 384] - - [25, 83.853] - - - [4224, 1153, 1, 384] - - [35, 63.2] - - - [12672, 2304, 1, 384] - - [25, 83.227] - - - [25728, 7297, 1, 384] - - [58, 88.005] - - - [5376, 1153, 1, 384] - - [34, 63.522] - - - [30720, 4992, 1, 384] - - [25, 89.461] - - - [27264, 7297, 1, 384] - - [29, 87.552] - - - [21504, 1920, 1, 384] - - [27, 87.192] - - - [11136, 11136, 1, 384] - - [58, 90.048] - - - [22656, 6144, 1, 384] - - [58, 89.491] - - - [26496, 13440, 1, 384] - - [395, 91.13] - - - [9216, 7296, 1, 384] - - [29, 87.749] - - - [17280, 7296, 1, 384] - - [37, 90.133] - - - [23040, 13441, 1, 384] - - [29, 89.079] - - - [23808, 13441, 1, 384] - - [36, 89.697] - - - [30336, 4224, 1, 384] - - [29, 89.287] - - - [6144, 1920, 1, 384] - - [55, 76.159] - - - [11904, 11904, 1, 384] - - [58, 90.052] - - - [30336, 13441, 1, 384] - - [25, 89.106] - - - [11904, 1536, 1, 384] - - [37, 78.25] - - - [24576, 9216, 1, 384] - - [32, 84.679] - - - [9984, 2304, 1, 384] - - [58, 83.947] - - - [18048, 4608, 1, 384] - - [53, 88.225] - - - [18432, 7297, 1, 384] - - [29, 87.867] - - - [11136, 3840, 1, 384] - - [27, 87.16] - - - [12288, 11904, 1, 384] - - [37, 89.561] - - - [19584, 7296, 1, 384] - - [27, 89.672] - - - [3072, 2689, 1, 384] - - [34, 71.091] - - - [2304, 2305, 1, 384] - - [55, 56.221] - - - [26496, 7297, 1, 384] - - [29, 88.044] - - - [15744, 1152, 1, 384] - - [25, 77.976] - - - [6912, 6912, 1, 384] - - [29, 87.066] - - - [4992, 3072, 1, 384] - - [27, 79.07] - - - [15744, 13440, 1, 384] - - [29, 90.395] - - - [2688, 2304, 1, 384] - - [34, 65.324] - - - [8448, 7297, 1, 384] - - [27, 86.977] - - - [25344, 11904, 1, 384] - - [27, 90.852] - - - [18432, 7296, 1, 384] - - [27, 89.726] - - - [8448, 8449, 1, 384] - - [29, 86.577] - - - [30720, 1536, 1, 384] - - [27, 85.263] - - - [9216, 1153, 1, 384] - - [44, 68.145] - - - [24192, 9216, 1, 384] - - [58, 90.538] - - - [25344, 2688, 1, 384] - - [58, 88.692] - - - [24576, 1153, 1, 384] - - [29, 73.377] - - - [14208, 7297, 1, 384] - - [53, 87.622] - - - [12672, 1920, 1, 384] - - [37, 82.978] - - - [4608, 4224, 1, 384] - - [35, 83.371] - - - [27264, 1536, 1, 384] - - [27, 84.468] - - - [24576, 13441, 1, 384] - - [25, 84.434] - - - [21504, 4992, 1, 384] - - [27, 89.361] - - - [21888, 4992, 1, 384] - - [50, 87.042] - - - [18432, 3072, 1, 384] - - [37, 86.579] - - - [19968, 6144, 1, 384] - - [27, 88.894] - - - [24192, 1536, 1, 384] - - [27, 85.442] - - - [9600, 7297, 1, 384] - - [27, 86.663] - - - [13824, 6528, 1, 384] - - [400, 90.979] - - - [2304, 2304, 1, 384] - - [49, 69.873] - - - [23424, 9984, 1, 384] - - [29, 90.684] - - - [18816, 1152, 1, 384] - - [27, 79.623] - - - [1152, 769, 1, 384] - - [171, 35.891] - - - [23424, 768, 1, 384] - - [53, 83.094] - - - [17280, 1153, 1, 384] - - [36, 72.434] - - - [9600, 2304, 1, 384] - - [27, 81.095] - - - [29184, 7297, 1, 384] - - [29, 88.119] - - - [26880, 3072, 1, 384] - - [29, 88.803] - - - [11520, 11520, 1, 384] - - [25, 90.049] - - - [23040, 6144, 1, 384] - - [58, 89.401] - - - [18048, 13440, 1, 384] - - [58, 90.675] - - - [30336, 1536, 1, 384] - - [27, 86.95] - - - [14976, 7680, 1, 384] - - [399, 92.364] - - - [14976, 1152, 1, 384] - - [25, 80.267] - - - [15360, 7680, 1, 384] - - [395, 92.178] - - - [28800, 13441, 1, 384] - - [25, 88.768] - - - [28032, 1920, 1, 384] - - [25, 88.185] - - - [16128, 2688, 1, 384] - - [58, 84.65] - - - [6144, 6145, 1, 384] - - [25, 81.33] - - - [10368, 7296, 1, 384] - - [27, 89.3] - - - [5760, 3072, 1, 384] - - [498, 82.938] - - - [24960, 9216, 1, 384] - - [36, 90.634] - - - [14592, 768, 1, 384] - - [60, 72.813] - - - [14208, 768, 1, 384] - - [40, 70.96] - - - [6912, 1153, 1, 384] - - [40, 68.662] - - - [21888, 13440, 1, 384] - - [28, 88.568] - - - [13056, 5760, 1, 384] - - [27, 89.082] - - - [12288, 1920, 1, 384] - - [27, 80.703] - - - [13056, 13056, 1, 384] - - [27, 90.324] - - - [6528, 1153, 1, 384] - - [34, 64.455] - - - [22272, 8448, 1, 384] - - [58, 90.311] - - - [7296, 1153, 1, 384] - - [60, 62.16] - - - [17280, 3456, 1, 384] - - [36, 87.076] - - - [27264, 13441, 1, 384] - - [395, 89.707] - - - [9216, 7297, 1, 384] - - [27, 87.23] - - - [4992, 4992, 1, 384] - - [37, 85.007] - - - [16128, 7297, 1, 384] - - [27, 87.628] - - - [20352, 13440, 1, 384] - - [25, 90.851] - - - [30336, 1153, 1, 384] - - [36, 76.888] - - - [13056, 7296, 1, 384] - - [29, 89.592] - - - [27648, 1152, 1, 384] - - [29, 85.05] - - - [13824, 6144, 1, 384] - - [25, 88.162] - - - [9216, 1920, 1, 384] - - [27, 81.957] - - - [17280, 13440, 1, 384] - - [27, 90.76] - - - [21888, 5376, 1, 384] - - [30, 85.822] - - - [3456, 3072, 1, 384] - - [34, 78.283] - - - [13440, 1153, 1, 384] - - [36, 71.698] - - - [24192, 7680, 1, 384] - - [25, 90.064] - - - [29952, 4224, 1, 384] - - [27, 89.448] - - - [8832, 3072, 1, 384] - - [29, 81.695] - - - [5760, 5760, 1, 384] - - [25, 85.038] - - - [23424, 6912, 1, 384] - - [27, 90.398] - - - [24192, 3072, 1, 384] - - [25, 87.871] - - - [18048, 3072, 1, 384] - - [25, 87.997] - - - [27264, 7296, 1, 384] - - [29, 90.021] - - - [11520, 3840, 1, 384] - - [29, 86.779] - - - [18432, 1536, 1, 384] - - [25, 84.098] - - - [11136, 10753, 1, 384] - - [27, 87.819] - - - [9600, 7296, 1, 384] - - [27, 89.09] - - - [26496, 13441, 1, 384] - - [27, 89.289] - - - [29568, 9216, 1, 384] - - [36, 89.74] - - - [25728, 7296, 1, 384] - - [27, 89.917] - - - [6528, 3072, 1, 384] - - [464, 82.255] - - - [18816, 9216, 1, 384] - - [407, 91.774] - - - [1920, 1153, 1, 384] - - [34, 42.154] - - - [1152, 1153, 1, 384] - - [181, 46.328] - - - [16896, 1153, 1, 384] - - [44, 75.524] - - - [4992, 1153, 1, 384] - - [60, 60.197] - - - [22656, 13441, 1, 384] - - [58, 89.502] - - - [9984, 1152, 1, 384] - - [34, 74.729] - - - [26496, 768, 1, 384] - - [36, 80.366] - - - [25344, 2304, 1, 384] - - [27, 87.575] - - - [14592, 6912, 1, 384] - - [400, 91.959] - - - [9216, 8833, 1, 384] - - [29, 87.479] - - - [19584, 7297, 1, 384] - - [25, 88.16] - - - [8448, 1153, 1, 384] - - [35, 71.078] - - - [21120, 7297, 1, 384] - - [47, 88.024] - - - [11520, 7297, 1, 384] - - [29, 87.053] - - - [12288, 7296, 1, 384] - - [29, 89.211] - - - [4224, 3841, 1, 384] - - [49, 76.561] - - - [9984, 9601, 1, 384] - - [27, 88.172] - - - [2304, 1152, 1, 384] - - [491, 63.971] - - - [21120, 7296, 1, 384] - - [27, 89.903] - - - [15360, 1153, 1, 384] - - [53, 74.403] - - - [27648, 3072, 1, 384] - - [29, 86.696] - - - [19200, 1153, 1, 384] - - [27, 74.708] - - - [28032, 1152, 1, 384] - - [58, 86.259] - - - [12672, 12288, 1, 384] - - [53, 90.041] - - - [22272, 5760, 1, 384] - - [25, 89.478] - - - [26496, 1152, 1, 384] - - [36, 86.187] - - - [26880, 7296, 1, 384] - - [27, 90.428] - - - [6528, 2304, 1, 384] - - [34, 77.781] - - - [9984, 7296, 1, 384] - - [27, 88.426] - - - [19968, 1152, 1, 384] - - [58, 83.787] - - - [10368, 9984, 1, 384] - - [29, 89.268] - - - [3840, 3840, 1, 384] - - [35, 76.876] - - - [5376, 1152, 1, 384] - - [35, 64.654] - - - [24192, 7296, 1, 384] - - [25, 90.513] - - - [14592, 3072, 1, 384] - - [29, 87.402] - - - [27648, 7297, 1, 384] - - [27, 87.86] - - - [23424, 1152, 1, 384] - - [27, 85.461] - - - [3456, 3457, 1, 384] - - [34, 76.449] - - - [13056, 2304, 1, 384] - - [29, 85.22] - - - [23808, 768, 1, 384] - - [36, 78.629] - - - [18048, 1152, 1, 384] - - [27, 81.926] - - - [28416, 9216, 1, 384] - - [53, 90.634] - - - [21888, 7297, 1, 384] - - [67, 83.447] - - - [25728, 12288, 1, 384] - - [58, 90.662] - - - [21120, 4224, 1, 384] - - [27, 88.889] - - - [20736, 3072, 1, 384] - - [29, 87.451] - - - [3840, 2688, 1, 384] - - [55, 76.533] - - - [29568, 7297, 1, 384] - - [47, 87.324] - - - [13824, 1153, 1, 384] - - [36, 73.584] - - - [15744, 1153, 1, 384] - - [53, 71.28] - - - [11136, 768, 1, 384] - - [40, 73.494] - - - [17664, 7297, 1, 384] - - [29, 88.339] - - - [24192, 7297, 1, 384] - - [29, 88.31] - - - [25344, 1153, 1, 384] - - [36, 77.647] - - - [30720, 4608, 1, 384] - - [25, 88.764] - - - [25728, 9216, 1, 384] - - [53, 90.516] - - - [29184, 1153, 1, 384] - - [44, 77.332] - - - [30336, 1152, 1, 384] - - [44, 84.968] - - - [24960, 13440, 1, 384] - - [44, 90.738] - - - [18432, 9216, 1, 384] - - [407, 91.744] - - - [15360, 13440, 1, 384] - - [27, 90.359] - - - [12288, 1536, 1, 384] - - [25, 80.308] - - - [8832, 8448, 1, 384] - - [27, 88.439] - - - [19968, 7297, 1, 384] - - [37, 88.003] - - - [19968, 3072, 1, 384] - - [37, 86.678] - - - [24960, 1920, 1, 384] - - [27, 87.186] - - - [15360, 1152, 1, 384] - - [27, 81.733] - - - [30720, 7296, 1, 384] - - [27, 90.195] - - - [14976, 1153, 1, 384] - - [53, 73.21] - - - [25344, 7296, 1, 384] - - [27, 90.358] - - - [16512, 8832, 1, 384] - - [41, 87.877] - - - [26112, 13441, 1, 384] - - [29, 89.566] - - - [22272, 1152, 1, 384] - - [29, 82.343] - - - [27648, 1536, 1, 384] - - [27, 85.661] - - - [15744, 1920, 1, 384] - - [25, 85.498] - - - [5760, 1153, 1, 384] - - [60, 57.969] - - - [29952, 13441, 1, 384] - - [27, 89.176] - - - [12672, 1153, 1, 384] - - [44, 74.098] - - - [13440, 2688, 1, 384] - - [58, 83.727] - - - [18816, 13440, 1, 384] - - [44, 90.572] - - - [22656, 9216, 1, 384] - - [36, 90.68] - - - [9216, 1152, 1, 384] - - [35, 77.636] - - - [20736, 1152, 1, 384] - - [25, 81.729] - - - [8832, 7296, 1, 384] - - [27, 88.738] - - - [15744, 7297, 1, 384] - - [25, 88.247] - - - [16512, 1153, 1, 384] - - [78, 70.083] - - - [29952, 7297, 1, 384] - - [37, 88.269] - - - [11136, 7297, 1, 384] - - [27, 87.578] - - - [9600, 3072, 1, 384] - - [25, 83.769] - - - [28800, 7297, 1, 384] - - [25, 87.828] - - - [27648, 13824, 1, 384] - - [395, 91.62] - - - [23808, 10368, 1, 384] - - [53, 90.899] - - - [13824, 13440, 1, 384] - - [27, 90.055] - - - [9216, 1536, 1, 384] - - [25, 80.124] - - - [23808, 1153, 1, 384] - - [44, 76.937] - - - [15360, 3072, 1, 384] - - [27, 85.325] - - - [12288, 3072, 1, 384] - - [27, 86.147] - - - [28416, 3072, 1, 384] - - [27, 88.406] - - - [30336, 13440, 1, 384] - - [25, 90.835] - - - [1152, 1152, 1, 384] - - [172, 53.392] - - - [21504, 3072, 1, 384] - - [37, 86.952] - - - [23040, 9216, 1, 384] - - [58, 90.364] - - - [22656, 7297, 1, 384] - - [37, 88.163] - - - [22656, 5760, 1, 384] - - [25, 89.68] - - - [12288, 11905, 1, 384] - - [27, 88.072] - - - [28032, 7296, 1, 384] - - [29, 90.255] - - - [29184, 3072, 1, 384] - - [29, 88.836] - - - [7680, 1152, 1, 384] - - [49, 75.841] - - - [16896, 7297, 1, 384] - - [37, 88.076] - - - [13056, 5376, 1, 384] - - [27, 89.113] - - - [5376, 4993, 1, 384] - - [27, 80.619] - - - [17280, 9216, 1, 384] - - [58, 89.968] - - - [8448, 8064, 1, 384] - - [25, 88.899] - - - [4608, 1153, 1, 384] - - [40, 55.973] - - - [19200, 9216, 1, 384] - - [407, 91.877] - - - [30720, 7297, 1, 384] - - [25, 88.048] - - - [13440, 5760, 1, 384] - - [25, 89.431] - - - [9984, 3072, 1, 384] - - [37, 82.616] - - - [29952, 15360, 1, 384] - - [25, 90.909] - - - [3840, 1152, 1, 384] - - [49, 60.113] - - - [10368, 9985, 1, 384] - - [27, 87.627] - - - [14592, 7297, 1, 384] - - [402, 89.442] - - - [3456, 3073, 1, 384] - - [27, 68.565] - - - [22272, 9216, 1, 384] - - [53, 90.666] - - - [8064, 8065, 1, 384] - - [37, 86.659] - - - [1536, 1536, 1, 384] - - [227, 53.023] - - - [30336, 4608, 1, 384] - - [53, 89.816] - - - [26112, 12288, 1, 384] - - [53, 90.415] - - - [11904, 11521, 1, 384] - - [53, 89.158] - - - [13440, 6144, 1, 384] - - [36, 89.227] - - - [19200, 13440, 1, 384] - - [37, 90.448] - - - [17280, 1152, 1, 384] - - [27, 78.728] - - - [23424, 3072, 1, 384] - - [27, 88.984] - - - [2304, 1921, 1, 384] - - [35, 59.305] - - - [12672, 7297, 1, 384] - - [25, 86.974] - - - [16896, 1152, 1, 384] - - [25, 82.505] - - - [18432, 1152, 1, 384] - - [27, 83.095] - - - [27264, 13824, 1, 384] - - [394, 91.359] - - - [10752, 1152, 1, 384] - - [34, 79.228] - - - [30336, 7296, 1, 384] - - [29, 90.3] - - - [11904, 3072, 1, 384] - - [37, 84.928] - - - [2304, 768, 1, 384] - - [422, 59.228] - - - [14592, 1152, 1, 384] - - [27, 78.634] - - - [20736, 13441, 1, 384] - - [53, 89.398] - - - [10752, 10752, 1, 384] - - [27, 89.21] - - - [23808, 13440, 1, 384] - - [25, 90.934] - - - [5376, 4992, 1, 384] - - [37, 85.527] - - - [10752, 3072, 1, 384] - - [25, 84.298] - - - [24576, 7296, 1, 384] - - [27, 87.228] - - - [7296, 7296, 1, 384] - - [37, 87.616] - - - [19200, 7296, 1, 384] - - [29, 89.94] - - - [25728, 8832, 1, 384] - - [29, 90.353] - - - [18048, 4224, 1, 384] - - [25, 88.169] - - - [4992, 1152, 1, 384] - - [55, 60.473] - - - [22272, 8832, 1, 384] - - [25, 90.167] - - - [21504, 1153, 1, 384] - - [37, 74.005] - - - [14208, 13440, 1, 384] - - [37, 90.637] - - - [10752, 7296, 1, 384] - - [37, 88.717] - - - [24192, 1152, 1, 384] - - [25, 83.698] - - - [7296, 1152, 1, 384] - - [34, 72.559] - - - [16128, 1153, 1, 384] - - [31, 72.093] - - - [19200, 7297, 1, 384] - - [27, 87.529] - - - [4992, 4993, 1, 384] - - [27, 79.64] - - - [12672, 12673, 1, 384] - - [27, 89.458] - - - [14208, 3072, 1, 384] - - [27, 85.44] - - - [23424, 6528, 1, 384] - - [25, 89.85] - - - [24576, 8064, 1, 384] - - [22, 87.204] - - - [6528, 6145, 1, 384] - - [37, 82.886] - - - [1920, 1537, 1, 384] - - [34, 55.069] - - - [21888, 8448, 1, 384] - - [28, 87.485] - - - [3072, 1536, 1, 384] - - [493, 65.798] - - - [7680, 7296, 1, 384] - - [27, 86.882] - - - [16896, 3072, 1, 384] - - [25, 87.3] - - - [24960, 11520, 1, 384] - - [27, 90.87] - - - [13824, 1152, 1, 384] - - [29, 81.264] - - - [25728, 1153, 1, 384] - - [27, 75.264] - - - [19968, 13441, 1, 384] - - [27, 89.301] - - - [13056, 13057, 1, 384] - - [25, 89.144] - - - [29184, 13440, 1, 384] - - [29, 90.843] - - - [23424, 7297, 1, 384] - - [30, 87.659] - - - [9216, 8832, 1, 384] - - [37, 88.217] - - - [11520, 1153, 1, 384] - - [34, 68.201] - - - [19968, 1153, 1, 384] - - [36, 73.427] - - - [14976, 13440, 1, 384] - - [25, 90.603] - - - [9216, 3072, 1, 384] - - [27, 84.6] - - - [24192, 10752, 1, 384] - - [27, 90.661] - - - [16128, 8832, 1, 384] - - [37, 89.401] - - - [9984, 1153, 1, 384] - - [58, 66.306] - - - [8064, 1153, 1, 384] - - [59, 67.1] - - - [12672, 12672, 1, 384] - - [27, 89.862] - - - [25728, 13441, 1, 384] - - [36, 89.082] - - - [11520, 1152, 1, 384] - - [27, 76.213] - - - [26496, 12672, 1, 384] - - [36, 90.847] - - - [1920, 768, 1, 384] - - [460, 51.55] - - - [20352, 1153, 1, 384] - - [53, 74.649] - - - [10368, 2688, 1, 384] - - [36, 83.77] - - - [6912, 2304, 1, 384] - - [44, 81.625] - - - [17664, 13440, 1, 384] - - [53, 90.735] - - - [17664, 9216, 1, 384] - - [53, 90.066] - - - [25728, 13440, 1, 384] - - [36, 90.911] - - - [10752, 3456, 1, 384] - - [25, 86.158] - - - [6144, 3072, 1, 384] - - [498, 82.19] - - - [9216, 9217, 1, 384] - - [28, 84.232] - - - [3840, 2304, 1, 384] - - [49, 76.562] - - - [12288, 12289, 1, 384] - - [54, 83.752] - - - [11136, 11137, 1, 384] - - [25, 88.387] - - - [11904, 7297, 1, 384] - - [27, 86.522] - - - [29568, 3072, 1, 384] - - [51, 87.919] - - - [12288, 1153, 1, 384] - - [27, 71.726] - - - [18816, 1920, 1, 384] - - [27, 83.941] - - - [13056, 1152, 1, 384] - - [27, 77.549] - - - [8448, 768, 1, 384] - - [40, 67.489] - - - [18816, 2304, 1, 384] - - [44, 85.066] - - - [5376, 3072, 1, 384] - - [27, 77.57] - - - [16512, 1152, 1, 384] - - [45, 75.119] - - - [27648, 7296, 1, 384] - - [37, 90.303] - - - [7296, 2688, 1, 384] - - [27, 83.159] - - - [29184, 15360, 1, 384] - - [27, 90.741] - - - [4608, 4609, 1, 384] - - [27, 77.982] - - - [7296, 7297, 1, 384] - - [25, 84.469] - - - [30720, 9216, 1, 384] - - [74, 88.322] - - - [16384, 3072, 1, 256] - - [22, 62.582] - - - [42496, 10240, 1, 256] - - [29, 74.992] - - - [20992, 7168, 1, 256] - - [27, 73.905] - - - [8960, 5632, 1, 256] - - [59, 72.47] - - - [4864, 256, 1, 256] - - [193, 41.505] - - - [23552, 3584, 1, 256] - - [25, 73.921] - - - [2560, 1281, 1, 256] - - [94, 50.321] - - - [7168, 1280, 1, 256] - - [29, 57.663] - - - [1536, 1153, 1, 384] - - [51, 33.947] - - - [18224, 256, 1, 256] - - [40, 46.948] - - - [13441, 128, 1, 384] - - [194, 47.394] - - - [10753, 128, 1, 384] - - [195, 40.289] - - - [12289, 128, 1, 384] - - [196, 42.921] - - - [385, 128, 1, 384] - - [160, 3.33] - - - [11136, 128, 1, 384] - - [182, 48.351] - - - [13440, 128, 1, 384] - - [181, 55.421] - - - [1153, 128, 1, 384] - - [132, 10.092] - - - [6145, 128, 1, 384] - - [148, 37.382] - - - [4225, 128, 1, 384] - - [151, 26.504] - - - [1537, 128, 1, 384] - - [132, 13.779] - - - [8064, 128, 1, 384] - - [108, 42.016] - - - [3072, 128, 1, 384] - - [121, 28.226] - - - [3457, 128, 1, 384] - - [118, 26.58] - - - [5760, 128, 1, 384] - - [108, 39.611] - - - [8449, 128, 1, 384] - - [140, 38.339] - - - [2305, 128, 1, 384] - - [150, 18.904] - - - [11520, 128, 1, 384] - - [183, 49.823] - - - [11521, 128, 1, 384] - - [197, 42.238] - - - [6528, 128, 1, 384] - - [108, 44.071] - - - [14208, 128, 1, 384] - - [23, 34.859] - - - [768, 128, 1, 384] - - [110, 7.268] - - - [12672, 128, 1, 384] - - [111, 53.145] - - - [9216, 128, 1, 384] - - [125, 45.249] - - - [8448, 128, 1, 384] - - [121, 44.434] - - - [6144, 128, 1, 384] - - [114, 41.479] - - - [2689, 128, 1, 384] - - [118, 21.972] - - - [4224, 128, 1, 384] - - [110, 29.979] - - - [9601, 128, 1, 384] - - [140, 41.933] - - - [13056, 128, 1, 384] - - [198, 54.343] - - - [8065, 128, 1, 384] - - [127, 36.521] - - - [2304, 128, 1, 384] - - [108, 21.347] - - - [8833, 128, 1, 384] - - [151, 39.514] - - - [13824, 128, 1, 384] - - [199, 56.065] - - - [7680, 128, 1, 384] - - [125, 38.65] - - - [3840, 128, 1, 384] - - [109, 27.254] - - - [1920, 128, 1, 384] - - [114, 18.247] - - - [5761, 128, 1, 384] - - [137, 35.338] - - - [7681, 128, 1, 384] - - [200, 34.854] - - - [4608, 128, 1, 384] - - [123, 32.496] - - - [10369, 128, 1, 384] - - [159, 44.325] - - - [3841, 128, 1, 384] - - [151, 24.094] - - - [7296, 128, 1, 384] - - [110, 38.194] - - - [7297, 128, 1, 384] - - [148, 33.523] - - - [10752, 128, 1, 384] - - [144, 47.522] - - - [1536, 128, 1, 384] - - [166, 14.724] - - - [11137, 128, 1, 384] - - [201, 40.762] - - - [2688, 128, 1, 384] - - [114, 24.495] - - - [4609, 128, 1, 384] - - [163, 28.912] - - - [6529, 128, 1, 384] - - [148, 39.286] - - - [11905, 128, 1, 384] - - [201, 43.004] - - - [6912, 128, 1, 384] - - [114, 46.101] - - - [769, 128, 1, 384] - - [128, 6.811] - - - [12288, 128, 1, 384] - - [191, 52.227] - - - [15360, 128, 1, 384] - - [57, 37.498] - - - [9600, 128, 1, 384] - - [114, 48.868] - - - [13057, 128, 1, 384] - - [202, 46.04] - - - [10368, 128, 1, 384] - - [114, 51.018] - - - [12673, 128, 1, 384] - - [135, 45.262] - - - [9217, 128, 1, 384] - - [200, 38.95] - - - [4993, 128, 1, 384] - - [141, 31.057] - - - [9984, 128, 1, 384] - - [137, 49.458] - - - [6913, 128, 1, 384] - - [163, 40.93] - - - [8832, 128, 1, 384] - - [172, 45.694] - - - [3073, 128, 1, 384] - - [203, 23.96] - - - [14976, 128, 1, 384] - - [34, 36.622] - - - [384, 128, 1, 384] - - [109, 3.649] - - - [5377, 128, 1, 384] - - [148, 33.352] - - - [1152, 128, 1, 384] - - [166, 10.996] - - - [9985, 128, 1, 384] - - [159, 42.849] - - - [14592, 128, 1, 384] - - [39, 35.35] - - - [4992, 128, 1, 384] - - [108, 34.871] - - - [3456, 128, 1, 384] - - [172, 30.612] - - - [1921, 128, 1, 384] - - [165, 15.697] - - - [5376, 128, 1, 384] - - [109, 37.318] - - - [11904, 128, 1, 384] - - [191, 50.594] - - - [44544, 2048, 1, 384] - - [53, 88.554] - - - [39552, 512, 1, 384] - - [44, 79.797] - - - [38016, 22145, 1, 384] - - [37, 89.555] - - - [39552, 23297, 1, 384] - - [37, 89.357] - - - [39552, 23681, 1, 384] - - [29, 89.45] - - - [36864, 2048, 1, 384] - - [37, 82.377] - - - [44544, 28673, 1, 384] - - [32, 87.03] - - - [43776, 512, 1, 384] - - [26, 79.402] - - - [43392, 1024, 1, 384] - - [36, 86.883] - - - [42240, 4096, 1, 384] - - [36, 90.427] - - - [42624, 26369, 1, 384] - - [51, 83.329] - - - [35328, 1024, 1, 384] - - [58, 83.966] - - - [36096, 384, 1, 384] - - [398, 79.689] - - - [38784, 4096, 1, 384] - - [36, 89.674] - - - [39552, 384, 1, 384] - - [399, 78.339] - - - [42240, 8192, 1, 384] - - [53, 91.007] - - - [42240, 25985, 1, 384] - - [25, 89.72] - - - [38016, 4096, 1, 384] - - [58, 89.941] - - - [39168, 4096, 1, 384] - - [53, 89.71] - - - [35328, 19457, 1, 384] - - [32, 87.338] - - - [43392, 2048, 1, 384] - - [36, 88.584] - - - [38400, 4096, 1, 384] - - [44, 88.905] - - - [35712, 1024, 1, 384] - - [53, 84.814] - - - [36480, 2048, 1, 384] - - [44, 87.979] - - - [40704, 512, 1, 384] - - [27, 81.464] - - - [36864, 20609, 1, 384] - - [37, 88.091] - - - [37632, 21761, 1, 384] - - [37, 89.191] - - - [38016, 2048, 1, 384] - - [44, 87.947] - - - [44160, 2048, 1, 384] - - [53, 88.431] - - - [35328, 384, 1, 384] - - [396, 78.358] - - - [43392, 384, 1, 384] - - [395, 78.525] - - - [39168, 512, 1, 384] - - [27, 79.257] - - - [38784, 1024, 1, 384] - - [44, 84.387] - - - [35328, 2048, 1, 384] - - [58, 87.335] - - - [44544, 8192, 1, 384] - - [25, 89.941] - - - [40704, 384, 1, 384] - - [399, 80.143] - - - [39936, 512, 1, 384] - - [27, 80.271] - - - [41472, 25217, 1, 384] - - [29, 89.363] - - - [42240, 2048, 1, 384] - - [58, 89.372] - - - [37632, 512, 1, 384] - - [36, 81.473] - - - [37248, 1024, 1, 384] - - [36, 84.781] - - - [42240, 26369, 1, 384] - - [27, 89.63] - - - [43776, 384, 1, 384] - - [403, 78.6] - - - [44160, 8192, 1, 384] - - [58, 90.909] - - - [39936, 1024, 1, 384] - - [58, 86.121] - - - [43392, 27137, 1, 384] - - [25, 88.823] - - - [39936, 384, 1, 384] - - [401, 78.888] - - - [41472, 25601, 1, 384] - - [32, 87.287] - - - [36864, 4096, 1, 384] - - [32, 84.148] - - - [43392, 8192, 1, 384] - - [58, 90.761] - - - [36096, 512, 1, 384] - - [27, 78.538] - - - [36480, 4096, 1, 384] - - [36, 90.149] - - - [40320, 512, 1, 384] - - [37, 80.915] - - - [41088, 4096, 1, 384] - - [74, 88.437] - - - [43776, 27521, 1, 384] - - [54, 88.034] - - - [35328, 19073, 1, 384] - - [29, 89.484] - - - [44160, 384, 1, 384] - - [395, 79.612] - - - [36864, 8192, 1, 384] - - [32, 87.044] - - - [41088, 2048, 1, 384] - - [61, 86.645] - - - [38016, 21761, 1, 384] - - [25, 89.316] - - - [41856, 1024, 1, 384] - - [36, 87.124] - - - [39552, 8192, 1, 384] - - [58, 90.847] - - - [37632, 4096, 1, 384] - - [44, 90.013] - - - [41856, 384, 1, 384] - - [395, 81.494] - - - [44160, 28289, 1, 384] - - [27, 89.054] - - - [43008, 26753, 1, 384] - - [54, 88.861] - - - [38400, 512, 1, 384] - - [36, 82.08] - - - [39168, 384, 1, 384] - - [399, 77.881] - - - [37632, 1024, 1, 384] - - [44, 85.403] - - - [44544, 4096, 1, 384] - - [74, 88.698] - - - [42240, 512, 1, 384] - - [36, 83.781] - - - [43008, 2048, 1, 384] - - [74, 86.187] - - - [36480, 20609, 1, 384] - - [53, 89.24] - - - [36864, 512, 1, 384] - - [53, 80.085] - - - [43008, 384, 1, 384] - - [400, 77.873] - - - [43392, 4096, 1, 384] - - [36, 90.249] - - - [38400, 22145, 1, 384] - - [37, 89.347] - - - [39936, 23681, 1, 384] - - [25, 89.105] - - - [36096, 19841, 1, 384] - - [54, 87.621] - - - [44544, 512, 1, 384] - - [27, 82.624] - - - [38400, 2048, 1, 384] - - [36, 88.327] - - - [41856, 25985, 1, 384] - - [27, 89.011] - - - [42624, 2048, 1, 384] - - [37, 83.262] - - - [38400, 1024, 1, 384] - - [53, 86.389] - - - [36480, 512, 1, 384] - - [27, 79.706] - - - [42624, 26753, 1, 384] - - [51, 83.126] - - - [43776, 27905, 1, 384] - - [32, 88.04] - - - [37248, 2048, 1, 384] - - [58, 87.981] - - - [35712, 19841, 1, 384] - - [29, 89.178] - - - [43392, 27521, 1, 384] - - [25, 89.004] - - - [43008, 1024, 1, 384] - - [36, 86.036] - - - [42624, 512, 1, 384] - - [27, 79.899] - - - [41472, 384, 1, 384] - - [389, 81.27] - - - [40704, 2048, 1, 384] - - [53, 88.243] - - - [36096, 2048, 1, 384] - - [62, 85.766] - - - [39936, 4096, 1, 384] - - [61, 87.649] - - - [40320, 2048, 1, 384] - - [58, 89.135] - - - [41088, 8192, 1, 384] - - [74, 89.419] - - - [35328, 8192, 1, 384] - - [29, 90.273] - - - [40320, 4096, 1, 384] - - [58, 90.213] - - - [41856, 512, 1, 384] - - [58, 83.581] - - - [39552, 4096, 1, 384] - - [58, 90.197] - - - [35712, 2048, 1, 384] - - [36, 88.452] - - - [39936, 24065, 1, 384] - - [37, 88.775] - - - [36480, 20225, 1, 384] - - [29, 89.099] - - - [38016, 1024, 1, 384] - - [36, 85.94] - - - [43008, 512, 1, 384] - - [25, 80.274] - - - [40704, 24833, 1, 384] - - [37, 88.984] - - - [37248, 4096, 1, 384] - - [53, 89.751] - - - [41856, 4096, 1, 384] - - [36, 89.878] - - - [41472, 512, 1, 384] - - [53, 82.889] - - - [39552, 2048, 1, 384] - - [58, 89.065] - - - [41088, 384, 1, 384] - - [402, 80.352] - - - [36480, 8192, 1, 384] - - [58, 90.778] - - - [37632, 2048, 1, 384] - - [53, 88.738] - - - [40704, 8192, 1, 384] - - [53, 90.642] - - - [36864, 20993, 1, 384] - - [37, 87.812] - - - [35328, 512, 1, 384] - - [44, 77.624] - - - [40320, 384, 1, 384] - - [397, 79.411] - - - [36096, 1024, 1, 384] - - [93, 83.824] - - - [42624, 8192, 1, 384] - - [25, 89.493] - - - [38784, 22529, 1, 384] - - [32, 87.014] - - - [44160, 4096, 1, 384] - - [36, 90.077] - - - [41472, 4096, 1, 384] - - [93, 88.798] - - - [36480, 1024, 1, 384] - - [36, 85.97] - - - [38784, 2048, 1, 384] - - [53, 87.715] - - - [44544, 1024, 1, 384] - - [36, 85.806] - - - [41088, 24833, 1, 384] - - [54, 88.119] - - - [36864, 384, 1, 384] - - [382, 80.185] - - - [43392, 512, 1, 384] - - [27, 81.03] - - - [39168, 8192, 1, 384] - - [27, 90.446] - - - [42624, 4096, 1, 384] - - [25, 87.761] - - - [40320, 24065, 1, 384] - - [29, 89.226] - - - [44160, 512, 1, 384] - - [58, 82.101] - - - [38016, 384, 1, 384] - - [395, 76.045] - - - [38016, 512, 1, 384] - - [25, 82.165] - - - [37248, 512, 1, 384] - - [25, 80.906] - - - [43776, 2048, 1, 384] - - [61, 86.412] - - - [35712, 8192, 1, 384] - - [25, 90.561] - - - [38400, 384, 1, 384] - - [385, 76.836] - - - [42240, 1024, 1, 384] - - [53, 87.472] - - - [35712, 19457, 1, 384] - - [32, 86.829] - - - [41856, 2048, 1, 384] - - [58, 88.873] - - - [41472, 1024, 1, 384] - - [44, 86.378] - - - [37632, 384, 1, 384] - - [400, 75.413] - - - [40704, 1024, 1, 384] - - [44, 84.888] - - - [43008, 27137, 1, 384] - - [32, 88.817] - - - [40704, 4096, 1, 384] - - [58, 90.082] - - - [36096, 20225, 1, 384] - - [32, 87.899] - - - [39936, 8192, 1, 384] - - [74, 88.219] - - - [38784, 384, 1, 384] - - [395, 77.186] - - - [38784, 8192, 1, 384] - - [36, 90.697] - - - [42624, 384, 1, 384] - - [399, 77.452] - - - [35712, 4096, 1, 384] - - [58, 89.523] - - - [37632, 8192, 1, 384] - - [53, 90.488] - - - [38784, 22913, 1, 384] - - [27, 89.099] - - - [36864, 1024, 1, 384] - - [44, 84.734] - - - [37248, 384, 1, 384] - - [398, 80.941] - - - [39168, 23297, 1, 384] - - [27, 89.091] - - - [40704, 24449, 1, 384] - - [27, 89.212] - - - [41472, 2048, 1, 384] - - [58, 88.036] - - - [44160, 27905, 1, 384] - - [25, 88.95] - - - [44160, 1024, 1, 384] - - [53, 85.41] - - - [36480, 384, 1, 384] - - [399, 79.846] - - - [42240, 384, 1, 384] - - [400, 82.169] - - - [44544, 28289, 1, 384] - - [32, 89.099] - - - [37248, 21377, 1, 384] - - [25, 88.605] - - - [36096, 4096, 1, 384] - - [74, 87.498] - - - [38784, 512, 1, 384] - - [25, 78.508] - - - [35712, 384, 1, 384] - - [397, 78.476] - - - [43776, 1024, 1, 384] - - [74, 84.107] - - - [41088, 25217, 1, 384] - - [32, 88.117] - - - [40320, 8192, 1, 384] - - [25, 90.614] - - - [39168, 22913, 1, 384] - - [29, 89.058] - - - [38400, 8192, 1, 384] - - [37, 90.376] - - - [41088, 512, 1, 384] - - [24, 79.633] - - - [42624, 1024, 1, 384] - - [25, 83.626] - - - [39168, 2048, 1, 384] - - [36, 88.606] - - - [43008, 4096, 1, 384] - - [61, 87.372] - - - [35712, 512, 1, 384] - - [29, 78.178] - - - [41856, 8192, 1, 384] - - [36, 90.813] - - - [43008, 8192, 1, 384] - - [54, 87.95] - - - [41472, 8192, 1, 384] - - [29, 90.544] - - - [41088, 1024, 1, 384] - - [39, 83.173] - - - [37248, 20993, 1, 384] - - [27, 88.42] - - - [44544, 384, 1, 384] - - [403, 79.871] - - - [36096, 8192, 1, 384] - - [51, 88.991] - - - [43776, 8192, 1, 384] - - [61, 89.499] - - - [41856, 25601, 1, 384] - - [32, 87.17] - - - [37632, 21377, 1, 384] - - [37, 89.33] - - - [40320, 24449, 1, 384] - - [25, 89.215] - - - [43776, 4096, 1, 384] - - [74, 88.614] - - - [35328, 4096, 1, 384] - - [56, 88.908] - - - [39552, 1024, 1, 384] - - [58, 85.957] - - - [38016, 8192, 1, 384] - - [44, 90.868] - - - [38400, 22529, 1, 384] - - [54, 87.23] - - - [39936, 2048, 1, 384] - - [74, 86.374] - - - [39168, 1024, 1, 384] - - [53, 85.143] - - - [37248, 8192, 1, 384] - - [29, 90.594] - - - [40320, 1024, 1, 384] - - [58, 87.081] - - - [26112, 1024, 1, 384] - - [58, 85.092] - - - [24192, 2048, 1, 384] - - [58, 86.935] - - - [13440, 5761, 1, 384] - - [25, 87.024] - - - [3456, 384, 1, 384] - - [346, 51.608] - - - [21888, 4096, 1, 384] - - [24, 84.551] - - - [384, 384, 1, 384] - - [331, 10.767] - - - [21120, 1024, 1, 384] - - [53, 83.912] - - - [30336, 4096, 1, 384] - - [58, 89.998] - - - [31488, 512, 1, 384] - - [36, 81.143] - - - [2304, 1793, 1, 384] - - [55, 55.398] - - - [16896, 9217, 1, 384] - - [53, 86.173] - - - [9216, 1024, 1, 384] - - [40, 70.126] - - - [29568, 1024, 1, 384] - - [56, 83.999] - - - [27264, 11393, 1, 384] - - [25, 88.608] - - - [33408, 17537, 1, 384] - - [25, 89.304] - - - [18816, 1024, 1, 384] - - [36, 81.935] - - - [5760, 1024, 1, 384] - - [53, 61.788] - - - [31104, 14849, 1, 384] - - [37, 89.596] - - - [18816, 4096, 1, 384] - - [44, 88.971] - - - [11136, 1024, 1, 384] - - [53, 73.552] - - - [17664, 9985, 1, 384] - - [36, 88.697] - - - [9216, 512, 1, 384] - - [35, 61.98] - - - [17664, 1024, 1, 384] - - [36, 78.042] - - - [17664, 512, 1, 384] - - [53, 67.758] - - - [31488, 384, 1, 384] - - [387, 78.426] - - - [15744, 8065, 1, 384] - - [37, 88.222] - - - [5760, 3841, 1, 384] - - [25, 80.878] - - - [24192, 1024, 1, 384] - - [44, 84.078] - - - [20352, 384, 1, 384] - - [378, 74.267] - - - [21888, 2048, 1, 384] - - [25, 80.794] - - - [7680, 2048, 1, 384] - - [40, 80.651] - - - [2688, 512, 1, 384] - - [183, 48.391] - - - [13056, 1024, 1, 384] - - [36, 76.455] - - - [22656, 14977, 1, 384] - - [53, 89.781] - - - [10752, 6785, 1, 384] - - [25, 86.162] - - - [6912, 2048, 1, 384] - - [44, 80.824] - - - [15360, 512, 1, 384] - - [34, 68.193] - - - [31104, 384, 1, 384] - - [393, 77.79] - - - [30720, 14465, 1, 384] - - [25, 89.054] - - - [17280, 2048, 1, 384] - - [58, 85.776] - - - [34176, 1024, 1, 384] - - [36, 84.779] - - - [16896, 2048, 1, 384] - - [53, 84.012] - - - [17664, 384, 1, 384] - - [374, 72.636] - - - [21504, 512, 1, 384] - - [53, 71.604] - - - [18048, 10369, 1, 384] - - [37, 88.443] - - - [15744, 1024, 1, 384] - - [53, 82.108] - - - [33408, 4096, 1, 384] - - [53, 89.946] - - - [11904, 4096, 1, 384] - - [36, 85.832] - - - [18816, 512, 1, 384] - - [23, 71.129] - - - [34944, 4096, 1, 384] - - [58, 89.626] - - - [13824, 2048, 1, 384] - - [44, 84.822] - - - [3840, 512, 1, 384] - - [53, 37.763] - - - [4992, 1024, 1, 384] - - [63, 67.094] - - - [11136, 7553, 1, 384] - - [27, 87.11] - - - [16512, 1024, 1, 384] - - [107, 72.635] - - - [17280, 9217, 1, 384] - - [93, 85.933] - - - [29184, 1024, 1, 384] - - [36, 84.597] - - - [18048, 512, 1, 384] - - [60, 68.959] - - - [6528, 384, 1, 384] - - [363, 60.501] - - - [28416, 1024, 1, 384] - - [53, 82.714] - - - [2688, 1153, 1, 384] - - [467, 61.295] - - - [34560, 18305, 1, 384] - - [37, 89.206] - - - [20736, 384, 1, 384] - - [376, 75.215] - - - [11520, 512, 1, 384] - - [34, 61.863] - - - [26112, 8192, 1, 384] - - [44, 90.445] - - - [31872, 384, 1, 384] - - [390, 78.822] - - - [24192, 512, 1, 384] - - [53, 78.806] - - - [19968, 2048, 1, 384] - - [36, 86.677] - - - [32256, 8192, 1, 384] - - [37, 90.325] - - - [11520, 384, 1, 384] - - [369, 69.974] - - - [1920, 1409, 1, 384] - - [35, 50.649] - - - [25728, 9857, 1, 384] - - [29, 88.779] - - - [9216, 5633, 1, 384] - - [37, 84.683] - - - [28032, 12161, 1, 384] - - [29, 89.361] - - - [28800, 8192, 1, 384] - - [58, 90.586] - - - [28416, 12161, 1, 384] - - [53, 89.491] - - - [23040, 15361, 1, 384] - - [54, 87.18] - - - [31488, 15617, 1, 384] - - [37, 89.332] - - - [22272, 14209, 1, 384] - - [58, 89.567] - - - [1536, 512, 1, 384] - - [109, 41.102] - - - [1152, 257, 1, 384] - - [171, 21.611] - - - [21120, 2048, 1, 384] - - [26, 86.878] - - - [32256, 16001, 1, 384] - - [27, 89.634] - - - [9600, 6017, 1, 384] - - [29, 86.235] - - - [32640, 384, 1, 384] - - [382, 79.683] - - - [34176, 512, 1, 384] - - [58, 80.817] - - - [10368, 512, 1, 384] - - [34, 68.48] - - - [21120, 384, 1, 384] - - [379, 76.068] - - - [29568, 4096, 1, 384] - - [26, 88.237] - - - [31872, 2048, 1, 384] - - [53, 87.371] - - - [8832, 384, 1, 384] - - [369, 66.166] - - - [4224, 384, 1, 384] - - [354, 55.472] - - - [33408, 8192, 1, 384] - - [27, 90.609] - - - [768, 257, 1, 384] - - [171, 14.592] - - - [10368, 6401, 1, 384] - - [27, 86.647] - - - [13824, 384, 1, 384] - - [370, 72.162] - - - [29568, 512, 1, 384] - - [44, 77.505] - - - [28032, 1024, 1, 384] - - [36, 85.796] - - - [19200, 384, 1, 384] - - [369, 71.557] - - - [23040, 2048, 1, 384] - - [36, 85.758] - - - [8448, 4481, 1, 384] - - [27, 83.925] - - - [22272, 14593, 1, 384] - - [58, 89.97] - - - [26496, 10241, 1, 384] - - [58, 85.965] - - - [19584, 384, 1, 384] - - [376, 72.581] - - - [4992, 3457, 1, 384] - - [27, 80.037] - - - [22656, 384, 1, 384] - - [383, 75.998] - - - [15360, 1024, 1, 384] - - [36, 80.098] - - - [7296, 2048, 1, 384] - - [25, 77.098] - - - [30720, 384, 1, 384] - - [386, 77.384] - - - [6144, 2177, 1, 384] - - [34, 75.794] - - - [30720, 14849, 1, 384] - - [37, 88.592] - - - [23424, 2048, 1, 384] - - [44, 85.898] - - - [5760, 384, 1, 384] - - [360, 55.312] - - - [6144, 2561, 1, 384] - - [29, 73.855] - - - [12672, 384, 1, 384] - - [369, 67.419] - - - [16128, 8065, 1, 384] - - [395, 90.412] - - - [10752, 7169, 1, 384] - - [93, 83.986] - - - [2304, 384, 1, 384] - - [331, 43.869] - - - [18816, 2048, 1, 384] - - [44, 85.312] - - - [22272, 4096, 1, 384] - - [58, 89.256] - - - [12672, 4993, 1, 384] - - [25, 86.778] - - - [12288, 512, 1, 384] - - [60, 65.525] - - - [13056, 4993, 1, 384] - - [37, 85.015] - - - [19584, 512, 1, 384] - - [53, 73.656] - - - [30336, 14465, 1, 384] - - [27, 89.258] - - - [5376, 3841, 1, 384] - - [27, 81.167] - - - [17664, 9601, 1, 384] - - [44, 88.671] - - - [29952, 2048, 1, 384] - - [58, 86.831] - - - [8832, 512, 1, 384] - - [34, 60.337] - - - [9984, 512, 1, 384] - - [34, 66.79] - - - [19200, 1024, 1, 384] - - [36, 82.967] - - - [24192, 8321, 1, 384] - - [58, 88.409] - - - [26112, 10241, 1, 384] - - [32, 86.108] - - - [17280, 9601, 1, 384] - - [58, 88.482] - - - [7296, 384, 1, 384] - - [365, 56.179] - - - [16512, 8449, 1, 384] - - [32, 85.255] - - - [11904, 4225, 1, 384] - - [53, 85.165] - - - [24576, 4096, 1, 384] - - [25, 82.017] - - - [6912, 2945, 1, 384] - - [27, 79.917] - - - [33024, 16769, 1, 384] - - [32, 88.675] - - - [24576, 8705, 1, 384] - - [29, 83.163] - - - [16128, 2048, 1, 384] - - [53, 83.245] - - - [13824, 6145, 1, 384] - - [29, 84.121] - - - [28800, 512, 1, 384] - - [27, 75.38] - - - [33792, 8192, 1, 384] - - [61, 88.394] - - - [27648, 11393, 1, 384] - - [27, 88.687] - - - [21888, 384, 1, 384] - - [381, 73.907] - - - [12672, 4096, 1, 384] - - [36, 87.969] - - - [23040, 14977, 1, 384] - - [25, 89.321] - - - [11904, 384, 1, 384] - - [369, 64.364] - - - [7680, 3713, 1, 384] - - [29, 80.768] - - - [24576, 8192, 1, 384] - - [32, 84.517] - - - [34176, 384, 1, 384] - - [394, 75.55] - - - [17664, 2048, 1, 384] - - [44, 84.228] - - - [29952, 4096, 1, 384] - - [58, 89.17] - - - [9984, 6017, 1, 384] - - [37, 84.842] - - - [33408, 2048, 1, 384] - - [58, 88.753] - - - [21120, 4096, 1, 384] - - [56, 88.935] - - - [34560, 4096, 1, 384] - - [36, 89.877] - - - [19200, 11521, 1, 384] - - [37, 88.647] - - - [21120, 13057, 1, 384] - - [27, 89.01] - - - [25728, 384, 1, 384] - - [369, 77.151] - - - [28800, 12929, 1, 384] - - [29, 88.677] - - - [20736, 1024, 1, 384] - - [36, 83.369] - - - [18816, 10753, 1, 384] - - [25, 89.03] - - - [34560, 8192, 1, 384] - - [25, 90.331] - - - [23040, 512, 1, 384] - - [53, 75.666] - - - [30336, 2048, 1, 384] - - [58, 87.643] - - - [17280, 512, 1, 384] - - [44, 75.461] - - - [19200, 2048, 1, 384] - - [53, 86.532] - - - [12288, 4225, 1, 384] - - [27, 84.633] - - - [15744, 7681, 1, 384] - - [433, 90.151] - - - [30720, 4096, 1, 384] - - [74, 86.995] - - - [10752, 384, 1, 384] - - [369, 67.277] - - - [15744, 512, 1, 384] - - [27, 69.785] - - - [24960, 384, 1, 384] - - [373, 76.014] - - - [768, 384, 1, 384] - - [334, 21.264] - - - [6912, 3329, 1, 384] - - [25, 78.713] - - - [8064, 512, 1, 384] - - [34, 55.111] - - - [26496, 384, 1, 384] - - [388, 75.859] - - - [24960, 4096, 1, 384] - - [53, 89.572] - - - [19584, 11905, 1, 384] - - [58, 89.153] - - - [16512, 8833, 1, 384] - - [54, 85.805] - - - [18816, 384, 1, 384] - - [366, 70.244] - - - [23808, 1024, 1, 384] - - [44, 83.004] - - - [16512, 384, 1, 384] - - [370, 71.856] - - - [8448, 4865, 1, 384] - - [37, 83.744] - - - [34944, 1024, 1, 384] - - [58, 86.21] - - - [29184, 4096, 1, 384] - - [44, 88.774] - - - [8832, 2048, 1, 384] - - [58, 77.616] - - - [9984, 1024, 1, 384] - - [63, 75.292] - - - [22272, 1024, 1, 384] - - [58, 83.249] - - - [14592, 6913, 1, 384] - - [400, 89.529] - - - [9216, 2048, 1, 384] - - [23, 78.405] - - - [7296, 1024, 1, 384] - - [40, 65.639] - - - [26880, 8192, 1, 384] - - [53, 90.358] - - - [26880, 10625, 1, 384] - - [44, 89.218] - - - [28800, 12545, 1, 384] - - [29, 88.809] - - - [18048, 1024, 1, 384] - - [36, 79.175] - - - [27264, 11009, 1, 384] - - [27, 88.429] - - - [12288, 2048, 1, 384] - - [27, 79.298] - - - [19200, 4096, 1, 384] - - [53, 88.577] - - - [32256, 384, 1, 384] - - [388, 79.89] - - - [9216, 5249, 1, 384] - - [37, 84.69] - - - [29952, 14081, 1, 384] - - [30, 89.172] - - - [7680, 384, 1, 384] - - [366, 58.932] - - - [19200, 11137, 1, 384] - - [47, 88.991] - - - [14976, 1024, 1, 384] - - [44, 78.752] - - - [25728, 1024, 1, 384] - - [58, 83.763] - - - [3456, 1921, 1, 384] - - [35, 68.429] - - - [21120, 13441, 1, 384] - - [47, 89.157] - - - [15360, 2048, 1, 384] - - [61, 82.111] - - - [34560, 512, 1, 384] - - [27, 81.541] - - - [31872, 8192, 1, 384] - - [29, 90.406] - - - [32640, 16769, 1, 384] - - [32, 87.83] - - - [26496, 1024, 1, 384] - - [53, 81.816] - - - [12672, 1024, 1, 384] - - [53, 74.602] - - - [3072, 384, 1, 384] - - [346, 46.928] - - - [31104, 4096, 1, 384] - - [53, 89.842] - - - [25344, 4096, 1, 384] - - [36, 89.495] - - - [4224, 2689, 1, 384] - - [35, 73.263] - - - [24576, 1024, 1, 384] - - [32, 78.371] - - - [8448, 512, 1, 384] - - [34, 58.027] - - - [1536, 1025, 1, 384] - - [204, 50.72] - - - [14208, 6145, 1, 384] - - [56, 85.307] - - - [27264, 384, 1, 384] - - [390, 77.856] - - - [34560, 1024, 1, 384] - - [53, 85.683] - - - [14976, 6913, 1, 384] - - [27, 87.539] - - - [21504, 2048, 1, 384] - - [61, 84.642] - - - [14208, 4096, 1, 384] - - [53, 87.288] - - - [14592, 4096, 1, 384] - - [58, 86.938] - - - [6528, 2561, 1, 384] - - [27, 77.922] - - - [34176, 18305, 1, 384] - - [51, 89.147] - - - [19968, 384, 1, 384] - - [377, 73.431] - - - [30720, 8192, 1, 384] - - [74, 87.502] - - - [14592, 512, 1, 384] - - [40, 65.731] - - - [25728, 2048, 1, 384] - - [53, 86.383] - - - [23424, 4096, 1, 384] - - [27, 87.192] - - - [27264, 2048, 1, 384] - - [53, 84.397] - - - [21504, 1024, 1, 384] - - [53, 80.566] - - - [30336, 384, 1, 384] - - [373, 78.286] - - - [2688, 1024, 1, 384] - - [59, 51.497] - - - [22656, 4096, 1, 384] - - [36, 89.102] - - - [20352, 2048, 1, 384] - - [53, 84.991] - - - [33408, 384, 1, 384] - - [390, 74.322] - - - [15360, 4096, 1, 384] - - [37, 86.111] - - - [22272, 512, 1, 384] - - [37, 73.8] - - - [14208, 384, 1, 384] - - [366, 66.866] - - - [32640, 512, 1, 384] - - [76, 72.894] - - - [23808, 512, 1, 384] - - [25, 78.073] - - - [24960, 1024, 1, 384] - - [58, 81.728] - - - [4608, 512, 1, 384] - - [40, 45.066] - - - [25344, 2048, 1, 384] - - [36, 87.652] - - - [11904, 1024, 1, 384] - - [53, 78.276] - - - [28416, 12545, 1, 384] - - [29, 89.264] - - - [14208, 6529, 1, 384] - - [25, 87.106] - - - [13824, 5761, 1, 384] - - [29, 85.86] - - - [26112, 9857, 1, 384] - - [29, 89.098] - - - [9600, 2048, 1, 384] - - [58, 83.301] - - - [33024, 1024, 1, 384] - - [61, 83.484] - - - [34944, 18689, 1, 384] - - [25, 89.205] - - - [13824, 512, 1, 384] - - [63, 73.084] - - - [26880, 384, 1, 384] - - [389, 76.793] - - - [15744, 384, 1, 384] - - [373, 71.954] - - - [29568, 8192, 1, 384] - - [53, 89.975] - - - [24960, 9089, 1, 384] - - [58, 89.011] - - - [28032, 2048, 1, 384] - - [53, 88.256] - - - [19968, 11905, 1, 384] - - [37, 89.202] - - - [6528, 2945, 1, 384] - - [27, 75.792] - - - [20352, 12289, 1, 384] - - [93, 87.032] - - - [5376, 512, 1, 384] - - [59, 51.87] - - - [5376, 3457, 1, 384] - - [35, 79.024] - - - [21504, 384, 1, 384] - - [380, 73.055] - - - [11520, 1024, 1, 384] - - [58, 76.232] - - - [3840, 1921, 1, 384] - - [495, 70.918] - - - [18432, 4096, 1, 384] - - [61, 85.855] - - - [28416, 2048, 1, 384] - - [58, 87.239] - - - [3456, 512, 1, 384] - - [183, 59.209] - - - [2688, 384, 1, 384] - - [346, 42.327] - - - [28032, 4096, 1, 384] - - [58, 89.737] - - - [16128, 384, 1, 384] - - [373, 72.906] - - - [33792, 17537, 1, 384] - - [25, 89.18] - - - [2688, 1793, 1, 384] - - [49, 63.259] - - - [27648, 1024, 1, 384] - - [53, 84.344] - - - [13440, 1024, 1, 384] - - [36, 78.549] - - - [28032, 8192, 1, 384] - - [58, 90.546] - - - [34560, 18689, 1, 384] - - [37, 89.298] - - - [16896, 512, 1, 384] - - [55, 74.187] - - - [13056, 2048, 1, 384] - - [53, 84.795] - - - [3072, 1537, 1, 384] - - [494, 64.951] - - - [3072, 512, 1, 384] - - [181, 53.669] - - - [25344, 9089, 1, 384] - - [51, 88.616] - - - [9600, 384, 1, 384] - - [370, 64.963] - - - [26880, 512, 1, 384] - - [58, 78.239] - - - [33024, 512, 1, 384] - - [55, 76.942] - - - [21888, 1024, 1, 384] - - [25, 78.92] - - - [18048, 384, 1, 384] - - [369, 73.352] - - - [16896, 4096, 1, 384] - - [36, 87.888] - - - [23808, 384, 1, 384] - - [386, 75.552] - - - [26496, 4096, 1, 384] - - [53, 89.126] - - - [20736, 13057, 1, 384] - - [37, 89.631] - - - [24576, 512, 1, 384] - - [37, 78.16] - - - [14592, 6529, 1, 384] - - [37, 87.597] - - - [6528, 512, 1, 384] - - [53, 61.445] - - - [22656, 14593, 1, 384] - - [58, 89.381] - - - [26112, 2048, 1, 384] - - [53, 87.556] - - - [25728, 9473, 1, 384] - - [25, 88.602] - - - [15744, 2048, 1, 384] - - [58, 86.199] - - - [31488, 1024, 1, 384] - - [36, 86.085] - - - [11136, 2048, 1, 384] - - [58, 83.043] - - - [4608, 2689, 1, 384] - - [55, 71.478] - - - [30720, 1024, 1, 384] - - [53, 84.187] - - - [1920, 512, 1, 384] - - [123, 40.881] - - - [25728, 8192, 1, 384] - - [58, 90.683] - - - [31104, 2048, 1, 384] - - [36, 87.584] - - - [3456, 1024, 1, 384] - - [80, 64.265] - - - [25344, 384, 1, 384] - - [373, 76.621] - - - [27264, 8192, 1, 384] - - [25, 90.393] - - - [16128, 4096, 1, 384] - - [58, 87.503] - - - [20736, 12673, 1, 384] - - [37, 89.313] - - - [4224, 2305, 1, 384] - - [34, 71.826] - - - [27648, 11777, 1, 384] - - [27, 88.558] - - - [6144, 512, 1, 384] - - [35, 58.515] - - - [24576, 2048, 1, 384] - - [29, 79.488] - - - [15360, 384, 1, 384] - - [369, 70.587] - - - [34944, 19073, 1, 384] - - [25, 89.106] - - - [33792, 384, 1, 384] - - [380, 75.202] - - - [15360, 7681, 1, 384] - - [395, 90.032] - - - [34176, 17921, 1, 384] - - [24, 88.898] - - - [10368, 1024, 1, 384] - - [53, 77.886] - - - [34176, 8192, 1, 384] - - [27, 90.199] - - - [34176, 2048, 1, 384] - - [93, 88.035] - - - [7680, 4097, 1, 384] - - [44, 82.31] - - - [10752, 1024, 1, 384] - - [36, 71.496] - - - [9984, 2048, 1, 384] - - [44, 80.888] - - - [5760, 2048, 1, 384] - - [23, 75.92] - - - [30336, 1024, 1, 384] - - [53, 83.496] - - - [23424, 384, 1, 384] - - [385, 77.634] - - - [13440, 5377, 1, 384] - - [29, 85.376] - - - [14592, 2048, 1, 384] - - [44, 84.533] - - - [31872, 4096, 1, 384] - - [53, 89.55] - - - [6528, 2048, 1, 384] - - [37, 76.208] - - - [8064, 384, 1, 384] - - [367, 61.594] - - - [31872, 16001, 1, 384] - - [29, 89.19] - - - [16896, 1024, 1, 384] - - [44, 80.371] - - - [15360, 7297, 1, 384] - - [395, 89.724] - - - [33792, 4096, 1, 384] - - [74, 87.283] - - - [16896, 384, 1, 384] - - [371, 73.144] - - - [29952, 1024, 1, 384] - - [36, 82.722] - - - [768, 512, 1, 384] - - [109, 28.582] - - - [24576, 384, 1, 384] - - [387, 77.179] - - - [9984, 384, 1, 384] - - [363, 66.812] - - - [28416, 4096, 1, 384] - - [53, 89.71] - - - [11904, 7937, 1, 384] - - [27, 87.153] - - - [22656, 512, 1, 384] - - [25, 74.849] - - - [32640, 16385, 1, 384] - - [32, 85.468] - - - [14592, 1024, 1, 384] - - [36, 77.259] - - - [29952, 13697, 1, 384] - - [37, 89.409] - - - [32640, 1024, 1, 384] - - [74, 79.399] - - - [24960, 512, 1, 384] - - [53, 73.643] - - - [24192, 384, 1, 384] - - [370, 76.515] - - - [10752, 512, 1, 384] - - [34, 58.246] - - - [25344, 8192, 1, 384] - - [36, 90.336] - - - [32256, 16385, 1, 384] - - [54, 85.964] - - - [18432, 10753, 1, 384] - - [37, 88.005] - - - [27648, 512, 1, 384] - - [29, 79.871] - - - [28800, 4096, 1, 384] - - [44, 89.288] - - - [13440, 512, 1, 384] - - [53, 71.196] - - - [22272, 2048, 1, 384] - - [53, 86.157] - - - [29184, 2048, 1, 384] - - [53, 86.734] - - - [29952, 8192, 1, 384] - - [58, 90.427] - - - [384, 385, 1, 384] - - [332, 10.485] - - - [33408, 17153, 1, 384] - - [27, 89.285] - - - [27264, 512, 1, 384] - - [53, 78.612] - - - [33792, 1024, 1, 384] - - [44, 83.815] - - - [12288, 384, 1, 384] - - [369, 66.011] - - - [4224, 1024, 1, 384] - - [34, 58.094] - - - [13056, 5377, 1, 384] - - [29, 86.897] - - - [9600, 5633, 1, 384] - - [58, 85.785] - - - [30336, 512, 1, 384] - - [27, 79.244] - - - [7680, 1024, 1, 384] - - [40, 68.745] - - - [14976, 384, 1, 384] - - [369, 69.592] - - - [11904, 512, 1, 384] - - [27, 63.746] - - - [16128, 512, 1, 384] - - [29, 70.223] - - - [16128, 8449, 1, 384] - - [29, 88.103] - - - [18432, 2048, 1, 384] - - [61, 83.507] - - - [32256, 1024, 1, 384] - - [53, 84.016] - - - [16896, 8833, 1, 384] - - [37, 89.018] - - - [11136, 7169, 1, 384] - - [62, 84.227] - - - [8832, 4865, 1, 384] - - [27, 84.203] - - - [13440, 4096, 1, 384] - - [53, 87.546] - - - [10752, 2048, 1, 384] - - [58, 80.532] - - - [27264, 1024, 1, 384] - - [27, 82.207] - - - [1536, 384, 1, 384] - - [339, 33.465] - - - [20352, 1024, 1, 384] - - [36, 81.987] - - - [30720, 512, 1, 384] - - [44, 80.004] - - - [16512, 512, 1, 384] - - [33, 67.144] - - - [20736, 4096, 1, 384] - - [53, 88.305] - - - [23424, 15745, 1, 384] - - [37, 89.763] - - - [24960, 2048, 1, 384] - - [44, 86.756] - - - [32256, 2048, 1, 384] - - [36, 87.968] - - - [10368, 384, 1, 384] - - [371, 68.69] - - - [14976, 7297, 1, 384] - - [395, 89.902] - - - [23040, 4096, 1, 384] - - [58, 88.538] - - - [16512, 4096, 1, 384] - - [76, 83.999] - - - [20736, 512, 1, 384] - - [37, 77.072] - - - [34560, 384, 1, 384] - - [388, 76.501] - - - [23040, 1024, 1, 384] - - [36, 80.485] - - - [5376, 384, 1, 384] - - [358, 52.152] - - - [11136, 512, 1, 384] - - [34, 60.033] - - - [19200, 512, 1, 384] - - [53, 72.362] - - - [19584, 11521, 1, 384] - - [27, 89.051] - - - [21504, 4096, 1, 384] - - [74, 86.954] - - - [25728, 4096, 1, 384] - - [53, 89.093] - - - [4992, 512, 1, 384] - - [59, 48.394] - - - [26880, 4096, 1, 384] - - [53, 88.944] - - - [31488, 15233, 1, 384] - - [24, 89.348] - - - [2304, 1409, 1, 384] - - [101, 59.428] - - - [28800, 1024, 1, 384] - - [36, 83.054] - - - [25344, 9473, 1, 384] - - [25, 89.006] - - - [13824, 4096, 1, 384] - - [58, 87.375] - - - [18048, 2048, 1, 384] - - [53, 85.495] - - - [13056, 512, 1, 384] - - [25, 69.355] - - - [31104, 8192, 1, 384] - - [27, 90.454] - - - [1152, 641, 1, 384] - - [114, 39.307] - - - [8064, 1024, 1, 384] - - [39, 70.051] - - - [7296, 512, 1, 384] - - [84, 50.843] - - - [12672, 4609, 1, 384] - - [25, 84.709] - - - [27264, 4096, 1, 384] - - [27, 87.94] - - - [11520, 2048, 1, 384] - - [53, 80.803] - - - [15744, 4096, 1, 384] - - [44, 88.574] - - - [19968, 512, 1, 384] - - [53, 75.065] - - - [5760, 2177, 1, 384] - - [35, 71.734] - - - [3840, 384, 1, 384] - - [353, 51.55] - - - [30336, 8192, 1, 384] - - [58, 90.763] - - - [28416, 8192, 1, 384] - - [53, 90.413] - - - [25344, 512, 1, 384] - - [44, 74.705] - - - [7296, 3713, 1, 384] - - [27, 80.855] - - - [28416, 384, 1, 384] - - [373, 75.08] - - - [19584, 2048, 1, 384] - - [53, 85.165] - - - [10368, 2048, 1, 384] - - [44, 83.477] - - - [33024, 4096, 1, 384] - - [53, 88.724] - - - [4224, 512, 1, 384] - - [34, 41.356] - - - [26496, 8192, 1, 384] - - [36, 90.525] - - - [768, 385, 1, 384] - - [331, 20.97] - - - [23040, 384, 1, 384] - - [384, 76.898] - - - [11520, 7937, 1, 384] - - [25, 87.554] - - - [28800, 384, 1, 384] - - [373, 75.824] - - - [8064, 4481, 1, 384] - - [27, 82.664] - - - [28032, 384, 1, 384] - - [389, 78.83] - - - [31104, 512, 1, 384] - - [27, 80.719] - - - [23808, 16129, 1, 384] - - [36, 90.108] - - - [29184, 384, 1, 384] - - [373, 76.231] - - - [9600, 512, 1, 384] - - [34, 64.587] - - - [26112, 512, 1, 384] - - [25, 76.532] - - - [31488, 8192, 1, 384] - - [29, 90.292] - - - [8448, 384, 1, 384] - - [369, 63.938] - - - [34944, 8192, 1, 384] - - [37, 90.6] - - - [4608, 3073, 1, 384] - - [27, 72.603] - - - [30720, 2048, 1, 384] - - [74, 85.262] - - - [34944, 512, 1, 384] - - [29, 82.01] - - - [27648, 8192, 1, 384] - - [74, 88.499] - - - [33024, 2048, 1, 384] - - [53, 87.021] - - - [26112, 4096, 1, 384] - - [36, 88.996] - - - [17280, 384, 1, 384] - - [370, 74.322] - - - [33024, 17153, 1, 384] - - [32, 88.628] - - - [14208, 2048, 1, 384] - - [53, 82.778] - - - [13440, 2048, 1, 384] - - [53, 82.8] - - - [1536, 641, 1, 384] - - [114, 39.611] - - - [8064, 4097, 1, 384] - - [56, 81.905] - - - [26496, 10625, 1, 384] - - [25, 88.98] - - - [33024, 384, 1, 384] - - [388, 73.665] - - - [26112, 384, 1, 384] - - [382, 75.246] - - - [23424, 15361, 1, 384] - - [56, 85.936] - - - [34944, 2048, 1, 384] - - [58, 88.407] - - - [32256, 512, 1, 384] - - [27, 77.446] - - - [23808, 15745, 1, 384] - - [53, 90.117] - - - [5760, 512, 1, 384] - - [27, 54.858] - - - [16128, 1024, 1, 384] - - [25, 76.093] - - - [31488, 4096, 1, 384] - - [58, 89.699] - - - [29568, 13313, 1, 384] - - [54, 86.436] - - - [18816, 11137, 1, 384] - - [25, 89.01] - - - [26496, 2048, 1, 384] - - [53, 86.456] - - - [1920, 384, 1, 384] - - [331, 37.415] - - - [31872, 1024, 1, 384] - - [36, 83.409] - - - [12672, 512, 1, 384] - - [53, 67.275] - - - [13056, 4096, 1, 384] - - [44, 87.556] - - - [17280, 1024, 1, 384] - - [44, 82.053] - - - [12288, 1024, 1, 384] - - [58, 78.16] - - - [1152, 512, 1, 384] - - [108, 32.6] - - - [31104, 15233, 1, 384] - - [25, 89.393] - - - [4608, 384, 1, 384] - - [353, 59.113] - - - [21888, 512, 1, 384] - - [29, 71.776] - - - [33408, 1024, 1, 384] - - [36, 86.352] - - - [8448, 2048, 1, 384] - - [58, 80.242] - - - [7296, 3329, 1, 384] - - [29, 82.434] - - - [10368, 6785, 1, 384] - - [25, 86.949] - - - [8832, 1024, 1, 384] - - [27, 67.434] - - - [31104, 1024, 1, 384] - - [53, 85.253] - - - [11520, 7553, 1, 384] - - [29, 86.704] - - - [34176, 4096, 1, 384] - - [58, 89.32] - - - [20352, 512, 1, 384] - - [44, 76.379] - - - [18432, 512, 1, 384] - - [23, 69.917] - - - [31488, 2048, 1, 384] - - [58, 88.404] - - - [9984, 6401, 1, 384] - - [27, 85.589] - - - [6144, 2048, 1, 384] - - [44, 78.784] - - - [22656, 2048, 1, 384] - - [53, 87.419] - - - [2304, 512, 1, 384] - - [108, 47.349] - - - [21504, 13441, 1, 384] - - [25, 88.86] - - - [1920, 1025, 1, 384] - - [35, 37.494] - - - [24960, 8705, 1, 384] - - [53, 88.641] - - - [16512, 2048, 1, 384] - - [61, 80.148] - - - [26880, 11009, 1, 384] - - [25, 89.218] - - - [32256, 4096, 1, 384] - - [26, 88.284] - - - [14976, 2048, 1, 384] - - [53, 82.793] - - - [21120, 512, 1, 384] - - [25, 78.172] - - - [31872, 512, 1, 384] - - [44, 76.429] - - - [8064, 2048, 1, 384] - - [29, 76.349] - - - [3072, 1024, 1, 384] - - [27, 58.286] - - - [23808, 2048, 1, 384] - - [44, 85.843] - - - [12672, 2048, 1, 384] - - [58, 83.079] - - - [19968, 4096, 1, 384] - - [58, 88.078] - - - [14976, 512, 1, 384] - - [25, 66.951] - - - [25344, 1024, 1, 384] - - [36, 82.986] - - - [31872, 15617, 1, 384] - - [37, 89.141] - - - [20352, 12673, 1, 384] - - [25, 89.358] - - - [11136, 384, 1, 384] - - [369, 68.837] - - - [32640, 8192, 1, 384] - - [32, 88.408] - - - [28800, 2048, 1, 384] - - [44, 87.385] - - - [22656, 1024, 1, 384] - - [58, 84.226] - - - [17280, 4096, 1, 384] - - [53, 87.951] - - - [17664, 4096, 1, 384] - - [58, 87.84] - - - [32640, 2048, 1, 384] - - [28, 83.394] - - - [28032, 11777, 1, 384] - - [58, 89.285] - - - [20352, 4096, 1, 384] - - [58, 88.403] - - - [33792, 512, 1, 384] - - [53, 80.242] - - - [24192, 4096, 1, 384] - - [44, 89.825] - - - [9216, 384, 1, 384] - - [369, 67.67] - - - [6912, 512, 1, 384] - - [78, 64.604] - - - [14208, 1024, 1, 384] - - [36, 75.175] - - - [26496, 512, 1, 384] - - [53, 77.278] - - - [4992, 384, 1, 384] - - [339, 57.491] - - - [33408, 512, 1, 384] - - [37, 79.233] - - - [3456, 1537, 1, 384] - - [373, 63.22] - - - [21888, 14209, 1, 384] - - [41, 84.605] - - - [24576, 8321, 1, 384] - - [25, 83.293] - - - [33792, 17921, 1, 384] - - [25, 88.783] - - - [13440, 384, 1, 384] - - [371, 70.825] - - - [18432, 384, 1, 384] - - [375, 74.867] - - - [6912, 1024, 1, 384] - - [40, 73.539] - - - [22272, 384, 1, 384] - - [382, 75.356] - - - [3840, 2305, 1, 384] - - [34, 66.085] - - - [6144, 1024, 1, 384] - - [53, 65.55] - - - [7680, 512, 1, 384] - - [55, 53.268] - - - [19584, 4096, 1, 384] - - [44, 88.664] - - - [23808, 4096, 1, 384] - - [44, 88.714] - - - [29568, 384, 1, 384] - - [373, 76.901] - - - [29184, 512, 1, 384] - - [27, 76.749] - - - [13056, 384, 1, 384] - - [372, 69.463] - - - [28032, 512, 1, 384] - - [36, 80.796] - - - [26880, 2048, 1, 384] - - [36, 87.507] - - - [18048, 9985, 1, 384] - - [27, 88.583] - - - [29952, 512, 1, 384] - - [37, 78.472] - - - [27648, 2048, 1, 384] - - [74, 85.281] - - - [29568, 13697, 1, 384] - - [30, 88.429] - - - [19584, 1024, 1, 384] - - [53, 79.278] - - - [27648, 384, 1, 384] - - [391, 78.481] - - - [6912, 384, 1, 384] - - [360, 63.006] - - - [26880, 1024, 1, 384] - - [53, 82.929] - - - [24960, 8192, 1, 384] - - [53, 90.373] - - - [13824, 1024, 1, 384] - - [53, 80.611] - - - [11904, 2048, 1, 384] - - [58, 83.021] - - - [34560, 2048, 1, 384] - - [53, 87.843] - - - [12288, 4609, 1, 384] - - [27, 83.753] - - - [21504, 13825, 1, 384] - - [29, 88.505] - - - [29184, 8192, 1, 384] - - [29, 90.304] - - - [12288, 4096, 1, 384] - - [25, 85.445] - - - [23424, 1024, 1, 384] - - [53, 81.746] - - - [14208, 512, 1, 384] - - [76, 63.839] - - - [25728, 512, 1, 384] - - [25, 75.456] - - - [29568, 2048, 1, 384] - - [93, 86.694] - - - [9600, 1024, 1, 384] - - [53, 72.644] - - - [29952, 384, 1, 384] - - [373, 77.657] - - - [18048, 4096, 1, 384] - - [53, 87.568] - - - [30336, 14081, 1, 384] - - [29, 89.07] - - - [24192, 8192, 1, 384] - - [53, 90.674] - - - [33792, 2048, 1, 384] - - [61, 85.928] - - - [6144, 384, 1, 384] - - [361, 57.994] - - - [8448, 1024, 1, 384] - - [40, 74.445] - - - [6528, 1024, 1, 384] - - [35, 69.453] - - - [18432, 10369, 1, 384] - - [29, 88.481] - - - [19968, 1024, 1, 384] - - [53, 80.777] - - - [23424, 512, 1, 384] - - [53, 76.757] - - - [20736, 2048, 1, 384] - - [58, 86.572] - - - [29184, 12929, 1, 384] - - [51, 89.169] - - - [3072, 1153, 1, 384] - - [371, 62.102] - - - [28416, 512, 1, 384] - - [58, 75.226] - - - [14592, 384, 1, 384] - - [366, 67.666] - - - [18432, 1024, 1, 384] - - [36, 80.413] - - - [29184, 13313, 1, 384] - - [54, 86.922] - - - [32640, 4096, 1, 384] - - [91, 86.09] - - - [21888, 13825, 1, 384] - - [59, 85.031] - - - [5376, 1024, 1, 384] - - [35, 57.879] - - - [4608, 1024, 1, 384] - - [35, 63.474] - - - [8832, 5249, 1, 384] - - [27, 84.238] - - - [14976, 4096, 1, 384] - - [58, 86.86] - - - [3840, 1024, 1, 384] - - [63, 53.498] - - - [24192, 16129, 1, 384] - - [58, 89.733] - - - [19968, 12289, 1, 384] - - [44, 86.208] - - - [1152, 384, 1, 384] - - [331, 30.499] - - - [27648, 4096, 1, 384] - - [61, 87.529] - - - [4992, 3073, 1, 384] - - [37, 77.085] - - - [33024, 8192, 1, 384] - - [62, 89.435] - - - [34944, 384, 1, 384] - - [395, 76.763] - - - [32, 28672, 1, 32] - - [7, 23.634] - - - [32, 24576, 1, 32] - - [6, 23.263] - - - [32, 16384, 1, 32] - - [3, 22.424] - - - [32, 20480, 1, 32] - - [5, 22.915] - - - [32, 12288, 1, 32] - - [3, 20.484] - - - [32, 8192, 1, 32] - - [4, 16.315] - - - [32, 4096, 1, 32] - - [2, 8.373] - - - [32, 32768, 1, 32] - - [1, 23.586] - - - [4224, 3840, 1, 4096] - - [15, 98.172] - - - [5376, 4096, 1, 4096] - - [16, 94.215] - - - [7040, 4096, 1, 384] - - [17, 86.756] - - - [7040, 4096, 1, 768] - - [15, 93.117] - - - [7040, 4096, 1, 1536] - - [18, 96.418] - - - [3840, 4224, 1, 4096] - - [8, 79.968] - - - [3840, 4224, 1, 4224] - - [9, 79.975] - - - [3840, 4224, 1, 4320] - - [10, 79.983] - - - [7680, 8448, 1, 8192] - - [11, 81.886] - - - [7680, 8448, 1, 8448] - - [11, 81.886] - - - [7680, 8448, 1, 8640] - - [11, 81.886] - - - [4096, 7169, 1, 512] - - [534, 88.167] - - - [4096, 7681, 1, 512] - - [402, 88.841] - - - [4096, 8193, 1, 512] - - [395, 87.86] - - - [4608, 512, 1, 512] - - [353, 61.028] - - - [4608, 8193, 1, 512] - - [528, 89.573] - - - [4608, 8705, 1, 512] - - [433, 89.242] - - - [4608, 9217, 1, 512] - - [529, 89.741] - - - [5120, 512, 1, 512] - - [354, 66.032] - - - [5120, 9217, 1, 512] - - [529, 90.059] - - - [5120, 9729, 1, 512] - - [395, 90.936] - - - [5120, 10241, 1, 512] - - [530, 90.535] - - - [5632, 512, 1, 512] - - [434, 60.928] - - - [5632, 10241, 1, 512] - - [529, 91.2] - - - [5632, 10753, 1, 512] - - [417, 91.858] - - - [5632, 11265, 1, 512] - - [533, 90.934] - - - [6144, 512, 1, 512] - - [366, 64.921] - - - [6144, 11265, 1, 512] - - [529, 91.689] - - - [6144, 11777, 1, 512] - - [402, 91.522] - - - [6144, 12289, 1, 512] - - [529, 91.486] - - - [6656, 512, 1, 512] - - [388, 68.733] - - - [6656, 12289, 1, 512] - - [529, 91.715] - - - [6656, 12801, 1, 512] - - [402, 92.349] - - - [6656, 13313, 1, 512] - - [529, 92.002] - - - [7168, 512, 1, 512] - - [435, 62.865] - - - [7168, 13313, 1, 512] - - [527, 92.196] - - - [7168, 13825, 1, 512] - - [532, 92.402] - - - [7168, 14337, 1, 512] - - [529, 92.768] - - - [7680, 512, 1, 512] - - [436, 66.322] - - - [7680, 14337, 1, 512] - - [529, 92.373] - - - [7680, 14849, 1, 512] - - [433, 92.877] - - - [7680, 15361, 1, 512] - - [530, 92.218] - - - [8192, 512, 1, 512] - - [369, 69.275] - - - [8192, 15361, 1, 512] - - [529, 92.959] - - - [8192, 15873, 1, 512] - - [408, 93.06] - - - [8192, 16385, 1, 512] - - [529, 93.033] - - - [8704, 512, 1, 512] - - [369, 71.705] - - - [8704, 16385, 1, 512] - - [529, 93.107] - - - [8704, 16897, 1, 512] - - [402, 93.129] - - - [8704, 17409, 1, 512] - - [529, 92.785] - - - [9216, 512, 1, 512] - - [533, 68.831] - - - [9216, 17409, 1, 512] - - [529, 93.08] - - - [9216, 17921, 1, 512] - - [402, 93.416] - - - [9216, 18433, 1, 512] - - [529, 92.976] - - - [9728, 512, 1, 512] - - [526, 72.309] - - - [9728, 18433, 1, 512] - - [528, 93.313] - - - [9728, 18945, 1, 512] - - [433, 93.384] - - - [9728, 19457, 1, 512] - - [526, 93.225] - - - [10240, 512, 1, 512] - - [437, 73.852] - - - [10240, 19457, 1, 512] - - [529, 93.374] - - - [10240, 19969, 1, 512] - - [438, 93.552] - - - [10240, 20481, 1, 512] - - [530, 93.181] - - - [10752, 512, 1, 512] - - [367, 69.447] - - - [10752, 20481, 1, 512] - - [530, 93.381] - - - [10752, 20993, 1, 512] - - [438, 93.522] - - - [10752, 21505, 1, 512] - - [526, 93.239] - - - [11264, 512, 1, 512] - - [369, 71.35] - - - [11264, 21505, 1, 512] - - [529, 93.509] - - - [11264, 22017, 1, 512] - - [433, 93.65] - - - [11264, 22529, 1, 512] - - [529, 93.548] - - - [11776, 512, 1, 512] - - [369, 73.606] - - - [11776, 22529, 1, 512] - - [533, 93.551] - - - [11776, 23041, 1, 512] - - [402, 93.761] - - - [11776, 23553, 1, 512] - - [529, 93.625] - - - [12288, 512, 1, 512] - - [373, 75.45] - - - [12288, 23553, 1, 512] - - [528, 93.541] - - - [12288, 24065, 1, 512] - - [438, 93.93] - - - [12288, 24577, 1, 512] - - [525, 93.635] - - - [12800, 512, 1, 512] - - [534, 74.983] - - - [12800, 24577, 1, 512] - - [529, 93.576] - - - [12800, 25089, 1, 512] - - [408, 93.881] - - - [12800, 25601, 1, 512] - - [530, 93.708] - - - [13312, 512, 1, 512] - - [439, 75.137] - - - [13312, 25601, 1, 512] - - [529, 93.705] - - - [13312, 26113, 1, 512] - - [438, 93.955] - - - [13312, 26625, 1, 512] - - [529, 93.722] - - - [13824, 512, 1, 512] - - [440, 77.693] - - - [13824, 26625, 1, 512] - - [529, 93.719] - - - [13824, 27137, 1, 512] - - [438, 94.152] - - - [13824, 27649, 1, 512] - - [528, 93.845] - - - [14336, 512, 1, 512] - - [373, 72.843] - - - [14336, 27649, 1, 512] - - [529, 93.819] - - - [14336, 28161, 1, 512] - - [395, 94.127] - - - [14336, 28673, 1, 512] - - [526, 93.934] - - - [14848, 512, 1, 512] - - [373, 74.864] - - - [14848, 28673, 1, 512] - - [526, 93.904] - - - [14848, 29185, 1, 512] - - [441, 94.19] - - - [14848, 29697, 1, 512] - - [526, 93.917] - - - [15360, 512, 1, 512] - - [373, 76.564] - - - [15360, 29697, 1, 512] - - [529, 93.898] - - - [15360, 30209, 1, 512] - - [441, 94.311] - - - [15360, 30721, 1, 512] - - [529, 93.948] - - - [15872, 512, 1, 512] - - [531, 78.265] - - - [15872, 30721, 1, 512] - - [530, 93.935] - - - [15872, 31233, 1, 512] - - [402, 94.373] - - - [15872, 31745, 1, 512] - - [529, 94.005] - - - [16384, 512, 1, 512] - - [529, 80.377] - - - [16384, 31745, 1, 512] - - [529, 93.967] - - - [16384, 32257, 1, 512] - - [438, 94.453] - - - [16384, 32769, 1, 512] - - [528, 93.942] - - - [16896, 512, 1, 512] - - [442, 79.05] - - - [16896, 32769, 1, 512] - - [526, 93.994] - - - [16896, 33281, 1, 512] - - [408, 94.381] - - - [16896, 33793, 1, 512] - - [525, 94.043] - - - [17408, 512, 1, 512] - - [394, 80.426] - - - [17408, 33793, 1, 512] - - [529, 94.061] - - - [17408, 34305, 1, 512] - - [438, 94.564] - - - [17408, 34817, 1, 512] - - [528, 94.06] - - - [17920, 512, 1, 512] - - [443, 77.178] - - - [17920, 34817, 1, 512] - - [529, 94.089] - - - [17920, 35329, 1, 512] - - [441, 94.567] - - - [17920, 35841, 1, 512] - - [526, 94.147] - - - [18432, 512, 1, 512] - - [532, 78.257] - - - [18432, 35841, 1, 512] - - [526, 94.155] - - - [18432, 36353, 1, 512] - - [438, 94.591] - - - [18432, 36865, 1, 512] - - [528, 94.111] - - - [18944, 512, 1, 512] - - [532, 80.276] - - - [18944, 36865, 1, 512] - - [528, 94.165] - - - [18944, 37377, 1, 512] - - [402, 94.622] - - - [18944, 37889, 1, 512] - - [529, 94.122] - - - [19456, 512, 1, 512] - - [530, 81.688] - - - [19456, 37889, 1, 512] - - [528, 94.133] - - - [19456, 38401, 1, 512] - - [402, 94.675] - - - [19456, 38913, 1, 512] - - [529, 94.15] - - - [19968, 512, 1, 512] - - [395, 79.452] - - - [19968, 38913, 1, 512] - - [529, 94.163] - - - [19968, 39425, 1, 512] - - [395, 94.652] - - - [19968, 39937, 1, 512] - - [529, 94.171] - - - [20480, 512, 1, 512] - - [444, 81.161] - - - [20480, 39937, 1, 512] - - [395, 94.182] - - - [20480, 40449, 1, 512] - - [395, 94.748] - - - [20480, 40961, 1, 512] - - [528, 94.182] - - - [20992, 512, 1, 512] - - [445, 82.746] - - - [20992, 40961, 1, 512] - - [528, 94.179] - - - [20992, 41473, 1, 512] - - [402, 94.748] - - - [20992, 41985, 1, 512] - - [526, 94.211] - - - [21504, 512, 1, 512] - - [525, 79.684] - - - [21504, 41985, 1, 512] - - [526, 94.229] - - - [21504, 42497, 1, 512] - - [395, 94.802] - - - [21504, 43009, 1, 512] - - [526, 94.283] - - - [22016, 512, 1, 512] - - [526, 81.34] - - - [22016, 43009, 1, 512] - - [526, 94.27] - - - [22016, 43521, 1, 512] - - [395, 94.793] - - - [22016, 44033, 1, 512] - - [526, 94.257] - - - [22528, 512, 1, 512] - - [525, 82.888] - - - [22528, 44033, 1, 512] - - [526, 94.263] - - - [22528, 44545, 1, 512] - - [438, 94.762] - - - [22528, 45057, 1, 512] - - [528, 94.261] - - - [23040, 512, 1, 512] - - [433, 79.889] - - - [23040, 45057, 1, 512] - - [533, 94.257] - - - [23040, 45569, 1, 512] - - [395, 94.882] - - - [23040, 46081, 1, 512] - - [528, 94.308] - - - [23552, 512, 1, 512] - - [408, 81.569] - - - [23552, 46081, 1, 512] - - [528, 94.276] - - - [23552, 46593, 1, 512] - - [395, 94.852] - - - [23552, 47105, 1, 512] - - [529, 94.291] - - - [24064, 512, 1, 512] - - [399, 83.075] - - - [24064, 47105, 1, 512] - - [529, 94.28] - - - [24064, 47617, 1, 512] - - [395, 94.911] - - - [24064, 48129, 1, 512] - - [529, 94.29] - - - [24576, 512, 1, 512] - - [399, 84.031] - - - [24576, 48129, 1, 512] - - [529, 94.283] - - - [24576, 48641, 1, 512] - - [395, 94.939] - - - [24576, 49153, 1, 512] - - [526, 94.311] - - - [25088, 512, 1, 512] - - [528, 82.17] - - - [25088, 49153, 1, 512] - - [528, 94.313] - - - [25088, 49665, 1, 512] - - [441, 94.917] - - - [25088, 50177, 1, 512] - - [526, 94.353] - - - [25600, 512, 1, 512] - - [530, 83.475] - - - [25600, 50177, 1, 512] - - [526, 94.346] - - - [25600, 50689, 1, 512] - - [441, 94.933] - - - [25600, 51201, 1, 512] - - [528, 94.34] - - - [26112, 512, 1, 512] - - [532, 84.967] - - - [26112, 51201, 1, 512] - - [529, 94.34] - - - [26112, 51713, 1, 512] - - [395, 94.958] - - - [26112, 52225, 1, 512] - - [526, 94.341] - - - [26624, 512, 1, 512] - - [433, 81.843] - - - [26624, 52225, 1, 512] - - [526, 94.342] - - - [26624, 52737, 1, 512] - - [395, 94.955] - - - [26624, 53249, 1, 512] - - [528, 94.363] - - - [27136, 512, 1, 512] - - [442, 83.179] - - - [27136, 53249, 1, 512] - - [528, 94.349] - - - [27136, 53761, 1, 512] - - [395, 94.96] - - - [27136, 54273, 1, 512] - - [528, 94.377] - - - [27648, 512, 1, 512] - - [399, 84.333] - - - [27648, 54273, 1, 512] - - [526, 94.371] - - - [27648, 54785, 1, 512] - - [395, 94.942] - - - [27648, 55297, 1, 512] - - [528, 94.393] - - - [28160, 512, 1, 512] - - [442, 85.343] - - - [28160, 55297, 1, 512] - - [528, 94.381] - - - [28160, 55809, 1, 512] - - [433, 94.674] - - - [28160, 56321, 1, 512] - - [526, 94.39] - - - [28672, 512, 1, 512] - - [530, 84.296] - - - [28672, 56321, 1, 512] - - [528, 94.375] - - - [28672, 56833, 1, 512] - - [395, 94.976] - - - [28672, 57345, 1, 512] - - [526, 94.387] - - - [29184, 512, 1, 512] - - [533, 85.42] - - - [29184, 57345, 1, 512] - - [526, 94.393] - - - [29184, 57857, 1, 512] - - [395, 94.988] - - - [29184, 58369, 1, 512] - - [528, 94.388] - - - [29696, 512, 1, 512] - - [402, 82.196] - - - [29696, 58369, 1, 512] - - [528, 94.395] - - - [29696, 58881, 1, 512] - - [395, 94.976] - - - [29696, 59393, 1, 512] - - [526, 94.406] - - - [30208, 512, 1, 512] - - [408, 83.301] - - - [30208, 59393, 1, 512] - - [526, 94.405] - - - [30208, 59905, 1, 512] - - [441, 94.955] - - - [30208, 60417, 1, 512] - - [526, 94.426] - - - [30720, 512, 1, 512] - - [402, 84.261] - - - [30720, 60417, 1, 512] - - [529, 94.409] - - - [30720, 60929, 1, 512] - - [395, 94.983] - - - [30720, 61441, 1, 512] - - [526, 94.417] - - - [31232, 512, 1, 512] - - [395, 85.605] - - - [31232, 61441, 1, 512] - - [528, 94.414] - - - [31232, 61953, 1, 512] - - [441, 94.923] - - - [31232, 62465, 1, 512] - - [526, 94.426] - - - [31744, 512, 1, 512] - - [532, 84.935] - - - [31744, 62465, 1, 512] - - [526, 94.415] - - - [31744, 62977, 1, 512] - - [395, 95.029] - - - [31744, 63489, 1, 512] - - [528, 94.416] - - - [32256, 512, 1, 512] - - [532, 85.881] - - - [32256, 63489, 1, 512] - - [526, 94.433] - - - [32256, 64001, 1, 512] - - [395, 95.023] - - - [32256, 64513, 1, 512] - - [528, 94.441] - - - [32768, 512, 1, 512] - - [533, 86.874] - - - [32768, 64513, 1, 512] - - [526, 94.447] - - - [32768, 65025, 1, 512] - - [441, 95.037] - - - [32768, 65537, 1, 512] - - [526, 94.427] - - - [33280, 512, 1, 512] - - [439, 83.067] - - - [33280, 65537, 1, 512] - - [526, 94.432] - - - [33280, 66049, 1, 512] - - [395, 94.996] - - - [33280, 66561, 1, 512] - - [526, 94.445] - - - [33792, 512, 1, 512] - - [407, 84.541] - - - [33792, 66561, 1, 512] - - [526, 94.465] - - - [33792, 67073, 1, 512] - - [408, 94.912] - - - [33792, 67585, 1, 512] - - [528, 94.45] - - - [34304, 512, 1, 512] - - [417, 85.298] - - - [34304, 67585, 1, 512] - - [528, 94.462] - - - [34304, 68097, 1, 512] - - [395, 94.973] - - - [34304, 68609, 1, 512] - - [526, 94.459] - - - [34816, 512, 1, 512] - - [417, 86.299] - - - [34816, 68609, 1, 512] - - [526, 94.47] - - - [34816, 69121, 1, 512] - - [402, 94.949] - - - [34816, 69633, 1, 512] - - [528, 94.467] - - - [35328, 512, 1, 512] - - [534, 86.267] - - - [35328, 69633, 1, 512] - - [528, 94.476] - - - [35328, 70145, 1, 512] - - [402, 94.951] - - - [35328, 70657, 1, 512] - - [528, 94.461] - - - [35840, 512, 1, 512] - - [533, 87.091] - - - [35840, 70657, 1, 512] - - [526, 94.486] - - - [35840, 71169, 1, 512] - - [402, 94.948] - - - [35840, 71681, 1, 512] - - [526, 94.501] - - - [36352, 512, 1, 512] - - [407, 83.258] - - - [36352, 71681, 1, 512] - - [526, 94.487] - - - [36352, 72193, 1, 512] - - [402, 94.927] - - - [36352, 72705, 1, 512] - - [526, 94.494] - - - [36864, 512, 1, 512] - - [407, 84.381] - - - [36864, 72705, 1, 512] - - [526, 94.484] - - - [36864, 73217, 1, 512] - - [408, 94.905] - - - [36864, 73729, 1, 512] - - [528, 94.489] - - - [37376, 512, 1, 512] - - [417, 85.141] - - - [37376, 73729, 1, 512] - - [528, 94.499] - - - [37376, 74241, 1, 512] - - [408, 94.923] - - - [37376, 74753, 1, 512] - - [528, 94.476] - - - [37888, 512, 1, 512] - - [438, 85.959] - - - [37888, 74753, 1, 512] - - [528, 94.492] - - - [37888, 75265, 1, 512] - - [402, 94.91] - - - [37888, 75777, 1, 512] - - [526, 94.498] - - - [38400, 512, 1, 512] - - [417, 87.188] - - - [38400, 75777, 1, 512] - - [526, 94.496] - - - [38400, 76289, 1, 512] - - [402, 94.918] - - - [38400, 76801, 1, 512] - - [448, 93.561] - - - [38912, 512, 1, 512] - - [531, 87.416] - - - [38912, 76801, 1, 512] - - [395, 93.586] - - - [38912, 77313, 1, 512] - - [402, 94.872] - - - [38912, 77825, 1, 512] - - [448, 93.387] - - - [39424, 512, 1, 512] - - [529, 83.464] - - - [39424, 77825, 1, 512] - - [448, 93.325] - - - [39424, 78337, 1, 512] - - [408, 94.835] - - - [39424, 78849, 1, 512] - - [448, 93.112] - - - [39936, 512, 1, 512] - - [533, 84.52] - - - [39936, 78849, 1, 512] - - [448, 93.438] - - - [39936, 79361, 1, 512] - - [408, 94.821] - - - [39936, 79873, 1, 512] - - [395, 93.36] - - - [40448, 512, 1, 512] - - [527, 85.041] - - - [40448, 79873, 1, 512] - - [395, 93.357] - - - [40448, 80385, 1, 512] - - [402, 94.832] - - - [40448, 80897, 1, 512] - - [448, 93.21] - - - [40960, 512, 1, 512] - - [526, 85.901] - - - [40960, 80897, 1, 512] - - [448, 93.29] - - - [40960, 81409, 1, 512] - - [408, 94.786] - - - [40960, 81921, 1, 512] - - [440, 91.125] - - - [41472, 512, 1, 512] - - [531, 86.786] - - - [41472, 81921, 1, 512] - - [440, 91.266] - - - [41472, 82433, 1, 512] - - [408, 94.779] - - - [41472, 82945, 1, 512] - - [448, 93.081] - - - [41984, 512, 1, 512] - - [533, 87.652] - - - [41984, 82945, 1, 512] - - [395, 92.87] - - - [41984, 83457, 1, 512] - - [450, 94.639] - - - [41984, 83969, 1, 512] - - [448, 92.824] - - - [42496, 512, 1, 512] - - [527, 88.572] - - - [42496, 83969, 1, 512] - - [395, 92.525] - - - [42496, 84481, 1, 512] - - [450, 94.621] - - - [42496, 84993, 1, 512] - - [448, 92.987] - - - [43008, 512, 1, 512] - - [529, 84.869] - - - [43008, 84993, 1, 512] - - [395, 92.72] - - - [43008, 85505, 1, 512] - - [450, 94.488] - - - [43008, 86017, 1, 512] - - [448, 92.774] - - - [43520, 512, 1, 512] - - [529, 85.38] - - - [43520, 86017, 1, 512] - - [448, 92.82] - - - [43520, 86529, 1, 512] - - [450, 94.487] - - - [43520, 87041, 1, 512] - - [395, 92.132] - - - [44032, 512, 1, 512] - - [530, 86.087] - - - [44032, 87041, 1, 512] - - [395, 92.595] - - - [44032, 87553, 1, 512] - - [450, 94.391] - - - [44032, 88065, 1, 512] - - [448, 92.292] - - - [44544, 512, 1, 512] - - [530, 86.966] - - - [44544, 88065, 1, 512] - - [448, 92.202] - - - [44544, 88577, 1, 512] - - [433, 94.274] - - - [44544, 89089, 1, 512] - - [433, 91.928] - - - [45056, 512, 1, 512] - - [531, 87.869] - - - [45056, 89089, 1, 512] - - [395, 92.193] - - - [45056, 89601, 1, 512] - - [433, 94.136] - - - [45056, 90113, 1, 512] - - [440, 91.652] - - - [45568, 512, 1, 512] - - [533, 88.702] - - - [45568, 90113, 1, 512] - - [440, 91.682] - - - [45568, 90625, 1, 512] - - [433, 94.145] - - - [45568, 91137, 1, 512] - - [433, 91.84] - - - [46080, 512, 1, 512] - - [526, 85.174] - - - [46080, 91137, 1, 512] - - [433, 91.935] - - - [46080, 91649, 1, 512] - - [433, 94.076] - - - [46080, 92161, 1, 512] - - [448, 91.813] - - - [46592, 512, 1, 512] - - [529, 85.815] - - - [46592, 92161, 1, 512] - - [395, 91.898] - - - [46592, 92673, 1, 512] - - [433, 93.999] - - - [46592, 93185, 1, 512] - - [440, 91.814] - - - [47104, 512, 1, 512] - - [525, 86.516] - - - [47104, 93185, 1, 512] - - [433, 91.757] - - - [47104, 93697, 1, 512] - - [433, 93.977] - - - [47104, 94209, 1, 512] - - [448, 91.808] - - - [47616, 512, 1, 512] - - [534, 87.277] - - - [47616, 94209, 1, 512] - - [440, 91.774] - - - [47616, 94721, 1, 512] - - [445, 93.91] - - - [47616, 95233, 1, 512] - - [433, 91.74] - - - [48128, 512, 1, 512] - - [529, 87.931] - - - [48128, 95233, 1, 512] - - [440, 91.733] - - - [48128, 95745, 1, 512] - - [433, 93.866] - - - [48128, 96257, 1, 512] - - [433, 91.803] - - - [48640, 512, 1, 512] - - [525, 88.711] - - - [48640, 96257, 1, 512] - - [433, 91.726] - - - [48640, 96769, 1, 512] - - [433, 93.801] - - - [48640, 97281, 1, 512] - - [415, 91.757] - - - [49152, 512, 1, 512] - - [417, 87.527] - - - [49152, 97281, 1, 512] - - [415, 91.678] - - - [49152, 97793, 1, 512] - - [445, 93.737] - - - [49152, 98305, 1, 512] - - [451, 90.818] - - - [49664, 512, 1, 512] - - [532, 86.176] - - - [49664, 98305, 1, 512] - - [451, 90.781] - - - [49664, 98817, 1, 512] - - [445, 93.678] - - - [49664, 99329, 1, 512] - - [433, 91.658] - - - [50176, 512, 1, 512] - - [532, 86.704] - - - [50176, 99329, 1, 512] - - [433, 91.653] - - - [50176, 99841, 1, 512] - - [445, 93.647] - - - [50176, 100353, 1, 512] - - [433, 91.585] - - - [50688, 512, 1, 512] - - [526, 87.298] - - - [50688, 100353, 1, 512] - - [433, 91.56] - - - [50688, 100865, 1, 512] - - [433, 93.567] - - - [50688, 101377, 1, 512] - - [433, 91.57] - - - [51200, 512, 1, 512] - - [526, 88.108] - - - [51200, 101377, 1, 512] - - [433, 91.563] - - - [51200, 101889, 1, 512] - - [445, 93.502] - - - [51200, 102401, 1, 512] - - [433, 91.443] - - - [51712, 512, 1, 512] - - [526, 88.879] - - - [51712, 102401, 1, 512] - - [433, 91.549] - - - [51712, 102913, 1, 512] - - [445, 93.47] - - - [51712, 103425, 1, 512] - - [433, 91.579] - - - [52224, 512, 1, 512] - - [533, 89.512] - - - [52224, 103425, 1, 512] - - [433, 91.494] - - - [52224, 103937, 1, 512] - - [445, 93.436] - - - [52224, 104449, 1, 512] - - [433, 91.41] - - - [52736, 512, 1, 512] - - [433, 88.847] - - - [52736, 104449, 1, 512] - - [433, 91.456] - - - [52736, 104961, 1, 512] - - [445, 93.35] - - - [52736, 105473, 1, 512] - - [433, 91.43] - - - [53248, 512, 1, 512] - - [531, 86.995] - - - [53248, 105473, 1, 512] - - [433, 91.37] - - - [53248, 105985, 1, 512] - - [445, 93.253] - - - [53248, 106497, 1, 512] - - [433, 91.059] - - - [53760, 512, 1, 512] - - [533, 87.557] - - - [53760, 106497, 1, 512] - - [433, 91.065] - - - [53760, 107009, 1, 512] - - [445, 93.249] - - - [53760, 107521, 1, 512] - - [433, 91.359] - - - [54272, 512, 1, 512] - - [527, 88.104] - - - [54272, 107521, 1, 512] - - [433, 91.276] - - - [54272, 108033, 1, 512] - - [445, 93.158] - - - [54272, 108545, 1, 512] - - [433, 91.189] - - - [54784, 512, 1, 512] - - [533, 88.877] - - - [54784, 108545, 1, 512] - - [433, 91.243] - - - [54784, 109057, 1, 512] - - [445, 93.098] - - - [54784, 109569, 1, 512] - - [433, 91.207] - - - [55296, 512, 1, 512] - - [526, 89.406] - - - [55296, 109569, 1, 512] - - [433, 91.139] - - - [55296, 110081, 1, 512] - - [445, 93.028] - - - [55296, 110593, 1, 512] - - [433, 91.032] - - - [55808, 512, 1, 512] - - [417, 88.484] - - - [55808, 110593, 1, 512] - - [433, 91.246] - - - [55808, 111105, 1, 512] - - [445, 92.994] - - - [55808, 111617, 1, 512] - - [433, 91.085] - - - [56320, 512, 1, 512] - - [433, 89.044] - - - [56320, 111617, 1, 512] - - [452, 90.97] - - - [56320, 112129, 1, 512] - - [445, 92.878] - - - [56320, 112641, 1, 512] - - [433, 90.893] - - - [56832, 512, 1, 512] - - [532, 87.658] - - - [56832, 112641, 1, 512] - - [452, 91.051] - - - [56832, 113153, 1, 512] - - [445, 92.866] - - - [56832, 113665, 1, 512] - - [452, 91.145] - - - [57344, 512, 1, 512] - - [532, 88.176] - - - [57344, 113665, 1, 512] - - [452, 91.103] - - - [57344, 114177, 1, 512] - - [445, 92.756] - - - [57344, 114689, 1, 512] - - [453, 89.228] - - - [57856, 512, 1, 512] - - [526, 88.821] - - - [57856, 114689, 1, 512] - - [454, 89.444] - - - [57856, 115201, 1, 512] - - [445, 92.668] - - - [57856, 115713, 1, 512] - - [452, 91.071] - - - [58368, 512, 1, 512] - - [532, 89.475] - - - [58368, 115713, 1, 512] - - [452, 90.936] - - - [58368, 116225, 1, 512] - - [445, 92.562] - - - [58368, 116737, 1, 512] - - [452, 90.921] - - - [58880, 512, 1, 512] - - [529, 89.984] - - - [58880, 116737, 1, 512] - - [452, 91.059] - - - [58880, 117249, 1, 512] - - [445, 92.463] - - - [58880, 117761, 1, 512] - - [452, 91.032] - - - [59392, 512, 1, 512] - - [433, 88.949] - - - [59392, 117761, 1, 512] - - [452, 90.935] - - - [59392, 118273, 1, 512] - - [445, 92.359] - - - [59392, 118785, 1, 512] - - [455, 90.655] - - - [59904, 512, 1, 512] - - [529, 87.843] - - - [59904, 118785, 1, 512] - - [455, 90.677] - - - [59904, 119297, 1, 512] - - [445, 92.282] - - - [59904, 119809, 1, 512] - - [452, 91.074] - - - [60416, 512, 1, 512] - - [532, 88.386] - - - [60416, 119809, 1, 512] - - [452, 91.021] - - - [60416, 120321, 1, 512] - - [445, 92.17] - - - [60416, 120833, 1, 512] - - [452, 90.865] - - - [60928, 512, 1, 512] - - [533, 88.947] - - - [60928, 120833, 1, 512] - - [452, 91.052] - - - [60928, 121345, 1, 512] - - [445, 92.033] - - - [60928, 121857, 1, 512] - - [452, 90.997] - - - [61440, 512, 1, 512] - - [525, 89.58] - - - [61440, 121857, 1, 512] - - [452, 91.095] - - - [61440, 122369, 1, 512] - - [445, 91.857] - - - [61440, 122881, 1, 512] - - [456, 90.354] - - - [61952, 512, 1, 512] - - [529, 90.063] - - - [61952, 122881, 1, 512] - - [456, 90.251] - - - [61952, 123393, 1, 512] - - [445, 91.691] - - - [61952, 123905, 1, 512] - - [452, 90.928] - - - [62464, 512, 1, 512] - - [417, 88.778] - - - [62464, 123905, 1, 512] - - [452, 91.029] - - - [62464, 124417, 1, 512] - - [445, 91.578] - - - [62464, 124929, 1, 512] - - [455, 90.749] - - - [62976, 512, 1, 512] - - [417, 89.013] - - - [62976, 124929, 1, 512] - - [452, 91.044] - - - [62976, 125441, 1, 512] - - [445, 91.493] - - - [62976, 125953, 1, 512] - - [452, 90.97] - - - [63488, 512, 1, 512] - - [531, 88.507] - - - [63488, 125953, 1, 512] - - [452, 91.013] - - - [63488, 126465, 1, 512] - - [457, 91.46] - - - [63488, 126977, 1, 512] - - [455, 90.619] - - - [64000, 512, 1, 512] - - [530, 88.992] - - - [64000, 126977, 1, 512] - - [455, 90.662] - - - [64000, 127489, 1, 512] - - [13, 84.172] - - - [64000, 128001, 1, 512] - - [12, 84.304] - - - [64512, 512, 1, 512] - - [531, 89.515] - - - [64512, 128001, 1, 512] - - [13, 84.344] - - - [4096, 4096, 1, 4128] - - [14, 87.241] - - - [25600, 25600, 1, 512] - - [526, 94.545] - - - [512, 512, 1, 512] - - [334, 21.149] - - - [1024, 512, 1, 512] - - [421, 33.417] - - - [1536, 512, 1, 512] - - [423, 43.936] - - - [1536, 1024, 1, 512] - - [460, 57.881] - - - [2048, 512, 1, 512] - - [427, 43.751] - - - [2048, 1024, 1, 512] - - [462, 55.811] - - - [2560, 512, 1, 512] - - [431, 51.346] - - - [2560, 1024, 1, 512] - - [361, 65.389] - - - [2560, 1536, 1, 512] - - [464, 70.881] - - - [3072, 512, 1, 512] - - [353, 58.252] - - - [3072, 1024, 1, 512] - - [465, 63.893] - - - [3072, 1536, 1, 512] - - [531, 68.831] - - - [3584, 512, 1, 512] - - [432, 55.564] - - - [3584, 1536, 1, 512] - - [367, 69.447] - - - [3584, 2048, 1, 512] - - [468, 73.443] - - - [4096, 512, 1, 512] - - [369, 55.619] - - - [4096, 1536, 1, 512] - - [369, 75.098] - - - [4096, 2048, 1, 512] - - [439, 76.685] - - - [4608, 2048, 1, 512] - - [533, 78.151] - - - [4608, 2560, 1, 512] - - [464, 81.688] - - - [5120, 2048, 1, 512] - - [407, 80.566] - - - [5120, 2560, 1, 512] - - [528, 83.759] - - - [5632, 2560, 1, 512] - - [448, 85.718] - - - [5632, 3072, 1, 512] - - [400, 84.22] - - - [6144, 2560, 1, 512] - - [394, 84.556] - - - [6144, 3072, 1, 512] - - [468, 84.809] - - - [6656, 3072, 1, 512] - - [529, 84.291] - - - [6656, 3584, 1, 512] - - [472, 86.435] - - - [7168, 3072, 1, 512] - - [529, 84.829] - - - [7168, 3584, 1, 512] - - [529, 86.932] - - - [7680, 3584, 1, 512] - - [534, 87.832] - - - [7680, 4096, 1, 512] - - [529, 89.747] - - - [8192, 3584, 1, 512] - - [533, 88.285] - - - [8192, 4096, 1, 512] - - [400, 88.705] - - - [8704, 4096, 1, 512] - - [407, 89.258] - - - [8704, 4608, 1, 512] - - [527, 91.333] - - - [9216, 4096, 1, 512] - - [439, 89.904] - - - [9216, 4608, 1, 512] - - [408, 89.962] - - - [9728, 4608, 1, 512] - - [526, 91.195] - - - [9728, 5120, 1, 512] - - [529, 91.586] - - - [10240, 4608, 1, 512] - - [528, 90.0] - - - [10240, 5120, 1, 512] - - [529, 90.725] - - - [10752, 5120, 1, 512] - - [529, 92.107] - - - [10752, 5632, 1, 512] - - [531, 91.088] - - - [11264, 5120, 1, 512] - - [407, 91.594] - - - [11264, 5632, 1, 512] - - [529, 92.54] - - - [11776, 5632, 1, 512] - - [530, 92.147] - - - [11776, 6144, 1, 512] - - [532, 91.717] - - - [12288, 5632, 1, 512] - - [530, 91.84] - - - [12288, 6144, 1, 512] - - [473, 91.959] - - - [12800, 6144, 1, 512] - - [528, 93.082] - - - [12800, 6656, 1, 512] - - [526, 93.145] - - - [13312, 6144, 1, 512] - - [529, 92.998] - - - [13312, 6656, 1, 512] - - [529, 93.233] - - - [13824, 6656, 1, 512] - - [526, 93.357] - - - [13824, 7168, 1, 512] - - [407, 92.689] - - - [14336, 6656, 1, 512] - - [526, 93.435] - - - [14336, 7168, 1, 512] - - [529, 92.858] - - - [14848, 7168, 1, 512] - - [529, 93.13] - - - [14848, 7680, 1, 512] - - [526, 92.822] - - - [15360, 7168, 1, 512] - - [529, 93.404] - - - [15360, 7680, 1, 512] - - [529, 93.201] - - - [15872, 7680, 1, 512] - - [533, 93.575] - - - [15872, 8192, 1, 512] - - [529, 93.537] - - - [16384, 7680, 1, 512] - - [408, 93.552] - - - [16384, 8192, 1, 512] - - [525, 93.168] - - - [16896, 8192, 1, 512] - - [528, 93.64] - - - [16896, 8704, 1, 512] - - [526, 93.837] - - - [17408, 8192, 1, 512] - - [531, 93.259] - - - [17408, 8704, 1, 512] - - [528, 93.583] - - - [17920, 8704, 1, 512] - - [402, 93.555] - - - [17920, 9216, 1, 512] - - [529, 93.722] - - - [18432, 8704, 1, 512] - - [529, 93.9] - - - [18432, 9216, 1, 512] - - [528, 93.63] - - - [18944, 9216, 1, 512] - - [528, 93.558] - - - [18944, 9728, 1, 512] - - [529, 93.48] - - - [19456, 9216, 1, 512] - - [535, 93.504] - - - [19456, 9728, 1, 512] - - [526, 93.502] - - - [19968, 9728, 1, 512] - - [528, 93.528] - - - [19968, 10240, 1, 512] - - [526, 93.622] - - - [20480, 9728, 1, 512] - - [528, 93.597] - - - [20480, 10240, 1, 512] - - [526, 93.733] - - - [20992, 10240, 1, 512] - - [535, 93.785] - - - [20992, 10752, 1, 512] - - [528, 93.981] - - - [21504, 10240, 1, 512] - - [528, 93.883] - - - [21504, 10752, 1, 512] - - [526, 93.762] - - - [22016, 10752, 1, 512] - - [535, 93.917] - - - [22016, 11264, 1, 512] - - [531, 93.852] - - - [22528, 10752, 1, 512] - - [529, 94.0] - - - [22528, 11264, 1, 512] - - [528, 93.977] - - - [23040, 11264, 1, 512] - - [525, 93.814] - - - [23040, 11776, 1, 512] - - [533, 93.827] - - - [23552, 11264, 1, 512] - - [528, 93.992] - - - [23552, 11776, 1, 512] - - [525, 94.036] - - - [24064, 11776, 1, 512] - - [535, 93.97] - - - [24064, 12288, 1, 512] - - [530, 94.06] - - - [24576, 11776, 1, 512] - - [529, 94.176] - - - [24576, 12288, 1, 512] - - [526, 94.082] - - - [25088, 12288, 1, 512] - - [529, 93.986] - - - [25088, 12800, 1, 512] - - [526, 94.136] - - - [25600, 12288, 1, 512] - - [526, 94.169] - - - [25600, 12800, 1, 512] - - [529, 94.153] - - - [26112, 12800, 1, 512] - - [525, 94.229] - - - [26112, 13312, 1, 512] - - [525, 94.275] - - - [26624, 12800, 1, 512] - - [529, 94.139] - - - [26624, 13312, 1, 512] - - [528, 94.186] - - - [27136, 13312, 1, 512] - - [526, 94.213] - - - [27136, 13824, 1, 512] - - [528, 94.152] - - - [27648, 13312, 1, 512] - - [526, 94.25] - - - [27648, 13824, 1, 512] - - [535, 94.193] - - - [28160, 13824, 1, 512] - - [530, 94.246] - - - [28160, 14336, 1, 512] - - [526, 94.2] - - - [28672, 13824, 1, 512] - - [528, 94.25] - - - [28672, 14336, 1, 512] - - [526, 94.274] - - - [29184, 14336, 1, 512] - - [526, 94.206] - - - [29184, 14848, 1, 512] - - [526, 94.304] - - - [29696, 14336, 1, 512] - - [526, 94.304] - - - [29696, 14848, 1, 512] - - [526, 94.304] - - - [30208, 14848, 1, 512] - - [528, 94.254] - - - [30208, 15360, 1, 512] - - [528, 94.306] - - - [30720, 14848, 1, 512] - - [529, 94.315] - - - [30720, 15360, 1, 512] - - [526, 94.352] - - - [31232, 15360, 1, 512] - - [526, 94.348] - - - [31232, 15872, 1, 512] - - [525, 94.338] - - - [31744, 15360, 1, 512] - - [530, 94.328] - - - [31744, 15872, 1, 512] - - [528, 94.351] - - - [32256, 15872, 1, 512] - - [526, 94.384] - - - [32256, 16384, 1, 512] - - [530, 94.384] - - - [32768, 15872, 1, 512] - - [526, 94.318] - - - [32768, 16384, 1, 512] - - [529, 94.258] - - - [33280, 16384, 1, 512] - - [530, 94.375] - - - [33280, 16896, 1, 512] - - [526, 94.383] - - - [33792, 16384, 1, 512] - - [533, 94.387] - - - [33792, 16896, 1, 512] - - [526, 94.409] - - - [34304, 16896, 1, 512] - - [526, 94.439] - - - [34304, 17408, 1, 512] - - [529, 94.467] - - - [34816, 16896, 1, 512] - - [528, 94.455] - - - [34816, 17408, 1, 512] - - [525, 94.424] - - - [35328, 17408, 1, 512] - - [526, 94.462] - - - [35328, 17920, 1, 512] - - [526, 94.458] - - - [35840, 17408, 1, 512] - - [526, 94.454] - - - [35840, 17920, 1, 512] - - [526, 94.441] - - - [36352, 17920, 1, 512] - - [526, 94.464] - - - [36352, 18432, 1, 512] - - [526, 94.478] - - - [36864, 17920, 1, 512] - - [526, 94.48] - - - [36864, 18432, 1, 512] - - [526, 94.445] - - - [37376, 18432, 1, 512] - - [526, 94.467] - - - [37376, 18944, 1, 512] - - [528, 94.465] - - - [37888, 18432, 1, 512] - - [526, 94.525] - - - [37888, 18944, 1, 512] - - [526, 94.461] - - - [38400, 18944, 1, 512] - - [530, 94.474] - - - [38400, 19456, 1, 512] - - [529, 94.517] - - - [38912, 18944, 1, 512] - - [529, 94.495] - - - [38912, 19456, 1, 512] - - [526, 94.527] - - - [39424, 19456, 1, 512] - - [528, 94.516] - - - [39424, 19968, 1, 512] - - [526, 94.511] - - - [39936, 19456, 1, 512] - - [528, 94.518] - - - [39936, 19968, 1, 512] - - [530, 94.504] - - - [40448, 19968, 1, 512] - - [526, 94.509] - - - [40448, 20480, 1, 512] - - [526, 94.523] - - - [40960, 19968, 1, 512] - - [526, 94.47] - - - [40960, 20480, 1, 512] - - [526, 94.513] - - - [41472, 20480, 1, 512] - - [528, 94.524] - - - [41472, 20992, 1, 512] - - [526, 94.533] - - - [41984, 20480, 1, 512] - - [529, 94.523] - - - [41984, 20992, 1, 512] - - [529, 94.521] - - - [42496, 20992, 1, 512] - - [530, 94.51] - - - [42496, 21504, 1, 512] - - [526, 94.55] - - - [43008, 20992, 1, 512] - - [529, 94.539] - - - [43008, 21504, 1, 512] - - [528, 94.537] - - - [43520, 21504, 1, 512] - - [526, 94.587] - - - [43520, 22016, 1, 512] - - [526, 94.549] - - - [44032, 21504, 1, 512] - - [526, 94.552] - - - [44032, 22016, 1, 512] - - [526, 94.554] - - - [44544, 22016, 1, 512] - - [526, 94.561] - - - [44544, 22528, 1, 512] - - [529, 94.558] - - - [45056, 22016, 1, 512] - - [529, 94.546] - - - [45056, 22528, 1, 512] - - [530, 94.557] - - - [45568, 22528, 1, 512] - - [529, 94.555] - - - [45568, 23040, 1, 512] - - [526, 94.562] - - - [46080, 22528, 1, 512] - - [526, 94.552] - - - [46080, 23040, 1, 512] - - [526, 94.577] - - - [46592, 23040, 1, 512] - - [526, 94.58] - - - [46592, 23552, 1, 512] - - [526, 94.595] - - - [47104, 23040, 1, 512] - - [528, 94.569] - - - [47104, 23552, 1, 512] - - [529, 94.582] - - - [47616, 23552, 1, 512] - - [528, 94.581] - - - [47616, 24064, 1, 512] - - [528, 94.57] - - - [48128, 23552, 1, 512] - - [529, 94.568] - - - [48128, 24064, 1, 512] - - [526, 94.578] - - - [48640, 24064, 1, 512] - - [528, 94.576] - - - [48640, 24576, 1, 512] - - [529, 94.566] - - - [49152, 24064, 1, 512] - - [528, 94.554] - - - [49152, 24576, 1, 512] - - [529, 94.536] - - - [49664, 24576, 1, 512] - - [526, 94.569] - - - [49664, 25088, 1, 512] - - [526, 94.591] - - - [50176, 24576, 1, 512] - - [529, 94.581] - - - [50176, 25088, 1, 512] - - [526, 94.602] - - - [50688, 25088, 1, 512] - - [528, 94.586] - - - [50688, 25600, 1, 512] - - [526, 94.588] - - - [51200, 25088, 1, 512] - - [526, 94.586] - - - [51200, 25600, 1, 512] - - [526, 94.582] - - - [51712, 25600, 1, 512] - - [530, 94.564] - - - [51712, 26112, 1, 512] - - [528, 94.588] - - - [52224, 25600, 1, 512] - - [526, 94.582] - - - [52224, 26112, 1, 512] - - [526, 94.586] - - - [52736, 26112, 1, 512] - - [528, 94.576] - - - [52736, 26624, 1, 512] - - [526, 94.576] - - - [53248, 26112, 1, 512] - - [528, 94.565] - - - [53248, 26624, 1, 512] - - [526, 94.583] - - - [53760, 26624, 1, 512] - - [526, 94.58] - - - [53760, 27136, 1, 512] - - [528, 94.578] - - - [54272, 26624, 1, 512] - - [529, 94.597] - - - [54272, 27136, 1, 512] - - [526, 94.586] - - - [54784, 27136, 1, 512] - - [526, 94.574] - - - [54784, 27648, 1, 512] - - [526, 94.582] - - - [55296, 27136, 1, 512] - - [438, 94.572] - - - [55296, 27648, 1, 512] - - [528, 94.598] - - - [55808, 27648, 1, 512] - - [526, 94.605] - - - [55808, 28160, 1, 512] - - [526, 94.581] - - - [56320, 27648, 1, 512] - - [528, 94.589] - - - [56320, 28160, 1, 512] - - [526, 94.58] - - - [56832, 28160, 1, 512] - - [526, 94.59] - - - [56832, 28672, 1, 512] - - [526, 94.595] - - - [57344, 28160, 1, 512] - - [533, 94.574] - - - [57344, 28672, 1, 512] - - [526, 94.595] - - - [57856, 28672, 1, 512] - - [526, 94.602] - - - [57856, 29184, 1, 512] - - [526, 94.592] - - - [58368, 28672, 1, 512] - - [526, 94.614] - - - [58368, 29184, 1, 512] - - [526, 94.609] - - - [58880, 29184, 1, 512] - - [526, 94.593] - - - [58880, 29696, 1, 512] - - [528, 94.589] - - - [59392, 29184, 1, 512] - - [526, 94.611] - - - [59392, 29696, 1, 512] - - [528, 94.602] - - - [59904, 29696, 1, 512] - - [526, 94.588] - - - [59904, 30208, 1, 512] - - [526, 94.602] - - - [60416, 29696, 1, 512] - - [526, 94.584] - - - [60416, 30208, 1, 512] - - [526, 94.606] - - - [60928, 30208, 1, 512] - - [526, 94.602] - - - [60928, 30720, 1, 512] - - [528, 94.613] - - - [61440, 30208, 1, 512] - - [526, 94.59] - - - [61440, 30720, 1, 512] - - [526, 94.602] - - - [61952, 30720, 1, 512] - - [528, 94.614] - - - [61952, 31232, 1, 512] - - [526, 94.617] - - - [62464, 30720, 1, 512] - - [526, 94.594] - - - [62464, 31232, 1, 512] - - [526, 94.615] - - - [62976, 31232, 1, 512] - - [526, 94.611] - - - [62976, 31744, 1, 512] - - [526, 94.618] - - - [63488, 31232, 1, 512] - - [526, 94.602] - - - [63488, 31744, 1, 512] - - [528, 94.595] - - - [64000, 31744, 1, 512] - - [528, 94.601] - - - [64000, 32256, 1, 512] - - [526, 94.621] - - - [64512, 31744, 1, 512] - - [528, 94.603] - - - [64512, 32256, 1, 512] - - [528, 94.621] - - - [65024, 512, 1, 512] - - [531, 90.116] - - - [65024, 32256, 1, 512] - - [526, 94.622] - - - [65024, 32768, 1, 512] - - [526, 94.609] - - - [65536, 512, 1, 512] - - [525, 87.59] - - - [65536, 32256, 1, 512] - - [528, 94.521] - - - [65536, 32768, 1, 512] - - [528, 94.491] - - - [66048, 512, 1, 512] - - [402, 88.777] - - - [66048, 32768, 1, 512] - - [526, 94.601] - - - [66048, 33280, 1, 512] - - [528, 94.607] - - - [66560, 512, 1, 512] - - [402, 89.14] - - - [66560, 32768, 1, 512] - - [526, 94.6] - - - [66560, 33280, 1, 512] - - [528, 94.607] - - - [67072, 512, 1, 512] - - [530, 88.977] - - - [67072, 33280, 1, 512] - - [528, 94.63] - - - [67072, 33792, 1, 512] - - [526, 94.621] - - - [67584, 512, 1, 512] - - [527, 89.599] - - - [67584, 33280, 1, 512] - - [528, 94.612] - - - [67584, 33792, 1, 512] - - [526, 94.614] - - - [68096, 512, 1, 512] - - [533, 90.192] - - - [68096, 33792, 1, 512] - - [526, 94.605] - - - [68096, 34304, 1, 512] - - [528, 94.617] - - - [68608, 512, 1, 512] - - [535, 90.362] - - - [68608, 33792, 1, 512] - - [526, 94.62] - - - [68608, 34304, 1, 512] - - [528, 94.624] - - - [69120, 512, 1, 512] - - [402, 88.486] - - - [69120, 34304, 1, 512] - - [526, 94.625] - - - [69120, 34816, 1, 512] - - [533, 94.598] - - - [69632, 512, 1, 512] - - [417, 89.267] - - - [69632, 34304, 1, 512] - - [528, 94.603] - - - [69632, 34816, 1, 512] - - [526, 94.62] - - - [70144, 512, 1, 512] - - [417, 89.313] - - - [70144, 34816, 1, 512] - - [526, 94.621] - - - [70144, 35328, 1, 512] - - [526, 94.606] - - - [70656, 512, 1, 512] - - [526, 89.566] - - - [70656, 34816, 1, 512] - - [526, 94.626] - - - [70656, 35328, 1, 512] - - [526, 94.628] - - - [71168, 512, 1, 512] - - [529, 90.033] - - - [71168, 35328, 1, 512] - - [528, 94.633] - - - [71168, 35840, 1, 512] - - [526, 94.618] - - - [71680, 512, 1, 512] - - [531, 90.415] - - - [71680, 35328, 1, 512] - - [528, 94.624] - - - [71680, 35840, 1, 512] - - [526, 94.638] - - - [72192, 512, 1, 512] - - [402, 88.487] - - - [72192, 35840, 1, 512] - - [526, 94.629] - - - [72192, 36352, 1, 512] - - [526, 94.616] - - - [72704, 512, 1, 512] - - [417, 89.06] - - - [72704, 35840, 1, 512] - - [526, 94.642] - - - [72704, 36352, 1, 512] - - [526, 94.629] - - - [73216, 512, 1, 512] - - [534, 89.236] - - - [73216, 36352, 1, 512] - - [528, 94.615] - - - [73216, 36864, 1, 512] - - [528, 94.626] - - - [73728, 512, 1, 512] - - [532, 89.589] - - - [73728, 36352, 1, 512] - - [526, 94.616] - - - [73728, 36864, 1, 512] - - [526, 94.605] - - - [74240, 512, 1, 512] - - [529, 90.036] - - - [74240, 36864, 1, 512] - - [526, 94.627] - - - [74240, 37376, 1, 512] - - [526, 94.62] - - - [74752, 512, 1, 512] - - [530, 90.525] - - - [74752, 36864, 1, 512] - - [528, 94.623] - - - [74752, 37376, 1, 512] - - [526, 94.621] - - - [75264, 512, 1, 512] - - [525, 90.916] - - - [75264, 37376, 1, 512] - - [526, 94.642] - - - [75264, 37888, 1, 512] - - [528, 94.621] - - - [75776, 512, 1, 512] - - [534, 88.879] - - - [75776, 37376, 1, 512] - - [526, 94.634] - - - [75776, 37888, 1, 512] - - [526, 94.61] - - - [76288, 512, 1, 512] - - [526, 89.203] - - - [76288, 37888, 1, 512] - - [526, 94.621] - - - [76288, 38400, 1, 512] - - [526, 94.637] - - - [76800, 512, 1, 512] - - [433, 89.705] - - - [76800, 37888, 1, 512] - - [526, 94.625] - - - [76800, 38400, 1, 512] - - [526, 94.633] - - - [77312, 512, 1, 512] - - [526, 90.047] - - - [77312, 38400, 1, 512] - - [526, 94.636] - - - [77312, 38912, 1, 512] - - [402, 94.093] - - - [77824, 512, 1, 512] - - [533, 90.491] - - - [77824, 38400, 1, 512] - - [526, 94.636] - - - [77824, 38912, 1, 512] - - [402, 94.041] - - - [78336, 512, 1, 512] - - [533, 90.977] - - - [78336, 38912, 1, 512] - - [402, 93.989] - - - [78336, 39424, 1, 512] - - [402, 94.056] - - - [78848, 512, 1, 512] - - [533, 88.894] - - - [78848, 38912, 1, 512] - - [402, 94.017] - - - [78848, 39424, 1, 512] - - [408, 94.354] - - - [79360, 512, 1, 512] - - [530, 89.302] - - - [79360, 39424, 1, 512] - - [438, 94.236] - - - [79360, 39936, 1, 512] - - [402, 93.803] - - - [79872, 512, 1, 512] - - [528, 89.676] - - - [79872, 39424, 1, 512] - - [408, 94.402] - - - [79872, 39936, 1, 512] - - [402, 93.715] - - - [80384, 512, 1, 512] - - [532, 90.162] - - - [80384, 39936, 1, 512] - - [402, 93.674] - - - [80384, 40448, 1, 512] - - [438, 94.177] - - - [80896, 512, 1, 512] - - [528, 90.566] - - - [80896, 39936, 1, 512] - - [402, 93.532] - - - [80896, 40448, 1, 512] - - [408, 94.323] - - - [81408, 512, 1, 512] - - [530, 90.895] - - - [81408, 40448, 1, 512] - - [438, 94.148] - - - [81408, 40960, 1, 512] - - [402, 93.922] - - - [81920, 512, 1, 512] - - [533, 88.877] - - - [81920, 40448, 1, 512] - - [402, 92.848] - - - [81920, 40960, 1, 512] - - [402, 92.602] - - - [82432, 512, 1, 512] - - [531, 89.332] - - - [82432, 40960, 1, 512] - - [402, 93.839] - - - [82432, 41472, 1, 512] - - [402, 94.069] - - - [82944, 512, 1, 512] - - [529, 89.77] - - - [82944, 40960, 1, 512] - - [402, 93.864] - - - [82944, 41472, 1, 512] - - [408, 94.266] - - - [83456, 512, 1, 512] - - [525, 90.137] - - - [83456, 41472, 1, 512] - - [408, 94.113] - - - [83456, 41984, 1, 512] - - [402, 93.797] - - - [83968, 512, 1, 512] - - [532, 90.465] - - - [83968, 41472, 1, 512] - - [408, 94.3] - - - [83968, 41984, 1, 512] - - [402, 93.696] - - - [84480, 512, 1, 512] - - [530, 90.826] - - - [84480, 41984, 1, 512] - - [402, 93.796] - - - [84480, 42496, 1, 512] - - [408, 94.105] - - - [84992, 512, 1, 512] - - [531, 90.923] - - - [84992, 41984, 1, 512] - - [469, 93.504] - - - [84992, 42496, 1, 512] - - [408, 94.229] - - - [85504, 512, 1, 512] - - [532, 89.464] - - - [85504, 42496, 1, 512] - - [402, 94.033] - - - [85504, 43008, 1, 512] - - [402, 93.547] - - - [86016, 512, 1, 512] - - [528, 89.7] - - - [86016, 42496, 1, 512] - - [408, 94.189] - - - [86016, 43008, 1, 512] - - [407, 93.236] - - - [86528, 512, 1, 512] - - [525, 90.068] - - - [86528, 43008, 1, 512] - - [402, 93.514] - - - [86528, 43520, 1, 512] - - [408, 94.028] - - - [87040, 512, 1, 512] - - [528, 90.435] - - - [87040, 43008, 1, 512] - - [469, 93.326] - - - [87040, 43520, 1, 512] - - [408, 94.139] - - - [87552, 512, 1, 512] - - [526, 90.906] - - - [87552, 43520, 1, 512] - - [408, 94.017] - - - [87552, 44032, 1, 512] - - [402, 93.791] - - - [88064, 512, 1, 512] - - [526, 91.173] - - - [88064, 43520, 1, 512] - - [408, 94.215] - - - [88064, 44032, 1, 512] - - [469, 93.594] - - - [88576, 512, 1, 512] - - [526, 89.551] - - - [88576, 44032, 1, 512] - - [402, 93.707] - - - [88576, 44544, 1, 512] - - [408, 94.007] - - - [89088, 512, 1, 512] - - [532, 89.779] - - - [89088, 44032, 1, 512] - - [407, 93.361] - - - [89088, 44544, 1, 512] - - [408, 94.087] - - - [89600, 512, 1, 512] - - [528, 90.113] - - - [89600, 44544, 1, 512] - - [402, 93.985] - - - [89600, 45056, 1, 512] - - [469, 93.377] - - - [90112, 512, 1, 512] - - [531, 90.445] - - - [90112, 44544, 1, 512] - - [402, 93.721] - - - [90112, 45056, 1, 512] - - [430, 92.682] - - - [90624, 512, 1, 512] - - [525, 90.923] - - - [90624, 45056, 1, 512] - - [469, 93.247] - - - [90624, 45568, 1, 512] - - [408, 93.884] - - - [91136, 512, 1, 512] - - [526, 91.049] - - - [91136, 45056, 1, 512] - - [469, 93.106] - - - [91136, 45568, 1, 512] - - [408, 94.047] - - - [91648, 512, 1, 512] - - [533, 89.5] - - - [91648, 45568, 1, 512] - - [408, 93.914] - - - [91648, 46080, 1, 512] - - [402, 93.506] - - - [92160, 512, 1, 512] - - [528, 89.909] - - - [92160, 45568, 1, 512] - - [408, 94.064] - - - [92160, 46080, 1, 512] - - [430, 93.318] - - - [92672, 512, 1, 512] - - [529, 90.183] - - - [92672, 46080, 1, 512] - - [402, 93.595] - - - [92672, 46592, 1, 512] - - [402, 93.903] - - - [93184, 512, 1, 512] - - [526, 90.455] - - - [93184, 46080, 1, 512] - - [430, 93.272] - - - [93184, 46592, 1, 512] - - [408, 93.942] - - - [93696, 512, 1, 512] - - [530, 90.861] - - - [93696, 46592, 1, 512] - - [402, 93.845] - - - [93696, 47104, 1, 512] - - [430, 93.108] - - - [94208, 512, 1, 512] - - [526, 91.188] - - - [94208, 46592, 1, 512] - - [408, 93.99] - - - [94208, 47104, 1, 512] - - [473, 92.897] - - - [94720, 512, 1, 512] - - [532, 91.555] - - - [94720, 47104, 1, 512] - - [469, 93.353] - - - [94720, 47616, 1, 512] - - [408, 93.811] - - - [95232, 512, 1, 512] - - [529, 89.922] - - - [95232, 47104, 1, 512] - - [473, 93.035] - - - [95232, 47616, 1, 512] - - [408, 93.915] - - - [95744, 512, 1, 512] - - [527, 90.208] - - - [95744, 47616, 1, 512] - - [408, 93.776] - - - [95744, 48128, 1, 512] - - [469, 93.291] - - - [96256, 512, 1, 512] - - [534, 90.547] - - - [96256, 47616, 1, 512] - - [408, 93.921] - - - [96256, 48128, 1, 512] - - [473, 92.916] - - - [96768, 512, 1, 512] - - [532, 90.775] - - - [96768, 48128, 1, 512] - - [469, 93.238] - - - [96768, 48640, 1, 512] - - [408, 93.718] - - - [97280, 512, 1, 512] - - [525, 91.098] - - - [97280, 48128, 1, 512] - - [473, 92.915] - - - [97280, 48640, 1, 512] - - [408, 93.74] - - - [97792, 512, 1, 512] - - [534, 91.468] - - - [97792, 48640, 1, 512] - - [402, 93.704] - - - [97792, 49152, 1, 512] - - [473, 92.662] - - - [98304, 512, 1, 512] - - [534, 89.848] - - - [98304, 48640, 1, 512] - - [402, 92.133] - - - [98304, 49152, 1, 512] - - [402, 91.717] - - - [98816, 512, 1, 512] - - [528, 90.204] - - - [98816, 49152, 1, 512] - - [447, 92.902] - - - [98816, 49664, 1, 512] - - [469, 93.685] - - - [99328, 512, 1, 512] - - [528, 90.467] - - - [99328, 49152, 1, 512] - - [473, 92.597] - - - [99328, 49664, 1, 512] - - [469, 93.637] - - - [99840, 512, 1, 512] - - [533, 90.893] - - - [99840, 49664, 1, 512] - - [450, 93.611] - - - [99840, 50176, 1, 512] - - [473, 92.745] - - - [100352, 512, 1, 512] - - [531, 91.081] - - - [100352, 49664, 1, 512] - - [450, 93.663] - - - [100352, 50176, 1, 512] - - [473, 92.79] - - - [100864, 512, 1, 512] - - [528, 91.446] - - - [100864, 50176, 1, 512] - - [473, 92.852] - - - [100864, 50688, 1, 512] - - [469, 93.521] - - - [101376, 512, 1, 512] - - [529, 91.704] - - - [101376, 50176, 1, 512] - - [473, 92.619] - - - [101376, 50688, 1, 512] - - [450, 93.594] - - - [101888, 512, 1, 512] - - [527, 90.163] - - - [101888, 50688, 1, 512] - - [450, 93.509] - - - [101888, 51200, 1, 512] - - [447, 92.706] - - - [102400, 512, 1, 512] - - [533, 90.571] - - - [102400, 50688, 1, 512] - - [450, 93.572] - - - [102400, 51200, 1, 512] - - [447, 92.687] - - - [102912, 512, 1, 512] - - [534, 90.85] - - - [102912, 51200, 1, 512] - - [447, 92.752] - - - [102912, 51712, 1, 512] - - [450, 93.489] - - - [103424, 512, 1, 512] - - [532, 91.276] - - - [103424, 51200, 1, 512] - - [473, 92.553] - - - [103424, 51712, 1, 512] - - [450, 93.526] - - - [103936, 512, 1, 512] - - [531, 91.361] - - - [103936, 51712, 1, 512] - - [469, 93.49] - - - [103936, 52224, 1, 512] - - [473, 92.549] - - - [104448, 512, 1, 512] - - [532, 91.502] - - - [104448, 51712, 1, 512] - - [450, 93.515] - - - [104448, 52224, 1, 512] - - [473, 92.539] - - - [104960, 512, 1, 512] - - [530, 90.242] - - - [104960, 52224, 1, 512] - - [473, 92.49] - - - [104960, 52736, 1, 512] - - [469, 93.387] - - - [105472, 512, 1, 512] - - [534, 90.532] - - - [105472, 52224, 1, 512] - - [473, 92.457] - - - [105472, 52736, 1, 512] - - [450, 93.388] - - - [105984, 512, 1, 512] - - [534, 90.791] - - - [105984, 52736, 1, 512] - - [450, 93.347] - - - [105984, 53248, 1, 512] - - [473, 92.514] - - - [106496, 512, 1, 512] - - [534, 91.036] - - - [106496, 52736, 1, 512] - - [469, 92.302] - - - [106496, 53248, 1, 512] - - [447, 91.45] - - - [107008, 512, 1, 512] - - [528, 91.33] - - - [107008, 53248, 1, 512] - - [473, 92.245] - - - [107008, 53760, 1, 512] - - [469, 93.317] - - - [107520, 512, 1, 512] - - [534, 91.636] - - - [107520, 53248, 1, 512] - - [473, 92.282] - - - [107520, 53760, 1, 512] - - [479, 92.371] - - - [108032, 512, 1, 512] - - [534, 90.244] - - - [108032, 53760, 1, 512] - - [450, 93.24] - - - [108032, 54272, 1, 512] - - [473, 92.339] - - - [108544, 512, 1, 512] - - [534, 90.563] - - - [108544, 53760, 1, 512] - - [479, 92.74] - - - [108544, 54272, 1, 512] - - [473, 91.992] - - - [109056, 512, 1, 512] - - [533, 90.85] - - - [109056, 54272, 1, 512] - - [473, 92.303] - - - [109056, 54784, 1, 512] - - [450, 93.11] - - - [109568, 512, 1, 512] - - [532, 91.107] - - - [109568, 54272, 1, 512] - - [473, 91.699] - - - [109568, 54784, 1, 512] - - [473, 91.884] - - - [110080, 512, 1, 512] - - [529, 91.417] - - - [110080, 54784, 1, 512] - - [450, 93.117] - - - [110080, 55296, 1, 512] - - [473, 92.089] - - - [110592, 512, 1, 512] - - [527, 91.605] - - - [110592, 54784, 1, 512] - - [450, 92.612] - - - [110592, 55296, 1, 512] - - [473, 91.688] - - - [111104, 512, 1, 512] - - [525, 90.677] - - - [111104, 55296, 1, 512] - - [473, 91.987] - - - [111104, 55808, 1, 512] - - [450, 92.947] - - - [111616, 512, 1, 512] - - [532, 90.532] - - - [111616, 55296, 1, 512] - - [473, 91.507] - - - [111616, 55808, 1, 512] - - [480, 91.232] - - - [112128, 512, 1, 512] - - [533, 90.865] - - - [112128, 55808, 1, 512] - - [450, 92.944] - - - [112128, 56320, 1, 512] - - [473, 91.869] - - - [112640, 512, 1, 512] - - [527, 91.156] - - - [112640, 55808, 1, 512] - - [473, 91.385] - - - [112640, 56320, 1, 512] - - [473, 91.134] - - - [113152, 512, 1, 512] - - [529, 91.381] - - - [113152, 56320, 1, 512] - - [473, 91.787] - - - [113152, 56832, 1, 512] - - [473, 91.745] - - - [113664, 512, 1, 512] - - [529, 91.528] - - - [113664, 56320, 1, 512] - - [473, 91.155] - - - [113664, 56832, 1, 512] - - [473, 91.51] - - - [114176, 512, 1, 512] - - [529, 91.745] - - - [114176, 56832, 1, 512] - - [473, 91.822] - - - [114176, 57344, 1, 512] - - [473, 91.335] - - - [114688, 512, 1, 512] - - [528, 90.487] - - - [114688, 56832, 1, 512] - - [402, 90.139] - - - [114688, 57344, 1, 512] - - [430, 88.626] - - - [115200, 512, 1, 512] - - [531, 90.839] - - - [115200, 57344, 1, 512] - - [473, 91.233] - - - [115200, 57856, 1, 512] - - [479, 91.741] - - - [115712, 512, 1, 512] - - [530, 91.088] - - - [115712, 57344, 1, 512] - - [473, 90.773] - - - [115712, 57856, 1, 512] - - [481, 91.291] - - - [116224, 512, 1, 512] - - [525, 91.341] - - - [116224, 57856, 1, 512] - - [481, 91.302] - - - [116224, 58368, 1, 512] - - [473, 90.986] - - - [116736, 512, 1, 512] - - [528, 91.64] - - - [116736, 57856, 1, 512] - - [481, 91.219] - - - [116736, 58368, 1, 512] - - [475, 90.79] - - - [117248, 512, 1, 512] - - [531, 91.914] - - - [117248, 58368, 1, 512] - - [484, 90.917] - - - [117248, 58880, 1, 512] - - [479, 91.904] - - - [117760, 512, 1, 512] - - [531, 90.627] - - - [117760, 58368, 1, 512] - - [485, 90.759] - - - [117760, 58880, 1, 512] - - [481, 91.234] - - - [118272, 512, 1, 512] - - [526, 90.786] - - - [118272, 58880, 1, 512] - - [481, 91.26] - - - [118272, 59392, 1, 512] - - [484, 91.152] - - - [118784, 512, 1, 512] - - [533, 91.096] - - - [118784, 58880, 1, 512] - - [481, 91.219] - - - [118784, 59392, 1, 512] - - [484, 90.634] - - - [119296, 512, 1, 512] - - [526, 91.393] - - - [119296, 59392, 1, 512] - - [484, 91.147] - - - [119296, 59904, 1, 512] - - [481, 91.26] - - - [119808, 512, 1, 512] - - [534, 91.6] - - - [119808, 59392, 1, 512] - - [484, 90.798] - - - [119808, 59904, 1, 512] - - [481, 91.232] - - - [120320, 512, 1, 512] - - [527, 91.94] - - - [120320, 59904, 1, 512] - - [481, 91.214] - - - [120320, 60416, 1, 512] - - [484, 90.921] - - - [120832, 512, 1, 512] - - [534, 91.955] - - - [120832, 59904, 1, 512] - - [481, 91.222] - - - [120832, 60416, 1, 512] - - [485, 90.758] - - - [121344, 512, 1, 512] - - [527, 90.894] - - - [121344, 60416, 1, 512] - - [475, 90.933] - - - [121344, 60928, 1, 512] - - [481, 91.238] - - - [121856, 512, 1, 512] - - [534, 91.114] - - - [121856, 60416, 1, 512] - - [485, 90.759] - - - [121856, 60928, 1, 512] - - [481, 91.213] - - - [122368, 512, 1, 512] - - [529, 91.387] - - - [122368, 60928, 1, 512] - - [481, 91.239] - - - [122368, 61440, 1, 512] - - [484, 91.101] - - - [122880, 512, 1, 512] - - [532, 91.633] - - - [122880, 60928, 1, 512] - - [483, 89.826] - - - [122880, 61440, 1, 512] - - [473, 89.109] - - - [123392, 512, 1, 512] - - [528, 91.927] - - - [123392, 61440, 1, 512] - - [484, 91.075] - - - [123392, 61952, 1, 512] - - [481, 91.252] - - - [123904, 512, 1, 512] - - [531, 92.017] - - - [123904, 61440, 1, 512] - - [484, 90.717] - - - [123904, 61952, 1, 512] - - [481, 91.176] - - - [124416, 512, 1, 512] - - [530, 90.885] - - - [124416, 61952, 1, 512] - - [481, 91.237] - - - [124416, 62464, 1, 512] - - [475, 90.967] - - - [124928, 512, 1, 512] - - [528, 91.195] - - - [124928, 61952, 1, 512] - - [480, 91.213] - - - [124928, 62464, 1, 512] - - [485, 90.761] - - - [125440, 512, 1, 512] - - [525, 91.377] - - - [125440, 62464, 1, 512] - - [484, 90.989] - - - [125440, 62976, 1, 512] - - [481, 91.247] - - - [125952, 512, 1, 512] - - [528, 91.563] - - - [125952, 62464, 1, 512] - - [475, 90.743] - - - [125952, 62976, 1, 512] - - [480, 91.21] - - - [126464, 512, 1, 512] - - [525, 91.839] - - - [126464, 62976, 1, 512] - - [481, 91.283] - - - [126464, 63488, 1, 512] - - [484, 90.926] - - - [126976, 512, 1, 512] - - [528, 92.029] - - - [126976, 62976, 1, 512] - - [481, 91.185] - - - [126976, 63488, 1, 512] - - [488, 90.545] - - - [127488, 512, 1, 512] - - [526, 92.208] - - - [127488, 63488, 1, 512] - - [475, 90.905] - - - [127488, 64000, 1, 512] - - [481, 91.224] - - - [128000, 512, 1, 512] - - [527, 91.075] - - - [128000, 63488, 1, 512] - - [488, 90.743] - - - [128000, 64000, 1, 512] - - [481, 91.229] - - - [4096, 1537, 1, 512] - - [468, 72.921] - - - [4096, 2049, 1, 512] - - [430, 74.623] - - - [4608, 2049, 1, 512] - - [529, 77.951] - - - [5120, 2049, 1, 512] - - [535, 76.654] - - - [5120, 2561, 1, 512] - - [471, 80.786] - - - [5632, 2561, 1, 512] - - [470, 81.849] - - - [6144, 2561, 1, 512] - - [535, 82.634] - - - [6144, 3073, 1, 512] - - [529, 83.913] - - - [6656, 3073, 1, 512] - - [529, 84.734] - - - [7168, 3073, 1, 512] - - [535, 84.829] - - - [7168, 3585, 1, 512] - - [534, 86.884] - - - [7680, 3585, 1, 512] - - [535, 87.65] - - - [8192, 3585, 1, 512] - - [417, 85.559] - - - [8192, 4097, 1, 512] - - [535, 87.668] - - - [8704, 4097, 1, 512] - - [529, 88.912] - - - [9216, 4097, 1, 512] - - [535, 87.103] - - - [9216, 4609, 1, 512] - - [535, 87.501] - - - [9728, 4609, 1, 512] - - [528, 88.982] - - - [10240, 4609, 1, 512] - - [535, 89.872] - - - [10240, 5121, 1, 512] - - [529, 88.816] - - - [10752, 5121, 1, 512] - - [529, 90.323] - - - [11264, 5121, 1, 512] - - [535, 89.538] - - - [11264, 5633, 1, 512] - - [533, 90.841] - - - [11776, 5633, 1, 512] - - [533, 90.555] - - - [12288, 5633, 1, 512] - - [535, 90.165] - - - [12288, 6145, 1, 512] - - [529, 90.195] - - - [12800, 6145, 1, 512] - - [525, 91.687] - - - [13312, 6145, 1, 512] - - [529, 91.638] - - - [13312, 6657, 1, 512] - - [535, 91.896] - - - [13824, 6657, 1, 512] - - [535, 90.698] - - - [14336, 6657, 1, 512] - - [535, 90.868] - - - [14336, 7169, 1, 512] - - [529, 91.642] - - - [14848, 7169, 1, 512] - - [529, 91.988] - - - [15360, 7169, 1, 512] - - [529, 92.319] - - - [15360, 7681, 1, 512] - - [531, 92.153] - - - [15872, 7681, 1, 512] - - [535, 91.613] - - - [16384, 7681, 1, 512] - - [530, 91.971] - - - [16384, 8193, 1, 512] - - [529, 92.219] - - - [16896, 8193, 1, 512] - - [529, 92.594] - - - [17408, 8193, 1, 512] - - [529, 92.311] - - - [17408, 8705, 1, 512] - - [535, 92.649] - - - [17920, 8705, 1, 512] - - [535, 92.388] - - - [18432, 8705, 1, 512] - - [535, 92.306] - - - [18432, 9217, 1, 512] - - [529, 92.837] - - - [18944, 9217, 1, 512] - - [528, 92.762] - - - [19456, 9217, 1, 512] - - [528, 92.708] - - - [19456, 9729, 1, 512] - - [535, 92.632] - - - [19968, 9729, 1, 512] - - [535, 92.705] - - - [20480, 9729, 1, 512] - - [535, 92.705] - - - [20480, 10241, 1, 512] - - [529, 92.93] - - - [20992, 10241, 1, 512] - - [531, 92.653] - - - [21504, 10241, 1, 512] - - [535, 92.763] - - - [21504, 10753, 1, 512] - - [526, 93.012] - - - [22016, 10753, 1, 512] - - [526, 92.808] - - - [22528, 10753, 1, 512] - - [526, 92.926] - - - [22528, 11265, 1, 512] - - [535, 92.935] - - - [23040, 11265, 1, 512] - - [529, 93.121] - - - [23552, 11265, 1, 512] - - [535, 92.979] - - - [23552, 11777, 1, 512] - - [526, 93.013] - - - [24064, 11777, 1, 512] - - [535, 92.934] - - - [24576, 11777, 1, 512] - - [535, 93.215] - - - [24576, 12289, 1, 512] - - [525, 93.108] - - - [25088, 12289, 1, 512] - - [535, 93.072] - - - [25600, 12289, 1, 512] - - [529, 93.355] - - - [25600, 12801, 1, 512] - - [530, 93.26] - - - [26112, 12801, 1, 512] - - [530, 93.34] - - - [26624, 12801, 1, 512] - - [530, 93.3] - - - [26624, 13313, 1, 512] - - [529, 93.398] - - - [27136, 13313, 1, 512] - - [529, 93.424] - - - [27648, 13313, 1, 512] - - [529, 93.361] - - - [27648, 13825, 1, 512] - - [525, 93.361] - - - [28160, 13825, 1, 512] - - [525, 93.439] - - - [28672, 13825, 1, 512] - - [528, 93.334] - - - [28672, 14337, 1, 512] - - [526, 93.547] - - - [29184, 14337, 1, 512] - - [526, 93.465] - - - [29696, 14337, 1, 512] - - [526, 93.549] - - - [29696, 14849, 1, 512] - - [533, 93.517] - - - [30208, 14849, 1, 512] - - [528, 93.496] - - - [30720, 14849, 1, 512] - - [526, 93.501] - - - [30720, 15361, 1, 512] - - [525, 93.593] - - - [31232, 15361, 1, 512] - - [529, 93.573] - - - [31744, 15361, 1, 512] - - [526, 93.639] - - - [31744, 15873, 1, 512] - - [526, 93.585] - - - [32256, 15873, 1, 512] - - [535, 93.581] - - - [32768, 15873, 1, 512] - - [526, 93.518] - - - [32768, 16385, 1, 512] - - [529, 93.646] - - - [33280, 16385, 1, 512] - - [529, 93.707] - - - [33792, 16385, 1, 512] - - [529, 93.698] - - - [33792, 16897, 1, 512] - - [533, 93.756] - - - [34304, 16897, 1, 512] - - [526, 93.703] - - - [34816, 16897, 1, 512] - - [525, 93.724] - - - [34816, 17409, 1, 512] - - [529, 93.781] - - - [35328, 17409, 1, 512] - - [529, 93.771] - - - [35840, 17409, 1, 512] - - [529, 93.816] - - - [35840, 17921, 1, 512] - - [530, 93.827] - - - [36352, 17921, 1, 512] - - [526, 93.828] - - - [36864, 17921, 1, 512] - - [526, 93.841] - - - [36864, 18433, 1, 512] - - [529, 93.836] - - - [37376, 18433, 1, 512] - - [529, 93.875] - - - [37888, 18433, 1, 512] - - [529, 93.857] - - - [37888, 18945, 1, 512] - - [529, 93.824] - - - [38400, 18945, 1, 512] - - [533, 93.843] - - - [38912, 18945, 1, 512] - - [526, 93.884] - - - [38912, 19457, 1, 512] - - [529, 93.931] - - - [39424, 19457, 1, 512] - - [529, 93.914] - - - [39936, 19457, 1, 512] - - [529, 93.935] - - - [39936, 19969, 1, 512] - - [526, 93.88] - - - [40448, 19969, 1, 512] - - [526, 93.91] - - - [40960, 19969, 1, 512] - - [526, 93.873] - - - [40960, 20481, 1, 512] - - [530, 93.961] - - - [41472, 20481, 1, 512] - - [529, 93.952] - - - [41984, 20481, 1, 512] - - [529, 93.964] - - - [41984, 20993, 1, 512] - - [529, 93.942] - - - [42496, 20993, 1, 512] - - [528, 93.941] - - - [43008, 20993, 1, 512] - - [529, 93.932] - - - [43008, 21505, 1, 512] - - [526, 94.033] - - - [43520, 21505, 1, 512] - - [526, 94.02] - - - [44032, 21505, 1, 512] - - [526, 94.021] - - - [44032, 22017, 1, 512] - - [528, 93.998] - - - [44544, 22017, 1, 512] - - [526, 94.002] - - - [45056, 22017, 1, 512] - - [528, 94.014] - - - [45056, 22529, 1, 512] - - [529, 94.046] - - - [45568, 22529, 1, 512] - - [529, 94.041] - - - [46080, 22529, 1, 512] - - [529, 94.05] - - - [46080, 23041, 1, 512] - - [528, 94.066] - - - [46592, 23041, 1, 512] - - [528, 94.086] - - - [47104, 23041, 1, 512] - - [528, 94.077] - - - [47104, 23553, 1, 512] - - [529, 94.063] - - - [47616, 23553, 1, 512] - - [529, 94.072] - - - [48128, 23553, 1, 512] - - [529, 94.066] - - - [48128, 24065, 1, 512] - - [526, 94.079] - - - [48640, 24065, 1, 512] - - [530, 94.034] - - - [49152, 24065, 1, 512] - - [529, 94.043] - - - [49152, 24577, 1, 512] - - [529, 94.063] - - - [49664, 24577, 1, 512] - - [526, 94.066] - - - [50176, 24577, 1, 512] - - [529, 94.095] - - - [50176, 25089, 1, 512] - - [526, 94.099] - - - [50688, 25089, 1, 512] - - [526, 94.124] - - - [51200, 25089, 1, 512] - - [526, 94.144] - - - [51200, 25601, 1, 512] - - [530, 94.128] - - - [51712, 25601, 1, 512] - - [529, 94.139] - - - [52224, 25601, 1, 512] - - [526, 94.136] - - - [52224, 26113, 1, 512] - - [526, 94.152] - - - [52736, 26113, 1, 512] - - [526, 94.102] - - - [53248, 26113, 1, 512] - - [533, 94.08] - - - [53248, 26625, 1, 512] - - [529, 94.144] - - - [53760, 26625, 1, 512] - - [529, 94.122] - - - [54272, 26625, 1, 512] - - [529, 94.133] - - - [54272, 27137, 1, 512] - - [526, 94.131] - - - [54784, 27137, 1, 512] - - [528, 94.109] - - - [55296, 27137, 1, 512] - - [529, 94.135] - - - [55296, 27649, 1, 512] - - [526, 94.162] - - - [55808, 27649, 1, 512] - - [528, 94.173] - - - [56320, 27649, 1, 512] - - [528, 94.159] - - - [56320, 28161, 1, 512] - - [526, 94.146] - - - [56832, 28161, 1, 512] - - [528, 94.141] - - - [57344, 28161, 1, 512] - - [530, 94.146] - - - [57344, 28673, 1, 512] - - [526, 94.186] - - - [57856, 28673, 1, 512] - - [526, 94.194] - - - [58368, 28673, 1, 512] - - [526, 94.195] - - - [58368, 29185, 1, 512] - - [526, 94.181] - - - [58880, 29185, 1, 512] - - [526, 94.176] - - - [59392, 29185, 1, 512] - - [526, 94.193] - - - [59392, 29697, 1, 512] - - [529, 94.176] - - - [59904, 29697, 1, 512] - - [528, 94.158] - - - [60416, 29697, 1, 512] - - [529, 94.188] - - - [60416, 30209, 1, 512] - - [528, 94.2] - - - [60928, 30209, 1, 512] - - [528, 94.18] - - - [61440, 30209, 1, 512] - - [529, 94.18] - - - [61440, 30721, 1, 512] - - [529, 94.201] - - - [61952, 30721, 1, 512] - - [529, 94.198] - - - [62464, 30721, 1, 512] - - [528, 94.223] - - - [62464, 31233, 1, 512] - - [533, 94.189] - - - [62976, 31233, 1, 512] - - [526, 94.222] - - - [63488, 31233, 1, 512] - - [526, 94.226] - - - [63488, 31745, 1, 512] - - [526, 94.214] - - - [64000, 31745, 1, 512] - - [529, 94.226] - - - [64512, 31745, 1, 512] - - [528, 94.219] - - - [64512, 32257, 1, 512] - - [526, 94.262] - - - [65024, 32257, 1, 512] - - [528, 94.258] - - - [65536, 32257, 1, 512] - - [528, 94.138] - - - [65536, 32769, 1, 512] - - [528, 94.156] - - - [66048, 32769, 1, 512] - - [526, 94.24] - - - [66560, 32769, 1, 512] - - [526, 94.235] - - - [66560, 33281, 1, 512] - - [528, 94.254] - - - [67072, 33281, 1, 512] - - [528, 94.221] - - - [67584, 33281, 1, 512] - - [528, 94.24] - - - [67584, 33793, 1, 512] - - [529, 94.253] - - - [68096, 33793, 1, 512] - - [526, 94.224] - - - [68608, 33793, 1, 512] - - [529, 94.248] - - - [68608, 34305, 1, 512] - - [526, 94.251] - - - [69120, 34305, 1, 512] - - [526, 94.246] - - - [69632, 34305, 1, 512] - - [526, 94.255] - - - [69632, 34817, 1, 512] - - [528, 94.251] - - - [70144, 34817, 1, 512] - - [526, 94.254] - - - [70656, 34817, 1, 512] - - [526, 94.258] - - - [70656, 35329, 1, 512] - - [528, 94.267] - - - [71168, 35329, 1, 512] - - [528, 94.25] - - - [71680, 35329, 1, 512] - - [526, 94.251] - - - [71680, 35841, 1, 512] - - [526, 94.294] - - - [72192, 35841, 1, 512] - - [526, 94.293] - - - [72704, 35841, 1, 512] - - [526, 94.303] - - - [72704, 36353, 1, 512] - - [526, 94.281] - - - [73216, 36353, 1, 512] - - [526, 94.279] - - - [73728, 36353, 1, 512] - - [526, 94.279] - - - [73728, 36865, 1, 512] - - [528, 94.283] - - - [74240, 36865, 1, 512] - - [526, 94.288] - - - [74752, 36865, 1, 512] - - [528, 94.288] - - - [74752, 37377, 1, 512] - - [528, 94.284] - - - [75264, 37377, 1, 512] - - [528, 94.298] - - - [75776, 37377, 1, 512] - - [533, 94.283] - - - [75776, 37889, 1, 512] - - [528, 94.308] - - - [76288, 37889, 1, 512] - - [528, 94.303] - - - [76800, 37889, 1, 512] - - [526, 94.299] - - - [76800, 38401, 1, 512] - - [408, 93.644] - - - [77312, 38401, 1, 512] - - [438, 93.507] - - - [77824, 38401, 1, 512] - - [408, 93.693] - - - [77824, 38913, 1, 512] - - [430, 91.448] - - - [78336, 38913, 1, 512] - - [402, 91.909] - - - [78848, 38913, 1, 512] - - [469, 91.037] - - - [78848, 39425, 1, 512] - - [408, 93.621] - - - [79360, 39425, 1, 512] - - [402, 93.44] - - - [79872, 39425, 1, 512] - - [408, 93.696] - - - [79872, 39937, 1, 512] - - [448, 90.877] - - - [80384, 39937, 1, 512] - - [402, 91.695] - - - [80896, 39937, 1, 512] - - [395, 90.885] - - - [80896, 40449, 1, 512] - - [408, 93.617] - - - [81408, 40449, 1, 512] - - [402, 93.427] - - - [81920, 40449, 1, 512] - - [408, 91.712] - - - [81920, 40961, 1, 512] - - [402, 90.511] - - - [82432, 40961, 1, 512] - - [430, 91.602] - - - [82944, 40961, 1, 512] - - [430, 91.358] - - - [82944, 41473, 1, 512] - - [408, 93.573] - - - [83456, 41473, 1, 512] - - [408, 93.308] - - - [83968, 41473, 1, 512] - - [408, 93.623] - - - [83968, 41985, 1, 512] - - [394, 90.156] - - - [84480, 41985, 1, 512] - - [430, 91.627] - - - [84992, 41985, 1, 512] - - [395, 90.837] - - - [84992, 42497, 1, 512] - - [408, 93.555] - - - [85504, 42497, 1, 512] - - [402, 93.266] - - - [86016, 42497, 1, 512] - - [408, 93.468] - - - [86016, 43009, 1, 512] - - [395, 90.573] - - - [86528, 43009, 1, 512] - - [430, 91.509] - - - [87040, 43009, 1, 512] - - [469, 90.146] - - - [87040, 43521, 1, 512] - - [408, 93.44] - - - [87552, 43521, 1, 512] - - [402, 93.241] - - - [88064, 43521, 1, 512] - - [408, 93.529] - - - [88064, 44033, 1, 512] - - [395, 90.845] - - - [88576, 44033, 1, 512] - - [430, 91.526] - - - [89088, 44033, 1, 512] - - [395, 90.752] - - - [89088, 44545, 1, 512] - - [408, 93.431] - - - [89600, 44545, 1, 512] - - [402, 93.252] - - - [90112, 44545, 1, 512] - - [402, 92.881] - - - [90112, 45057, 1, 512] - - [469, 90.15] - - - [90624, 45057, 1, 512] - - [447, 90.922] - - - [91136, 45057, 1, 512] - - [395, 89.698] - - - [91136, 45569, 1, 512] - - [408, 93.37] - - - [91648, 45569, 1, 512] - - [408, 93.132] - - - [92160, 45569, 1, 512] - - [408, 93.414] - - - [92160, 46081, 1, 512] - - [395, 90.737] - - - [92672, 46081, 1, 512] - - [430, 91.146] - - - [93184, 46081, 1, 512] - - [395, 89.816] - - - [93184, 46593, 1, 512] - - [408, 93.287] - - - [93696, 46593, 1, 512] - - [402, 93.15] - - - [94208, 46593, 1, 512] - - [408, 93.33] - - - [94208, 47105, 1, 512] - - [447, 90.024] - - - [94720, 47105, 1, 512] - - [430, 91.146] - - - [95232, 47105, 1, 512] - - [395, 89.821] - - - [95232, 47617, 1, 512] - - [408, 93.268] - - - [95744, 47617, 1, 512] - - [402, 93.09] - - - [96256, 47617, 1, 512] - - [408, 93.273] - - - [96256, 48129, 1, 512] - - [395, 89.51] - - - [96768, 48129, 1, 512] - - [430, 91.09] - - - [97280, 48129, 1, 512] - - [474, 89.477] - - - [97280, 48641, 1, 512] - - [408, 93.124] - - - [97792, 48641, 1, 512] - - [402, 93.002] - - - [98304, 48641, 1, 512] - - [402, 90.881] - - - [98304, 49153, 1, 512] - - [402, 89.65] - - - [98816, 49153, 1, 512] - - [447, 90.759] - - - [99328, 49153, 1, 512] - - [440, 88.096] - - - [99328, 49665, 1, 512] - - [408, 93.094] - - - [99840, 49665, 1, 512] - - [408, 92.832] - - - [100352, 49665, 1, 512] - - [408, 92.984] - - - [100352, 50177, 1, 512] - - [475, 89.171] - - - [100864, 50177, 1, 512] - - [469, 91.014] - - - [101376, 50177, 1, 512] - - [476, 88.937] - - - [101376, 50689, 1, 512] - - [450, 92.814] - - - [101888, 50689, 1, 512] - - [469, 92.728] - - - [102400, 50689, 1, 512] - - [450, 92.708] - - - [102400, 51201, 1, 512] - - [451, 89.1] - - - [102912, 51201, 1, 512] - - [447, 90.032] - - - [103424, 51201, 1, 512] - - [475, 89.017] - - - [103424, 51713, 1, 512] - - [450, 92.701] - - - [103936, 51713, 1, 512] - - [469, 92.663] - - - [104448, 51713, 1, 512] - - [450, 92.685] - - - [104448, 52225, 1, 512] - - [477, 89.025] - - - [104960, 52225, 1, 512] - - [478, 89.973] - - - [105472, 52225, 1, 512] - - [477, 89.085] - - - [105472, 52737, 1, 512] - - [479, 92.166] - - - [105984, 52737, 1, 512] - - [469, 92.58] - - - [106496, 52737, 1, 512] - - [403, 91.727] - - - [106496, 53249, 1, 512] - - [447, 89.65] - - - [107008, 53249, 1, 512] - - [478, 89.721] - - - [107520, 53249, 1, 512] - - [477, 88.422] - - - [107520, 53761, 1, 512] - - [450, 92.495] - - - [108032, 53761, 1, 512] - - [469, 92.46] - - - [108544, 53761, 1, 512] - - [450, 92.463] - - - [108544, 54273, 1, 512] - - [477, 89.015] - - - [109056, 54273, 1, 512] - - [478, 89.897] - - - [109568, 54273, 1, 512] - - [477, 88.893] - - - [109568, 54785, 1, 512] - - [479, 91.957] - - - [110080, 54785, 1, 512] - - [469, 92.337] - - - [110592, 54785, 1, 512] - - [479, 91.851] - - - [110592, 55297, 1, 512] - - [477, 88.458] - - - [111104, 55297, 1, 512] - - [478, 89.939] - - - [111616, 55297, 1, 512] - - [477, 88.902] - - - [111616, 55809, 1, 512] - - [479, 91.636] - - - [112128, 55809, 1, 512] - - [469, 92.221] - - - [112640, 55809, 1, 512] - - [481, 90.918] - - - [112640, 56321, 1, 512] - - [475, 88.692] - - - [113152, 56321, 1, 512] - - [478, 89.927] - - - [113664, 56321, 1, 512] - - [475, 88.997] - - - [113664, 56833, 1, 512] - - [481, 90.952] - - - [114176, 56833, 1, 512] - - [479, 91.61] - - - [114688, 56833, 1, 512] - - [402, 89.574] - - - [114688, 57345, 1, 512] - - [430, 87.847] - - - [115200, 57345, 1, 512] - - [482, 89.34] - - - [115712, 57345, 1, 512] - - [453, 87.387] - - - [115712, 57857, 1, 512] - - [483, 90.838] - - - [116224, 57857, 1, 512] - - [479, 91.423] - - - [116736, 57857, 1, 512] - - [481, 90.883] - - - [116736, 58369, 1, 512] - - [475, 88.849] - - - [117248, 58369, 1, 512] - - [478, 89.793] - - - [117760, 58369, 1, 512] - - [475, 89.007] - - - [117760, 58881, 1, 512] - - [481, 90.924] - - - [118272, 58881, 1, 512] - - [479, 91.199] - - - [118784, 58881, 1, 512] - - [481, 90.844] - - - [118784, 59393, 1, 512] - - [477, 88.376] - - - [119296, 59393, 1, 512] - - [478, 89.884] - - - [119808, 59393, 1, 512] - - [475, 88.935] - - - [119808, 59905, 1, 512] - - [481, 90.889] - - - [120320, 59905, 1, 512] - - [481, 90.946] - - - [120832, 59905, 1, 512] - - [481, 90.899] - - - [120832, 60417, 1, 512] - - [475, 88.84] - - - [121344, 60417, 1, 512] - - [478, 89.913] - - - [121856, 60417, 1, 512] - - [477, 88.925] - - - [121856, 60929, 1, 512] - - [481, 90.869] - - - [122368, 60929, 1, 512] - - [481, 90.994] - - - [122880, 60929, 1, 512] - - [479, 88.807] - - - [122880, 61441, 1, 512] - - [486, 86.857] - - - [123392, 61441, 1, 512] - - [487, 89.659] - - - [123904, 61441, 1, 512] - - [477, 88.534] - - - [123904, 61953, 1, 512] - - [481, 90.867] - - - [124416, 61953, 1, 512] - - [481, 90.997] - - - [124928, 61953, 1, 512] - - [481, 90.907] - - - [124928, 62465, 1, 512] - - [475, 88.927] - - - [125440, 62465, 1, 512] - - [478, 89.782] - - - [125952, 62465, 1, 512] - - [475, 88.881] - - - [125952, 62977, 1, 512] - - [481, 90.919] - - - [126464, 62977, 1, 512] - - [481, 91.033] - - - [126976, 62977, 1, 512] - - [481, 90.82] - - - [126976, 63489, 1, 512] - - [477, 88.3] - - - [127488, 63489, 1, 512] - - [478, 89.893] - - - [128000, 63489, 1, 512] - - [477, 88.924] - - - [3584, 6657, 1, 512] - - [417, 85.708] - - - [3584, 6145, 1, 512] - - [430, 83.855] - - - [3072, 5633, 1, 512] - - [395, 85.17] - - - [3072, 5121, 1, 512] - - [394, 84.434] - - - [2560, 4609, 1, 512] - - [395, 81.218] - - - [2560, 4097, 1, 512] - - [402, 81.236] - - - [2048, 3585, 1, 512] - - [429, 73.114] - - - [2048, 3073, 1, 512] - - [428, 72.101] - - - [1536, 2561, 1, 512] - - [425, 70.909] - - - [1536, 2049, 1, 512] - - [424, 65.718] - - - [1024, 1537, 1, 512] - - [422, 58.197] - - - [1024, 1025, 1, 512] - - [420, 43.401] - - - [512, 513, 1, 512] - - [419, 21.264] - - - [512, 1, 1, 512] - - [458, 0.034] - - - [6656, 4096, 1, 512] - - [21, 87.137] - - - [6144, 4992, 1, 512] - - [21, 87.63] - - - [8192, 3328, 1, 512] - - [21, 87.21] - - - [8320, 4096, 1, 512] - - [21, 87.851] - - - [7040, 4096, 1, 512] - - [20, 91.956] - - - [7040, 4096, 1, 512] - - [15, 90.075] - - - [8448, 3840, 1, 512] - - [20, 92.567] - - - [8448, 3840, 1, 512] - - [19, 90.593] - - - [7680, 4224, 1, 512] - - [20, 92.491] - - - [7680, 4224, 1, 512] - - [19, 90.538] - - - [1024, 513, 1, 512] - - [459, 33.669] - - - [1536, 513, 1, 512] - - [460, 43.705] - - - [2048, 513, 1, 512] - - [461, 44.808] - - - [2048, 1025, 1, 512] - - [371, 55.673] - - - [2560, 1025, 1, 512] - - [463, 64.962] - - - [3072, 1025, 1, 512] - - [437, 62.686] - - - [3072, 1537, 1, 512] - - [466, 66.96] - - - [3584, 1537, 1, 512] - - [467, 66.406] - - - [1024, 1, 1, 512] - - [489, 0.082] - - - [1152, 385, 1, 384] - - [339, 25.247] - - - [1536, 385, 1, 384] - - [331, 30.548] - - - [1536, 769, 1, 384] - - [420, 45.323] - - - [1920, 769, 1, 384] - - [490, 51.306] - - - [2304, 769, 1, 384] - - [420, 50.762] - - - [3456, 1536, 1, 384] - - [495, 72.791] - - - [3840, 1536, 1, 384] - - [467, 70.93] - - - [3840, 1537, 1, 384] - - [467, 67.505] - - - [3840, 1920, 1, 384] - - [379, 71.961] - - - [4224, 1920, 1, 384] - - [464, 77.453] - - - [4224, 1921, 1, 384] - - [385, 71.533] - - - [4224, 2304, 1, 384] - - [369, 76.587] - - - [4608, 1920, 1, 384] - - [390, 77.994] - - - [4608, 1921, 1, 384] - - [464, 73.504] - - - [4608, 2304, 1, 384] - - [464, 79.328] - - - [4608, 2305, 1, 384] - - [468, 72.307] - - - [4992, 2304, 1, 384] - - [464, 78.546] - - - [4992, 2305, 1, 384] - - [379, 76.379] - - - [5376, 2304, 1, 384] - - [496, 80.46] - - - [5376, 2305, 1, 384] - - [379, 75.242] - - - [5376, 2689, 1, 384] - - [497, 75.669] - - - [5760, 2689, 1, 384] - - [497, 79.724] - - - [6144, 2689, 1, 384] - - [498, 78.35] - - - [6144, 3073, 1, 384] - - [499, 79.283] - - - [6528, 3073, 1, 384] - - [499, 78.275] - - - [6528, 3456, 1, 384] - - [464, 84.946] - - - [6912, 3073, 1, 384] - - [499, 77.891] - - - [6912, 3456, 1, 384] - - [464, 84.848] - - - [6912, 3457, 1, 384] - - [400, 82.06] - - - [7296, 3456, 1, 384] - - [399, 86.422] - - - [7296, 3457, 1, 384] - - [430, 81.31] - - - [7296, 3840, 1, 384] - - [464, 86.038] - - - [7680, 3456, 1, 384] - - [400, 85.949] - - - [7680, 3457, 1, 384] - - [441, 80.973] - - - [7680, 3840, 1, 384] - - [464, 86.09] - - - [7680, 3841, 1, 384] - - [395, 84.741] - - - [8064, 3840, 1, 384] - - [464, 86.513] - - - [8064, 3841, 1, 384] - - [395, 84.29] - - - [8064, 4224, 1, 384] - - [394, 88.004] - - - [8448, 3840, 1, 384] - - [399, 88.267] - - - [8448, 3841, 1, 384] - - [394, 84.415] - - - [8448, 4224, 1, 384] - - [394, 88.229] - - - [8448, 4225, 1, 384] - - [395, 84.403] - - - [8832, 4224, 1, 384] - - [395, 88.014] - - - [8832, 4225, 1, 384] - - [395, 84.344] - - - [8832, 4608, 1, 384] - - [439, 88.443] - - - [9216, 4224, 1, 384] - - [394, 87.945] - - - [9216, 4225, 1, 384] - - [417, 84.185] - - - [9216, 4608, 1, 384] - - [400, 88.531] - - - [9216, 4609, 1, 384] - - [433, 84.645] - - - [9600, 4608, 1, 384] - - [400, 88.745] - - - [9600, 4609, 1, 384] - - [407, 84.869] - - - [9600, 4992, 1, 384] - - [399, 89.231] - - - [9984, 4608, 1, 384] - - [399, 88.984] - - - [9984, 4609, 1, 384] - - [407, 85.085] - - - [9984, 4992, 1, 384] - - [399, 89.618] - - - [9984, 4993, 1, 384] - - [441, 86.111] - - - [10368, 4992, 1, 384] - - [399, 89.817] - - - [10368, 4993, 1, 384] - - [399, 86.747] - - - [10368, 5376, 1, 384] - - [400, 90.255] - - - [10752, 4992, 1, 384] - - [399, 90.178] - - - [10752, 4993, 1, 384] - - [400, 87.092] - - - [10752, 5376, 1, 384] - - [399, 88.993] - - - [10752, 5377, 1, 384] - - [395, 87.906] - - - [11136, 5376, 1, 384] - - [400, 89.472] - - - [11136, 5377, 1, 384] - - [473, 87.185] - - - [11136, 5760, 1, 384] - - [394, 90.428] - - - [11520, 5376, 1, 384] - - [400, 89.902] - - - [11520, 5377, 1, 384] - - [407, 86.901] - - - [11520, 5760, 1, 384] - - [400, 90.917] - - - [11520, 5761, 1, 384] - - [395, 88.229] - - - [11904, 5760, 1, 384] - - [399, 89.89] - - - [11904, 5761, 1, 384] - - [395, 88.307] - - - [11904, 6144, 1, 384] - - [433, 90.669] - - - [12288, 5760, 1, 384] - - [400, 90.44] - - - [12288, 5761, 1, 384] - - [395, 87.842] - - - [12288, 6144, 1, 384] - - [400, 91.197] - - - [12288, 6145, 1, 384] - - [394, 85.506] - - - [12672, 6144, 1, 384] - - [407, 90.509] - - - [12672, 6145, 1, 384] - - [500, 85.983] - - - [12672, 6528, 1, 384] - - [400, 91.553] - - - [13056, 6144, 1, 384] - - [407, 90.829] - - - [13056, 6145, 1, 384] - - [389, 85.712] - - - [13056, 6528, 1, 384] - - [400, 91.077] - - - [13056, 6529, 1, 384] - - [395, 88.884] - - - [13440, 6528, 1, 384] - - [400, 91.506] - - - [13440, 6529, 1, 384] - - [400, 89.408] - - - [13440, 6912, 1, 384] - - [399, 91.7] - - - [13824, 6529, 1, 384] - - [407, 88.936] - - - [13824, 6912, 1, 384] - - [407, 91.274] - - - [13824, 6913, 1, 384] - - [394, 89.693] - - - [14208, 6913, 1, 384] - - [395, 89.505] - - - [15744, 7680, 1, 384] - - [399, 92.015] - - - [16128, 7680, 1, 384] - - [399, 92.486] - - - [16128, 7681, 1, 384] - - [400, 90.194] - - - [16128, 8064, 1, 384] - - [399, 92.246] - - - [16512, 8064, 1, 384] - - [400, 92.676] - - - [16512, 8065, 1, 384] - - [433, 90.185] - - - [16512, 8448, 1, 384] - - [399, 92.297] - - - [16896, 8064, 1, 384] - - [395, 92.582] - - - [16896, 8065, 1, 384] - - [395, 90.64] - - - [16896, 8448, 1, 384] - - [399, 92.308] - - - [16896, 8449, 1, 384] - - [395, 90.752] - - - [17280, 8448, 1, 384] - - [400, 92.701] - - - [17280, 8449, 1, 384] - - [407, 90.566] - - - [17280, 8832, 1, 384] - - [395, 92.525] - - - [17664, 8448, 1, 384] - - [399, 92.61] - - - [17664, 8449, 1, 384] - - [433, 90.366] - - - [17664, 8832, 1, 384] - - [395, 92.311] - - - [17664, 8833, 1, 384] - - [395, 90.272] - - - [18048, 8832, 1, 384] - - [394, 92.224] - - - [18048, 8833, 1, 384] - - [395, 90.447] - - - [18432, 8832, 1, 384] - - [395, 92.388] - - - [18432, 8833, 1, 384] - - [395, 90.586] - - - [18432, 9217, 1, 384] - - [394, 87.476] - - - [18816, 9217, 1, 384] - - [391, 87.941] - - - [18816, 9600, 1, 384] - - [395, 91.77] - - - [19200, 9217, 1, 384] - - [382, 87.68] - - - [19200, 9600, 1, 384] - - [395, 91.9] - - - [19200, 9601, 1, 384] - - [441, 89.907] - - - [19584, 9600, 1, 384] - - [395, 92.176] - - - [19584, 9601, 1, 384] - - [395, 89.913] - - - [19584, 9984, 1, 384] - - [441, 91.614] - - - [19968, 9600, 1, 384] - - [395, 91.901] - - - [19968, 9601, 1, 384] - - [441, 90.193] - - - [19968, 9984, 1, 384] - - [441, 91.658] - - - [19968, 9985, 1, 384] - - [395, 89.945] - - - [20352, 9984, 1, 384] - - [441, 91.605] - - - [20352, 9985, 1, 384] - - [395, 90.077] - - - [20352, 10368, 1, 384] - - [441, 91.185] - - - [20736, 9984, 1, 384] - - [395, 91.566] - - - [20736, 9985, 1, 384] - - [395, 90.122] - - - [20736, 10368, 1, 384] - - [441, 91.18] - - - [20736, 10369, 1, 384] - - [395, 89.815] - - - [21120, 10368, 1, 384] - - [395, 91.153] - - - [21120, 10369, 1, 384] - - [395, 89.796] - - - [21120, 10752, 1, 384] - - [394, 91.279] - - - [21504, 10368, 1, 384] - - [395, 91.366] - - - [21504, 10369, 1, 384] - - [395, 90.032] - - - [21504, 10752, 1, 384] - - [394, 91.51] - - - [21504, 10753, 1, 384] - - [395, 89.923] - - - [21888, 10752, 1, 384] - - [417, 91.499] - - - [21888, 10753, 1, 384] - - [395, 89.938] - - - [21888, 11136, 1, 384] - - [395, 91.058] - - - [22272, 10752, 1, 384] - - [394, 91.447] - - - [22272, 10753, 1, 384] - - [395, 90.046] - - - [22272, 11136, 1, 384] - - [395, 91.191] - - - [22272, 11137, 1, 384] - - [395, 89.78] - - - [22656, 11136, 1, 384] - - [441, 91.097] - - - [22656, 11137, 1, 384] - - [400, 89.585] - - - [22656, 11520, 1, 384] - - [441, 91.124] - - - [23040, 11136, 1, 384] - - [395, 91.296] - - - [23040, 11137, 1, 384] - - [395, 89.887] - - - [23040, 11520, 1, 384] - - [395, 91.189] - - - [23040, 11521, 1, 384] - - [395, 89.813] - - - [23424, 11520, 1, 384] - - [441, 91.143] - - - [23424, 11521, 1, 384] - - [395, 89.614] - - - [23424, 11904, 1, 384] - - [394, 90.942] - - - [23808, 11520, 1, 384] - - [395, 91.298] - - - [23808, 11521, 1, 384] - - [395, 89.55] - - - [23808, 11904, 1, 384] - - [395, 91.072] - - - [23808, 11905, 1, 384] - - [395, 89.627] - - - [24192, 11904, 1, 384] - - [395, 91.168] - - - [24192, 11905, 1, 384] - - [395, 89.7] - - - [24192, 12288, 1, 384] - - [407, 90.654] - - - [24576, 11904, 1, 384] - - [395, 91.348] - - - [24576, 11905, 1, 384] - - [395, 89.899] - - - [24576, 12288, 1, 384] - - [407, 90.83] - - - [24576, 12289, 1, 384] - - [395, 86.891] - - - [24960, 12288, 1, 384] - - [407, 90.658] - - - [24960, 12289, 1, 384] - - [417, 87.054] - - - [24960, 12672, 1, 384] - - [395, 91.106] - - - [25344, 12288, 1, 384] - - [433, 90.638] - - - [25344, 12289, 1, 384] - - [417, 86.951] - - - [25344, 12672, 1, 384] - - [438, 90.39] - - - [25344, 12673, 1, 384] - - [441, 89.049] - - - [25728, 12672, 1, 384] - - [397, 90.787] - - - [25728, 12673, 1, 384] - - [441, 89.312] - - - [25728, 13056, 1, 384] - - [397, 90.906] - - - [26112, 12673, 1, 384] - - [395, 89.719] - - - [26112, 13056, 1, 384] - - [395, 91.096] - - - [26112, 13057, 1, 384] - - [395, 89.816] - - - [26496, 13057, 1, 384] - - [395, 89.639] - - - [26880, 13057, 1, 384] - - [395, 89.517] - - - [27648, 13825, 1, 384] - - [395, 90.253] - - - [28032, 13824, 1, 384] - - [394, 91.635] - - - [28032, 13825, 1, 384] - - [417, 90.037] - - - [28416, 13824, 1, 384] - - [395, 90.778] - - - [28416, 13825, 1, 384] - - [395, 89.362] - - - [28416, 14208, 1, 384] - - [399, 90.7] - - - [28416, 14209, 1, 384] - - [395, 89.409] - - - [28800, 14208, 1, 384] - - [395, 91.067] - - - [28800, 14209, 1, 384] - - [395, 89.958] - - - [28800, 14592, 1, 384] - - [395, 91.328] - - - [29184, 14208, 1, 384] - - [397, 91.283] - - - [29184, 14209, 1, 384] - - [395, 90.182] - - - [29184, 14592, 1, 384] - - [395, 91.462] - - - [29184, 14593, 1, 384] - - [395, 90.145] - - - [29568, 14592, 1, 384] - - [438, 91.223] - - - [29568, 14593, 1, 384] - - [395, 89.933] - - - [29568, 14976, 1, 384] - - [395, 91.216] - - - [29952, 14592, 1, 384] - - [395, 91.391] - - - [29952, 14593, 1, 384] - - [395, 90.049] - - - [29952, 14976, 1, 384] - - [395, 91.417] - - - [29952, 14977, 1, 384] - - [395, 90.116] - - - [30336, 14976, 1, 384] - - [395, 91.59] - - - [30336, 14977, 1, 384] - - [395, 90.319] - - - [30720, 14976, 1, 384] - - [395, 91.861] - - - [30720, 14977, 1, 384] - - [395, 90.472] - - - [30720, 15361, 1, 384] - - [417, 87.78] - - - [31104, 15361, 1, 384] - - [395, 87.786] - - - [31104, 15744, 1, 384] - - [397, 91.504] - - - [31488, 15361, 1, 384] - - [395, 87.911] - - - [31488, 15744, 1, 384] - - [441, 91.78] - - - [31488, 15745, 1, 384] - - [395, 90.298] - - - [31872, 15744, 1, 384] - - [394, 91.559] - - - [31872, 15745, 1, 384] - - [395, 90.275] - - - [31872, 16128, 1, 384] - - [395, 91.411] - - - [32256, 15744, 1, 384] - - [395, 91.873] - - - [32256, 15745, 1, 384] - - [395, 90.552] - - - [32256, 16128, 1, 384] - - [441, 91.599] - - - [32256, 16129, 1, 384] - - [395, 90.58] - - - [32640, 16128, 1, 384] - - [441, 91.339] - - - [32640, 16129, 1, 384] - - [395, 90.118] - - - [32640, 16512, 1, 384] - - [395, 91.364] - - - [33024, 16128, 1, 384] - - [395, 91.623] - - - [33024, 16129, 1, 384] - - [395, 90.74] - - - [33024, 16512, 1, 384] - - [395, 91.948] - - - [33024, 16513, 1, 384] - - [395, 90.728] - - - [33408, 16512, 1, 384] - - [395, 91.999] - - - [33408, 16513, 1, 384] - - [395, 90.727] - - - [33408, 16896, 1, 384] - - [394, 92.053] - - - [33792, 16512, 1, 384] - - [397, 91.688] - - - [33792, 16513, 1, 384] - - [441, 90.563] - - - [33792, 16896, 1, 384] - - [441, 91.969] - - - [33792, 16897, 1, 384] - - [441, 90.67] - - - [34176, 16896, 1, 384] - - [441, 91.996] - - - [34176, 16897, 1, 384] - - [441, 90.551] - - - [34176, 17280, 1, 384] - - [397, 91.649] - - - [34560, 16896, 1, 384] - - [397, 91.942] - - - [34560, 16897, 1, 384] - - [395, 90.658] - - - [34560, 17280, 1, 384] - - [395, 91.866] - - - [34560, 17281, 1, 384] - - [441, 90.688] - - - [34944, 17280, 1, 384] - - [441, 91.95] - - - [34944, 17281, 1, 384] - - [395, 90.649] - - - [34944, 17664, 1, 384] - - [441, 92.0] - - - [35328, 17280, 1, 384] - - [395, 92.039] - - - [35328, 17281, 1, 384] - - [395, 90.896] - - - [35328, 17664, 1, 384] - - [395, 92.015] - - - [35328, 17665, 1, 384] - - [395, 90.848] - - - [35712, 17664, 1, 384] - - [395, 92.006] - - - [35712, 17665, 1, 384] - - [395, 90.82] - - - [35712, 18048, 1, 384] - - [395, 91.873] - - - [36096, 17664, 1, 384] - - [395, 91.969] - - - [36096, 17665, 1, 384] - - [395, 90.832] - - - [36096, 18048, 1, 384] - - [395, 91.818] - - - [36096, 18049, 1, 384] - - [395, 90.912] - - - [36480, 18048, 1, 384] - - [395, 91.776] - - - [36480, 18049, 1, 384] - - [395, 90.668] - - - [36480, 18432, 1, 384] - - [417, 91.937] - - - [36864, 18048, 1, 384] - - [397, 91.789] - - - [36864, 18049, 1, 384] - - [441, 90.841] - - - [36864, 18432, 1, 384] - - [400, 91.846] - - - [36864, 18433, 1, 384] - - [395, 88.35] - - - [37248, 18432, 1, 384] - - [417, 91.988] - - - [37248, 18433, 1, 384] - - [394, 88.591] - - - [37248, 18816, 1, 384] - - [395, 91.964] - - - [37632, 18432, 1, 384] - - [395, 91.966] - - - [37632, 18433, 1, 384] - - [394, 88.577] - - - [37632, 18816, 1, 384] - - [395, 92.008] - - - [37632, 18817, 1, 384] - - [395, 90.943] - - - [38016, 18816, 1, 384] - - [441, 92.136] - - - [38016, 18817, 1, 384] - - [441, 90.899] - - - [38016, 19200, 1, 384] - - [441, 92.093] - - - [38400, 18816, 1, 384] - - [441, 92.317] - - - [38400, 18817, 1, 384] - - [395, 91.151] - - - [38400, 19200, 1, 384] - - [441, 92.183] - - - [38400, 19201, 1, 384] - - [395, 91.082] - - - [38784, 19200, 1, 384] - - [395, 92.27] - - - [38784, 19201, 1, 384] - - [395, 91.11] - - - [38784, 19584, 1, 384] - - [395, 92.215] - - - [39168, 19200, 1, 384] - - [395, 92.209] - - - [39168, 19201, 1, 384] - - [395, 90.963] - - - [39168, 19584, 1, 384] - - [395, 92.207] - - - [39168, 19585, 1, 384] - - [395, 90.933] - - - [39552, 19584, 1, 384] - - [441, 91.883] - - - [39552, 19585, 1, 384] - - [441, 90.819] - - - [39552, 19968, 1, 384] - - [397, 92.207] - - - [39936, 19584, 1, 384] - - [395, 92.285] - - - [39936, 19585, 1, 384] - - [441, 91.031] - - - [39936, 19968, 1, 384] - - [441, 92.277] - - - [39936, 19969, 1, 384] - - [395, 91.13] - - - [40320, 19968, 1, 384] - - [441, 92.306] - - - [40320, 19969, 1, 384] - - [395, 91.26] - - - [40320, 20352, 1, 384] - - [395, 92.286] - - - [40704, 19968, 1, 384] - - [441, 92.216] - - - [40704, 19969, 1, 384] - - [395, 91.19] - - - [40704, 20352, 1, 384] - - [395, 92.247] - - - [40704, 20353, 1, 384] - - [395, 91.179] - - - [41088, 20352, 1, 384] - - [395, 92.331] - - - [41088, 20353, 1, 384] - - [395, 91.267] - - - [41088, 20736, 1, 384] - - [395, 92.306] - - - [41472, 20352, 1, 384] - - [395, 92.498] - - - [41472, 20353, 1, 384] - - [395, 91.431] - - - [41472, 20736, 1, 384] - - [395, 92.428] - - - [41472, 20737, 1, 384] - - [395, 91.473] - - - [41856, 20736, 1, 384] - - [395, 92.475] - - - [41856, 20737, 1, 384] - - [395, 91.304] - - - [41856, 21120, 1, 384] - - [395, 92.535] - - - [42240, 20736, 1, 384] - - [395, 91.857] - - - [42240, 20737, 1, 384] - - [395, 91.012] - - - [42240, 21120, 1, 384] - - [394, 92.015] - - - [42240, 21121, 1, 384] - - [395, 91.032] - - - [42624, 21120, 1, 384] - - [400, 92.034] - - - [42624, 21121, 1, 384] - - [400, 90.974] - - - [42624, 21504, 1, 384] - - [400, 92.03] - - - [43008, 21120, 1, 384] - - [395, 92.49] - - - [43008, 21121, 1, 384] - - [395, 91.371] - - - [43008, 21504, 1, 384] - - [400, 92.208] - - - [43008, 21505, 1, 384] - - [395, 88.593] - - - [43392, 21504, 1, 384] - - [395, 92.288] - - - [43392, 21505, 1, 384] - - [417, 89.017] - - - [43392, 21888, 1, 384] - - [441, 92.286] - - - [43776, 21504, 1, 384] - - [417, 92.123] - - - [43776, 21505, 1, 384] - - [395, 88.788] - - - [43776, 21888, 1, 384] - - [441, 92.276] - - - [43776, 21889, 1, 384] - - [395, 91.315] - - - [44160, 21888, 1, 384] - - [395, 92.357] - - - [44160, 21889, 1, 384] - - [395, 91.468] - - - [44160, 22272, 1, 384] - - [395, 92.553] - - - [44544, 21888, 1, 384] - - [441, 92.487] - - - [44544, 21889, 1, 384] - - [395, 91.611] - - - [44544, 22272, 1, 384] - - [395, 92.593] - - - [44544, 22273, 1, 384] - - [395, 91.635] - - - [44928, 384, 1, 384] - - [399, 80.407] - - - [44928, 22272, 1, 384] - - [441, 92.472] - - - [44928, 22273, 1, 384] - - [395, 91.371] - - - [44928, 22656, 1, 384] - - [395, 92.335] - - - [45312, 384, 1, 384] - - [397, 80.964] - - - [45312, 22272, 1, 384] - - [441, 92.386] - - - [45312, 22273, 1, 384] - - [395, 91.18] - - - [45312, 22656, 1, 384] - - [441, 92.306] - - - [45312, 22657, 1, 384] - - [395, 91.31] - - - [45696, 384, 1, 384] - - [395, 81.584] - - - [45696, 22656, 1, 384] - - [395, 92.496] - - - [45696, 22657, 1, 384] - - [395, 91.451] - - - [45696, 23040, 1, 384] - - [395, 92.56] - - - [46080, 384, 1, 384] - - [395, 81.939] - - - [46080, 22656, 1, 384] - - [395, 92.599] - - - [46080, 22657, 1, 384] - - [395, 91.433] - - - [46080, 23040, 1, 384] - - [395, 92.569] - - - [46080, 23041, 1, 384] - - [395, 91.54] - - - [46464, 384, 1, 384] - - [395, 82.292] - - - [46464, 23040, 1, 384] - - [394, 92.528] - - - [46464, 23041, 1, 384] - - [395, 91.536] - - - [46464, 23424, 1, 384] - - [395, 92.516] - - - [46848, 384, 1, 384] - - [395, 82.906] - - - [46848, 23040, 1, 384] - - [395, 92.505] - - - [46848, 23041, 1, 384] - - [395, 91.521] - - - [46848, 23424, 1, 384] - - [395, 92.525] - - - [46848, 23425, 1, 384] - - [395, 91.518] - - - [47232, 384, 1, 384] - - [394, 78.27] - - - [47232, 23424, 1, 384] - - [395, 92.586] - - - [47232, 23425, 1, 384] - - [395, 91.54] - - - [47232, 23808, 1, 384] - - [441, 92.452] - - - [47616, 384, 1, 384] - - [395, 79.521] - - - [47616, 23424, 1, 384] - - [395, 92.67] - - - [47616, 23425, 1, 384] - - [395, 91.574] - - - [47616, 23808, 1, 384] - - [441, 92.452] - - - [47616, 23809, 1, 384] - - [395, 91.579] - - - [48000, 384, 1, 384] - - [395, 79.761] - - - [48000, 23808, 1, 384] - - [441, 92.454] - - - [48000, 23809, 1, 384] - - [395, 91.429] - - - [48000, 24192, 1, 384] - - [395, 92.35] - - - [48384, 384, 1, 384] - - [397, 80.359] - - - [48384, 23808, 1, 384] - - [441, 92.42] - - - [48384, 23809, 1, 384] - - [441, 91.371] - - - [48384, 24192, 1, 384] - - [395, 92.487] - - - [48384, 24193, 1, 384] - - [441, 91.461] - - - [48768, 384, 1, 384] - - [394, 80.594] - - - [48768, 24192, 1, 384] - - [395, 92.609] - - - [48768, 24193, 1, 384] - - [395, 91.595] - - - [48768, 24576, 1, 384] - - [400, 92.391] - - - [49152, 384, 1, 384] - - [395, 80.409] - - - [49152, 24192, 1, 384] - - [395, 92.328] - - - [49152, 24193, 1, 384] - - [395, 90.984] - - - [49152, 24576, 1, 384] - - [400, 92.011] - - - [49152, 24577, 1, 384] - - [395, 88.798] - - - [49536, 384, 1, 384] - - [399, 81.479] - - - [49536, 24576, 1, 384] - - [400, 92.355] - - - [49536, 24577, 1, 384] - - [395, 89.514] - - - [49536, 24960, 1, 384] - - [395, 92.66] - - - [49920, 384, 1, 384] - - [400, 81.948] - - - [49920, 24576, 1, 384] - - [395, 92.293] - - - [49920, 24577, 1, 384] - - [395, 89.224] - - - [49920, 24960, 1, 384] - - [395, 92.67] - - - [49920, 24961, 1, 384] - - [395, 91.668] - - - [50304, 384, 1, 384] - - [394, 82.476] - - - [50304, 24960, 1, 384] - - [395, 92.698] - - - [50304, 24961, 1, 384] - - [400, 91.555] - - - [50304, 25344, 1, 384] - - [395, 92.687] - - - [50688, 384, 1, 384] - - [407, 82.698] - - - [50688, 24960, 1, 384] - - [441, 92.651] - - - [50688, 24961, 1, 384] - - [441, 91.598] - - - [50688, 25344, 1, 384] - - [441, 92.558] - - - [50688, 25345, 1, 384] - - [395, 91.685] - - - [51072, 384, 1, 384] - - [399, 83.181] - - - [51072, 25344, 1, 384] - - [441, 92.545] - - - [51072, 25345, 1, 384] - - [395, 91.638] - - - [51072, 25728, 1, 384] - - [395, 92.385] - - - [51456, 384, 1, 384] - - [399, 83.642] - - - [51456, 25344, 1, 384] - - [395, 92.525] - - - [51456, 25345, 1, 384] - - [395, 91.616] - - - [51456, 25728, 1, 384] - - [395, 92.472] - - - [51456, 25729, 1, 384] - - [395, 91.676] - - - [51840, 384, 1, 384] - - [408, 79.411] - - - [51840, 25728, 1, 384] - - [395, 92.539] - - - [51840, 25729, 1, 384] - - [395, 91.742] - - - [51840, 26112, 1, 384] - - [395, 92.777] - - - [52224, 384, 1, 384] - - [395, 80.036] - - - [52224, 25728, 1, 384] - - [441, 92.613] - - - [52224, 25729, 1, 384] - - [395, 91.752] - - - [52224, 26112, 1, 384] - - [441, 92.759] - - - [52224, 26113, 1, 384] - - [395, 91.818] - - - [52608, 384, 1, 384] - - [397, 80.513] - - - [52608, 26112, 1, 384] - - [395, 92.771] - - - [52608, 26113, 1, 384] - - [395, 91.81] - - - [52608, 26496, 1, 384] - - [395, 92.671] - - - [52992, 384, 1, 384] - - [395, 81.138] - - - [52992, 26112, 1, 384] - - [395, 92.72] - - - [52992, 26113, 1, 384] - - [395, 91.736] - - - [52992, 26496, 1, 384] - - [395, 92.717] - - - [52992, 26497, 1, 384] - - [395, 91.665] - - - [53376, 384, 1, 384] - - [399, 81.183] - - - [53376, 26496, 1, 384] - - [395, 92.609] - - - [53376, 26497, 1, 384] - - [395, 91.683] - - - [53376, 26880, 1, 384] - - [395, 92.648] - - - [53760, 384, 1, 384] - - [395, 81.618] - - - [53760, 26496, 1, 384] - - [441, 92.711] - - - [53760, 26497, 1, 384] - - [395, 91.796] - - - [53760, 26880, 1, 384] - - [441, 92.622] - - - [53760, 26881, 1, 384] - - [395, 91.796] - - - [54144, 384, 1, 384] - - [399, 82.239] - - - [54144, 26880, 1, 384] - - [395, 92.697] - - - [54144, 26881, 1, 384] - - [395, 91.715] - - - [54144, 27264, 1, 384] - - [395, 92.689] - - - [54528, 384, 1, 384] - - [399, 82.482] - - - [54528, 26880, 1, 384] - - [395, 92.68] - - - [54528, 26881, 1, 384] - - [395, 91.757] - - - [54528, 27264, 1, 384] - - [395, 92.762] - - - [54528, 27265, 1, 384] - - [395, 91.775] - - - [54912, 384, 1, 384] - - [399, 82.988] - - - [54912, 27264, 1, 384] - - [395, 92.772] - - - [54912, 27265, 1, 384] - - [395, 91.788] - - - [54912, 27648, 1, 384] - - [400, 92.571] - - - [55296, 384, 1, 384] - - [400, 83.359] - - - [55296, 27264, 1, 384] - - [395, 92.852] - - - [55296, 27265, 1, 384] - - [395, 91.898] - - - [55296, 27648, 1, 384] - - [400, 92.673] - - - [55296, 27649, 1, 384] - - [395, 88.969] - - - [55680, 384, 1, 384] - - [395, 83.654] - - - [55680, 27648, 1, 384] - - [400, 92.638] - - - [55680, 27649, 1, 384] - - [417, 89.602] - - - [55680, 28032, 1, 384] - - [395, 92.854] - - - [56064, 384, 1, 384] - - [400, 84.136] - - - [56064, 27648, 1, 384] - - [400, 92.58] - - - [56064, 27649, 1, 384] - - [395, 89.475] - - - [56064, 28032, 1, 384] - - [395, 92.714] - - - [56064, 28033, 1, 384] - - [395, 91.806] - - - [56448, 384, 1, 384] - - [397, 80.477] - - - [56448, 28032, 1, 384] - - [395, 92.57] - - - [56448, 28033, 1, 384] - - [400, 91.727] - - - [56448, 28416, 1, 384] - - [399, 92.598] - - - [56832, 384, 1, 384] - - [395, 80.99] - - - [56832, 28032, 1, 384] - - [395, 92.733] - - - [56832, 28033, 1, 384] - - [400, 91.868] - - - [56832, 28416, 1, 384] - - [395, 92.706] - - - [56832, 28417, 1, 384] - - [400, 91.918] - - - [57216, 384, 1, 384] - - [400, 81.224] - - - [57216, 28416, 1, 384] - - [395, 92.779] - - - [57216, 28417, 1, 384] - - [395, 91.848] - - - [57216, 28800, 1, 384] - - [395, 92.806] - - - [57600, 384, 1, 384] - - [399, 81.716] - - - [57600, 28416, 1, 384] - - [395, 92.698] - - - [57600, 28417, 1, 384] - - [395, 91.899] - - - [57600, 28800, 1, 384] - - [395, 92.843] - - - [57600, 28801, 1, 384] - - [395, 91.921] - - - [57984, 384, 1, 384] - - [399, 82.086] - - - [57984, 28800, 1, 384] - - [395, 92.752] - - - [57984, 28801, 1, 384] - - [417, 91.77] - - - [57984, 29184, 1, 384] - - [395, 92.813] - - - [58368, 384, 1, 384] - - [400, 82.314] - - - [58368, 28800, 1, 384] - - [395, 92.908] - - - [58368, 28801, 1, 384] - - [395, 91.822] - - - [58368, 29184, 1, 384] - - [395, 92.863] - - - [58368, 29185, 1, 384] - - [395, 91.929] - - - [58752, 384, 1, 384] - - [395, 82.767] - - - [58752, 29184, 1, 384] - - [395, 92.872] - - - [58752, 29185, 1, 384] - - [395, 91.906] - - - [58752, 29568, 1, 384] - - [400, 92.739] - - - [59136, 384, 1, 384] - - [397, 83.238] - - - [59136, 29184, 1, 384] - - [441, 92.768] - - - [59136, 29185, 1, 384] - - [400, 91.899] - - - [59136, 29568, 1, 384] - - [441, 92.79] - - - [59136, 29569, 1, 384] - - [395, 91.872] - - - [59520, 384, 1, 384] - - [400, 83.478] - - - [59520, 29568, 1, 384] - - [400, 92.759] - - - [59520, 29569, 1, 384] - - [395, 91.919] - - - [59520, 29952, 1, 384] - - [395, 92.773] - - - [59904, 384, 1, 384] - - [395, 83.945] - - - [59904, 29568, 1, 384] - - [395, 92.768] - - - [59904, 29569, 1, 384] - - [395, 91.981] - - - [59904, 29952, 1, 384] - - [395, 92.85] - - - [59904, 29953, 1, 384] - - [400, 91.949] - - - [60288, 384, 1, 384] - - [415, 84.058] - - - [60288, 29952, 1, 384] - - [395, 92.915] - - - [60288, 29953, 1, 384] - - [395, 92.018] - - - [60288, 30336, 1, 384] - - [395, 92.859] - - - [60672, 384, 1, 384] - - [417, 84.593] - - - [60672, 29952, 1, 384] - - [395, 92.807] - - - [60672, 29953, 1, 384] - - [395, 92.01] - - - [60672, 30336, 1, 384] - - [395, 92.866] - - - [60672, 30337, 1, 384] - - [395, 92.048] - - - [61056, 384, 1, 384] - - [417, 81.438] - - - [61056, 30336, 1, 384] - - [395, 92.909] - - - [61056, 30337, 1, 384] - - [395, 92.036] - - - [61056, 30720, 1, 384] - - [395, 92.921] - - - [61440, 384, 1, 384] - - [395, 81.168] - - - [61440, 30336, 1, 384] - - [395, 92.906] - - - [61440, 30337, 1, 384] - - [441, 91.899] - - - [61440, 30720, 1, 384] - - [400, 92.692] - - - [61440, 30721, 1, 384] - - [395, 89.242] - - - [61824, 384, 1, 384] - - [399, 81.871] - - - [61824, 30720, 1, 384] - - [395, 92.832] - - - [61824, 30721, 1, 384] - - [395, 90.084] - - - [61824, 31104, 1, 384] - - [395, 92.818] - - - [62208, 384, 1, 384] - - [395, 82.61] - - - [62208, 30720, 1, 384] - - [395, 92.81] - - - [62208, 30721, 1, 384] - - [395, 89.954] - - - [62208, 31104, 1, 384] - - [395, 92.853] - - - [62208, 31105, 1, 384] - - [395, 92.041] - - - [62592, 384, 1, 384] - - [415, 82.707] - - - [62592, 31104, 1, 384] - - [400, 92.81] - - - [62592, 31105, 1, 384] - - [400, 91.997] - - - [62592, 31488, 1, 384] - - [400, 92.834] - - - [62976, 384, 1, 384] - - [395, 83.032] - - - [62976, 31104, 1, 384] - - [395, 92.934] - - - [62976, 31105, 1, 384] - - [395, 92.072] - - - [62976, 31488, 1, 384] - - [400, 92.826] - - - [62976, 31489, 1, 384] - - [395, 92.105] - - - [63360, 384, 1, 384] - - [400, 83.357] - - - [63360, 31488, 1, 384] - - [400, 92.821] - - - [63360, 31489, 1, 384] - - [400, 91.972] - - - [63360, 31872, 1, 384] - - [395, 92.845] - - - [63744, 384, 1, 384] - - [395, 83.63] - - - [63744, 31488, 1, 384] - - [400, 92.771] - - - [63744, 31489, 1, 384] - - [395, 92.051] - - - [63744, 31872, 1, 384] - - [395, 92.955] - - - [63744, 31873, 1, 384] - - [395, 92.08] - - - [64128, 384, 1, 384] - - [407, 84.051] - - - [64128, 31872, 1, 384] - - [395, 92.987] - - - [64128, 31873, 1, 384] - - [400, 92.03] - - - [64128, 32256, 1, 384] - - [395, 92.994] - - - [64512, 384, 1, 384] - - [395, 84.404] - - - [64512, 31872, 1, 384] - - [395, 92.857] - - - [64512, 31873, 1, 384] - - [395, 91.947] - - - [64512, 32256, 1, 384] - - [400, 92.907] - - - [64512, 32257, 1, 384] - - [395, 92.007] - - - [64896, 384, 1, 384] - - [397, 84.524] - - - [64896, 32256, 1, 384] - - [441, 92.959] - - - [64896, 32257, 1, 384] - - [395, 92.202] - - - [64896, 32640, 1, 384] - - [395, 92.866] - - - [65280, 384, 1, 384] - - [394, 84.907] - - - [65280, 32256, 1, 384] - - [400, 92.867] - - - [65280, 32257, 1, 384] - - [395, 92.085] - - - [65280, 32640, 1, 384] - - [395, 92.86] - - - [65280, 32641, 1, 384] - - [395, 92.07] - - - [65664, 384, 1, 384] - - [408, 85.074] - - - [65664, 32640, 1, 384] - - [400, 92.877] - - - [65664, 32641, 1, 384] - - [400, 92.08] - - - [65664, 33024, 1, 384] - - [400, 92.901] - - - [66048, 384, 1, 384] - - [417, 82.052] - - - [66048, 32640, 1, 384] - - [395, 93.003] - - - [66048, 32641, 1, 384] - - [395, 92.147] - - - [66048, 33024, 1, 384] - - [395, 93.038] - - - [66048, 33025, 1, 384] - - [395, 92.168] - - - [66432, 384, 1, 384] - - [399, 82.606] - - - [66432, 33024, 1, 384] - - [395, 93.004] - - - [66432, 33025, 1, 384] - - [395, 92.12] - - - [66432, 33408, 1, 384] - - [395, 92.871] - - - [66816, 384, 1, 384] - - [395, 83.022] - - - [66816, 33024, 1, 384] - - [395, 92.906] - - - [66816, 33025, 1, 384] - - [400, 91.964] - - - [66816, 33408, 1, 384] - - [395, 92.896] - - - [66816, 33409, 1, 384] - - [400, 92.155] - - - [67200, 384, 1, 384] - - [395, 83.499] - - - [67200, 33408, 1, 384] - - [400, 92.959] - - - [67200, 33409, 1, 384] - - [400, 92.197] - - - [67200, 33792, 1, 384] - - [400, 93.01] - - - [67584, 384, 1, 384] - - [395, 83.772] - - - [67584, 33408, 1, 384] - - [400, 93.003] - - - [67584, 33409, 1, 384] - - [395, 92.116] - - - [67584, 33792, 1, 384] - - [400, 92.985] - - - [67584, 33793, 1, 384] - - [395, 89.609] - - - [67968, 384, 1, 384] - - [395, 83.81] - - - [67968, 33792, 1, 384] - - [395, 92.992] - - - [67968, 33793, 1, 384] - - [395, 90.356] - - - [67968, 34176, 1, 384] - - [395, 93.013] - - - [68352, 384, 1, 384] - - [433, 84.159] - - - [68352, 33792, 1, 384] - - [400, 92.961] - - - [68352, 33793, 1, 384] - - [395, 90.159] - - - [68352, 34176, 1, 384] - - [400, 92.99] - - - [68352, 34177, 1, 384] - - [400, 92.182] - - - [68736, 384, 1, 384] - - [399, 84.257] - - - [68736, 34176, 1, 384] - - [400, 93.293] - - - [68736, 34177, 1, 384] - - [400, 92.561] - - - [68736, 34560, 1, 384] - - [399, 93.319] - - - [69120, 384, 1, 384] - - [403, 84.759] - - - [69120, 34176, 1, 384] - - [400, 93.387] - - - [69120, 34177, 1, 384] - - [400, 92.573] - - - [69120, 34560, 1, 384] - - [400, 93.349] - - - [69120, 34561, 1, 384] - - [400, 92.538] - - - [69504, 384, 1, 384] - - [400, 84.73] - - - [69504, 34560, 1, 384] - - [400, 93.382] - - - [69504, 34561, 1, 384] - - [400, 92.476] - - - [69504, 34944, 1, 384] - - [395, 93.324] - - - [69888, 384, 1, 384] - - [400, 85.214] - - - [69888, 34560, 1, 384] - - [399, 93.25] - - - [69888, 34561, 1, 384] - - [400, 92.533] - - - [69888, 34944, 1, 384] - - [395, 93.22] - - - [69888, 34945, 1, 384] - - [400, 92.505] - - - [70272, 384, 1, 384] - - [402, 85.384] - - - [70272, 34944, 1, 384] - - [395, 93.257] - - - [70272, 34945, 1, 384] - - [400, 92.503] - - - [70272, 35328, 1, 384] - - [400, 93.388] - - - [70656, 384, 1, 384] - - [438, 82.691] - - - [70656, 34944, 1, 384] - - [395, 93.221] - - - [70656, 34945, 1, 384] - - [400, 92.384] - - - [70656, 35328, 1, 384] - - [400, 93.293] - - - [70656, 35329, 1, 384] - - [400, 92.389] - - - [71040, 384, 1, 384] - - [407, 82.994] - - - [71040, 35328, 1, 384] - - [400, 93.373] - - - [71040, 35329, 1, 384] - - [400, 92.496] - - - [71040, 35712, 1, 384] - - [400, 93.306] - - - [71424, 384, 1, 384] - - [399, 83.457] - - - [71424, 35328, 1, 384] - - [400, 93.26] - - - [71424, 35329, 1, 384] - - [395, 92.429] - - - [71424, 35712, 1, 384] - - [400, 93.263] - - - [71424, 35713, 1, 384] - - [400, 92.412] - - - [71808, 384, 1, 384] - - [449, 83.773] - - - [71808, 35712, 1, 384] - - [400, 93.277] - - - [71808, 35713, 1, 384] - - [400, 92.489] - - - [71808, 36096, 1, 384] - - [400, 93.323] - - - [72192, 384, 1, 384] - - [417, 83.809] - - - [72192, 35712, 1, 384] - - [395, 93.297] - - - [72192, 35713, 1, 384] - - [400, 92.505] - - - [72192, 36096, 1, 384] - - [400, 93.265] - - - [72192, 36097, 1, 384] - - [400, 92.522] - - - [72576, 384, 1, 384] - - [399, 84.225] - - - [72576, 36096, 1, 384] - - [400, 93.277] - - - [72576, 36097, 1, 384] - - [400, 92.487] - - - [72576, 36480, 1, 384] - - [400, 93.289] - - - [72960, 384, 1, 384] - - [395, 84.48] - - - [72960, 36096, 1, 384] - - [400, 93.261] - - - [72960, 36097, 1, 384] - - [395, 92.432] - - - [72960, 36480, 1, 384] - - [400, 93.296] - - - [72960, 36481, 1, 384] - - [395, 92.439] - - - [73344, 384, 1, 384] - - [395, 84.556] - - - [73344, 36480, 1, 384] - - [395, 93.225] - - - [73344, 36481, 1, 384] - - [400, 92.421] - - - [73344, 36864, 1, 384] - - [400, 93.251] - - - [73728, 384, 1, 384] - - [417, 83.95] - - - [73728, 36480, 1, 384] - - [400, 92.849] - - - [73728, 36481, 1, 384] - - [400, 91.857] - - - [73728, 36864, 1, 384] - - [400, 92.839] - - - [73728, 36865, 1, 384] - - [395, 89.489] - - - [74112, 384, 1, 384] - - [400, 85.264] - - - [74112, 36864, 1, 384] - - [400, 93.324] - - - [74112, 36865, 1, 384] - - [395, 90.592] - - - [74112, 37248, 1, 384] - - [400, 93.274] - - - [74496, 384, 1, 384] - - [395, 85.292] - - - [74496, 36864, 1, 384] - - [400, 93.29] - - - [74496, 36865, 1, 384] - - [395, 90.42] - - - [74496, 37248, 1, 384] - - [400, 93.251] - - - [74496, 37249, 1, 384] - - [400, 92.48] - - - [74880, 384, 1, 384] - - [438, 85.54] - - - [74880, 37248, 1, 384] - - [400, 93.236] - - - [74880, 37249, 1, 384] - - [400, 92.455] - - - [74880, 37632, 1, 384] - - [395, 93.276] - - - [75264, 384, 1, 384] - - [402, 83.093] - - - [75264, 37248, 1, 384] - - [400, 93.204] - - - [75264, 37249, 1, 384] - - [400, 92.412] - - - [75264, 37632, 1, 384] - - [400, 93.187] - - - [75264, 37633, 1, 384] - - [400, 92.394] - - - [75648, 384, 1, 384] - - [417, 83.601] - - - [75648, 37632, 1, 384] - - [395, 93.231] - - - [75648, 37633, 1, 384] - - [400, 92.489] - - - [75648, 38016, 1, 384] - - [395, 93.163] - - - [76032, 384, 1, 384] - - [397, 84.025] - - - [76032, 37632, 1, 384] - - [400, 93.2] - - - [76032, 37633, 1, 384] - - [400, 92.438] - - - [76032, 38016, 1, 384] - - [395, 93.076] - - - [76032, 38017, 1, 384] - - [400, 92.405] - - - [76416, 384, 1, 384] - - [399, 83.988] - - - [76416, 38016, 1, 384] - - [395, 93.185] - - - [76416, 38017, 1, 384] - - [395, 92.391] - - - [76416, 38400, 1, 384] - - [395, 93.239] - - - [76800, 384, 1, 384] - - [399, 84.326] - - - [76800, 38016, 1, 384] - - [395, 93.08] - - - [76800, 38017, 1, 384] - - [395, 92.255] - - - [76800, 38400, 1, 384] - - [400, 93.067] - - - [76800, 38401, 1, 384] - - [400, 92.265] - - - [77184, 384, 1, 384] - - [399, 84.734] - - - [77184, 38400, 1, 384] - - [400, 93.227] - - - [77184, 38401, 1, 384] - - [395, 92.449] - - - [77184, 38784, 1, 384] - - [400, 93.166] - - - [77568, 384, 1, 384] - - [400, 84.889] - - - [77568, 38400, 1, 384] - - [400, 93.161] - - - [77568, 38401, 1, 384] - - [400, 92.448] - - - [77568, 38784, 1, 384] - - [400, 93.108] - - - [77568, 38785, 1, 384] - - [400, 92.284] - - - [77952, 384, 1, 384] - - [400, 84.862] - - - [77952, 38784, 1, 384] - - [400, 93.14] - - - [77952, 38785, 1, 384] - - [395, 92.261] - - - [77952, 39168, 1, 384] - - [400, 93.115] - - - [78336, 384, 1, 384] - - [395, 85.099] - - - [78336, 38784, 1, 384] - - [395, 93.115] - - - [78336, 38785, 1, 384] - - [400, 92.364] - - - [78336, 39168, 1, 384] - - [400, 92.987] - - - [78336, 39169, 1, 384] - - [395, 92.292] - - - [78720, 384, 1, 384] - - [438, 85.113] - - - [78720, 39168, 1, 384] - - [400, 93.066] - - - [78720, 39169, 1, 384] - - [395, 92.341] - - - [78720, 39552, 1, 384] - - [395, 93.11] - - - [79104, 384, 1, 384] - - [399, 85.514] - - - [79104, 39168, 1, 384] - - [400, 93.026] - - - [79104, 39169, 1, 384] - - [400, 92.331] - - - [79104, 39552, 1, 384] - - [400, 93.017] - - - [79104, 39553, 1, 384] - - [400, 92.267] - - - [79488, 384, 1, 384] - - [399, 85.803] - - - [79488, 39552, 1, 384] - - [400, 93.086] - - - [79488, 39553, 1, 384] - - [400, 92.355] - - - [79488, 39936, 1, 384] - - [400, 93.132] - - - [79872, 384, 1, 384] - - [408, 83.822] - - - [79872, 39552, 1, 384] - - [400, 93.062] - - - [79872, 39553, 1, 384] - - [395, 92.254] - - - [79872, 39936, 1, 384] - - [400, 92.874] - - - [79872, 39937, 1, 384] - - [395, 89.714] - - - [80256, 384, 1, 384] - - [417, 84.158] - - - [80256, 39936, 1, 384] - - [400, 93.044] - - - [80256, 39937, 1, 384] - - [395, 90.42] - - - [80256, 40320, 1, 384] - - [400, 93.059] - - - [80640, 384, 1, 384] - - [407, 84.255] - - - [80640, 39936, 1, 384] - - [400, 93.019] - - - [80640, 39937, 1, 384] - - [395, 90.379] - - - [80640, 40320, 1, 384] - - [400, 92.949] - - - [80640, 40321, 1, 384] - - [400, 92.218] - - - [81024, 384, 1, 384] - - [394, 84.55] - - - [81024, 40320, 1, 384] - - [400, 93.055] - - - [81024, 40321, 1, 384] - - [400, 92.284] - - - [81024, 40704, 1, 384] - - [400, 93.043] - - - [81408, 384, 1, 384] - - [400, 84.924] - - - [81408, 40320, 1, 384] - - [400, 93.024] - - - [81408, 40321, 1, 384] - - [395, 92.24] - - - [81408, 40704, 1, 384] - - [400, 92.926] - - - [81408, 40705, 1, 384] - - [395, 92.158] - - - [81792, 384, 1, 384] - - [417, 84.885] - - - [81792, 40704, 1, 384] - - [395, 93.022] - - - [81792, 40705, 1, 384] - - [395, 92.249] - - - [81792, 41088, 1, 384] - - [395, 92.901] - - - [82176, 384, 1, 384] - - [407, 84.885] - - - [82176, 40704, 1, 384] - - [400, 92.882] - - - [82176, 40705, 1, 384] - - [400, 92.172] - - - [82176, 41088, 1, 384] - - [400, 92.857] - - - [82176, 41089, 1, 384] - - [400, 92.159] - - - [82560, 384, 1, 384] - - [395, 85.428] - - - [82560, 41088, 1, 384] - - [400, 92.896] - - - [82560, 41089, 1, 384] - - [400, 92.244] - - - [82560, 41472, 1, 384] - - [400, 93.001] - - - [82944, 384, 1, 384] - - [400, 85.268] - - - [82944, 41088, 1, 384] - - [400, 92.841] - - - [82944, 41089, 1, 384] - - [395, 92.056] - - - [82944, 41472, 1, 384] - - [400, 92.819] - - - [82944, 41473, 1, 384] - - [395, 92.008] - - - [83328, 384, 1, 384] - - [395, 85.57] - - - [83328, 41472, 1, 384] - - [400, 92.98] - - - [83328, 41473, 1, 384] - - [395, 92.201] - - - [83328, 41856, 1, 384] - - [400, 92.891] - - - [83712, 384, 1, 384] - - [400, 85.62] - - - [83712, 41472, 1, 384] - - [400, 92.896] - - - [83712, 41473, 1, 384] - - [400, 92.161] - - - [83712, 41856, 1, 384] - - [400, 92.813] - - - [83712, 41857, 1, 384] - - [400, 92.061] - - - [84096, 384, 1, 384] - - [417, 85.802] - - - [84096, 41856, 1, 384] - - [395, 92.879] - - - [84096, 41857, 1, 384] - - [400, 92.137] - - - [84096, 42240, 1, 384] - - [395, 92.852] - - - [84480, 384, 1, 384] - - [438, 85.785] - - - [84480, 41856, 1, 384] - - [400, 92.877] - - - [84480, 41857, 1, 384] - - [400, 92.105] - - - [84480, 42240, 1, 384] - - [400, 92.735] - - - [84480, 42241, 1, 384] - - [395, 92.007] - - - [84864, 384, 1, 384] - - [399, 84.383] - - - [84864, 42240, 1, 384] - - [395, 92.853] - - - [84864, 42241, 1, 384] - - [395, 92.101] - - - [84864, 42624, 1, 384] - - [395, 92.8] - - - [85248, 384, 1, 384] - - [408, 84.6] - - - [85248, 42240, 1, 384] - - [395, 92.729] - - - [85248, 42241, 1, 384] - - [395, 92.041] - - - [85248, 42624, 1, 384] - - [400, 92.705] - - - [85248, 42625, 1, 384] - - [400, 91.991] - - - [85632, 384, 1, 384] - - [399, 84.64] - - - [85632, 42624, 1, 384] - - [395, 92.799] - - - [85632, 42625, 1, 384] - - [400, 92.058] - - - [85632, 43008, 1, 384] - - [400, 92.754] - - - [86016, 384, 1, 384] - - [400, 84.981] - - - [86016, 42624, 1, 384] - - [395, 92.705] - - - [86016, 42625, 1, 384] - - [395, 91.928] - - - [86016, 43008, 1, 384] - - [400, 92.53] - - - [86016, 43009, 1, 384] - - [395, 89.555] - - - [86400, 384, 1, 384] - - [400, 85.335] - - - [86400, 43008, 1, 384] - - [400, 92.722] - - - [86400, 43009, 1, 384] - - [395, 90.219] - - - [86400, 43392, 1, 384] - - [400, 92.676] - - - [86784, 384, 1, 384] - - [400, 85.372] - - - [86784, 43008, 1, 384] - - [400, 92.684] - - - [86784, 43009, 1, 384] - - [395, 90.15] - - - [86784, 43392, 1, 384] - - [400, 92.57] - - - [86784, 43393, 1, 384] - - [400, 91.889] - - - [87168, 384, 1, 384] - - [417, 85.547] - - - [87168, 43392, 1, 384] - - [395, 92.659] - - - [87168, 43393, 1, 384] - - [395, 91.95] - - - [87168, 43776, 1, 384] - - [395, 92.562] - - - [87552, 384, 1, 384] - - [400, 85.558] - - - [87552, 43392, 1, 384] - - [395, 92.603] - - - [87552, 43393, 1, 384] - - [395, 91.913] - - - [87552, 43776, 1, 384] - - [400, 92.691] - - - [87552, 43777, 1, 384] - - [395, 92.01] - - - [87936, 384, 1, 384] - - [395, 85.631] - - - [87936, 43776, 1, 384] - - [400, 92.644] - - - [87936, 43777, 1, 384] - - [395, 91.946] - - - [87936, 44160, 1, 384] - - [395, 92.667] - - - [88320, 384, 1, 384] - - [438, 85.728] - - - [88320, 43776, 1, 384] - - [395, 92.599] - - - [88320, 43777, 1, 384] - - [400, 91.936] - - - [88320, 44160, 1, 384] - - [395, 92.554] - - - [88320, 44161, 1, 384] - - [400, 91.84] - - - [88704, 384, 1, 384] - - [446, 85.85] - - - [88704, 44160, 1, 384] - - [395, 92.674] - - - [88704, 44161, 1, 384] - - [395, 91.923] - - - [88704, 44544, 1, 384] - - [439, 92.604] - - - [89088, 384, 1, 384] - - [400, 85.71] - - - [89088, 44160, 1, 384] - - [395, 92.542] - - - [89088, 44161, 1, 384] - - [395, 91.767] - - - [89088, 44544, 1, 384] - - [400, 92.487] - - - [89088, 44545, 1, 384] - - [400, 91.739] - - - [89472, 384, 1, 384] - - [433, 84.73] - - - [89472, 44544, 1, 384] - - [400, 92.645] - - - [89472, 44545, 1, 384] - - [395, 91.956] - - - [89472, 44928, 1, 384] - - [400, 92.562] - - - [89856, 384, 1, 384] - - [438, 84.9] - - - [89856, 44544, 1, 384] - - [439, 92.527] - - - [89856, 44545, 1, 384] - - [400, 91.855] - - - [89856, 44928, 1, 384] - - [400, 92.445] - - - [89856, 44929, 1, 384] - - [400, 91.784] - - - [90240, 384, 1, 384] - - [407, 84.913] - - - [90240, 44928, 1, 384] - - [400, 92.532] - - - [90240, 44929, 1, 384] - - [400, 91.795] - - - [90240, 45312, 1, 384] - - [395, 92.524] - - - [90624, 384, 1, 384] - - [395, 85.444] - - - [90624, 44928, 1, 384] - - [400, 92.527] - - - [90624, 44929, 1, 384] - - [395, 91.83] - - - [90624, 45312, 1, 384] - - [395, 92.441] - - - [90624, 45313, 1, 384] - - [395, 91.775] - - - [91008, 384, 1, 384] - - [395, 85.696] - - - [91008, 45312, 1, 384] - - [395, 92.513] - - - [91008, 45313, 1, 384] - - [395, 91.795] - - - [91008, 45696, 1, 384] - - [400, 92.476] - - - [91392, 384, 1, 384] - - [417, 85.778] - - - [91392, 45312, 1, 384] - - [395, 92.421] - - - [91392, 45313, 1, 384] - - [395, 91.749] - - - [91392, 45696, 1, 384] - - [400, 92.37] - - - [91392, 45697, 1, 384] - - [395, 91.674] - - - [91776, 384, 1, 384] - - [402, 85.607] - - - [91776, 45696, 1, 384] - - [400, 92.379] - - - [91776, 45697, 1, 384] - - [400, 91.674] - - - [91776, 46080, 1, 384] - - [439, 92.42] - - - [92160, 384, 1, 384] - - [395, 85.856] - - - [92160, 45696, 1, 384] - - [400, 92.373] - - - [92160, 45697, 1, 384] - - [395, 91.624] - - - [92160, 46080, 1, 384] - - [400, 92.245] - - - [92160, 46081, 1, 384] - - [395, 89.53] - - - [92544, 384, 1, 384] - - [395, 85.793] - - - [92544, 46080, 1, 384] - - [439, 92.469] - - - [92544, 46081, 1, 384] - - [395, 90.234] - - - [92544, 46464, 1, 384] - - [395, 92.34] - - - [92928, 384, 1, 384] - - [408, 85.766] - - - [92928, 46080, 1, 384] - - [439, 92.435] - - - [92928, 46081, 1, 384] - - [395, 90.168] - - - [92928, 46464, 1, 384] - - [439, 92.313] - - - [92928, 46465, 1, 384] - - [395, 91.607] - - - [93312, 384, 1, 384] - - [395, 86.085] - - - [93312, 46464, 1, 384] - - [395, 92.345] - - - [93312, 46465, 1, 384] - - [400, 91.61] - - - [93312, 46848, 1, 384] - - [501, 92.327] - - - [93696, 384, 1, 384] - - [438, 86.104] - - - [93696, 46464, 1, 384] - - [439, 92.376] - - - [93696, 46465, 1, 384] - - [400, 91.604] - - - [93696, 46848, 1, 384] - - [439, 92.291] - - - [93696, 46849, 1, 384] - - [395, 91.547] - - - [94080, 384, 1, 384] - - [417, 85.057] - - - [94080, 46848, 1, 384] - - [439, 92.315] - - - [94080, 46849, 1, 384] - - [395, 91.594] - - - [94080, 47232, 1, 384] - - [448, 92.251] - - - [94464, 384, 1, 384] - - [395, 85.335] - - - [94464, 46848, 1, 384] - - [439, 92.291] - - - [94464, 46849, 1, 384] - - [395, 91.527] - - - [94464, 47232, 1, 384] - - [448, 92.204] - - - [94464, 47233, 1, 384] - - [395, 91.484] - - - [94848, 384, 1, 384] - - [417, 85.438] - - - [94848, 47232, 1, 384] - - [448, 92.265] - - - [94848, 47233, 1, 384] - - [395, 91.512] - - - [94848, 47616, 1, 384] - - [439, 92.368] - - - [95232, 384, 1, 384] - - [395, 85.517] - - - [95232, 47232, 1, 384] - - [395, 92.123] - - - [95232, 47233, 1, 384] - - [395, 91.376] - - - [95232, 47616, 1, 384] - - [439, 92.09] - - - [95232, 47617, 1, 384] - - [395, 91.374] - - - [95616, 384, 1, 384] - - [438, 85.688] - - - [95616, 47616, 1, 384] - - [439, 92.354] - - - [95616, 47617, 1, 384] - - [395, 91.487] - - - [95616, 48000, 1, 384] - - [439, 92.264] - - - [96000, 384, 1, 384] - - [402, 85.525] - - - [96000, 47616, 1, 384] - - [439, 92.296] - - - [96000, 47617, 1, 384] - - [439, 91.458] - - - [96000, 48000, 1, 384] - - [439, 92.25] - - - [96000, 48001, 1, 384] - - [439, 91.381] - - - [96384, 384, 1, 384] - - [394, 85.971] - - - [96384, 48000, 1, 384] - - [439, 92.173] - - - [96384, 48001, 1, 384] - - [400, 91.307] - - - [96384, 48384, 1, 384] - - [439, 92.169] - - - [96768, 384, 1, 384] - - [395, 86.094] - - - [96768, 48000, 1, 384] - - [439, 92.225] - - - [96768, 48001, 1, 384] - - [439, 91.378] - - - [96768, 48384, 1, 384] - - [439, 92.13] - - - [96768, 48385, 1, 384] - - [439, 91.359] - - - [97152, 384, 1, 384] - - [395, 86.067] - - - [97152, 48384, 1, 384] - - [439, 92.157] - - - [97152, 48385, 1, 384] - - [395, 91.313] - - - [97152, 48768, 1, 384] - - [439, 92.119] - - - [97536, 384, 1, 384] - - [395, 85.971] - - - [97536, 48384, 1, 384] - - [439, 92.149] - - - [97536, 48385, 1, 384] - - [439, 91.311] - - - [97536, 48768, 1, 384] - - [439, 92.152] - - - [97536, 48769, 1, 384] - - [439, 91.338] - - - [97920, 384, 1, 384] - - [402, 85.934] - - - [97920, 48768, 1, 384] - - [439, 91.965] - - - [97920, 48769, 1, 384] - - [395, 91.209] - - - [97920, 49152, 1, 384] - - [439, 91.911] - - - [98304, 384, 1, 384] - - [395, 83.088] - - - [98304, 48768, 1, 384] - - [400, 90.529] - - - [98304, 48769, 1, 384] - - [400, 89.38] - - - [98304, 49152, 1, 384] - - [400, 90.415] - - - [98304, 49153, 1, 384] - - [400, 88.259] - - - [98688, 384, 1, 384] - - [433, 85.423] - - - [98688, 49152, 1, 384] - - [439, 92.242] - - - [98688, 49153, 1, 384] - - [395, 89.64] - - - [98688, 49536, 1, 384] - - [439, 92.209] - - - [99072, 384, 1, 384] - - [395, 85.3] - - - [99072, 49152, 1, 384] - - [439, 92.029] - - - [99072, 49153, 1, 384] - - [395, 89.533] - - - [99072, 49536, 1, 384] - - [439, 92.069] - - - [99072, 49537, 1, 384] - - [439, 91.256] - - - [99456, 384, 1, 384] - - [394, 85.653] - - - [99456, 49536, 1, 384] - - [439, 92.083] - - - [99456, 49537, 1, 384] - - [439, 91.199] - - - [99456, 49920, 1, 384] - - [439, 92.093] - - - [99840, 384, 1, 384] - - [395, 85.95] - - - [99840, 49536, 1, 384] - - [439, 92.038] - - - [99840, 49537, 1, 384] - - [439, 91.205] - - - [99840, 49920, 1, 384] - - [439, 91.964] - - - [99840, 49921, 1, 384] - - [439, 91.191] - - - [100224, 384, 1, 384] - - [395, 86.158] - - - [100224, 49920, 1, 384] - - [439, 92.003] - - - [100224, 49921, 1, 384] - - [439, 91.196] - - - [100224, 50304, 1, 384] - - [448, 91.89] - - - [100608, 384, 1, 384] - - [395, 85.965] - - - [100608, 49920, 1, 384] - - [439, 91.931] - - - [100608, 49921, 1, 384] - - [439, 91.143] - - - [100608, 50304, 1, 384] - - [448, 91.88] - - - [100608, 50305, 1, 384] - - [439, 91.127] - - - [100992, 384, 1, 384] - - [395, 86.238] - - - [100992, 50304, 1, 384] - - [439, 91.849] - - - [100992, 50305, 1, 384] - - [439, 91.061] - - - [100992, 50688, 1, 384] - - [501, 91.964] - - - [101376, 384, 1, 384] - - [395, 85.991] - - - [101376, 50304, 1, 384] - - [439, 91.645] - - - [101376, 50305, 1, 384] - - [439, 90.706] - - - [101376, 50688, 1, 384] - - [439, 91.686] - - - [101376, 50689, 1, 384] - - [439, 90.729] - - - [101760, 384, 1, 384] - - [395, 86.229] - - - [101760, 50688, 1, 384] - - [439, 91.893] - - - [101760, 50689, 1, 384] - - [439, 91.067] - - - [101760, 51072, 1, 384] - - [439, 91.814] - - - [102144, 384, 1, 384] - - [400, 86.093] - - - [102144, 50688, 1, 384] - - [439, 91.848] - - - [102144, 50689, 1, 384] - - [439, 91.071] - - - [102144, 51072, 1, 384] - - [439, 91.826] - - - [102144, 51073, 1, 384] - - [439, 91.037] - - - [102528, 384, 1, 384] - - [395, 85.905] - - - [102528, 51072, 1, 384] - - [439, 91.777] - - - [102528, 51073, 1, 384] - - [439, 90.938] - - - [102528, 51456, 1, 384] - - [439, 91.773] - - - [102912, 384, 1, 384] - - [449, 86.074] - - - [102912, 51072, 1, 384] - - [439, 91.76] - - - [102912, 51073, 1, 384] - - [439, 90.948] - - - [102912, 51456, 1, 384] - - [439, 91.7] - - - [102912, 51457, 1, 384] - - [439, 90.918] - - - [103296, 384, 1, 384] - - [438, 85.545] - - - [103296, 51456, 1, 384] - - [439, 91.67] - - - [103296, 51457, 1, 384] - - [439, 90.841] - - - [103296, 51840, 1, 384] - - [439, 91.649] - - - [103680, 384, 1, 384] - - [402, 85.767] - - - [103680, 51456, 1, 384] - - [439, 91.615] - - - [103680, 51457, 1, 384] - - [439, 90.854] - - - [103680, 51840, 1, 384] - - [439, 91.681] - - - [103680, 51841, 1, 384] - - [439, 90.862] - - - [104064, 384, 1, 384] - - [400, 85.796] - - - [104064, 51840, 1, 384] - - [439, 91.625] - - - [104064, 51841, 1, 384] - - [439, 90.789] - - - [104064, 52224, 1, 384] - - [439, 91.638] - - - [104448, 384, 1, 384] - - [395, 85.91] - - - [104448, 51840, 1, 384] - - [439, 91.423] - - - [104448, 51841, 1, 384] - - [439, 90.442] - - - [104448, 52224, 1, 384] - - [439, 91.356] - - - [104448, 52225, 1, 384] - - [439, 89.319] - - - [104832, 384, 1, 384] - - [438, 86.237] - - - [104832, 52224, 1, 384] - - [439, 91.544] - - - [104832, 52225, 1, 384] - - [448, 89.134] - - - [104832, 52608, 1, 384] - - [439, 91.464] - - - [105216, 384, 1, 384] - - [400, 86.435] - - - [105216, 52224, 1, 384] - - [439, 91.546] - - - [105216, 52225, 1, 384] - - [448, 89.086] - - - [105216, 52608, 1, 384] - - [439, 91.48] - - - [105216, 52609, 1, 384] - - [439, 90.705] - - - [105600, 384, 1, 384] - - [438, 86.154] - - - [105600, 52608, 1, 384] - - [439, 91.435] - - - [105600, 52609, 1, 384] - - [439, 90.608] - - - [105600, 52992, 1, 384] - - [439, 91.462] - - - [105984, 384, 1, 384] - - [400, 86.234] - - - [105984, 52608, 1, 384] - - [439, 91.46] - - - [105984, 52609, 1, 384] - - [439, 90.635] - - - [105984, 52992, 1, 384] - - [439, 91.424] - - - [105984, 52993, 1, 384] - - [439, 90.635] - - - [106368, 384, 1, 384] - - [402, 86.031] - - - [106368, 52992, 1, 384] - - [501, 91.416] - - - [106368, 52993, 1, 384] - - [439, 90.619] - - - [106368, 53376, 1, 384] - - [439, 91.312] - - - [106752, 384, 1, 384] - - [399, 85.996] - - - [106752, 52992, 1, 384] - - [439, 91.346] - - - [106752, 52993, 1, 384] - - [439, 90.582] - - - [106752, 53376, 1, 384] - - [439, 91.264] - - - [106752, 53377, 1, 384] - - [439, 90.539] - - - [107136, 384, 1, 384] - - [449, 85.651] - - - [107136, 53376, 1, 384] - - [439, 91.229] - - - [107136, 53377, 1, 384] - - [439, 90.459] - - - [107136, 53760, 1, 384] - - [501, 91.374] - - - [107520, 384, 1, 384] - - [395, 85.71] - - - [107520, 53376, 1, 384] - - [439, 91.028] - - - [107520, 53377, 1, 384] - - [439, 90.13] - - - [107520, 53760, 1, 384] - - [439, 91.077] - - - [107520, 53761, 1, 384] - - [439, 90.139] - - - [107904, 384, 1, 384] - - [430, 85.985] - - - [107904, 53760, 1, 384] - - [501, 91.337] - - - [107904, 53761, 1, 384] - - [439, 90.502] - - - [107904, 54144, 1, 384] - - [439, 91.226] - - - [108288, 384, 1, 384] - - [399, 85.849] - - - [108288, 53760, 1, 384] - - [439, 91.24] - - - [108288, 53761, 1, 384] - - [439, 90.48] - - - [108288, 54144, 1, 384] - - [439, 91.2] - - - [108288, 54145, 1, 384] - - [439, 90.388] - - - [108672, 384, 1, 384] - - [394, 86.03] - - - [108672, 54144, 1, 384] - - [439, 91.138] - - - [108672, 54145, 1, 384] - - [439, 90.33] - - - [108672, 54528, 1, 384] - - [501, 91.222] - - - [109056, 384, 1, 384] - - [395, 86.283] - - - [109056, 54144, 1, 384] - - [439, 91.125] - - - [109056, 54145, 1, 384] - - [439, 90.304] - - - [109056, 54528, 1, 384] - - [439, 91.003] - - - [109056, 54529, 1, 384] - - [439, 90.259] - - - [109440, 384, 1, 384] - - [400, 86.279] - - - [109440, 54528, 1, 384] - - [501, 91.138] - - - [109440, 54529, 1, 384] - - [439, 90.211] - - - [109440, 54912, 1, 384] - - [439, 91.093] - - - [109824, 384, 1, 384] - - [395, 86.509] - - - [109824, 54528, 1, 384] - - [501, 91.06] - - - [109824, 54529, 1, 384] - - [439, 90.262] - - - [109824, 54912, 1, 384] - - [439, 90.972] - - - [109824, 54913, 1, 384] - - [439, 90.191] - - - [110208, 384, 1, 384] - - [395, 86.545] - - - [110208, 54912, 1, 384] - - [439, 90.96] - - - [110208, 54913, 1, 384] - - [439, 90.183] - - - [110208, 55296, 1, 384] - - [439, 90.979] - - - [110592, 384, 1, 384] - - [400, 86.144] - - - [110592, 54912, 1, 384] - - [439, 90.716] - - - [110592, 54913, 1, 384] - - [439, 89.759] - - - [110592, 55296, 1, 384] - - [439, 90.609] - - - [110592, 55297, 1, 384] - - [439, 88.908] - - - [110976, 384, 1, 384] - - [395, 86.413] - - - [110976, 55296, 1, 384] - - [439, 90.869] - - - [110976, 55297, 1, 384] - - [448, 88.698] - - - [110976, 55680, 1, 384] - - [439, 90.861] - - - [111360, 384, 1, 384] - - [394, 86.157] - - - [111360, 55296, 1, 384] - - [439, 90.804] - - - [111360, 55297, 1, 384] - - [448, 88.707] - - - [111360, 55680, 1, 384] - - [439, 90.783] - - - [111360, 55681, 1, 384] - - [439, 90.009] - - - [111744, 384, 1, 384] - - [402, 85.775] - - - [111744, 55680, 1, 384] - - [439, 90.744] - - - [111744, 55681, 1, 384] - - [439, 89.952] - - - [111744, 56064, 1, 384] - - [501, 90.793] - - - [112128, 384, 1, 384] - - [400, 85.91] - - - [112128, 55680, 1, 384] - - [439, 90.717] - - - [112128, 55681, 1, 384] - - [439, 89.943] - - - [112128, 56064, 1, 384] - - [501, 90.637] - - - [112128, 56065, 1, 384] - - [439, 89.85] - - - [112512, 384, 1, 384] - - [446, 85.898] - - - [112512, 56064, 1, 384] - - [501, 90.689] - - - [112512, 56065, 1, 384] - - [439, 89.837] - - - [112512, 56448, 1, 384] - - [439, 90.582] - - - [112896, 384, 1, 384] - - [395, 85.934] - - - [112896, 56064, 1, 384] - - [501, 90.592] - - - [112896, 56065, 1, 384] - - [439, 89.825] - - - [112896, 56448, 1, 384] - - [439, 90.518] - - - [112896, 56449, 1, 384] - - [439, 89.752] - - - [113280, 384, 1, 384] - - [407, 86.148] - - - [113280, 56448, 1, 384] - - [501, 90.474] - - - [113280, 56449, 1, 384] - - [439, 89.73] - - - [113280, 56832, 1, 384] - - [501, 90.603] - - - [113664, 384, 1, 384] - - [395, 86.301] - - - [113664, 56448, 1, 384] - - [439, 90.29] - - - [113664, 56449, 1, 384] - - [439, 89.419] - - - [113664, 56832, 1, 384] - - [501, 90.28] - - - [113664, 56833, 1, 384] - - [439, 89.377] - - - [114048, 384, 1, 384] - - [438, 86.415] - - - [114048, 56832, 1, 384] - - [501, 90.525] - - - [114048, 56833, 1, 384] - - [439, 89.653] - - - [114048, 57216, 1, 384] - - [502, 90.427] - - - [114432, 384, 1, 384] - - [400, 86.588] - - - [114432, 56832, 1, 384] - - [501, 90.441] - - - [114432, 56833, 1, 384] - - [439, 89.587] - - - [114432, 57216, 1, 384] - - [439, 90.348] - - - [114432, 57217, 1, 384] - - [439, 89.538] - - - [114816, 384, 1, 384] - - [402, 86.259] - - - [114816, 57216, 1, 384] - - [502, 90.473] - - - [114816, 57217, 1, 384] - - [439, 89.361] - - - [114816, 57600, 1, 384] - - [501, 90.242] - - - [115200, 384, 1, 384] - - [438, 86.304] - - - [115200, 57216, 1, 384] - - [439, 90.202] - - - [115200, 57217, 1, 384] - - [439, 89.376] - - - [115200, 57600, 1, 384] - - [501, 90.129] - - - [115200, 57601, 1, 384] - - [448, 89.271] - - - [115584, 384, 1, 384] - - [399, 86.126] - - - [115584, 57600, 1, 384] - - [501, 90.196] - - - [115584, 57601, 1, 384] - - [439, 89.331] - - - [115584, 57984, 1, 384] - - [502, 90.21] - - - [115968, 384, 1, 384] - - [395, 86.074] - - - [115968, 57600, 1, 384] - - [501, 90.1] - - - [115968, 57601, 1, 384] - - [439, 89.325] - - - [115968, 57984, 1, 384] - - [502, 90.093] - - - [115968, 57985, 1, 384] - - [448, 89.167] - - - [116352, 384, 1, 384] - - [446, 85.736] - - - [116352, 57984, 1, 384] - - [502, 90.236] - - - [116352, 57985, 1, 384] - - [439, 89.217] - - - [116352, 58368, 1, 384] - - [502, 90.124] - - - [116736, 384, 1, 384] - - [395, 85.64] - - - [116736, 57984, 1, 384] - - [439, 89.847] - - - [116736, 57985, 1, 384] - - [439, 88.975] - - - [116736, 58368, 1, 384] - - [439, 89.712] - - - [116736, 58369, 1, 384] - - [439, 88.209] - - - [117120, 384, 1, 384] - - [400, 85.95] - - - [117120, 58368, 1, 384] - - [502, 90.117] - - - [117120, 58369, 1, 384] - - [439, 88.082] - - - [117120, 58752, 1, 384] - - [502, 90.026] - - - [117504, 384, 1, 384] - - [395, 85.956] - - - [117504, 58368, 1, 384] - - [502, 89.982] - - - [117504, 58369, 1, 384] - - [448, 88.005] - - - [117504, 58752, 1, 384] - - [502, 89.953] - - - [117504, 58753, 1, 384] - - [448, 88.921] - - - [117888, 384, 1, 384] - - [395, 86.237] - - - [117888, 58752, 1, 384] - - [502, 90.067] - - - [117888, 58753, 1, 384] - - [448, 88.884] - - - [117888, 59136, 1, 384] - - [502, 89.937] - - - [118272, 384, 1, 384] - - [400, 86.366] - - - [118272, 58752, 1, 384] - - [502, 89.796] - - - [118272, 58753, 1, 384] - - [439, 88.931] - - - [118272, 59136, 1, 384] - - [502, 89.745] - - - [118272, 59137, 1, 384] - - [439, 88.873] - - - [118656, 384, 1, 384] - - [400, 86.58] - - - [118656, 59136, 1, 384] - - [502, 89.958] - - - [118656, 59137, 1, 384] - - [502, 88.873] - - - [118656, 59520, 1, 384] - - [502, 89.96] - - - [119040, 384, 1, 384] - - [395, 86.688] - - - [119040, 59136, 1, 384] - - [502, 89.935] - - - [119040, 59137, 1, 384] - - [501, 88.771] - - - [119040, 59520, 1, 384] - - [502, 89.886] - - - [119040, 59521, 1, 384] - - [502, 88.825] - - - [119424, 384, 1, 384] - - [395, 86.589] - - - [119424, 59520, 1, 384] - - [502, 89.932] - - - [119424, 59521, 1, 384] - - [502, 88.728] - - - [119424, 59904, 1, 384] - - [502, 89.941] - - - [119808, 384, 1, 384] - - [395, 86.424] - - - [119808, 59520, 1, 384] - - [501, 89.317] - - - [119808, 59521, 1, 384] - - [439, 88.467] - - - [119808, 59904, 1, 384] - - [501, 89.37] - - - [119808, 59905, 1, 384] - - [439, 88.456] - - - [120192, 384, 1, 384] - - [438, 86.028] - - - [120192, 59904, 1, 384] - - [502, 89.864] - - - [120192, 59905, 1, 384] - - [502, 88.689] - - - [120192, 60288, 1, 384] - - [502, 89.891] - - - [120576, 384, 1, 384] - - [400, 86.007] - - - [120576, 59904, 1, 384] - - [502, 89.859] - - - [120576, 59905, 1, 384] - - [439, 88.654] - - - [120576, 60288, 1, 384] - - [502, 89.833] - - - [120576, 60289, 1, 384] - - [502, 88.734] - - - [120960, 384, 1, 384] - - [400, 86.068] - - - [120960, 60288, 1, 384] - - [502, 89.911] - - - [120960, 60289, 1, 384] - - [502, 88.696] - - - [120960, 60672, 1, 384] - - [502, 89.712] - - - [121344, 384, 1, 384] - - [394, 85.718] - - - [121344, 60288, 1, 384] - - [502, 89.674] - - - [121344, 60289, 1, 384] - - [502, 88.514] - - - [121344, 60672, 1, 384] - - [502, 89.561] - - - [121344, 60673, 1, 384] - - [439, 88.345] - - - [121728, 384, 1, 384] - - [417, 85.97] - - - [121728, 60672, 1, 384] - - [502, 89.683] - - - [121728, 60673, 1, 384] - - [502, 88.674] - - - [121728, 61056, 1, 384] - - [502, 89.845] - - - [122112, 384, 1, 384] - - [400, 85.967] - - - [122112, 60672, 1, 384] - - [502, 89.619] - - - [122112, 60673, 1, 384] - - [502, 88.544] - - - [122112, 61056, 1, 384] - - [502, 89.747] - - - [122112, 61057, 1, 384] - - [502, 88.614] - - - [122496, 384, 1, 384] - - [400, 86.11] - - - [122496, 61056, 1, 384] - - [502, 89.801] - - - [122496, 61057, 1, 384] - - [502, 88.637] - - - [122496, 61440, 1, 384] - - [502, 89.814] - - - [122880, 384, 1, 384] - - [395, 85.61] - - - [122880, 61056, 1, 384] - - [501, 88.112] - - - [122880, 61057, 1, 384] - - [439, 87.07] - - - [122880, 61440, 1, 384] - - [439, 87.958] - - - [122880, 61441, 1, 384] - - [439, 86.846] - - - [123264, 384, 1, 384] - - [395, 86.604] - - - [123264, 61440, 1, 384] - - [502, 89.954] - - - [123264, 61441, 1, 384] - - [448, 87.219] - - - [123264, 61824, 1, 384] - - [502, 89.819] - - - [123648, 384, 1, 384] - - [395, 86.563] - - - [123648, 61440, 1, 384] - - [502, 89.723] - - - [123648, 61441, 1, 384] - - [448, 87.212] - - - [123648, 61824, 1, 384] - - [502, 89.78] - - - [123648, 61825, 1, 384] - - [502, 88.655] - - - [124032, 384, 1, 384] - - [395, 86.54] - - - [124032, 61824, 1, 384] - - [502, 89.751] - - - [124032, 61825, 1, 384] - - [502, 88.56] - - - [124032, 62208, 1, 384] - - [502, 89.665] - - - [124416, 384, 1, 384] - - [400, 86.59] - - - [124416, 61824, 1, 384] - - [502, 89.473] - - - [124416, 61825, 1, 384] - - [502, 88.332] - - - [124416, 62208, 1, 384] - - [502, 89.445] - - - [124416, 62209, 1, 384] - - [502, 88.124] - - - [124800, 384, 1, 384] - - [395, 86.117] - - - [124800, 62208, 1, 384] - - [502, 89.66] - - - [124800, 62209, 1, 384] - - [502, 88.568] - - - [124800, 62592, 1, 384] - - [502, 89.653] - - - [125184, 384, 1, 384] - - [395, 85.98] - - - [125184, 62208, 1, 384] - - [502, 89.659] - - - [125184, 62209, 1, 384] - - [502, 88.411] - - - [125184, 62592, 1, 384] - - [502, 89.58] - - - [125184, 62593, 1, 384] - - [502, 88.531] - - - [125568, 384, 1, 384] - - [433, 85.854] - - - [125568, 62592, 1, 384] - - [502, 89.625] - - - [125568, 62593, 1, 384] - - [502, 88.456] - - - [125568, 62976, 1, 384] - - [502, 89.539] - - - [125952, 384, 1, 384] - - [395, 85.737] - - - [125952, 62592, 1, 384] - - [502, 88.72] - - - [125952, 62593, 1, 384] - - [383, 87.36] - - - [125952, 62976, 1, 384] - - [502, 88.617] - - - [125952, 62977, 1, 384] - - [448, 87.21] - - - [126336, 384, 1, 384] - - [394, 85.814] - - - [126336, 62976, 1, 384] - - [502, 89.525] - - - [126336, 62977, 1, 384] - - [502, 88.393] - - - [126336, 63360, 1, 384] - - [502, 89.631] - - - [126720, 384, 1, 384] - - [395, 86.198] - - - [126720, 62976, 1, 384] - - [502, 89.554] - - - [126720, 62977, 1, 384] - - [502, 88.354] - - - [126720, 63360, 1, 384] - - [502, 89.587] - - - [126720, 63361, 1, 384] - - [502, 88.496] - - - [127104, 384, 1, 384] - - [395, 86.089] - - - [127104, 63360, 1, 384] - - [502, 89.608] - - - [127104, 63361, 1, 384] - - [502, 88.441] - - - [127104, 63744, 1, 384] - - [502, 89.373] - - - [127488, 384, 1, 384] - - [400, 86.384] - - - [127488, 63360, 1, 384] - - [502, 89.338] - - - [127488, 63361, 1, 384] - - [502, 88.193] - - - [127488, 63744, 1, 384] - - [502, 89.139] - - - [127488, 63745, 1, 384] - - [502, 87.803] - - - [127872, 384, 1, 384] - - [400, 86.187] - - - [127872, 63744, 1, 384] - - [502, 89.362] - - - [127872, 63745, 1, 384] - - [502, 88.345] - - - [127872, 64128, 1, 384] - - [502, 89.537] - - - [128256, 384, 1, 384] - - [400, 86.71] - - - [128256, 63744, 1, 384] - - [502, 89.328] - - - [128256, 63745, 1, 384] - - [502, 88.21] - - - [128256, 64128, 1, 384] - - [502, 89.482] - - - [768, 1, 1, 384] - - [418, 0.057] - - - [64128, 127489, 1, 384] - - [414, 87.906] - - - [63744, 126721, 1, 384] - - [414, 88.217] - - - [63744, 127105, 1, 384] - - [412, 88.569] - - - [63744, 127489, 1, 384] - - [414, 87.876] - - - [63360, 125953, 1, 384] - - [416, 86.829] - - - [63360, 126337, 1, 384] - - [412, 88.503] - - - [63360, 126721, 1, 384] - - [414, 88.199] - - - [62976, 125185, 1, 384] - - [414, 88.213] - - - [62976, 125569, 1, 384] - - [412, 88.472] - - - [62976, 125953, 1, 384] - - [416, 87.09] - - - [62592, 124417, 1, 384] - - [414, 87.862] - - - [62592, 124801, 1, 384] - - [412, 88.561] - - - [62592, 125185, 1, 384] - - [413, 88.243] - - - [62208, 123649, 1, 384] - - [414, 88.247] - - - [62208, 124033, 1, 384] - - [412, 88.527] - - - [62208, 124417, 1, 384] - - [414, 87.884] - - - [61824, 122881, 1, 384] - - [416, 86.013] - - - [61824, 123265, 1, 384] - - [412, 88.575] - - - [61824, 123649, 1, 384] - - [416, 88.261] - - - [61440, 122113, 1, 384] - - [414, 88.273] - - - [61440, 122497, 1, 384] - - [412, 88.604] - - - [61440, 122881, 1, 384] - - [416, 86.332] - - - [61056, 121345, 1, 384] - - [414, 87.916] - - - [61056, 121729, 1, 384] - - [412, 88.538] - - - [61056, 122113, 1, 384] - - [413, 88.222] - - - [60672, 120577, 1, 384] - - [414, 88.305] - - - [60672, 120961, 1, 384] - - [412, 88.537] - - - [60672, 121345, 1, 384] - - [414, 87.91] - - - [60288, 119809, 1, 384] - - [392, 86.905] - - - [60288, 120193, 1, 384] - - [412, 88.56] - - - [60288, 120577, 1, 384] - - [416, 88.274] - - - [59904, 119041, 1, 384] - - [414, 88.277] - - - [59904, 119425, 1, 384] - - [412, 88.599] - - - [59904, 119809, 1, 384] - - [411, 86.727] - - - [59520, 118273, 1, 384] - - [409, 88.073] - - - [59520, 118657, 1, 384] - - [412, 88.554] - - - [59520, 119041, 1, 384] - - [414, 88.258] - - - [59136, 117505, 1, 384] - - [414, 88.313] - - - [59136, 117889, 1, 384] - - [412, 88.564] - - - [59136, 118273, 1, 384] - - [409, 88.067] - - - [58752, 116737, 1, 384] - - [411, 87.014] - - - [58752, 117121, 1, 384] - - [412, 88.577] - - - [58752, 117505, 1, 384] - - [413, 88.295] - - - [58368, 115969, 1, 384] - - [409, 88.33] - - - [58368, 116353, 1, 384] - - [412, 88.55] - - - [58368, 116737, 1, 384] - - [411, 86.991] - - - [57984, 115201, 1, 384] - - [409, 88.424] - - - [57984, 115585, 1, 384] - - [412, 88.532] - - - [57984, 115969, 1, 384] - - [409, 88.3] - - - [57600, 114433, 1, 384] - - [409, 88.561] - - - [57600, 114817, 1, 384] - - [412, 88.556] - - - [57600, 115201, 1, 384] - - [409, 88.409] - - - [57216, 113665, 1, 384] - - [352, 87.566] - - - [57216, 114049, 1, 384] - - [410, 88.717] - - - [57216, 114433, 1, 384] - - [409, 88.521] - - - [56832, 112897, 1, 384] - - [409, 88.696] - - - [56832, 113281, 1, 384] - - [410, 88.789] - - - [56832, 113665, 1, 384] - - [411, 87.345] - - - [56448, 112129, 1, 384] - - [406, 88.902] - - - [56448, 112513, 1, 384] - - [409, 88.823] - - - [56448, 112897, 1, 384] - - [409, 88.698] - - - [56064, 111361, 1, 384] - - [409, 88.83] - - - [56064, 111745, 1, 384] - - [409, 88.895] - - - [56064, 112129, 1, 384] - - [409, 88.747] - - - [55680, 110593, 1, 384] - - [352, 88.423] - - - [55680, 110977, 1, 384] - - [409, 88.963] - - - [55680, 111361, 1, 384] - - [409, 88.892] - - - [55296, 109825, 1, 384] - - [409, 89.014] - - - [55296, 110209, 1, 384] - - [410, 89.093] - - - [55296, 110593, 1, 384] - - [352, 88.376] - - - [54912, 109057, 1, 384] - - [409, 89.052] - - - [54912, 109441, 1, 384] - - [410, 89.159] - - - [54912, 109825, 1, 384] - - [409, 89.018] - - - [54528, 108289, 1, 384] - - [406, 89.208] - - - [54528, 108673, 1, 384] - - [368, 89.184] - - - [54528, 109057, 1, 384] - - [409, 89.063] - - - [54144, 107521, 1, 384] - - [352, 88.661] - - - [54144, 107905, 1, 384] - - [406, 89.295] - - - [54144, 108289, 1, 384] - - [406, 89.425] - - - [53760, 106753, 1, 384] - - [406, 89.927] - - - [53760, 107137, 1, 384] - - [405, 89.437] - - - [53760, 107521, 1, 384] - - [352, 88.639] - - - [53376, 105985, 1, 384] - - [406, 89.716] - - - [53376, 106369, 1, 384] - - [409, 89.379] - - - [53376, 106753, 1, 384] - - [405, 89.402] - - - [52992, 105217, 1, 384] - - [406, 90.36] - - - [52992, 105601, 1, 384] - - [406, 89.771] - - - [52992, 105985, 1, 384] - - [406, 89.415] - - - [52608, 104449, 1, 384] - - [352, 88.823] - - - [52608, 104833, 1, 384] - - [406, 89.845] - - - [52608, 105217, 1, 384] - - [406, 90.402] - - - [52224, 103681, 1, 384] - - [406, 90.518] - - - [52224, 104065, 1, 384] - - [406, 90.477] - - - [52224, 104449, 1, 384] - - [352, 88.823] - - - [51840, 102913, 1, 384] - - [352, 90.027] - - - [51840, 103297, 1, 384] - - [352, 90.135] - - - [51840, 103681, 1, 384] - - [352, 90.59] - - - [51456, 102145, 1, 384] - - [352, 90.722] - - - [51456, 102529, 1, 384] - - [406, 90.57] - - - [51456, 102913, 1, 384] - - [406, 90.578] - - - [51072, 101377, 1, 384] - - [352, 89.038] - - - [51072, 101761, 1, 384] - - [352, 90.688] - - - [51072, 102145, 1, 384] - - [352, 90.767] - - - [50688, 100609, 1, 384] - - [352, 90.857] - - - [50688, 100993, 1, 384] - - [352, 90.842] - - - [50688, 101377, 1, 384] - - [352, 89.034] - - - [50304, 99841, 1, 384] - - [352, 90.898] - - - [50304, 100225, 1, 384] - - [352, 90.881] - - - [50304, 100609, 1, 384] - - [406, 90.763] - - - [49920, 99073, 1, 384] - - [352, 90.932] - - - [49920, 99457, 1, 384] - - [352, 90.987] - - - [49920, 99841, 1, 384] - - [352, 90.887] - - - [49536, 98305, 1, 384] - - [404, 88.163] - - - [49536, 98689, 1, 384] - - [352, 91.126] - - - [49536, 99073, 1, 384] - - [406, 90.812] - - - [49152, 97537, 1, 384] - - [352, 91.076] - - - [49152, 97921, 1, 384] - - [406, 90.675] - - - [49152, 98305, 1, 384] - - [404, 88.177] - - - [48768, 96769, 1, 384] - - [352, 91.084] - - - [48768, 97153, 1, 384] - - [352, 91.01] - - - [48768, 97537, 1, 384] - - [352, 91.053] - - - [48384, 96001, 1, 384] - - [405, 91.203] - - - [48384, 96385, 1, 384] - - [405, 91.253] - - - [48384, 96769, 1, 384] - - [405, 91.031] - - - [48000, 95233, 1, 384] - - [352, 89.293] - - - [48000, 95617, 1, 384] - - [405, 91.33] - - - [48000, 96001, 1, 384] - - [405, 91.211] - - - [47616, 94465, 1, 384] - - [352, 91.255] - - - [47616, 94849, 1, 384] - - [405, 91.271] - - - [47616, 95233, 1, 384] - - [352, 89.247] - - - [47232, 93697, 1, 384] - - [352, 91.254] - - - [47232, 94081, 1, 384] - - [405, 91.4] - - - [47232, 94465, 1, 384] - - [405, 91.328] - - - [46848, 92929, 1, 384] - - [352, 91.344] - - - [46848, 93313, 1, 384] - - [362, 91.348] - - - [46848, 93697, 1, 384] - - [352, 91.286] - - - [46464, 92161, 1, 384] - - [404, 89.524] - - - [46464, 92545, 1, 384] - - [362, 91.592] - - - [46464, 92929, 1, 384] - - [362, 91.483] - - - [46080, 91393, 1, 384] - - [362, 91.7] - - - [46080, 91777, 1, 384] - - [362, 91.495] - - - [46080, 92161, 1, 384] - - [404, 89.557] - - - [45696, 90625, 1, 384] - - [362, 91.59] - - - [45696, 91009, 1, 384] - - [362, 91.729] - - - [45696, 91393, 1, 384] - - [362, 91.753] - - - [45312, 89857, 1, 384] - - [362, 91.802] - - - [45312, 90241, 1, 384] - - [362, 91.805] - - - [45312, 90625, 1, 384] - - [362, 91.703] - - - [44928, 89089, 1, 384] - - [351, 89.762] - - - [44928, 89473, 1, 384] - - [362, 91.831] - - - [44928, 89857, 1, 384] - - [362, 91.841] - - - [44544, 88321, 1, 384] - - [362, 91.997] - - - [44544, 88705, 1, 384] - - [362, 91.915] - - - [44544, 89089, 1, 384] - - [351, 89.802] - - - [44160, 87553, 1, 384] - - [362, 92.037] - - - [44160, 87937, 1, 384] - - [362, 91.991] - - - [44160, 88321, 1, 384] - - [362, 91.997] - - - [43776, 86785, 1, 384] - - [362, 92.115] - - - [43776, 87169, 1, 384] - - [362, 92.101] - - - [43776, 87553, 1, 384] - - [362, 92.083] - - - [43392, 86017, 1, 384] - - [351, 90.273] - - - [43392, 86401, 1, 384] - - [362, 92.074] - - - [43392, 86785, 1, 384] - - [362, 92.127] - - - [43008, 85249, 1, 384] - - [362, 92.17] - - - [43008, 85633, 1, 384] - - [362, 92.15] - - - [43008, 86017, 1, 384] - - [351, 90.173] - - - [42624, 84481, 1, 384] - - [362, 92.233] - - - [42624, 84865, 1, 384] - - [362, 92.193] - - - [42624, 85249, 1, 384] - - [362, 92.168] - - - [42240, 83713, 1, 384] - - [362, 92.167] - - - [42240, 84097, 1, 384] - - [362, 92.151] - - - [42240, 84481, 1, 384] - - [362, 92.173] - - - [41856, 82945, 1, 384] - - [351, 90.565] - - - [41856, 83329, 1, 384] - - [362, 92.184] - - - [41856, 83713, 1, 384] - - [362, 92.176] - - - [41472, 82177, 1, 384] - - [362, 92.283] - - - [41472, 82561, 1, 384] - - [362, 92.305] - - - [41472, 82945, 1, 384] - - [351, 90.349] - - - [41088, 81409, 1, 384] - - [362, 92.421] - - - [41088, 81793, 1, 384] - - [362, 92.373] - - - [41088, 82177, 1, 384] - - [362, 92.307] - - - [40704, 80641, 1, 384] - - [362, 92.286] - - - [40704, 81025, 1, 384] - - [362, 92.24] - - - [40704, 81409, 1, 384] - - [362, 92.261] - - - [40320, 79873, 1, 384] - - [351, 90.704] - - - [40320, 80257, 1, 384] - - [362, 92.267] - - - [40320, 80641, 1, 384] - - [362, 92.264] - - - [39936, 79105, 1, 384] - - [362, 92.328] - - - [39936, 79489, 1, 384] - - [362, 92.253] - - - [39936, 79873, 1, 384] - - [351, 90.739] - - - [39552, 78337, 1, 384] - - [362, 92.33] - - - [39552, 78721, 1, 384] - - [362, 92.237] - - - [39552, 79105, 1, 384] - - [362, 92.263] - - - [39168, 77569, 1, 384] - - [362, 92.269] - - - [39168, 77953, 1, 384] - - [362, 92.191] - - - [39168, 78337, 1, 384] - - [362, 92.247] - - - [38784, 76801, 1, 384] - - [351, 90.727] - - - [38784, 77185, 1, 384] - - [362, 92.376] - - - [38784, 77569, 1, 384] - - [362, 92.402] - - - [38400, 76033, 1, 384] - - [362, 92.349] - - - [38400, 76417, 1, 384] - - [362, 92.308] - - - [38400, 76801, 1, 384] - - [351, 90.722] - - - [38016, 75265, 1, 384] - - [351, 92.346] - - - [38016, 75649, 1, 384] - - [351, 92.27] - - - [38016, 76033, 1, 384] - - [362, 92.286] - - - [37632, 74497, 1, 384] - - [351, 92.343] - - - [37632, 74881, 1, 384] - - [351, 92.312] - - - [37632, 75265, 1, 384] - - [351, 92.305] - - - [37248, 73729, 1, 384] - - [351, 90.613] - - - [37248, 74113, 1, 384] - - [351, 92.356] - - - [37248, 74497, 1, 384] - - [351, 92.352] - - - [36864, 72961, 1, 384] - - [351, 92.345] - - - [36864, 73345, 1, 384] - - [351, 92.302] - - - [36864, 73729, 1, 384] - - [351, 90.66] - - - [36480, 72193, 1, 384] - - [351, 92.295] - - - [36480, 72577, 1, 384] - - [351, 92.269] - - - [36480, 72961, 1, 384] - - [351, 92.266] - - - [36096, 71425, 1, 384] - - [351, 92.449] - - - [36096, 71809, 1, 384] - - [351, 92.443] - - - [36096, 72193, 1, 384] - - [351, 92.442] - - - [35712, 70657, 1, 384] - - [351, 90.934] - - - [35712, 71041, 1, 384] - - [351, 92.474] - - - [35712, 71425, 1, 384] - - [351, 92.48] - - - [35328, 69889, 1, 384] - - [351, 92.472] - - - [35328, 70273, 1, 384] - - [351, 92.462] - - - [35328, 70657, 1, 384] - - [351, 90.908] - - - [34944, 69121, 1, 384] - - [351, 92.457] - - - [34944, 69505, 1, 384] - - [351, 92.411] - - - [34944, 69889, 1, 384] - - [351, 92.417] - - - [34560, 68353, 1, 384] - - [351, 92.44] - - - [34560, 68737, 1, 384] - - [351, 92.403] - - - [34560, 69121, 1, 384] - - [351, 92.451] - - - [34176, 67585, 1, 384] - - [351, 90.857] - - - [34176, 67969, 1, 384] - - [351, 92.379] - - - [34176, 68353, 1, 384] - - [351, 92.445] - - - [33792, 66817, 1, 384] - - [350, 92.368] - - - [33792, 67201, 1, 384] - - [350, 92.353] - - - [33792, 67585, 1, 384] - - [351, 90.792] - - - [33408, 66049, 1, 384] - - [351, 92.542] - - - [33408, 66433, 1, 384] - - [351, 92.454] - - - [33408, 66817, 1, 384] - - [351, 92.52] - - - [33024, 65281, 1, 384] - - [351, 92.601] - - - [33024, 65665, 1, 384] - - [351, 92.527] - - - [33024, 66049, 1, 384] - - [351, 92.579] - - - [32640, 64513, 1, 384] - - [351, 91.045] - - - [32640, 64897, 1, 384] - - [351, 92.498] - - - [32640, 65281, 1, 384] - - [351, 92.558] - - - [32256, 63745, 1, 384] - - [351, 92.485] - - - [32256, 64129, 1, 384] - - [351, 92.465] - - - [32256, 64513, 1, 384] - - [351, 90.976] - - - [31872, 62977, 1, 384] - - [351, 92.407] - - - [31872, 63361, 1, 384] - - [351, 92.411] - - - [31872, 63745, 1, 384] - - [351, 92.405] - - - [31488, 62209, 1, 384] - - [351, 92.401] - - - [31488, 62593, 1, 384] - - [351, 92.34] - - - [31488, 62977, 1, 384] - - [351, 92.389] - - - [31104, 61441, 1, 384] - - [351, 90.783] - - - [31104, 61825, 1, 384] - - [350, 92.353] - - - [31104, 62209, 1, 384] - - [350, 92.344] - - - [30720, 60673, 1, 384] - - [350, 92.405] - - - [30720, 61057, 1, 384] - - [350, 92.375] - - - [30720, 61441, 1, 384] - - [351, 90.62] - - - [30336, 59905, 1, 384] - - [351, 92.573] - - - [30336, 60289, 1, 384] - - [351, 92.498] - - - [30336, 60673, 1, 384] - - [351, 92.502] - - - [29952, 59137, 1, 384] - - [351, 92.508] - - - [29952, 59521, 1, 384] - - [351, 92.485] - - - [29952, 59905, 1, 384] - - [351, 92.511] - - - [29568, 58369, 1, 384] - - [351, 90.828] - - - [29568, 58753, 1, 384] - - [351, 92.445] - - - [29568, 59137, 1, 384] - - [351, 92.41] - - - [29184, 57601, 1, 384] - - [351, 92.369] - - - [29184, 57985, 1, 384] - - [351, 92.371] - - - [29184, 58369, 1, 384] - - [351, 90.768] - - - [28800, 56833, 1, 384] - - [351, 92.255] - - - [28800, 57217, 1, 384] - - [351, 92.185] - - - [28800, 57601, 1, 384] - - [351, 92.21] - - - [28416, 56065, 1, 384] - - [392, 92.184] - - - [28416, 56449, 1, 384] - - [368, 92.169] - - - [28416, 56833, 1, 384] - - [351, 92.168] - - - [28032, 55297, 1, 384] - - [351, 90.529] - - - [28032, 55681, 1, 384] - - [368, 92.235] - - - [28032, 56065, 1, 384] - - [368, 92.161] - - - [27648, 54529, 1, 384] - - [351, 92.451] - - - [27648, 54913, 1, 384] - - [351, 92.45] - - - [27648, 55297, 1, 384] - - [351, 90.69] - - - [27264, 53761, 1, 384] - - [351, 92.424] - - - [27264, 54145, 1, 384] - - [351, 92.392] - - - [27264, 54529, 1, 384] - - [351, 92.434] - - - [26880, 52993, 1, 384] - - [351, 92.398] - - - [26880, 53377, 1, 384] - - [351, 92.365] - - - [26880, 53761, 1, 384] - - [351, 92.41] - - - [26496, 52225, 1, 384] - - [351, 90.707] - - - [26496, 52609, 1, 384] - - [351, 92.304] - - - [26496, 52993, 1, 384] - - [351, 92.352] - - - [26112, 51457, 1, 384] - - [351, 92.162] - - - [26112, 51841, 1, 384] - - [364, 92.33] - - - [26112, 52225, 1, 384] - - [351, 90.68] - - - [25728, 50689, 1, 384] - - [351, 92.071] - - - [25728, 51073, 1, 384] - - [364, 92.193] - - - [25728, 51457, 1, 384] - - [350, 92.103] - - - [25344, 49921, 1, 384] - - [364, 92.271] - - - [25344, 50305, 1, 384] - - [364, 92.341] - - - [25344, 50689, 1, 384] - - [350, 91.781] - - - [24960, 49153, 1, 384] - - [351, 90.127] - - - [24960, 49537, 1, 384] - - [351, 92.344] - - - [24960, 49921, 1, 384] - - [351, 92.382] - - - [24576, 48385, 1, 384] - - [351, 92.354] - - - [24576, 48769, 1, 384] - - [351, 92.298] - - - [24576, 49153, 1, 384] - - [351, 90.599] - - - [24192, 47617, 1, 384] - - [351, 92.288] - - - [24192, 48001, 1, 384] - - [351, 92.26] - - - [24192, 48385, 1, 384] - - [351, 92.308] - - - [23808, 46849, 1, 384] - - [350, 92.225] - - - [23808, 47233, 1, 384] - - [364, 92.235] - - - [23808, 47617, 1, 384] - - [350, 92.197] - - - [23424, 46081, 1, 384] - - [351, 90.613] - - - [23424, 46465, 1, 384] - - [364, 92.124] - - - [23424, 46849, 1, 384] - - [351, 92.114] - - - [23040, 45313, 1, 384] - - [364, 92.046] - - - [23040, 45697, 1, 384] - - [364, 92.096] - - - [23040, 46081, 1, 384] - - [351, 90.361] - - - [22656, 44545, 1, 384] - - [350, 91.887] - - - [22656, 44929, 1, 384] - - [364, 92.062] - - - [22656, 45313, 1, 384] - - [364, 92.088] - - - [22272, 43777, 1, 384] - - [351, 92.231] - - - [22272, 44161, 1, 384] - - [351, 92.158] - - - [22272, 44545, 1, 384] - - [351, 92.183] - - - [21888, 43009, 1, 384] - - [351, 90.509] - - - [21888, 43393, 1, 384] - - [351, 92.143] - - - [21888, 43777, 1, 384] - - [351, 92.204] - - - [21504, 42241, 1, 384] - - [351, 92.069] - - - [21504, 42625, 1, 384] - - [351, 92.031] - - - [21504, 43009, 1, 384] - - [351, 90.457] - - - [21120, 41473, 1, 384] - - [357, 91.82] - - - [21120, 41857, 1, 384] - - [351, 91.749] - - - [21120, 42241, 1, 384] - - [364, 91.798] - - - [20736, 40705, 1, 384] - - [350, 92.079] - - - [20736, 41089, 1, 384] - - [350, 92.017] - - - [20736, 41473, 1, 384] - - [350, 92.031] - - - [20352, 39937, 1, 384] - - [351, 90.3] - - - [20352, 40321, 1, 384] - - [350, 91.997] - - - [20352, 40705, 1, 384] - - [350, 91.947] - - - [19968, 39169, 1, 384] - - [364, 91.859] - - - [19968, 39553, 1, 384] - - [364, 91.829] - - - [19968, 39937, 1, 384] - - [351, 89.971] - - - [19584, 38401, 1, 384] - - [351, 91.756] - - - [19584, 38785, 1, 384] - - [351, 91.79] - - - [19584, 39169, 1, 384] - - [351, 91.893] - - - [19200, 37633, 1, 384] - - [351, 92.021] - - - [19200, 38017, 1, 384] - - [351, 91.964] - - - [19200, 38401, 1, 384] - - [351, 91.996] - - - [18816, 36865, 1, 384] - - [351, 90.208] - - - [18816, 37249, 1, 384] - - [351, 91.826] - - - [18816, 37633, 1, 384] - - [351, 91.92] - - - [18432, 36097, 1, 384] - - [351, 91.745] - - - [18432, 36481, 1, 384] - - [351, 91.7] - - - [18432, 36865, 1, 384] - - [351, 90.171] - - - [18048, 35329, 1, 384] - - [351, 91.71] - - - [18048, 35713, 1, 384] - - [351, 91.78] - - - [18048, 36097, 1, 384] - - [351, 91.763] - - - [17664, 34561, 1, 384] - - [351, 91.516] - - - [17664, 34945, 1, 384] - - [351, 91.496] - - - [17664, 35329, 1, 384] - - [351, 91.537] - - - [17280, 33793, 1, 384] - - [351, 89.78] - - - [17280, 34177, 1, 384] - - [350, 91.699] - - - [17280, 34561, 1, 384] - - [350, 91.589] - - - [16896, 33025, 1, 384] - - [350, 91.515] - - - [16896, 33409, 1, 384] - - [350, 91.488] - - - [16896, 33793, 1, 384] - - [352, 88.965] - - - [16512, 32257, 1, 384] - - [351, 91.746] - - - [16512, 32641, 1, 384] - - [351, 91.652] - - - [16512, 33025, 1, 384] - - [351, 91.637] - - - [16128, 31489, 1, 384] - - [351, 91.455] - - - [16128, 31873, 1, 384] - - [351, 91.455] - - - [16128, 32257, 1, 384] - - [351, 91.474] - - - [15744, 30721, 1, 384] - - [351, 89.905] - - - [15744, 31105, 1, 384] - - [351, 91.293] - - - [15744, 31489, 1, 384] - - [351, 91.334] - - - [15360, 29953, 1, 384] - - [357, 91.275] - - - [15360, 30337, 1, 384] - - [357, 91.209] - - - [15360, 30721, 1, 384] - - [351, 89.767] - - - [14976, 29185, 1, 384] - - [351, 91.141] - - - [14976, 29569, 1, 384] - - [364, 91.07] - - - [14976, 29953, 1, 384] - - [351, 91.176] - - - [14592, 28417, 1, 384] - - [351, 90.799] - - - [14592, 28801, 1, 384] - - [351, 90.745] - - - [14592, 29185, 1, 384] - - [364, 90.825] - - - [14208, 27649, 1, 384] - - [352, 88.273] - - - [14208, 28033, 1, 384] - - [364, 90.529] - - - [14208, 28417, 1, 384] - - [364, 90.527] - - - [13824, 26881, 1, 384] - - [351, 91.192] - - - [13824, 27265, 1, 384] - - [351, 91.19] - - - [13824, 27649, 1, 384] - - [357, 89.333] - - - [13440, 26113, 1, 384] - - [350, 91.029] - - - [13440, 26497, 1, 384] - - [350, 90.923] - - - [13440, 26881, 1, 384] - - [350, 90.944] - - - [13056, 25345, 1, 384] - - [351, 90.78] - - - [13056, 25729, 1, 384] - - [351, 90.728] - - - [13056, 26113, 1, 384] - - [351, 90.836] - - - [12672, 24577, 1, 384] - - [351, 88.815] - - - [12672, 24961, 1, 384] - - [350, 90.484] - - - [12672, 25345, 1, 384] - - [350, 90.465] - - - [12288, 23809, 1, 384] - - [351, 90.521] - - - [12288, 24193, 1, 384] - - [351, 90.523] - - - [12288, 24577, 1, 384] - - [351, 89.249] - - - [11904, 23041, 1, 384] - - [350, 90.548] - - - [11904, 23425, 1, 384] - - [351, 90.459] - - - [11904, 23809, 1, 384] - - [350, 90.505] - - - [11520, 22273, 1, 384] - - [350, 90.629] - - - [11520, 22657, 1, 384] - - [350, 90.558] - - - [11520, 23041, 1, 384] - - [350, 90.457] - - - [11136, 21505, 1, 384] - - [357, 88.68] - - - [11136, 21889, 1, 384] - - [351, 90.601] - - - [11136, 22273, 1, 384] - - [351, 90.885] - - - [10752, 20737, 1, 384] - - [351, 90.592] - - - [10752, 21121, 1, 384] - - [351, 90.587] - - - [10752, 21505, 1, 384] - - [351, 89.124] - - - [10368, 19969, 1, 384] - - [364, 90.75] - - - [10368, 20353, 1, 384] - - [350, 90.789] - - - [10368, 20737, 1, 384] - - [351, 90.426] - - - [9984, 19201, 1, 384] - - [364, 90.553] - - - [9984, 19585, 1, 384] - - [351, 90.613] - - - [9984, 19969, 1, 384] - - [350, 90.613] - - - [9600, 18433, 1, 384] - - [351, 89.325] - - - [9600, 18817, 1, 384] - - [364, 90.628] - - - [9600, 19201, 1, 384] - - [364, 90.656] - - - [9216, 17665, 1, 384] - - [351, 90.672] - - - [9216, 18049, 1, 384] - - [351, 90.502] - - - [9216, 18433, 1, 384] - - [351, 89.1] - - - [8832, 16897, 1, 384] - - [351, 90.805] - - - [8832, 17281, 1, 384] - - [351, 90.554] - - - [8832, 17665, 1, 384] - - [351, 90.539] - - - [8448, 16129, 1, 384] - - [364, 90.54] - - - [8448, 16513, 1, 384] - - [364, 90.768] - - - [8448, 16897, 1, 384] - - [368, 90.557] - - - [8064, 15361, 1, 384] - - [351, 89.079] - - - [8064, 15745, 1, 384] - - [368, 90.445] - - - [8064, 16129, 1, 384] - - [364, 90.472] - - - [7680, 14593, 1, 384] - - [351, 90.144] - - - [7680, 14977, 1, 384] - - [364, 90.612] - - - [7680, 15361, 1, 384] - - [351, 88.932] - - - [7296, 13825, 1, 384] - - [351, 89.674] - - - [7296, 14209, 1, 384] - - [351, 89.682] - - - [7296, 14593, 1, 384] - - [351, 89.842] - - - [6912, 13057, 1, 384] - - [364, 89.608] - - - [6912, 13441, 1, 384] - - [351, 89.826] - - - [6912, 13825, 1, 384] - - [351, 89.872] - - - [6528, 12289, 1, 384] - - [351, 87.482] - - - [6528, 12673, 1, 384] - - [351, 89.165] - - - [6528, 13057, 1, 384] - - [351, 89.319] - - - [6144, 11521, 1, 384] - - [351, 88.629] - - - [6144, 11905, 1, 384] - - [362, 88.133] - - - [6144, 12289, 1, 384] - - [351, 87.106] - - - [5760, 10753, 1, 384] - - [350, 88.042] - - - [5760, 11137, 1, 384] - - [351, 88.117] - - - [5760, 11521, 1, 384] - - [351, 88.432] - - - [5376, 9985, 1, 384] - - [352, 86.702] - - - [5376, 10369, 1, 384] - - [357, 87.062] - - - [5376, 10753, 1, 384] - - [359, 87.59] - - - [4992, 9217, 1, 384] - - [351, 85.149] - - - [4992, 9601, 1, 384] - - [351, 87.416] - - - [4992, 9985, 1, 384] - - [355, 86.83] - - - [4608, 8449, 1, 384] - - [351, 85.804] - - - [4608, 8833, 1, 384] - - [356, 86.023] - - - [4608, 9217, 1, 384] - - [357, 84.861] - - - [4224, 7681, 1, 384] - - [351, 84.569] - - - [4224, 8065, 1, 384] - - [351, 83.93] - - - [4224, 8449, 1, 384] - - [355, 83.752] - - - [3840, 6913, 1, 384] - - [351, 83.92] - - - [3840, 7297, 1, 384] - - [351, 83.602] - - - [3840, 7681, 1, 384] - - [351, 84.06] - - - [3456, 6145, 1, 384] - - [352, 77.566] - - - [3456, 6529, 1, 384] - - [351, 81.809] - - - [3456, 6913, 1, 384] - - [350, 81.482] - - - [3072, 5377, 1, 384] - - [350, 76.944] - - - [3072, 5761, 1, 384] - - [351, 77.299] - - - [3072, 6145, 1, 384] - - [351, 79.328] - - - [2688, 4609, 1, 384] - - [347, 74.133] - - - [2688, 4993, 1, 384] - - [348, 77.161] - - - [2688, 5377, 1, 384] - - [349, 77.197] - - - [2304, 3841, 1, 384] - - [345, 73.591] - - - [2304, 4225, 1, 384] - - [341, 73.981] - - - [2304, 4609, 1, 384] - - [345, 74.009] - - - [1920, 3073, 1, 384] - - [343, 66.43] - - - [1920, 3457, 1, 384] - - [344, 70.294] - - - [1920, 3841, 1, 384] - - [345, 70.159] - - - [1536, 2305, 1, 384] - - [340, 61.452] - - - [1536, 2689, 1, 384] - - [341, 65.936] - - - [1536, 3073, 1, 384] - - [342, 64.165] - - - [1152, 1537, 1, 384] - - [335, 48.178] - - - [1152, 1921, 1, 384] - - [337, 54.787] - - - [1152, 2305, 1, 384] - - [338, 55.296] - - - [768, 1153, 1, 384] - - [336, 34.205] - - - [768, 1537, 1, 384] - - [335, 44.797] - - - [384, 769, 1, 384] - - [333, 20.523] - - - [512, 1025, 1, 512] - - [420, 32.377] - - - [1024, 2049, 1, 512] - - [360, 55.968] - - - [1536, 3073, 1, 512] - - [426, 69.017] - - - [2048, 4097, 1, 512] - - [430, 77.471] - - - [2560, 5121, 1, 512] - - [402, 79.143] - - - [3072, 6145, 1, 512] - - [395, 84.135] - - - [3584, 7169, 1, 512] - - [526, 86.716] - - - [1024, 1024, 8, 1024] - - [224, 37.66] - - - [2048, 2048, 4, 2048] - - [225, 55.362] - - - [4096, 4096, 2, 4096] - - [226, 96.942] - - - [8192, 8192, 1, 8192] - - [226, 97.575] - - - [16384, 16384, 1, 16384] - - [226, 99.795] - - - [768, 768, 1, 768] - - [230, 43.067] - - - [1152, 1152, 1, 1152] - - [228, 70.518] - - - [1536, 1536, 1, 1536] - - [231, 76.206] - - - [1920, 1920, 1, 1920] - - [232, 82.25] - - - [2304, 2304, 1, 2304] - - [233, 92.123] - - - [2688, 2688, 1, 2688] - - [234, 87.132] - - - [3072, 3072, 1, 3072] - - [235, 95.558] - - - [3456, 3456, 1, 3456] - - [236, 95.674] - - - [3840, 3840, 1, 3840] - - [237, 97.134] - - - [4224, 4224, 1, 4224] - - [238, 97.539] - - - [4992, 4992, 1, 4992] - - [233, 97.764] - - - [5376, 5376, 1, 5376] - - [232, 96.198] - - - [6144, 6144, 1, 6144] - - [233, 99.046] - - - [6528, 6528, 1, 6528] - - [240, 97.946] - - - [6912, 6912, 1, 6912] - - [238, 97.667] - - - [7296, 7296, 1, 7296] - - [240, 97.983] - - - [7680, 7680, 1, 7680] - - [238, 98.739] - - - [1024, 1024, 1, 2048] - - [241, 64.821] - - - [1024, 1024, 1, 3072] - - [242, 68.93] - - - [1024, 2048, 1, 11264] - - [243, 76.197] - - - [1024, 2048, 1, 15360] - - [244, 76.452] - - - [1024, 2048, 1, 3072] - - [243, 74.102] - - - [1024, 2048, 1, 7168] - - [243, 75.65] - - - [1024, 4096, 1, 13312] - - [245, 91.746] - - - [1024, 4096, 1, 5120] - - [246, 90.65] - - - [1024, 8192, 1, 9216] - - [247, 92.392] - - - [2048, 2048, 1, 4096] - - [248, 83.954] - - - [2048, 2048, 1, 5120] - - [249, 86.155] - - - [2048, 2048, 1, 6144] - - [250, 84.986] - - - [2048, 2048, 1, 7168] - - [251, 83.173] - - - [2048, 4096, 1, 14336] - - [252, 92.588] - - - [2048, 4096, 1, 6144] - - [253, 92.096] - - - [2048, 8192, 1, 10240] - - [241, 94.987] - - - [256, 256, 1, 512] - - [254, 6.691] - - - [3072, 4096, 1, 15360] - - [255, 99.323] - - - [3072, 4096, 1, 7168] - - [255, 98.993] - - - [3072, 8192, 1, 11264] - - [253, 99.396] - - - [4096, 4096, 1, 10240] - - [256, 94.88] - - - [4096, 4096, 1, 11264] - - [257, 94.974] - - - [4096, 4096, 1, 12288] - - [258, 94.911] - - - [4096, 4096, 1, 13312] - - [257, 95.034] - - - [4096, 4096, 1, 14336] - - [258, 95.025] - - - [4096, 4096, 1, 15360] - - [257, 95.07] - - - [4096, 4096, 1, 8192] - - [259, 94.473] - - - [4096, 4096, 1, 9216] - - [257, 94.893] - - - [4096, 8192, 1, 12288] - - [260, 97.713] - - - [512, 512, 1, 1024] - - [261, 31.259] - - - [5120, 8192, 1, 13312] - - [253, 96.746] - - - [6144, 8192, 1, 14336] - - [253, 99.532] - - - [7168, 8192, 1, 15360] - - [253, 98.548] - - - [8192, 8192, 1, 16384] - - [262, 97.775] - - - [1024, 1024, 2, 4096] - - [241, 74.316] - - - [1024, 1024, 2, 5120] - - [243, 75.126] - - - [128, 128, 2, 512] - - [254, 3.259] - - - [2048, 2048, 2, 10240] - - [255, 92.451] - - - [2048, 2048, 2, 11264] - - [263, 92.491] - - - [2048, 2048, 2, 8192] - - [264, 91.95] - - - [2048, 2048, 2, 9216] - - [255, 92.394] - - - [256, 256, 2, 1024] - - [254, 15.589] - - - [4096, 4096, 2, 16384] - - [262, 97.694] - - - [512, 512, 2, 2048] - - [265, 45.997] - - - [1024, 1024, 3, 6144] - - [247, 85.459] - - - [1024, 1024, 3, 7168] - - [247, 85.68] - - - [2048, 2048, 3, 12288] - - [263, 99.208] - - - [2048, 2048, 3, 13312] - - [253, 99.259] - - - [2048, 2048, 3, 14336] - - [266, 99.225] - - - [2048, 2048, 3, 15360] - - [267, 99.235] - - - [512, 512, 3, 3072] - - [268, 55.21] - - - [1024, 1024, 4, 8192] - - [269, 80.106] - - - [1024, 1024, 4, 9216] - - [270, 91.434] - - - [128, 128, 4, 1024] - - [271, 7.695] - - - [2048, 2048, 4, 16384] - - [272, 44.087] - - - [256, 256, 4, 2048] - - [273, 25.766] - - - [512, 512, 4, 4096] - - [274, 67.207] - - - [64, 64, 4, 512] - - [275, 1.575] - - - [1024, 1024, 5, 10240] - - [253, 95.725] - - - [1024, 1024, 5, 11264] - - [276, 95.541] - - - [512, 512, 5, 5120] - - [277, 68.408] - - - [1024, 1024, 6, 12288] - - [278, 90.231] - - - [1024, 1024, 6, 13312] - - [279, 90.527] - - - [256, 256, 6, 3072] - - [280, 31.43] - - - [512, 512, 6, 6144] - - [260, 80.948] - - - [1024, 1024, 7, 14336] - - [259, 88.813] - - - [1024, 1024, 7, 15360] - - [281, 88.875] - - - [512, 512, 7, 7168] - - [282, 52.676] - - - [1024, 1024, 8, 16384] - - [283, 53.337] - - - [128, 128, 8, 2048] - - [284, 8.022] - - - [256, 256, 8, 4096] - - [285, 27.223] - - - [32, 32, 8, 512] - - [261, 0.67] - - - [512, 512, 8, 8192] - - [286, 17.468] - - - [64, 64, 8, 1024] - - [261, 2.24] - - - [512, 512, 9, 9216] - - [287, 61.208] - - - [256, 256, 10, 5120] - - [288, 38.321] - - - [512, 512, 10, 10240] - - [289, 67.597] - - - [512, 512, 11, 11264] - - [290, 71.195] - - - [128, 128, 12, 3072] - - [291, 9.673] - - - [256, 256, 12, 6144] - - [292, 36.782] - - - [512, 512, 12, 12288] - - [293, 54.771] - - - [512, 512, 13, 13312] - - [294, 66.377] - - - [256, 256, 14, 7168] - - [293, 36.06] - - - [512, 512, 14, 14336] - - [295, 47.973] - - - [512, 512, 15, 15360] - - [296, 52.736] - - - [128, 128, 16, 4096] - - [297, 6.562] - - - [256, 256, 16, 8192] - - [298, 9.909] - - - [32, 32, 16, 1024] - - [299, 0.584] - - - [512, 512, 16, 16384] - - [300, 91.972] - - - [64, 64, 16, 2048] - - [301, 2.019] - - - [256, 256, 18, 9216] - - [302, 20.658] - - - [128, 128, 20, 5120] - - [303, 9.686] - - - [256, 256, 20, 10240] - - [304, 18.355] - - - [256, 256, 22, 11264] - - [255, 19.873] - - - [128, 128, 24, 6144] - - [305, 9.309] - - - [256, 256, 24, 12288] - - [306, 13.467] - - - [64, 64, 24, 3072] - - [307, 2.474] - - - [256, 256, 26, 13312] - - [308, 92.025] - - - [128, 128, 28, 7168] - - [309, 9.665] - - - [256, 256, 28, 14336] - - [279, 66.832] - - - [256, 256, 30, 15360] - - [310, 71.605] - - - [128, 128, 32, 8192] - - [311, 4.094] - - - [256, 256, 32, 16384] - - [274, 76.406] - - - [32, 32, 32, 2048] - - [312, 0.535] - - - [64, 64, 32, 4096] - - [313, 1.682] - - - [128, 128, 36, 9216] - - [247, 9.831] - - - [128, 128, 40, 10240] - - [314, 9.459] - - - [64, 64, 40, 5120] - - [315, 2.522] - - - [128, 128, 44, 11264] - - [270, 73.615] - - - [128, 128, 48, 12288] - - [243, 77.025] - - - [32, 32, 48, 3072] - - [316, 0.625] - - - [64, 64, 48, 6144] - - [317, 2.39] - - - [128, 128, 52, 13312] - - [243, 77.699] - - - [128, 128, 56, 14336] - - [318, 49.512] - - - [64, 64, 56, 7168] - - [280, 2.526] - - - [128, 128, 60, 15360] - - [319, 53.128] - - - [128, 128, 64, 16384] - - [320, 56.728] - - - [32, 32, 64, 4096] - - [321, 0.42] - - - [64, 64, 64, 8192] - - [322, 36.465] - - - [64, 64, 72, 9216] - - [261, 37.434] - - - [32, 32, 80, 5120] - - [323, 0.628] - - - [64, 64, 80, 10240] - - [261, 37.946] - - - [64, 64, 88, 11264] - - [261, 38.19] - - - [32, 32, 96, 6144] - - [324, 0.592] - - - [64, 64, 96, 12288] - - [261, 38.343] - - - [64, 64, 104, 13312] - - [261, 38.234] - - - [32, 32, 112, 7168] - - [325, 9.154] - - - [64, 64, 112, 14336] - - [261, 26.486] - - - [64, 64, 120, 15360] - - [326, 28.89] - - - [32, 32, 128, 8192] - - [325, 10.568] - - - [64, 64, 128, 16384] - - [327, 30.662] - - - [32, 32, 144, 9216] - - [328, 11.343] - - - [32, 32, 160, 10240] - - [325, 12.66] - - - [32, 32, 176, 11264] - - [329, 13.317] - - - [32, 32, 192, 12288] - - [254, 14.273] - - - [32, 32, 208, 13312] - - [261, 15.262] - - - [32, 32, 224, 14336] - - [261, 12.166] - - - [32, 32, 240, 15360] - - [326, 13.234] - - - [32, 32, 256, 16384] - - [261, 14.13] - - - [512, 512, 11, 512] - - [330, 53.606] - - - [512, 512, 21, 512] - - [330, 58.098] - - - [512, 512, 31, 512] - - [330, 70.428] - - - [512, 512, 41, 512] - - [330, 78.047] - - - [512, 512, 51, 512] - - [330, 76.725] - - - [512, 512, 61, 512] - - [330, 81.402] - - - [512, 512, 71, 512] - - [330, 79.37] - - - [512, 512, 81, 512] - - [330, 83.325] - - - [512, 512, 91, 512] - - [330, 81.232] - - - [3840, 4223, 1, 4096] - - [503, 97.857] - - - [3840, 4225, 1, 4096] - - [504, 92.088] - - - [3840, 4223, 1, 4224] - - [503, 97.953] - - - [3840, 4225, 1, 4224] - - [504, 92.221] - - - [3840, 4223, 1, 4320] - - [505, 97.97] - - - [3840, 4225, 1, 4320] - - [504, 92.232] - - - [7680, 8447, 1, 8192] - - [506, 99.471] - - - [7680, 8449, 1, 8192] - - [503, 97.449] - - - [7680, 8447, 1, 8448] - - [506, 99.483] - - - [7680, 8449, 1, 8448] - - [503, 97.466] - - - [7680, 8447, 1, 8640] - - [506, 99.497] - - - [7680, 8449, 1, 8640] - - [503, 97.475] - - - [3840, 4224, 1, 4095] - - [507, 97.732] - - - [3840, 4224, 1, 4097] - - [508, 97.732] - - - [3840, 4224, 1, 4223] - - [506, 97.78] - - - [3840, 4224, 1, 4225] - - [509, 97.792] - - - [3840, 4224, 1, 4319] - - [506, 97.852] - - - [3840, 4224, 1, 4321] - - [510, 97.9] - - - [7680, 8448, 1, 8191] - - [508, 99.431] - - - [7680, 8448, 1, 8193] - - [508, 99.413] - - - [7680, 8448, 1, 8447] - - [507, 99.439] - - - [7680, 8448, 1, 8449] - - [507, 99.447] - - - [7680, 8448, 1, 8639] - - [506, 99.455] - - - [7680, 8448, 1, 8641] - - [507, 99.455] - - - [3839, 4224, 1, 4096] - - [511, 97.689] - - - [3841, 4224, 1, 4096] - - [512, 91.004] - - - [3839, 4224, 1, 4224] - - [511, 97.728] - - - [3841, 4224, 1, 4224] - - [513, 91.11] - - - [3839, 4224, 1, 4320] - - [511, 97.806] - - - [3841, 4224, 1, 4320] - - [512, 91.18] - - - [7679, 8448, 1, 8192] - - [514, 99.402] - - - [7681, 8448, 1, 8192] - - [511, 97.351] - - - [7679, 8448, 1, 8448] - - [514, 99.422] - - - [7681, 8448, 1, 8448] - - [511, 97.373] - - - [7679, 8448, 1, 8640] - - [514, 99.436] - - - [7681, 8448, 1, 8640] - - [511, 97.395] - - - [100, 50, 1, 11776] - - [515, 6.072] - - - [100, 50, 1, 5888] - - [516, 3.716] - - - [50, 25, 1, 11776] - - [517, 1.688] - - - [50, 25, 1, 5888] - - [518, 0.961] - - - [5939, 5939, 1, 1009] - - [519, 83.29] - - - [10789, 10789, 1, 2211] - - [520, 91.365] - - - [15957, 15957, 1, 1382] - - [520, 91.576] - - - [20613, 20613, 1, 2189] - - [521, 92.613] - - - [25429, 25429, 1, 2404] - - [522, 93.043] - - - [31985, 31985, 1, 1573] - - [522, 92.685] - - - [37053, 37053, 1, 3873] - - [520, 93.873] - - - [43909, 43909, 1, 1995] - - [523, 92.982] - - - [56549, 56549, 1, 2278] - - [520, 93.264] - - - [62002, 62002, 1, 2408] - - [524, 93.472] - - - [127488, 38400, 1, 512] - - [525, 94.593] - - - [128000, 38400, 1, 512] - - [526, 94.637] - - - [127488, 25088, 1, 512] - - [526, 94.585] - - - [128000, 25088, 1, 512] - - [526, 94.625] - - - [127488, 25600, 1, 512] - - [526, 94.615] - - - [128000, 25600, 1, 512] - - [526, 94.617] - - - [127488, 25089, 1, 512] - - [526, 94.113] - - - [128000, 25089, 1, 512] - - [526, 94.14] - - - [126976, 38400, 1, 512] - - [526, 94.635] - - - [126976, 25088, 1, 512] - - [526, 94.609] - - - [126976, 25089, 1, 512] - - [526, 94.144] - - - [126976, 24577, 1, 512] - - [526, 94.053] - - - [126464, 38400, 1, 512] - - [526, 94.632] - - - [126464, 24576, 1, 512] - - [526, 94.618] - - - [126976, 24576, 1, 512] - - [526, 94.606] - - - [126464, 25088, 1, 512] - - [526, 94.635] - - - [126464, 24577, 1, 512] - - [529, 94.091] - - - [125952, 38400, 1, 512] - - [526, 94.626] - - - [125952, 24576, 1, 512] - - [526, 94.592] - - - [125952, 24577, 1, 512] - - [530, 94.076] - - - [125952, 24065, 1, 512] - - [526, 94.108] - - - [125440, 38400, 1, 512] - - [526, 94.62] - - - [125440, 24064, 1, 512] - - [526, 94.626] - - - [125952, 24064, 1, 512] - - [526, 94.613] - - - [125440, 24576, 1, 512] - - [526, 94.585] - - - [125440, 24065, 1, 512] - - [526, 94.099] - - - [124928, 38400, 1, 512] - - [526, 94.611] - - - [124928, 24064, 1, 512] - - [526, 94.612] - - - [124928, 24065, 1, 512] - - [526, 94.11] - - - [124928, 23553, 1, 512] - - [526, 94.076] - - - [124416, 38400, 1, 512] - - [526, 94.637] - - - [124416, 23552, 1, 512] - - [526, 94.605] - - - [124928, 23552, 1, 512] - - [526, 94.596] - - - [124416, 24064, 1, 512] - - [526, 94.624] - - - [124416, 23553, 1, 512] - - [526, 94.092] - - - [123904, 38400, 1, 512] - - [526, 94.64] - - - [123904, 23552, 1, 512] - - [526, 94.589] - - - [123904, 23553, 1, 512] - - [526, 94.079] - - - [123904, 23041, 1, 512] - - [528, 94.076] - - - [123392, 38400, 1, 512] - - [526, 94.647] - - - [123904, 23040, 1, 512] - - [526, 94.609] - - - [123392, 23040, 1, 512] - - [526, 94.617] - - - [123392, 23552, 1, 512] - - [526, 94.626] - - - [123392, 23041, 1, 512] - - [528, 94.075] - - - [122880, 38400, 1, 512] - - [526, 94.624] - - - [122880, 23040, 1, 512] - - [526, 94.605] - - - [122880, 23041, 1, 512] - - [530, 94.05] - - - [122880, 22529, 1, 512] - - [529, 94.021] - - - [122368, 38400, 1, 512] - - [530, 94.577] - - - [122368, 22528, 1, 512] - - [526, 94.603] - - - [122880, 22528, 1, 512] - - [528, 94.571] - - - [122368, 23040, 1, 512] - - [526, 94.63] - - - [122368, 22529, 1, 512] - - [533, 94.059] - - - [121856, 38400, 1, 512] - - [526, 94.629] - - - [121856, 22528, 1, 512] - - [526, 94.604] - - - [121856, 22529, 1, 512] - - [529, 94.046] - - - [121856, 22017, 1, 512] - - [526, 94.058] - - - [121344, 38400, 1, 512] - - [526, 94.651] - - - [121344, 22016, 1, 512] - - [526, 94.618] - - - [121856, 22016, 1, 512] - - [526, 94.59] - - - [121344, 22528, 1, 512] - - [526, 94.611] - - - [121344, 22017, 1, 512] - - [526, 94.047] - - - [120832, 38400, 1, 512] - - [526, 94.632] - - - [120832, 22016, 1, 512] - - [526, 94.607] - - - [120832, 22017, 1, 512] - - [526, 94.048] - - - [120832, 21505, 1, 512] - - [526, 94.058] - - - [120320, 38400, 1, 512] - - [526, 94.631] - - - [120832, 21504, 1, 512] - - [526, 94.61] - - - [120320, 21504, 1, 512] - - [529, 94.612] - - - [120320, 22016, 1, 512] - - [533, 94.596] - - - [120320, 21505, 1, 512] - - [526, 94.057] - - - [119808, 38400, 1, 512] - - [526, 94.64] - - - [119808, 21504, 1, 512] - - [526, 94.619] - - - [119808, 21505, 1, 512] - - [526, 94.064] - - - [119808, 20993, 1, 512] - - [528, 93.976] - - - [119296, 38400, 1, 512] - - [526, 94.641] - - - [119296, 20992, 1, 512] - - [526, 94.584] - - - [119808, 20992, 1, 512] - - [526, 94.595] - - - [119296, 21504, 1, 512] - - [526, 94.606] - - - [119296, 20993, 1, 512] - - [526, 93.987] - - - [118784, 38400, 1, 512] - - [526, 94.635] - - - [118784, 20992, 1, 512] - - [526, 94.572] - - - [118784, 20993, 1, 512] - - [528, 93.986] - - - [118784, 20481, 1, 512] - - [530, 93.983] - - - [118272, 38400, 1, 512] - - [526, 94.634] - - - [118272, 20480, 1, 512] - - [526, 94.613] - - - [118784, 20480, 1, 512] - - [529, 94.574] - - - [118272, 20992, 1, 512] - - [526, 94.603] - - - [118272, 20481, 1, 512] - - [526, 94.013] - - - [117760, 38400, 1, 512] - - [526, 94.634] - - - [117760, 20480, 1, 512] - - [528, 94.603] - - - [117760, 20481, 1, 512] - - [529, 94.01] - - - [117760, 19969, 1, 512] - - [526, 93.975] - - - [117248, 38400, 1, 512] - - [526, 94.641] - - - [117760, 19968, 1, 512] - - [526, 94.584] - - - [117248, 19968, 1, 512] - - [526, 94.597] - - - [117248, 20480, 1, 512] - - [526, 94.614] - - - [117248, 19969, 1, 512] - - [526, 93.991] - - - [116736, 38400, 1, 512] - - [526, 94.634] - - - [116736, 19968, 1, 512] - - [526, 94.586] - - - [116736, 19969, 1, 512] - - [526, 93.992] - - - [116736, 19457, 1, 512] - - [528, 93.937] - - - [116224, 38400, 1, 512] - - [526, 94.637] - - - [116224, 19456, 1, 512] - - [526, 94.614] - - - [116736, 19456, 1, 512] - - [528, 94.6] - - - [116224, 19968, 1, 512] - - [526, 94.6] - - - [116224, 19457, 1, 512] - - [529, 93.945] - - - [115712, 38400, 1, 512] - - [526, 94.632] - - - [115712, 19456, 1, 512] - - [526, 94.6] - - - [115712, 19457, 1, 512] - - [529, 93.949] - - - [115712, 18945, 1, 512] - - [528, 93.927] - - - [115200, 38400, 1, 512] - - [526, 94.638] - - - [115200, 18944, 1, 512] - - [533, 94.571] - - - [115712, 18944, 1, 512] - - [530, 94.583] - - - [115200, 19456, 1, 512] - - [526, 94.6] - - - [115200, 18945, 1, 512] - - [529, 93.908] - - - [114688, 38400, 1, 512] - - [528, 94.619] - - - [114688, 18944, 1, 512] - - [528, 94.566] - - - [114688, 18945, 1, 512] - - [528, 93.922] - - - [114688, 18433, 1, 512] - - [529, 93.902] - - - [114176, 38400, 1, 512] - - [526, 94.645] - - - [114176, 18432, 1, 512] - - [526, 94.592] - - - [114688, 18432, 1, 512] - - [528, 94.587] - - - [114176, 18944, 1, 512] - - [530, 94.556] - - - [114176, 18433, 1, 512] - - [528, 93.951] - - - [113664, 38400, 1, 512] - - [526, 94.632] - - - [113664, 18432, 1, 512] - - [528, 94.6] - - - [113664, 18433, 1, 512] - - [529, 93.951] - - - [113664, 17921, 1, 512] - - [526, 93.938] - - - [113152, 38400, 1, 512] - - [526, 94.643] - - - [113152, 17920, 1, 512] - - [526, 94.602] - - - [113664, 17920, 1, 512] - - [526, 94.606] - - - [113152, 18432, 1, 512] - - [526, 94.597] - - - [113152, 17921, 1, 512] - - [526, 93.922] - - - [112640, 38400, 1, 512] - - [526, 94.64] - - - [112640, 17920, 1, 512] - - [526, 94.59] - - - [112640, 17921, 1, 512] - - [526, 93.928] - - - [112640, 17409, 1, 512] - - [529, 93.845] - - - [112128, 38400, 1, 512] - - [526, 94.639] - - - [112128, 17408, 1, 512] - - [526, 94.577] - - - [112640, 17408, 1, 512] - - [526, 94.578] - - - [112128, 17409, 1, 512] - - [529, 93.864] - - - [112128, 17920, 1, 512] - - [526, 94.593] - - - [111616, 38400, 1, 512] - - [528, 94.613] - - - [111616, 17408, 1, 512] - - [526, 94.582] - - - [111616, 17409, 1, 512] - - [529, 93.901] - - - [111616, 16897, 1, 512] - - [526, 93.886] - - - [111104, 38400, 1, 512] - - [526, 94.624] - - - [111104, 16896, 1, 512] - - [526, 94.589] - - - [111616, 16896, 1, 512] - - [526, 94.573] - - - [111104, 16897, 1, 512] - - [526, 93.866] - - - [111104, 17408, 1, 512] - - [529, 94.565] - - - [110592, 38400, 1, 512] - - [526, 94.63] - - - [110592, 16896, 1, 512] - - [526, 94.582] - - - [110592, 16897, 1, 512] - - [526, 93.848] - - - [110592, 16385, 1, 512] - - [526, 93.812] - - - [110080, 38400, 1, 512] - - [526, 94.636] - - - [110080, 16384, 1, 512] - - [526, 94.575] - - - [110592, 16384, 1, 512] - - [526, 94.562] - - - [110080, 16896, 1, 512] - - [526, 94.573] - - - [110080, 16385, 1, 512] - - [529, 93.84] - - - [109568, 38400, 1, 512] - - [526, 94.615] - - - [109568, 16384, 1, 512] - - [533, 94.55] - - - [109568, 16385, 1, 512] - - [529, 93.798] - - - [109568, 15873, 1, 512] - - [529, 93.732] - - - [109056, 38400, 1, 512] - - [526, 94.622] - - - [109056, 15872, 1, 512] - - [526, 94.566] - - - [109568, 15872, 1, 512] - - [526, 94.563] - - - [109056, 16384, 1, 512] - - [530, 94.553] - - - [109056, 15873, 1, 512] - - [526, 93.768] - - - [108544, 38400, 1, 512] - - [526, 94.613] - - - [108544, 15872, 1, 512] - - [526, 94.55] - - - [108544, 15873, 1, 512] - - [526, 93.754] - - - [108544, 15361, 1, 512] - - [530, 93.761] - - - [108032, 38400, 1, 512] - - [526, 94.639] - - - [108032, 15360, 1, 512] - - [526, 94.546] - - - [108544, 15360, 1, 512] - - [526, 94.542] - - - [108032, 15361, 1, 512] - - [530, 93.728] - - - [108032, 15872, 1, 512] - - [526, 94.57] - - - [107520, 38400, 1, 512] - - [526, 94.616] - - - [107520, 15360, 1, 512] - - [526, 94.547] - - - [107520, 15361, 1, 512] - - [530, 93.758] - - - [107520, 14849, 1, 512] - - [533, 93.684] - - - [107008, 38400, 1, 512] - - [526, 94.626] - - - [107008, 14848, 1, 512] - - [528, 94.54] - - - [107520, 14848, 1, 512] - - [526, 94.564] - - - [107008, 15360, 1, 512] - - [528, 94.533] - - - [107008, 14849, 1, 512] - - [526, 93.727] - - - [106496, 38400, 1, 512] - - [526, 94.617] - - - [106496, 14848, 1, 512] - - [526, 94.537] - - - [106496, 14849, 1, 512] - - [533, 93.718] - - - [106496, 14337, 1, 512] - - [526, 93.719] - - - [105984, 38400, 1, 512] - - [526, 94.627] - - - [105984, 14336, 1, 512] - - [526, 94.59] - - - [106496, 14336, 1, 512] - - [526, 94.56] - - - [105984, 14848, 1, 512] - - [526, 94.55] - - - [105984, 14337, 1, 512] - - [526, 93.735] - - - [105472, 38400, 1, 512] - - [526, 94.641] - - - [105472, 14336, 1, 512] - - [526, 94.565] - - - [105472, 14337, 1, 512] - - [526, 93.72] - - - [105472, 13825, 1, 512] - - [530, 93.641] - - - [104960, 38400, 1, 512] - - [526, 94.618] - - - [104960, 13824, 1, 512] - - [528, 94.563] - - - [105472, 13824, 1, 512] - - [530, 94.528] - - - [104960, 14336, 1, 512] - - [526, 94.557] - - - [104960, 13825, 1, 512] - - [528, 93.7] - - - [104448, 38400, 1, 512] - - [526, 94.635] - - - [104448, 13824, 1, 512] - - [528, 94.544] - - - [104448, 13825, 1, 512] - - [528, 93.673] - - - [104448, 13313, 1, 512] - - [529, 93.67] - - - [103936, 38400, 1, 512] - - [526, 94.634] - - - [103936, 13312, 1, 512] - - [526, 94.556] - - - [104448, 13312, 1, 512] - - [526, 94.557] - - - [103936, 13313, 1, 512] - - [529, 93.667] - - - [103936, 13824, 1, 512] - - [528, 94.553] - - - [103424, 38400, 1, 512] - - [526, 94.631] - - - [103424, 13312, 1, 512] - - [526, 94.56] - - - [103424, 13313, 1, 512] - - [529, 93.653] - - - [103424, 12801, 1, 512] - - [530, 93.573] - - - [102912, 38400, 1, 512] - - [526, 94.622] - - - [102912, 12800, 1, 512] - - [530, 94.494] - - - [103424, 12800, 1, 512] - - [530, 94.486] - - - [102912, 13312, 1, 512] - - [526, 94.559] - - - [102912, 12801, 1, 512] - - [526, 93.582] - - - [102400, 38400, 1, 512] - - [526, 94.609] - - - [102400, 12800, 1, 512] - - [526, 94.524] - - - [102400, 12801, 1, 512] - - [530, 93.602] - - - [102400, 12289, 1, 512] - - [529, 93.542] - - - [101888, 38400, 1, 512] - - [526, 94.627] - - - [101888, 12288, 1, 512] - - [526, 94.551] - - - [102400, 12288, 1, 512] - - [529, 94.516] - - - [101888, 12800, 1, 512] - - [526, 94.534] - - - [101888, 12289, 1, 512] - - [529, 93.567] - - - [101376, 38400, 1, 512] - - [526, 94.638] - - - [101376, 12288, 1, 512] - - [528, 94.533] - - - [101376, 12289, 1, 512] - - [529, 93.537] - - - [101376, 11777, 1, 512] - - [528, 93.462] - - - [100864, 38400, 1, 512] - - [526, 94.613] - - - [100864, 11776, 1, 512] - - [526, 94.495] - - - [101376, 11776, 1, 512] - - [529, 94.515] - - - [100864, 12288, 1, 512] - - [526, 94.542] - - - [100864, 11777, 1, 512] - - [533, 93.444] - - - [100352, 38400, 1, 512] - - [526, 94.642] - - - [100352, 11776, 1, 512] - - [529, 94.508] - - - [100352, 11777, 1, 512] - - [529, 93.472] - - - [100352, 11265, 1, 512] - - [529, 93.452] - - - [99840, 38400, 1, 512] - - [526, 94.636] - - - [99840, 11264, 1, 512] - - [528, 94.526] - - - [100352, 11264, 1, 512] - - [529, 94.528] - - - [99840, 11265, 1, 512] - - [529, 93.46] - - - [99840, 11776, 1, 512] - - [530, 94.515] - - - [99328, 38400, 1, 512] - - [526, 94.633] - - - [99328, 11264, 1, 512] - - [529, 94.552] - - - [99328, 11265, 1, 512] - - [529, 93.507] - - - [99328, 10753, 1, 512] - - [526, 93.426] - - - [98816, 38400, 1, 512] - - [526, 94.632] - - - [98816, 10752, 1, 512] - - [526, 94.544] - - - [99328, 10752, 1, 512] - - [533, 94.526] - - - [98816, 10753, 1, 512] - - [526, 93.42] - - - [98816, 11264, 1, 512] - - [529, 94.522] - - - [98304, 38400, 1, 512] - - [526, 94.565] - - - [98304, 10752, 1, 512] - - [526, 94.471] - - - [98304, 10753, 1, 512] - - [526, 93.345] - - - [98304, 10241, 1, 512] - - [530, 93.274] - - - [97792, 38400, 1, 512] - - [526, 94.627] - - - [97792, 10240, 1, 512] - - [528, 94.531] - - - [98304, 10240, 1, 512] - - [528, 94.476] - - - [97792, 10752, 1, 512] - - [526, 94.485] - - - [97792, 10241, 1, 512] - - [530, 93.372] - - - [97280, 38400, 1, 512] - - [526, 94.637] - - - [97280, 10240, 1, 512] - - [530, 94.534] - - - [97280, 10241, 1, 512] - - [529, 93.375] - - - [97280, 9729, 1, 512] - - [526, 93.282] - - - [96768, 38400, 1, 512] - - [526, 94.637] - - - [96768, 9728, 1, 512] - - [528, 94.491] - - - [97280, 9728, 1, 512] - - [526, 94.492] - - - [96768, 10240, 1, 512] - - [528, 94.525] - - - [96768, 9729, 1, 512] - - [526, 93.278] - - - [96256, 38400, 1, 512] - - [526, 94.64] - - - [96256, 9728, 1, 512] - - [526, 94.508] - - - [96256, 9729, 1, 512] - - [530, 93.238] - - - [96256, 9217, 1, 512] - - [529, 93.233] - - - [95744, 38400, 1, 512] - - [526, 94.648] - - - [95744, 9216, 1, 512] - - [528, 94.5] - - - [96256, 9216, 1, 512] - - [528, 94.535] - - - [95744, 9217, 1, 512] - - [529, 93.24] - - - [95744, 9728, 1, 512] - - [530, 94.482] - - - [95232, 38400, 1, 512] - - [526, 94.627] - - - [95232, 9216, 1, 512] - - [529, 94.514] - - - [95232, 9217, 1, 512] - - [528, 93.269] - - - [95232, 8705, 1, 512] - - [530, 93.139] - - - [94720, 38400, 1, 512] - - [526, 94.64] - - - [94720, 8704, 1, 512] - - [526, 94.491] - - - [95232, 8704, 1, 512] - - [530, 94.545] - - - [94720, 8705, 1, 512] - - [530, 93.151] - - - [94720, 9216, 1, 512] - - [529, 94.53] - - - [94208, 38400, 1, 512] - - [529, 94.6] - - - [94208, 8704, 1, 512] - - [526, 94.524] - - - [94208, 8705, 1, 512] - - [530, 93.126] - - - [94208, 8193, 1, 512] - - [530, 93.054] - - - [93696, 38400, 1, 512] - - [526, 94.638] - - - [93696, 8192, 1, 512] - - [529, 94.517] - - - [94208, 8192, 1, 512] - - [533, 94.523] - - - [93696, 8704, 1, 512] - - [530, 94.518] - - - [93696, 8193, 1, 512] - - [529, 93.056] - - - [93184, 38400, 1, 512] - - [526, 94.632] - - - [93184, 8192, 1, 512] - - [533, 94.516] - - - [93184, 8193, 1, 512] - - [529, 93.109] - - - [93184, 7681, 1, 512] - - [530, 92.922] - - - [92672, 38400, 1, 512] - - [526, 94.635] - - - [92672, 7680, 1, 512] - - [530, 94.526] - - - [93184, 7680, 1, 512] - - [533, 94.479] - - - [92672, 8192, 1, 512] - - [533, 94.494] - - - [92672, 7681, 1, 512] - - [528, 92.94] - - - [92160, 38400, 1, 512] - - [526, 94.635] - - - [92160, 7680, 1, 512] - - [528, 94.488] - - - [92160, 7681, 1, 512] - - [528, 92.923] - - - [92160, 7169, 1, 512] - - [529, 92.808] - - - [91648, 38400, 1, 512] - - [526, 94.619] - - - [91648, 7168, 1, 512] - - [526, 94.495] - - - [92160, 7168, 1, 512] - - [529, 94.475] - - - [91648, 7680, 1, 512] - - [533, 94.52] - - - [91648, 7169, 1, 512] - - [526, 92.895] - - - [91136, 38400, 1, 512] - - [526, 94.632] - - - [91136, 7168, 1, 512] - - [526, 94.504] - - - [91136, 7169, 1, 512] - - [529, 92.871] - - - [91136, 6657, 1, 512] - - [533, 92.641] - - - [90624, 38400, 1, 512] - - [526, 94.632] - - - [90624, 6656, 1, 512] - - [533, 94.505] - - - [91136, 6656, 1, 512] - - [533, 94.506] - - - [90624, 7168, 1, 512] - - [526, 94.503] - - - [90624, 6657, 1, 512] - - [527, 92.672] - - - [90112, 38400, 1, 512] - - [526, 94.628] - - - [90112, 6656, 1, 512] - - [528, 94.477] - - - [90112, 6657, 1, 512] - - [527, 92.665] - - - [90112, 6145, 1, 512] - - [532, 92.475] - - - [89600, 38400, 1, 512] - - [526, 94.631] - - - [89600, 6144, 1, 512] - - [529, 94.421] - - - [90112, 6144, 1, 512] - - [527, 94.442] - - - [89600, 6656, 1, 512] - - [533, 94.49] - - - [89600, 6145, 1, 512] - - [529, 92.511] - - - [89088, 38400, 1, 512] - - [526, 94.626] - - - [89088, 6144, 1, 512] - - [529, 94.443] - - - [89088, 6145, 1, 512] - - [529, 92.488] - - - [89088, 5633, 1, 512] - - [533, 92.38] - - - [88576, 38400, 1, 512] - - [526, 94.638] - - - [88576, 5632, 1, 512] - - [528, 94.389] - - - [89088, 5632, 1, 512] - - [533, 94.493] - - - [88576, 6144, 1, 512] - - [529, 94.441] - - - [88576, 5633, 1, 512] - - [533, 92.348] - - - [88064, 38400, 1, 512] - - [526, 94.641] - - - [88064, 5632, 1, 512] - - [533, 94.407] - - - [88064, 5633, 1, 512] - - [533, 92.274] - - - [88064, 5121, 1, 512] - - [529, 92.095] - - - [87552, 38400, 1, 512] - - [526, 94.618] - - - [88064, 5120, 1, 512] - - [533, 94.392] - - - [87552, 5120, 1, 512] - - [529, 94.292] - - - [87552, 5121, 1, 512] - - [529, 92.133] - - - [87552, 5632, 1, 512] - - [533, 94.487] - - - [87040, 38400, 1, 512] - - [526, 94.634] - - - [87040, 5120, 1, 512] - - [529, 94.37] - - - [87040, 5121, 1, 512] - - [529, 92.035] - - - [87040, 4609, 1, 512] - - [528, 91.824] - - - [86528, 38400, 1, 512] - - [526, 94.623] - - - [86528, 4608, 1, 512] - - [530, 94.299] - - - [87040, 4608, 1, 512] - - [530, 94.383] - - - [86528, 5120, 1, 512] - - [530, 94.438] - - - [86528, 4609, 1, 512] - - [532, 91.732] - - - [86016, 38400, 1, 512] - - [526, 94.64] - - - [86016, 4608, 1, 512] - - [528, 94.394] - - - [86016, 4609, 1, 512] - - [528, 91.79] - - - [86016, 4097, 1, 512] - - [529, 91.416] - - - [85504, 38400, 1, 512] - - [526, 94.637] - - - [85504, 4096, 1, 512] - - [529, 94.236] - - - [86016, 4096, 1, 512] - - [532, 94.298] - - - [85504, 4608, 1, 512] - - [528, 94.344] - - - [85504, 4097, 1, 512] - - [529, 91.453] - - - [84992, 38400, 1, 512] - - [526, 94.642] - - - [84992, 4096, 1, 512] - - [528, 94.185] - - - [84992, 4097, 1, 512] - - [529, 91.502] - - - [84992, 3585, 1, 512] - - [526, 90.995] - - - [84480, 38400, 1, 512] - - [526, 94.632] - - - [84480, 3584, 1, 512] - - [533, 94.16] - - - [84992, 3584, 1, 512] - - [530, 94.045] - - - [84480, 4096, 1, 512] - - [533, 94.319] - - - [84480, 3585, 1, 512] - - [526, 90.961] - - - [83968, 38400, 1, 512] - - [526, 94.63] - - - [83968, 3584, 1, 512] - - [530, 94.272] - - - [83968, 3585, 1, 512] - - [526, 90.958] - - - [83968, 3073, 1, 512] - - [529, 90.431] - - - [83456, 38400, 1, 512] - - [526, 94.647] - - - [83456, 3072, 1, 512] - - [529, 94.002] - - - [83968, 3072, 1, 512] - - [529, 93.942] - - - [83456, 3584, 1, 512] - - [533, 94.25] - - - [83456, 3073, 1, 512] - - [529, 90.428] - - - [82944, 38400, 1, 512] - - [526, 94.639] - - - [82944, 3072, 1, 512] - - [529, 94.031] - - - [82944, 3073, 1, 512] - - [529, 90.434] - - - [82944, 2561, 1, 512] - - [530, 89.378] - - - [82432, 38400, 1, 512] - - [526, 94.624] - - - [82432, 2560, 1, 512] - - [530, 93.692] - - - [82944, 2560, 1, 512] - - [530, 94.044] - - - [82432, 2561, 1, 512] - - [530, 89.46] - - - [82432, 3072, 1, 512] - - [529, 94.018] - - - [81920, 38400, 1, 512] - - [526, 94.59] - - - [81920, 2560, 1, 512] - - [530, 93.826] - - - [81920, 2561, 1, 512] - - [530, 89.507] - - - [81920, 2049, 1, 512] - - [529, 88.229] - - - [81408, 38400, 1, 512] - - [526, 94.64] - - - [81408, 2048, 1, 512] - - [529, 93.75] - - - [81920, 2048, 1, 512] - - [529, 93.515] - - - [81408, 2049, 1, 512] - - [529, 88.407] - - - [81408, 2560, 1, 512] - - [533, 93.933] - - - [80896, 38400, 1, 512] - - [526, 94.637] - - - [80896, 2048, 1, 512] - - [529, 93.836] - - - [80896, 2049, 1, 512] - - [529, 88.058] - - - [80896, 1537, 1, 512] - - [525, 86.143] - - - [80384, 38400, 1, 512] - - [526, 94.643] - - - [80384, 1536, 1, 512] - - [534, 93.445] - - - [80896, 1536, 1, 512] - - [526, 93.033] - - - [80384, 2048, 1, 512] - - [529, 93.566] - - - [80384, 1537, 1, 512] - - [532, 86.556] - - - [79872, 38400, 1, 512] - - [526, 94.642] - - - [79872, 1536, 1, 512] - - [525, 93.031] - - - [79872, 1537, 1, 512] - - [532, 86.141] - - - [79872, 1025, 1, 512] - - [532, 82.691] - - - [79360, 38400, 1, 512] - - [526, 94.637] - - - [79360, 1024, 1, 512] - - [533, 92.51] - - - [79872, 1024, 1, 512] - - [533, 92.899] - - - [79360, 1536, 1, 512] - - [529, 93.482] - - - [79360, 1025, 1, 512] - - [529, 82.625] - - - [78848, 38400, 1, 512] - - [526, 94.644] - - - [78848, 1024, 1, 512] - - [533, 92.072] - - - [78848, 1025, 1, 512] - - [529, 82.236] - - - [78848, 513, 1, 512] - - [535, 72.143] - - - [78336, 38400, 1, 512] - - [526, 94.627] - - - [78336, 513, 1, 512] - - [525, 73.146] - - - [78336, 1024, 1, 512] - - [532, 92.964] - - - [77824, 513, 1, 512] - - [526, 72.656] - - - [90880, 512, 1, 512] - - [527, 91.158] - - - [90880, 54272, 1, 512] - - [526, 94.647] - - - [90624, 54272, 1, 512] - - [526, 94.658] - - - [90624, 36352, 1, 512] - - [526, 94.653] - - - [90880, 36352, 1, 512] - - [526, 94.636] - - - [90624, 36097, 1, 512] - - [526, 94.303] - - - [90880, 36097, 1, 512] - - [526, 94.286] - - - [90368, 512, 1, 512] - - [530, 90.659] - - - [90368, 54272, 1, 512] - - [526, 94.639] - - - [90624, 35840, 1, 512] - - [526, 94.648] - - - [90368, 35840, 1, 512] - - [526, 94.65] - - - [90368, 36097, 1, 512] - - [526, 94.293] - - - [90112, 54272, 1, 512] - - [526, 94.647] - - - [90112, 35840, 1, 512] - - [526, 94.647] - - - [90112, 35585, 1, 512] - - [528, 94.314] - - - [90368, 35585, 1, 512] - - [526, 94.271] - - - [89856, 54272, 1, 512] - - [526, 94.65] - - - [89856, 512, 1, 512] - - [531, 90.29] - - - [90112, 35328, 1, 512] - - [526, 94.637] - - - [89856, 35328, 1, 512] - - [526, 94.622] - - - [89856, 35585, 1, 512] - - [528, 94.306] - - - [89600, 54272, 1, 512] - - [526, 94.658] - - - [89600, 35328, 1, 512] - - [526, 94.643] - - - [89856, 35073, 1, 512] - - [528, 94.26] - - - [89600, 35073, 1, 512] - - [528, 94.276] - - - [89344, 512, 1, 512] - - [533, 89.935] - - - [89344, 54272, 1, 512] - - [526, 94.662] - - - [89344, 34816, 1, 512] - - [526, 94.624] - - - [89600, 34816, 1, 512] - - [526, 94.642] - - - [89344, 35073, 1, 512] - - [533, 94.261] - - - [89088, 54272, 1, 512] - - [526, 94.655] - - - [89088, 34816, 1, 512] - - [526, 94.654] - - - [89344, 34561, 1, 512] - - [528, 94.292] - - - [89088, 34561, 1, 512] - - [526, 94.299] - - - [88832, 54272, 1, 512] - - [526, 94.642] - - - [88832, 512, 1, 512] - - [526, 89.673] - - - [88832, 34304, 1, 512] - - [528, 94.634] - - - [89088, 34304, 1, 512] - - [526, 94.649] - - - [88832, 34561, 1, 512] - - [526, 94.287] - - - [88576, 54272, 1, 512] - - [526, 94.655] - - - [88576, 34304, 1, 512] - - [528, 94.645] - - - [88576, 34049, 1, 512] - - [526, 94.299] - - - [88832, 34049, 1, 512] - - [526, 94.293] - - - [88320, 54272, 1, 512] - - [526, 94.642] - - - [88320, 512, 1, 512] - - [526, 91.226] - - - [88576, 33792, 1, 512] - - [526, 94.65] - - - [88320, 33792, 1, 512] - - [526, 94.636] - - - [88320, 34049, 1, 512] - - [526, 94.295] - - - [88064, 54272, 1, 512] - - [528, 94.656] - - - [88064, 33792, 1, 512] - - [526, 94.653] - - - [88320, 33537, 1, 512] - - [526, 94.232] - - - [88064, 33537, 1, 512] - - [528, 94.254] - - - [87808, 54272, 1, 512] - - [526, 94.643] - - - [87808, 512, 1, 512] - - [535, 91.089] - - - [87808, 33280, 1, 512] - - [528, 94.619] - - - [88064, 33280, 1, 512] - - [526, 94.629] - - - [87808, 33537, 1, 512] - - [526, 94.248] - - - [87552, 54272, 1, 512] - - [526, 94.651] - - - [87552, 33280, 1, 512] - - [528, 94.627] - - - [87808, 33025, 1, 512] - - [526, 94.251] - - - [87552, 33025, 1, 512] - - [528, 94.283] - - - [87296, 54272, 1, 512] - - [526, 94.65] - - - [87296, 512, 1, 512] - - [531, 90.596] - - - [87552, 32768, 1, 512] - - [526, 94.633] - - - [87296, 32768, 1, 512] - - [528, 94.623] - - - [87296, 33025, 1, 512] - - [526, 94.253] - - - [87040, 54272, 1, 512] - - [526, 94.652] - - - [87040, 32768, 1, 512] - - [526, 94.624] - - - [87040, 32513, 1, 512] - - [528, 94.242] - - - [87296, 32513, 1, 512] - - [526, 94.226] - - - [86784, 54272, 1, 512] - - [526, 94.656] - - - [86784, 512, 1, 512] - - [533, 90.267] - - - [87040, 32256, 1, 512] - - [526, 94.654] - - - [86784, 32256, 1, 512] - - [526, 94.635] - - - [86784, 32513, 1, 512] - - [528, 94.256] - - - [86528, 54272, 1, 512] - - [526, 94.661] - - - [86528, 32256, 1, 512] - - [528, 94.65] - - - [86784, 32001, 1, 512] - - [526, 94.214] - - - [86528, 32001, 1, 512] - - [526, 94.237] - - - [86272, 54272, 1, 512] - - [526, 94.654] - - - [86272, 512, 1, 512] - - [530, 89.914] - - - [86528, 31744, 1, 512] - - [526, 94.64] - - - [86272, 31744, 1, 512] - - [526, 94.643] - - - [86272, 32001, 1, 512] - - [526, 94.233] - - - [86016, 54272, 1, 512] - - [526, 94.654] - - - [86016, 31744, 1, 512] - - [526, 94.64] - - - [86016, 31489, 1, 512] - - [533, 94.22] - - - [86272, 31489, 1, 512] - - [526, 94.222] - - - [85760, 54272, 1, 512] - - [526, 94.656] - - - [85760, 512, 1, 512] - - [525, 89.589] - - - [86016, 31232, 1, 512] - - [526, 94.648] - - - [85760, 31232, 1, 512] - - [526, 94.629] - - - [85760, 31489, 1, 512] - - [528, 94.217] - - - [85504, 54272, 1, 512] - - [526, 94.659] - - - [85504, 31232, 1, 512] - - [526, 94.64] - - - [85504, 30977, 1, 512] - - [526, 94.237] - - - [85760, 30977, 1, 512] - - [526, 94.243] - - - [85248, 54272, 1, 512] - - [526, 94.659] - - - [85248, 512, 1, 512] - - [531, 89.136] - - - [85504, 30720, 1, 512] - - [526, 94.638] - - - [85248, 30720, 1, 512] - - [526, 94.623] - - - [85248, 30977, 1, 512] - - [526, 94.236] - - - [84992, 54272, 1, 512] - - [526, 94.653] - - - [84992, 30720, 1, 512] - - [526, 94.623] - - - [84992, 30465, 1, 512] - - [526, 94.265] - - - [85248, 30465, 1, 512] - - [526, 94.249] - - - [84736, 54272, 1, 512] - - [526, 94.656] - - - [84736, 512, 1, 512] - - [534, 91.086] - - - [84992, 30208, 1, 512] - - [526, 94.639] - - - [84736, 30208, 1, 512] - - [526, 94.642] - - - [84736, 30465, 1, 512] - - [526, 94.241] - - - [84480, 54272, 1, 512] - - [526, 94.655] - - - [84480, 30208, 1, 512] - - [526, 94.643] - - - [84480, 29953, 1, 512] - - [528, 94.237] - - - [84736, 29953, 1, 512] - - [528, 94.235] - - - [84224, 54272, 1, 512] - - [526, 94.64] - - - [84224, 512, 1, 512] - - [527, 90.637] - - - [84480, 29696, 1, 512] - - [528, 94.633] - - - [84224, 29696, 1, 512] - - [528, 94.637] - - - [84224, 29953, 1, 512] - - [528, 94.216] - - - [83968, 54272, 1, 512] - - [526, 94.654] - - - [83968, 29696, 1, 512] - - [528, 94.648] - - - [83968, 29441, 1, 512] - - [526, 94.222] - - - [84224, 29441, 1, 512] - - [526, 94.201] - - - [83712, 54272, 1, 512] - - [526, 94.657] - - - [83712, 512, 1, 512] - - [532, 90.236] - - - [83968, 29184, 1, 512] - - [526, 94.654] - - - [83712, 29184, 1, 512] - - [526, 94.635] - - - [83712, 29441, 1, 512] - - [526, 94.203] - - - [83456, 54272, 1, 512] - - [526, 94.65] - - - [83456, 29184, 1, 512] - - [526, 94.63] - - - [83456, 28929, 1, 512] - - [526, 94.197] - - - [83712, 28929, 1, 512] - - [526, 94.191] - - - [83200, 54272, 1, 512] - - [526, 94.651] - - - [83200, 512, 1, 512] - - [525, 89.86] - - - [83456, 28672, 1, 512] - - [526, 94.64] - - - [83200, 28672, 1, 512] - - [526, 94.637] - - - [83200, 28929, 1, 512] - - [526, 94.214] - - - [82944, 54272, 1, 512] - - [526, 94.678] - - - [82944, 28417, 1, 512] - - [528, 94.206] - - - [83200, 28417, 1, 512] - - [528, 94.18] - - - [82944, 28672, 1, 512] - - [528, 94.649] - - - [82688, 54272, 1, 512] - - [526, 94.654] - - - [82688, 512, 1, 512] - - [534, 89.501] - - - [82944, 28160, 1, 512] - - [526, 94.608] - - - [82688, 28160, 1, 512] - - [526, 94.611] - - - [82688, 28417, 1, 512] - - [528, 94.172] - - - [82432, 54272, 1, 512] - - [526, 94.657] - - - [82432, 27905, 1, 512] - - [528, 94.175] - - - [82688, 27905, 1, 512] - - [528, 94.174] - - - [82432, 28160, 1, 512] - - [526, 94.621] - - - [82176, 54272, 1, 512] - - [526, 94.65] - - - [82176, 512, 1, 512] - - [527, 89.139] - - - [82432, 27648, 1, 512] - - [526, 94.649] - - - [82176, 27648, 1, 512] - - [526, 94.619] - - - [82176, 27905, 1, 512] - - [528, 94.165] - - - [81920, 54272, 1, 512] - - [533, 94.642] - - - [82176, 27393, 1, 512] - - [526, 94.168] - - - [81920, 27393, 1, 512] - - [526, 94.132] - - - [81920, 27648, 1, 512] - - [533, 94.617] - - - [81664, 54272, 1, 512] - - [526, 94.661] - - - [81664, 512, 1, 512] - - [534, 90.888] - - - [81920, 27136, 1, 512] - - [533, 94.599] - - - [81664, 27136, 1, 512] - - [526, 94.626] - - - [81664, 27393, 1, 512] - - [528, 94.179] - - - [81408, 54272, 1, 512] - - [526, 94.659] - - - [81408, 26881, 1, 512] - - [526, 94.19] - - - [81664, 26881, 1, 512] - - [526, 94.192] - - - [81408, 27136, 1, 512] - - [526, 94.624] - - - [81152, 54272, 1, 512] - - [526, 94.655] - - - [81152, 512, 1, 512] - - [525, 90.73] - - - [81408, 26624, 1, 512] - - [526, 94.636] - - - [81152, 26624, 1, 512] - - [526, 94.634] - - - [81152, 26881, 1, 512] - - [526, 94.188] - - - [80896, 54272, 1, 512] - - [526, 94.665] - - - [81152, 26369, 1, 512] - - [528, 94.168] - - - [80896, 26369, 1, 512] - - [529, 94.15] - - - [80896, 26624, 1, 512] - - [526, 94.631] - - - [80640, 54272, 1, 512] - - [526, 94.658] - - - [80640, 512, 1, 512] - - [527, 90.36] - - - [80896, 26112, 1, 512] - - [526, 94.618] - - - [80640, 26112, 1, 512] - - [528, 94.613] - - - [80640, 26369, 1, 512] - - [528, 94.137] - - - [80384, 54272, 1, 512] - - [526, 94.667] - - - [80384, 25857, 1, 512] - - [526, 94.166] - - - [80640, 25857, 1, 512] - - [526, 94.135] - - - [80384, 26112, 1, 512] - - [528, 94.62] - - - [80128, 54272, 1, 512] - - [526, 94.643] - - - [80128, 512, 1, 512] - - [531, 89.811] - - - [80128, 25600, 1, 512] - - [526, 94.601] - - - [80384, 25600, 1, 512] - - [526, 94.617] - - - [80128, 25857, 1, 512] - - [526, 94.155] - - - [79872, 54272, 1, 512] - - [526, 94.656] - - - [79872, 25345, 1, 512] - - [528, 94.136] - - - [80128, 25345, 1, 512] - - [526, 94.129] - - - [79872, 25600, 1, 512] - - [526, 94.613] - - - [79616, 54272, 1, 512] - - [526, 94.662] - - - [79616, 512, 1, 512] - - [528, 89.445] - - - [79872, 25088, 1, 512] - - [526, 94.642] - - - [79616, 25088, 1, 512] - - [526, 94.624] - - - [79616, 25345, 1, 512] - - [528, 94.156] - - - [79360, 54272, 1, 512] - - [526, 94.667] - - - [79360, 24833, 1, 512] - - [526, 94.109] - - - [79616, 24833, 1, 512] - - [526, 94.105] - - - [79360, 25088, 1, 512] - - [526, 94.632] - - - [79104, 54272, 1, 512] - - [526, 94.645] - - - [79104, 512, 1, 512] - - [532, 89.134] - - - [79360, 24576, 1, 512] - - [526, 94.621] - - - [79104, 24576, 1, 512] - - [529, 94.602] - - - [79104, 24833, 1, 512] - - [526, 94.101] - - - [78848, 54272, 1, 512] - - [526, 94.669] - - - [79104, 24321, 1, 512] - - [529, 94.072] - - - [78848, 24321, 1, 512] - - [526, 94.095] - - - [78848, 24576, 1, 512] - - [528, 94.615] - - - [78592, 54272, 1, 512] - - [526, 94.655] - - - [78592, 512, 1, 512] - - [530, 88.653] - - - [78848, 24064, 1, 512] - - [526, 94.632] - - - [78592, 24064, 1, 512] - - [526, 94.608] - - - [78592, 24321, 1, 512] - - [526, 94.087] - - - [78336, 54272, 1, 512] - - [526, 94.66] - - - [78592, 23809, 1, 512] - - [526, 94.097] - - - [78336, 23809, 1, 512] - - [526, 94.134] - - - [78336, 24064, 1, 512] - - [526, 94.637] - - - [78080, 54272, 1, 512] - - [526, 94.659] - - - [78080, 512, 1, 512] - - [526, 90.62] - - - [78336, 23552, 1, 512] - - [526, 94.621] - - - [78080, 23552, 1, 512] - - [526, 94.602] - - - [78080, 23809, 1, 512] - - [526, 94.094] - - - [77824, 54272, 1, 512] - - [526, 94.668] - - - [77824, 23297, 1, 512] - - [526, 94.12] - - - [78080, 23297, 1, 512] - - [526, 94.097] - - - [77824, 23552, 1, 512] - - [528, 94.603] - - - [77568, 54272, 1, 512] - - [526, 94.649] - - - [77568, 512, 1, 512] - - [533, 90.32] - - - [77824, 23040, 1, 512] - - [528, 94.602] - - - [77568, 23040, 1, 512] - - [528, 94.602] - - - [77568, 23297, 1, 512] - - [526, 94.122] - - - [77312, 54272, 1, 512] - - [526, 94.654] - - - [77312, 22785, 1, 512] - - [526, 94.024] - - - [77568, 22785, 1, 512] - - [526, 94.055] - - - [77312, 23040, 1, 512] - - [528, 94.63] - - - [77056, 54272, 1, 512] - - [526, 94.653] - - - [77056, 512, 1, 512] - - [534, 89.732] - - - [77056, 22528, 1, 512] - - [528, 94.597] - - - [77312, 22528, 1, 512] - - [528, 94.598] - - - [77056, 22785, 1, 512] - - [528, 94.05] - - - [76800, 54272, 1, 512] - - [526, 94.662] - - - [76800, 22273, 1, 512] - - [526, 94.07] - - - [77056, 22273, 1, 512] - - [526, 94.057] - - - [76800, 22528, 1, 512] - - [526, 94.599] - - - [76544, 54272, 1, 512] - - [526, 94.664] - - - [76544, 512, 1, 512] - - [533, 89.344] - - - [76800, 22016, 1, 512] - - [528, 94.593] - - - [76544, 22016, 1, 512] - - [528, 94.59] - - - [76544, 22273, 1, 512] - - [526, 94.055] - - - [76288, 54272, 1, 512] - - [526, 94.655] - - - [76288, 21761, 1, 512] - - [528, 94.039] - - - [76544, 21761, 1, 512] - - [528, 94.058] - - - [76288, 22016, 1, 512] - - [526, 94.615] - - - [76032, 54272, 1, 512] - - [528, 94.619] - - - [76032, 512, 1, 512] - - [525, 89.062] - - - [76288, 21504, 1, 512] - - [528, 94.594] - - - [76032, 21504, 1, 512] - - [526, 94.621] - - - [76032, 21761, 1, 512] - - [526, 94.03] - - - [75776, 54272, 1, 512] - - [526, 94.66] - - - [75776, 21249, 1, 512] - - [528, 94.005] - - - [76032, 21249, 1, 512] - - [529, 93.988] - - - [75776, 21504, 1, 512] - - [528, 94.61] - - - [75520, 54272, 1, 512] - - [526, 94.649] - - - [75520, 512, 1, 512] - - [526, 88.529] - - - [75776, 20992, 1, 512] - - [526, 94.576] - - - [75520, 20992, 1, 512] - - [526, 94.57] - - - [75520, 21249, 1, 512] - - [528, 94.034] - - - [75264, 54272, 1, 512] - - [526, 94.658] - - - [75264, 20737, 1, 512] - - [528, 94.008] - - - [75520, 20737, 1, 512] - - [528, 94.007] - - - [75264, 20992, 1, 512] - - [528, 94.598] - - - [75008, 54272, 1, 512] - - [526, 94.663] - - - [75008, 512, 1, 512] - - [530, 90.729] - - - [75264, 20480, 1, 512] - - [528, 94.612] - - - [75008, 20480, 1, 512] - - [526, 94.618] - - - [75008, 20737, 1, 512] - - [528, 94.022] - - - [74752, 54272, 1, 512] - - [526, 94.661] - - - [74752, 20225, 1, 512] - - [526, 93.995] - - - [75008, 20225, 1, 512] - - [529, 93.977] - - - [74752, 20480, 1, 512] - - [526, 94.61] - - - [74496, 54272, 1, 512] - - [526, 94.659] - - - [74496, 512, 1, 512] - - [529, 90.407] - - - [74752, 19968, 1, 512] - - [526, 94.58] - - - [74496, 19968, 1, 512] - - [526, 94.6] - - - [74496, 20225, 1, 512] - - [526, 93.978] - - - [74240, 54272, 1, 512] - - [526, 94.656] - - - [74240, 19713, 1, 512] - - [526, 94.02] - - - [74496, 19713, 1, 512] - - [526, 93.999] - - - [74240, 19968, 1, 512] - - [529, 94.586] - - - [73984, 54272, 1, 512] - - [526, 94.657] - - - [73984, 512, 1, 512] - - [529, 89.742] - - - [74240, 19456, 1, 512] - - [528, 94.593] - - - [73984, 19456, 1, 512] - - [528, 94.597] - - - [73984, 19713, 1, 512] - - [526, 93.982] - - - [73728, 54272, 1, 512] - - [526, 94.652] - - - [73984, 19201, 1, 512] - - [530, 93.948] - - - [73728, 19201, 1, 512] - - [530, 93.949] - - - [73728, 19456, 1, 512] - - [526, 94.599] - - - [73472, 54272, 1, 512] - - [526, 94.653] - - - [73472, 512, 1, 512] - - [525, 89.33] - - - [73472, 18944, 1, 512] - - [526, 94.556] - - - [73728, 18944, 1, 512] - - [528, 94.559] - - - [73472, 19201, 1, 512] - - [530, 93.944] - - - [73216, 54272, 1, 512] - - [526, 94.654] - - - [73216, 18689, 1, 512] - - [526, 93.953] - - - [73472, 18689, 1, 512] - - [526, 93.939] - - - [73216, 18944, 1, 512] - - [528, 94.557] - - - [72960, 54272, 1, 512] - - [526, 94.665] - - - [72960, 512, 1, 512] - - [528, 88.889] - - - [72960, 18432, 1, 512] - - [526, 94.581] - - - [73216, 18432, 1, 512] - - [528, 94.622] - - - [72960, 18689, 1, 512] - - [526, 93.927] - - - [72704, 54272, 1, 512] - - [526, 94.664] - - - [72960, 18177, 1, 512] - - [526, 93.896] - - - [72704, 18177, 1, 512] - - [529, 93.889] - - - [72704, 18432, 1, 512] - - [529, 94.579] - - - [72448, 54272, 1, 512] - - [526, 94.661] - - - [72448, 512, 1, 512] - - [531, 88.515] - - - [72704, 17920, 1, 512] - - [526, 94.595] - - - [72448, 17920, 1, 512] - - [526, 94.579] - - - [72448, 18177, 1, 512] - - [526, 93.928] - - - [72192, 54272, 1, 512] - - [526, 94.666] - - - [72192, 17665, 1, 512] - - [525, 93.863] - - - [72448, 17665, 1, 512] - - [526, 93.884] - - - [72192, 17920, 1, 512] - - [528, 94.591] - - - [71936, 54272, 1, 512] - - [526, 94.671] - - - [71936, 512, 1, 512] - - [528, 90.564] - - - [71936, 17408, 1, 512] - - [526, 94.569] - - - [72192, 17408, 1, 512] - - [526, 94.555] - - - [71936, 17665, 1, 512] - - [526, 93.868] - - - [71680, 54272, 1, 512] - - [526, 94.656] - - - [71680, 17153, 1, 512] - - [528, 93.861] - - - [71936, 17153, 1, 512] - - [526, 93.842] - - - [71680, 17408, 1, 512] - - [530, 94.568] - - - [71424, 54272, 1, 512] - - [526, 94.663] - - - [71424, 512, 1, 512] - - [525, 90.192] - - - [71680, 16896, 1, 512] - - [528, 94.589] - - - [71424, 16896, 1, 512] - - [528, 94.577] - - - [71424, 17153, 1, 512] - - [528, 93.847] - - - [71168, 54272, 1, 512] - - [528, 94.657] - - - [71424, 16641, 1, 512] - - [526, 93.837] - - - [71168, 16641, 1, 512] - - [530, 93.811] - - - [71168, 16896, 1, 512] - - [528, 94.574] - - - [70912, 54272, 1, 512] - - [526, 94.667] - - - [70912, 512, 1, 512] - - [531, 89.763] - - - [71168, 16384, 1, 512] - - [529, 94.553] - - - [70912, 16384, 1, 512] - - [529, 94.58] - - - [70912, 16641, 1, 512] - - [528, 93.83] - - - [70656, 54272, 1, 512] - - [526, 94.668] - - - [70656, 16129, 1, 512] - - [528, 93.841] - - - [70912, 16129, 1, 512] - - [526, 93.825] - - - [70656, 16384, 1, 512] - - [529, 94.58] - - - [70400, 54272, 1, 512] - - [526, 94.672] - - - [70400, 512, 1, 512] - - [528, 89.278] - - - [70656, 15872, 1, 512] - - [526, 94.545] - - - [70400, 15872, 1, 512] - - [528, 94.574] - - - [70400, 16129, 1, 512] - - [526, 93.862] - - - [70144, 54272, 1, 512] - - [526, 94.668] - - - [70144, 15617, 1, 512] - - [529, 93.693] - - - [70400, 15617, 1, 512] - - [526, 93.761] - - - [70144, 15872, 1, 512] - - [526, 94.563] - - - [69888, 54272, 1, 512] - - [526, 94.669] - - - [69888, 512, 1, 512] - - [533, 88.935] - - - [69888, 15360, 1, 512] - - [528, 94.561] - - - [70144, 15360, 1, 512] - - [529, 94.593] - - - [69888, 15617, 1, 512] - - [526, 93.728] - - - [69632, 54272, 1, 512] - - [526, 94.659] - - - [69632, 15105, 1, 512] - - [526, 93.731] - - - [69888, 15105, 1, 512] - - [526, 93.756] - - - [69632, 15360, 1, 512] - - [528, 94.54] - - - [69376, 54272, 1, 512] - - [526, 94.664] - - - [69376, 512, 1, 512] - - [529, 88.473] - - - [69376, 14848, 1, 512] - - [526, 94.556] - - - [69632, 14848, 1, 512] - - [528, 94.563] - - - [69376, 15105, 1, 512] - - [526, 93.769] - - - [69120, 54272, 1, 512] - - [526, 94.659] - - - [69120, 14593, 1, 512] - - [530, 93.718] - - - [69376, 14593, 1, 512] - - [525, 93.697] - - - [69120, 14848, 1, 512] - - [526, 94.574] - - - [68864, 54272, 1, 512] - - [526, 94.676] - - - [68864, 512, 1, 512] - - [526, 87.937] - - - [68864, 14336, 1, 512] - - [526, 94.593] - - - [69120, 14336, 1, 512] - - [526, 94.595] - - - [68864, 14593, 1, 512] - - [528, 93.721] - - - [68608, 54272, 1, 512] - - [526, 94.657] - - - [68608, 14081, 1, 512] - - [530, 93.702] - - - [68864, 14081, 1, 512] - - [530, 93.71] - - - [68608, 14336, 1, 512] - - [526, 94.564] - - - [68352, 54272, 1, 512] - - [526, 94.679] - - - [68352, 512, 1, 512] - - [527, 90.186] - - - [68352, 13824, 1, 512] - - [533, 94.569] - - - [68608, 13824, 1, 512] - - [528, 94.54] - - - [68352, 14081, 1, 512] - - [530, 93.712] - - - [68096, 54272, 1, 512] - - [526, 94.666] - - - [68096, 13569, 1, 512] - - [528, 93.633] - - - [68352, 13569, 1, 512] - - [528, 93.601] - - - [68096, 13824, 1, 512] - - [528, 94.558] - - - [67840, 54272, 1, 512] - - [526, 94.672] - - - [67840, 512, 1, 512] - - [527, 89.729] - - - [68096, 13312, 1, 512] - - [529, 94.553] - - - [67840, 13312, 1, 512] - - [528, 94.557] - - - [67840, 13569, 1, 512] - - [530, 93.612] - - - [67584, 54272, 1, 512] - - [528, 94.656] - - - [67584, 13057, 1, 512] - - [526, 93.616] - - - [67840, 13057, 1, 512] - - [529, 93.596] - - - [67584, 13312, 1, 512] - - [529, 94.566] - - - [67328, 54272, 1, 512] - - [526, 94.678] - - - [67328, 512, 1, 512] - - [529, 89.241] - - - [67328, 12800, 1, 512] - - [530, 94.557] - - - [67584, 12800, 1, 512] - - [530, 94.511] - - - [67328, 13057, 1, 512] - - [526, 93.637] - - - [67072, 54272, 1, 512] - - [526, 94.674] - - - [67072, 12545, 1, 512] - - [526, 93.596] - - - [67328, 12545, 1, 512] - - [526, 93.558] - - - [67072, 12800, 1, 512] - - [530, 94.541] - - - [66816, 54272, 1, 512] - - [526, 94.669] - - - [66816, 512, 1, 512] - - [531, 88.778] - - - [66816, 12288, 1, 512] - - [529, 94.569] - - - [67072, 12288, 1, 512] - - [526, 94.585] - - - [66816, 12545, 1, 512] - - [526, 93.622] - - - [66560, 54272, 1, 512] - - [526, 94.678] - - - [66560, 12033, 1, 512] - - [533, 93.519] - - - [66816, 12033, 1, 512] - - [529, 93.543] - - - [66560, 12288, 1, 512] - - [528, 94.532] - - - [66304, 54272, 1, 512] - - [526, 94.67] - - - [66304, 512, 1, 512] - - [533, 88.323] - - - [66304, 11776, 1, 512] - - [532, 94.534] - - - [66560, 11776, 1, 512] - - [526, 94.543] - - - [66304, 12033, 1, 512] - - [528, 93.521] - - - [66048, 54272, 1, 512] - - [526, 94.655] - - - [66048, 11521, 1, 512] - - [530, 93.484] - - - [66304, 11521, 1, 512] - - [530, 93.557] - - - [66048, 11776, 1, 512] - - [529, 94.498] - - - [65792, 54272, 1, 512] - - [526, 94.678] - - - [65792, 512, 1, 512] - - [525, 87.866] - - - [65792, 11264, 1, 512] - - [529, 94.55] - - - [66048, 11264, 1, 512] - - [530, 94.532] - - - [65792, 11521, 1, 512] - - [528, 93.528] - - - [65536, 54272, 1, 512] - - [533, 94.56] - - - [65536, 11009, 1, 512] - - [529, 93.334] - - - [65792, 11009, 1, 512] - - [530, 93.419] - - - [65536, 11264, 1, 512] - - [533, 94.456] - - - [65280, 54272, 1, 512] - - [526, 94.674] - - - [65280, 512, 1, 512] - - [532, 90.071] - - - [65536, 10752, 1, 512] - - [530, 94.44] - - - [65280, 10752, 1, 512] - - [526, 94.544] - - - [65280, 11009, 1, 512] - - [530, 93.401] - - - [65024, 54272, 1, 512] - - [526, 94.68] - - - [65280, 10497, 1, 512] - - [532, 93.299] - - - [65024, 10497, 1, 512] - - [528, 93.318] - - - [65024, 10752, 1, 512] - - [530, 94.515] - - - [64768, 54272, 1, 512] - - [526, 94.672] - - - [64768, 512, 1, 512] - - [532, 89.681] - - - [65024, 10240, 1, 512] - - [530, 94.49] - - - [64768, 10240, 1, 512] - - [528, 94.531] - - - [64768, 10497, 1, 512] - - [530, 93.344] - - - [64512, 54272, 1, 512] - - [526, 94.673] - - - [64512, 9985, 1, 512] - - [528, 93.297] - - - [64768, 9985, 1, 512] - - [528, 93.291] - - - [64512, 10240, 1, 512] - - [530, 94.526] - - - [64256, 54272, 1, 512] - - [526, 94.674] - - - [64256, 512, 1, 512] - - [528, 89.239] - - - [64256, 9728, 1, 512] - - [526, 94.552] - - - [64512, 9728, 1, 512] - - [528, 94.488] - - - [64256, 9985, 1, 512] - - [526, 93.305] - - - [64000, 54272, 1, 512] - - [526, 94.664] - - - [64000, 9473, 1, 512] - - [526, 93.164] - - - [64256, 9473, 1, 512] - - [530, 93.225] - - - [64000, 9728, 1, 512] - - [526, 94.48] - - - [63744, 54272, 1, 512] - - [526, 94.678] - - - [63744, 512, 1, 512] - - [529, 88.706] - - - [63744, 9216, 1, 512] - - [529, 94.476] - - - [64000, 9216, 1, 512] - - [529, 94.508] - - - [63744, 9473, 1, 512] - - [533, 93.255] - - - [63488, 54272, 1, 512] - - [526, 94.671] - - - [63488, 8961, 1, 512] - - [530, 93.154] - - - [63744, 8961, 1, 512] - - [530, 93.179] - - - [63488, 9216, 1, 512] - - [529, 94.522] - - - [63232, 54272, 1, 512] - - [526, 94.675] - - - [63232, 512, 1, 512] - - [526, 88.121] - - - [63488, 8704, 1, 512] - - [530, 94.471] - - - [63232, 8704, 1, 512] - - [533, 94.433] - - - [63232, 8961, 1, 512] - - [530, 93.186] - - - [62976, 54272, 1, 512] - - [526, 94.671] - - - [62976, 8449, 1, 512] - - [533, 93.069] - - - [63232, 8449, 1, 512] - - [533, 92.991] - - - [62976, 8704, 1, 512] - - [526, 94.482] - - - [62720, 54272, 1, 512] - - [526, 94.675] - - - [62720, 512, 1, 512] - - [529, 87.681] - - - [62720, 8192, 1, 512] - - [530, 94.4] - - - [62976, 8192, 1, 512] - - [529, 94.463] - - - [62720, 8449, 1, 512] - - [533, 93.085] - - - [62464, 54272, 1, 512] - - [526, 94.675] - - - [62464, 7937, 1, 512] - - [532, 92.876] - - - [62720, 7937, 1, 512] - - [529, 92.93] - - - [62464, 8192, 1, 512] - - [533, 94.484] - - - [62208, 54272, 1, 512] - - [526, 94.677] - - - [62208, 512, 1, 512] - - [528, 88.389] - - - [62208, 7680, 1, 512] - - [530, 94.426] - - - [62464, 7680, 1, 512] - - [526, 94.46] - - - [62208, 7937, 1, 512] - - [526, 92.955] - - - [61952, 54272, 1, 512] - - [526, 94.672] - - - [61952, 7425, 1, 512] - - [530, 92.804] - - - [62208, 7425, 1, 512] - - [530, 92.812] - - - [61952, 7680, 1, 512] - - [528, 94.405] - - - [61696, 54272, 1, 512] - - [526, 94.679] - - - [61696, 512, 1, 512] - - [531, 89.754] - - - [61696, 7168, 1, 512] - - [529, 94.409] - - - [61952, 7168, 1, 512] - - [529, 94.403] - - - [61696, 7425, 1, 512] - - [529, 92.793] - - - [61440, 54272, 1, 512] - - [526, 94.674] - - - [61440, 6913, 1, 512] - - [528, 92.711] - - - [61696, 6913, 1, 512] - - [528, 92.728] - - - [61440, 7168, 1, 512] - - [529, 94.38] - - - [61184, 54272, 1, 512] - - [526, 94.667] - - - [61184, 512, 1, 512] - - [533, 89.217] - - - [61184, 6656, 1, 512] - - [528, 94.362] - - - [61440, 6656, 1, 512] - - [528, 94.357] - - - [61184, 6913, 1, 512] - - [528, 92.715] - - - [60928, 54272, 1, 512] - - [526, 94.676] - - - [60928, 6401, 1, 512] - - [530, 92.548] - - - [61184, 6401, 1, 512] - - [530, 92.588] - - - [60928, 6656, 1, 512] - - [533, 94.327] - - - [60672, 54272, 1, 512] - - [526, 94.672] - - - [60672, 512, 1, 512] - - [531, 88.625] - - - [60672, 6144, 1, 512] - - [530, 94.375] - - - [60928, 6144, 1, 512] - - [529, 94.382] - - - [60672, 6401, 1, 512] - - [530, 92.588] - - - [60416, 54272, 1, 512] - - [526, 94.669] - - - [60416, 5889, 1, 512] - - [528, 92.179] - - - [60672, 5889, 1, 512] - - [525, 92.153] - - - [60416, 6144, 1, 512] - - [530, 94.369] - - - [60160, 54272, 1, 512] - - [526, 94.671] - - - [60160, 512, 1, 512] - - [526, 88.114] - - - [60160, 5632, 1, 512] - - [528, 94.275] - - - [60416, 5632, 1, 512] - - [528, 94.24] - - - [60160, 5889, 1, 512] - - [529, 92.208] - - - [59904, 54272, 1, 512] - - [526, 94.664] - - - [60160, 5377, 1, 512] - - [526, 92.227] - - - [59904, 5377, 1, 512] - - [526, 91.976] - - - [59904, 5632, 1, 512] - - [528, 94.314] - - - [59648, 54272, 1, 512] - - [526, 94.674] - - - [59648, 512, 1, 512] - - [533, 87.642] - - - [59648, 5120, 1, 512] - - [529, 94.244] - - - [59904, 5120, 1, 512] - - [528, 94.156] - - - [59648, 5377, 1, 512] - - [526, 92.057] - - - [59392, 54272, 1, 512] - - [526, 94.669] - - - [59392, 4865, 1, 512] - - [528, 91.624] - - - [59648, 4865, 1, 512] - - [530, 91.765] - - - [59392, 5120, 1, 512] - - [529, 94.277] - - - [59136, 54272, 1, 512] - - [526, 94.672] - - - [59136, 512, 1, 512] - - [533, 87.043] - - - [59136, 4608, 1, 512] - - [528, 94.211] - - - [59392, 4608, 1, 512] - - [528, 94.159] - - - [59136, 4865, 1, 512] - - [527, 91.669] - - - [58880, 54272, 1, 512] - - [526, 94.663] - - - [58880, 4353, 1, 512] - - [533, 91.446] - - - [59136, 4353, 1, 512] - - [529, 91.249] - - - [58880, 4608, 1, 512] - - [528, 94.018] - - - [58624, 54272, 1, 512] - - [526, 94.68] - - - [58624, 512, 1, 512] - - [529, 89.747] - - - [58880, 4096, 1, 512] - - [528, 93.862] - - - [58624, 4096, 1, 512] - - [529, 94.055] - - - [58624, 4353, 1, 512] - - [526, 91.5] - - - [58368, 54272, 1, 512] - - [528, 94.667] - - - [58368, 3841, 1, 512] - - [530, 90.983] - - - [58624, 3841, 1, 512] - - [533, 90.971] - - - [58368, 4096, 1, 512] - - [529, 94.179] - - - [58112, 54272, 1, 512] - - [526, 94.673] - - - [58112, 512, 1, 512] - - [528, 89.094] - - - [58112, 3584, 1, 512] - - [530, 93.898] - - - [58368, 3584, 1, 512] - - [526, 93.883] - - - [58112, 3841, 1, 512] - - [530, 91.083] - - - [57856, 54272, 1, 512] - - [526, 94.68] - - - [58112, 3329, 1, 512] - - [530, 90.442] - - - [57856, 3329, 1, 512] - - [527, 90.374] - - - [57856, 3584, 1, 512] - - [526, 94.081] - - - [57600, 54272, 1, 512] - - [526, 94.68] - - - [57600, 512, 1, 512] - - [528, 88.417] - - - [57856, 3072, 1, 512] - - [528, 93.58] - - - [57600, 3072, 1, 512] - - [529, 93.897] - - - [57600, 3329, 1, 512] - - [530, 90.539] - - - [57344, 54272, 1, 512] - - [526, 94.667] - - - [57344, 2817, 1, 512] - - [526, 89.776] - - - [57600, 2817, 1, 512] - - [533, 89.547] - - - [57344, 3072, 1, 512] - - [528, 93.626] - - - [57088, 54272, 1, 512] - - [526, 94.663] - - - [57088, 512, 1, 512] - - [527, 87.967] - - - [57088, 2560, 1, 512] - - [529, 93.398] - - - [57344, 2560, 1, 512] - - [533, 93.668] - - - [57088, 2817, 1, 512] - - [533, 89.6] - - - [56832, 54272, 1, 512] - - [526, 94.675] - - - [56832, 2305, 1, 512] - - [528, 88.705] - - - [57088, 2305, 1, 512] - - [525, 88.836] - - - [56832, 2560, 1, 512] - - [529, 93.812] - - - [56576, 54272, 1, 512] - - [526, 94.672] - - - [56576, 512, 1, 512] - - [531, 87.339] - - - [56576, 2048, 1, 512] - - [529, 92.906] - - - [56832, 2048, 1, 512] - - [528, 93.173] - - - [56576, 2305, 1, 512] - - [528, 88.401] - - - [56320, 54272, 1, 512] - - [526, 94.671] - - - [56576, 1793, 1, 512] - - [526, 87.332] - - - [56320, 1793, 1, 512] - - [534, 87.002] - - - [56320, 2048, 1, 512] - - [528, 93.511] - - - [56064, 54272, 1, 512] - - [526, 94.675] - - - [56064, 512, 1, 512] - - [526, 86.989] - - - [56064, 1536, 1, 512] - - [534, 92.446] - - - [56320, 1536, 1, 512] - - [526, 92.791] - - - [56064, 1793, 1, 512] - - [534, 86.634] - - - [55808, 54272, 1, 512] - - [526, 94.678] - - - [55808, 1281, 1, 512] - - [530, 84.755] - - - [56064, 1281, 1, 512] - - [530, 83.624] - - - [55808, 1536, 1, 512] - - [529, 92.17] - - - [55552, 54272, 1, 512] - - [526, 94.681] - - - [55552, 512, 1, 512] - - [526, 89.669] - - - [55808, 1024, 1, 512] - - [533, 90.754] - - - [55552, 1024, 1, 512] - - [530, 92.224] - - - [55552, 1281, 1, 512] - - [530, 84.432] - - - [55296, 54272, 1, 512] - - [526, 94.662] - - - [55296, 769, 1, 512] - - [525, 78.344] - - - [55552, 769, 1, 512] - - [525, 78.588] - - - [55296, 1024, 1, 512] - - [530, 91.995] - - - [55040, 54272, 1, 512] - - [526, 94.689] - - - [55040, 512, 1, 512] - - [527, 89.119] - - - [55040, 769, 1, 512] - - [525, 77.922] - - - [54784, 54272, 1, 512] - - [526, 94.674] - - - [54784, 257, 1, 512] - - [528, 57.105] - - - [55040, 257, 1, 512] - - [535, 57.324] - - - [54528, 54272, 1, 512] - - [526, 94.67] - - - [54528, 512, 1, 512] - - [534, 88.531] - - - [54528, 257, 1, 512] - - [528, 56.697] - - - [54528, 54017, 1, 512] - - [526, 94.455] - - - [54272, 54017, 1, 512] - - [526, 94.446] - - - [54272, 54272, 1, 512] - - [526, 94.676] - - - [54016, 54017, 1, 512] - - [526, 94.453] - - - [54016, 512, 1, 512] - - [533, 87.779] - - - [54016, 53760, 1, 512] - - [526, 94.69] - - - [54272, 53760, 1, 512] - - [526, 94.685] - - - [53760, 53505, 1, 512] - - [528, 94.444] - - - [54016, 53505, 1, 512] - - [526, 94.461] - - - [53760, 53760, 1, 512] - - [526, 94.691] - - - [53504, 53505, 1, 512] - - [526, 94.454] - - - [53504, 512, 1, 512] - - [527, 87.151] - - - [53504, 53248, 1, 512] - - [526, 94.686] - - - [53760, 53248, 1, 512] - - [526, 94.683] - - - [53248, 52993, 1, 512] - - [528, 94.446] - - - [53504, 52993, 1, 512] - - [526, 94.459] - - - [53248, 53248, 1, 512] - - [526, 94.674] - - - [52992, 52993, 1, 512] - - [528, 94.439] - - - [52992, 512, 1, 512] - - [527, 86.611] - - - [52992, 52736, 1, 512] - - [526, 94.691] - - - [53248, 52736, 1, 512] - - [526, 94.682] - - - [52992, 52481, 1, 512] - - [526, 94.45] - - - [52736, 52481, 1, 512] - - [526, 94.447] - - - [52736, 52736, 1, 512] - - [526, 94.687] - - - [52480, 52481, 1, 512] - - [526, 94.458] - - - [52480, 512, 1, 512] - - [527, 86.157] - - - [52480, 52224, 1, 512] - - [528, 94.682] - - - [52736, 52224, 1, 512] - - [526, 94.687] - - - [52480, 51969, 1, 512] - - [526, 94.464] - - - [52224, 51969, 1, 512] - - [526, 94.485] - - - [52224, 52224, 1, 512] - - [526, 94.691] - - - [51968, 51969, 1, 512] - - [526, 94.472] - - - [51968, 512, 1, 512] - - [527, 89.037] - - - [52224, 51712, 1, 512] - - [526, 94.702] - - - [51968, 51712, 1, 512] - - [526, 94.693] - - - [51968, 51457, 1, 512] - - [526, 94.441] - - - [51712, 51457, 1, 512] - - [526, 94.445] - - - [51712, 51712, 1, 512] - - [526, 94.692] - - - [51456, 51457, 1, 512] - - [528, 94.438] - - - [51456, 512, 1, 512] - - [526, 88.354] - - - [51712, 51200, 1, 512] - - [526, 94.685] - - - [51456, 51200, 1, 512] - - [528, 94.661] - - - [51200, 50945, 1, 512] - - [526, 94.458] - - - [51456, 50945, 1, 512] - - [526, 94.432] - - - [51200, 51200, 1, 512] - - [526, 94.687] - - - [50944, 50945, 1, 512] - - [526, 94.447] - - - [50944, 512, 1, 512] - - [531, 87.667] - - - [50944, 50688, 1, 512] - - [526, 94.682] - - - [51200, 50688, 1, 512] - - [526, 94.686] - - - [50944, 50433, 1, 512] - - [526, 94.452] - - - [50688, 50433, 1, 512] - - [526, 94.444] - - - [50688, 50688, 1, 512] - - [526, 94.69] - - - [50432, 50433, 1, 512] - - [526, 94.434] - - - [50432, 512, 1, 512] - - [526, 87.002] - - - [50432, 50176, 1, 512] - - [528, 94.684] - - - [50688, 50176, 1, 512] - - [526, 94.691] - - - [50176, 49921, 1, 512] - - [526, 94.442] - - - [50432, 49921, 1, 512] - - [528, 94.424] - - - [50176, 50176, 1, 512] - - [526, 94.702] - - - [49920, 49921, 1, 512] - - [526, 94.43] - - - [49920, 512, 1, 512] - - [533, 86.369] - - - [49920, 49664, 1, 512] - - [526, 94.679] - - - [50176, 49664, 1, 512] - - [526, 94.682] - - - [49664, 49409, 1, 512] - - [526, 94.436] - - - [49920, 49409, 1, 512] - - [526, 94.434] - - - [49664, 49664, 1, 512] - - [526, 94.695] - - - [49408, 49409, 1, 512] - - [526, 94.43] - - - [49408, 512, 1, 512] - - [531, 85.685] - - - [49408, 49152, 1, 512] - - [526, 94.678] - - - [49664, 49152, 1, 512] - - [526, 94.693] - - - [49408, 48897, 1, 512] - - [526, 94.41] - - - [49152, 48897, 1, 512] - - [528, 94.409] - - - [49152, 49152, 1, 512] - - [528, 94.65] - - - [48896, 48897, 1, 512] - - [526, 94.446] - - - [48896, 512, 1, 512] - - [532, 88.996] - - - [48896, 48640, 1, 512] - - [526, 94.677] - - - [49152, 48640, 1, 512] - - [528, 94.65] - - - [48640, 48385, 1, 512] - - [526, 94.454] - - - [48896, 48385, 1, 512] - - [526, 94.465] - - - [48640, 48640, 1, 512] - - [526, 94.679] - - - [48384, 48385, 1, 512] - - [526, 94.456] - - - [48384, 512, 1, 512] - - [531, 88.296] - - - [48384, 48128, 1, 512] - - [526, 94.695] - - - [48640, 48128, 1, 512] - - [526, 94.707] - - - [48128, 47873, 1, 512] - - [526, 94.43] - - - [48384, 47873, 1, 512] - - [528, 94.417] - - - [48128, 48128, 1, 512] - - [526, 94.695] - - - [47872, 47873, 1, 512] - - [526, 94.429] - - - [47872, 512, 1, 512] - - [529, 87.451] - - - [47872, 47616, 1, 512] - - [526, 94.695] - - - [48128, 47616, 1, 512] - - [526, 94.686] - - - [47616, 47361, 1, 512] - - [526, 94.431] - - - [47872, 47361, 1, 512] - - [526, 94.435] - - - [47616, 47616, 1, 512] - - [526, 94.709] - - - [47360, 47361, 1, 512] - - [526, 94.439] - - - [47360, 512, 1, 512] - - [528, 86.757] - - - [47360, 47104, 1, 512] - - [526, 94.705] - - - [47616, 47104, 1, 512] - - [526, 94.696] - - - [47104, 46849, 1, 512] - - [526, 94.434] - - - [47360, 46849, 1, 512] - - [526, 94.437] - - - [47104, 47104, 1, 512] - - [526, 94.711] - - - [46848, 46849, 1, 512] - - [526, 94.433] - - - [46848, 512, 1, 512] - - [529, 86.097] - - - [46848, 46592, 1, 512] - - [526, 94.687] - - - [47104, 46592, 1, 512] - - [526, 94.711] - - - [46848, 46337, 1, 512] - - [526, 94.431] - - - [46592, 46337, 1, 512] - - [526, 94.438] - - - [46592, 46592, 1, 512] - - [526, 94.713] - - - [46336, 46337, 1, 512] - - [528, 94.418] - - - [46336, 512, 1, 512] - - [534, 85.52] - - - [46336, 46080, 1, 512] - - [526, 94.689] - - - [46592, 46080, 1, 512] - - [526, 94.711] - - - [46336, 45825, 1, 512] - - [529, 94.41] - - - [46080, 45825, 1, 512] - - [526, 94.424] - - - [46080, 46080, 1, 512] - - [526, 94.698] - - - [45824, 45825, 1, 512] - - [526, 94.406] - - - [45824, 512, 1, 512] - - [534, 84.964] - - - [45824, 45568, 1, 512] - - [526, 94.688] - - - [46080, 45568, 1, 512] - - [526, 94.7] - - - [45568, 45313, 1, 512] - - [526, 94.419] - - - [45824, 45313, 1, 512] - - [526, 94.402] - - - [45568, 45568, 1, 512] - - [526, 94.701] - - - [45312, 45313, 1, 512] - - [526, 94.41] - - - [45312, 512, 1, 512] - - [525, 88.259] - - - [45312, 45056, 1, 512] - - [526, 94.688] - - - [45568, 45056, 1, 512] - - [526, 94.693] - - - [45056, 44801, 1, 512] - - [526, 94.431] - - - [45312, 44801, 1, 512] - - [526, 94.447] - - - [45056, 45056, 1, 512] - - [526, 94.683] - - - [44800, 44801, 1, 512] - - [526, 94.461] - - - [44800, 512, 1, 512] - - [529, 87.289] - - - [44800, 44544, 1, 512] - - [526, 94.706] - - - [45056, 44544, 1, 512] - - [526, 94.719] - - - [44544, 44289, 1, 512] - - [529, 94.396] - - - [44800, 44289, 1, 512] - - [526, 94.413] - - - [44544, 44544, 1, 512] - - [526, 94.708] - - - [44288, 44289, 1, 512] - - [526, 94.41] - - - [44288, 512, 1, 512] - - [528, 86.52] - - - [44288, 44032, 1, 512] - - [526, 94.691] - - - [44544, 44032, 1, 512] - - [526, 94.698] - - - [44032, 43777, 1, 512] - - [528, 94.416] - - - [44288, 43777, 1, 512] - - [528, 94.428] - - - [44032, 44032, 1, 512] - - [526, 94.694] - - - [43776, 43777, 1, 512] - - [528, 94.414] - - - [43776, 512, 1, 512] - - [533, 85.761] - - - [43776, 43520, 1, 512] - - [526, 94.69] - - - [44032, 43520, 1, 512] - - [526, 94.697] - - - [43520, 43265, 1, 512] - - [528, 94.392] - - - [43776, 43265, 1, 512] - - [526, 94.407] - - - [43520, 43520, 1, 512] - - [526, 94.695] - - - [43264, 43265, 1, 512] - - [526, 94.409] - - - [43264, 512, 1, 512] - - [534, 84.985] - - - [43264, 43008, 1, 512] - - [526, 94.689] - - - [43520, 43008, 1, 512] - - [526, 94.703] - - - [43008, 42753, 1, 512] - - [528, 94.39] - - - [43264, 42753, 1, 512] - - [526, 94.397] - - - [43008, 43008, 1, 512] - - [528, 94.68] - - - [42752, 42753, 1, 512] - - [528, 94.385] - - - [42752, 512, 1, 512] - - [526, 84.324] - - - [42752, 42496, 1, 512] - - [526, 94.698] - - - [43008, 42496, 1, 512] - - [528, 94.675] - - - [42496, 42241, 1, 512] - - [526, 94.409] - - - [42752, 42241, 1, 512] - - [526, 94.413] - - - [42496, 42496, 1, 512] - - [526, 94.685] - - - [42240, 42241, 1, 512] - - [526, 94.409] - - - [42240, 512, 1, 512] - - [532, 88.156] - - - [42240, 41984, 1, 512] - - [526, 94.691] - - - [42496, 41984, 1, 512] - - [526, 94.699] - - - [41984, 41729, 1, 512] - - [526, 94.394] - - - [42240, 41729, 1, 512] - - [526, 94.38] - - - [41984, 41984, 1, 512] - - [526, 94.701] - - - [41728, 41729, 1, 512] - - [526, 94.381] - - - [41728, 512, 1, 512] - - [531, 87.102] - - - [41728, 41472, 1, 512] - - [526, 94.677] - - - [41984, 41472, 1, 512] - - [526, 94.693] - - - [41472, 41217, 1, 512] - - [526, 94.408] - - - [41728, 41217, 1, 512] - - [526, 94.398] - - - [41472, 41472, 1, 512] - - [526, 94.686] - - - [41216, 41217, 1, 512] - - [526, 94.383] - - - [41216, 512, 1, 512] - - [535, 86.149] - - - [41216, 40960, 1, 512] - - [526, 94.666] - - - [41472, 40960, 1, 512] - - [526, 94.68] - - - [40960, 40705, 1, 512] - - [530, 94.334] - - - [41216, 40705, 1, 512] - - [526, 94.334] - - - [40960, 40960, 1, 512] - - [526, 94.668] - - - [40704, 40705, 1, 512] - - [528, 94.368] - - - [40704, 512, 1, 512] - - [527, 85.451] - - - [40704, 40448, 1, 512] - - [526, 94.677] - - - [40960, 40448, 1, 512] - - [526, 94.658] - - - [40448, 40193, 1, 512] - - [526, 94.385] - - - [40704, 40193, 1, 512] - - [526, 94.386] - - - [40448, 40448, 1, 512] - - [526, 94.688] - - - [40192, 40193, 1, 512] - - [526, 94.364] - - - [40192, 512, 1, 512] - - [529, 84.574] - - - [40192, 39936, 1, 512] - - [526, 94.668] - - - [40448, 39936, 1, 512] - - [526, 94.679] - - - [39936, 39936, 1, 512] - - [526, 94.685] - - - [40192, 39681, 1, 512] - - [526, 94.377] - - - [39936, 39681, 1, 512] - - [530, 94.362] - - - [39680, 39681, 1, 512] - - [526, 94.378] - - - [39680, 512, 1, 512] - - [526, 84.021] - - - [39680, 39424, 1, 512] - - [526, 94.69] - - - [39936, 39424, 1, 512] - - [526, 94.685] - - - [39424, 39424, 1, 512] - - [526, 94.689] - - - [39680, 39169, 1, 512] - - [528, 94.378] - - - [39424, 39169, 1, 512] - - [528, 94.383] - - - [39168, 39169, 1, 512] - - [528, 94.392] - - - [39168, 512, 1, 512] - - [530, 87.849] - - - [39168, 38912, 1, 512] - - [526, 94.703] - - - [39424, 38912, 1, 512] - - [526, 94.678] - - - [38912, 38912, 1, 512] - - [526, 94.703] - - - [38912, 38657, 1, 512] - - [526, 94.375] - - - [39168, 38657, 1, 512] - - [526, 94.389] - - - [38656, 38657, 1, 512] - - [526, 94.386] - - - [38656, 512, 1, 512] - - [534, 86.904] - - - [38656, 38400, 1, 512] - - [526, 94.686] - - - [38912, 38400, 1, 512] - - [526, 94.688] - - - [38400, 38400, 1, 512] - - [526, 94.68] - - - [38400, 38145, 1, 512] - - [526, 94.349] - - - [38656, 38145, 1, 512] - - [526, 94.371] - - - [38144, 38145, 1, 512] - - [526, 94.356] - - - [38144, 512, 1, 512] - - [530, 86.047] - - - [38144, 37888, 1, 512] - - [526, 94.67] - - - [38400, 37888, 1, 512] - - [528, 94.665] - - - [37888, 37888, 1, 512] - - [526, 94.669] - - - [38144, 37633, 1, 512] - - [526, 94.386] - - - [37888, 37633, 1, 512] - - [526, 94.379] - - - [37632, 37633, 1, 512] - - [526, 94.362] - - - [37632, 512, 1, 512] - - [534, 85.014] - - - [37632, 37376, 1, 512] - - [526, 94.68] - - - [37888, 37376, 1, 512] - - [526, 94.667] - - - [37376, 37376, 1, 512] - - [526, 94.669] - - - [37376, 37121, 1, 512] - - [526, 94.336] - - - [37632, 37121, 1, 512] - - [530, 94.332] - - - [37120, 37121, 1, 512] - - [529, 94.308] - - - [37120, 512, 1, 512] - - [529, 84.315] - - - [37120, 36864, 1, 512] - - [526, 94.671] - - - [37376, 36864, 1, 512] - - [528, 94.663] - - - [36864, 36864, 1, 512] - - [526, 94.659] - - - [36864, 36609, 1, 512] - - [526, 94.344] - - - [37120, 36609, 1, 512] - - [526, 94.325] - - - [36608, 36609, 1, 512] - - [526, 94.336] - - - [36608, 512, 1, 512] - - [531, 83.288] - - - [36608, 36352, 1, 512] - - [526, 94.666] - - - [36864, 36352, 1, 512] - - [526, 94.657] - - - [36352, 36352, 1, 512] - - [526, 94.674] - - - [36352, 36097, 1, 512] - - [526, 94.318] - - - [36608, 36097, 1, 512] - - [526, 94.331] - - - [36096, 36097, 1, 512] - - [526, 94.344] - - - [36096, 512, 1, 512] - - [529, 82.514] - - - [36096, 35840, 1, 512] - - [526, 94.672] - - - [36352, 35840, 1, 512] - - [526, 94.668] - - - [35840, 35840, 1, 512] - - [526, 94.669] - - - [35840, 35585, 1, 512] - - [526, 94.295] - - - [36096, 35585, 1, 512] - - [526, 94.313] - - - [35584, 35585, 1, 512] - - [526, 94.309] - - - [35584, 512, 1, 512] - - [531, 86.503] - - - [35584, 35328, 1, 512] - - [526, 94.66] - - - [35840, 35328, 1, 512] - - [526, 94.666] - - - [35328, 35328, 1, 512] - - [526, 94.67] - - - [35328, 35073, 1, 512] - - [526, 94.336] - - - [35584, 35073, 1, 512] - - [526, 94.319] - - - [35072, 35073, 1, 512] - - [528, 94.3] - - - [35072, 512, 1, 512] - - [532, 85.725] - - - [35072, 34816, 1, 512] - - [526, 94.655] - - - [35328, 34816, 1, 512] - - [526, 94.701] - - - [34816, 34816, 1, 512] - - [526, 94.683] - - - [34816, 34561, 1, 512] - - [528, 94.337] - - - [35072, 34561, 1, 512] - - [526, 94.311] - - - [34560, 34561, 1, 512] - - [528, 94.318] - - - [34560, 512, 1, 512] - - [531, 84.573] - - - [34560, 34304, 1, 512] - - [526, 94.67] - - - [34816, 34304, 1, 512] - - [526, 94.665] - - - [34304, 34304, 1, 512] - - [526, 94.669] - - - [34304, 34049, 1, 512] - - [526, 94.319] - - - [34560, 34049, 1, 512] - - [526, 94.299] - - - [34048, 34049, 1, 512] - - [526, 94.332] - - - [34048, 512, 1, 512] - - [531, 83.631] - - - [34048, 33792, 1, 512] - - [526, 94.649] - - - [34304, 33792, 1, 512] - - [529, 94.641] - - - [33792, 33792, 1, 512] - - [526, 94.682] - - - [33792, 33537, 1, 512] - - [526, 94.292] - - - [34048, 33537, 1, 512] - - [528, 94.296] - - - [33536, 33537, 1, 512] - - [528, 94.29] - - - [33536, 512, 1, 512] - - [532, 82.863] - - - [33536, 33280, 1, 512] - - [526, 94.659] - - - [33792, 33280, 1, 512] - - [526, 94.642] - - - [33280, 33280, 1, 512] - - [526, 94.632] - - - [33536, 33025, 1, 512] - - [526, 94.294] - - - [33280, 33025, 1, 512] - - [526, 94.277] - - - [33024, 33025, 1, 512] - - [525, 94.295] - - - [33024, 512, 1, 512] - - [527, 81.921] - - - [33024, 32768, 1, 512] - - [526, 94.616] - - - [33280, 32768, 1, 512] - - [526, 94.626] - - - [32768, 32768, 1, 512] - - [526, 94.541] - - - [32768, 32513, 1, 512] - - [526, 94.152] - - - [33024, 32513, 1, 512] - - [526, 94.253] - - - [32512, 32513, 1, 512] - - [526, 94.136] - - - [32512, 512, 1, 512] - - [533, 86.122] - - - [32512, 32256, 1, 512] - - [526, 94.558] - - - [32768, 32256, 1, 512] - - [526, 94.559] - - - [32256, 32256, 1, 512] - - [526, 94.633] - - - [32256, 32001, 1, 512] - - [526, 94.235] - - - [32512, 32001, 1, 512] - - [533, 94.097] - - - [32000, 32001, 1, 512] - - [530, 94.223] - - - [32000, 512, 1, 512] - - [525, 85.309] - - - [32000, 31744, 1, 512] - - [526, 94.616] - - - [32256, 31744, 1, 512] - - [526, 94.624] - - - [31744, 31744, 1, 512] - - [526, 94.614] - - - [31744, 31489, 1, 512] - - [526, 94.309] - - - [32000, 31489, 1, 512] - - [526, 94.227] - - - [31488, 31489, 1, 512] - - [526, 94.242] - - - [31488, 512, 1, 512] - - [532, 84.124] - - - [31488, 31232, 1, 512] - - [526, 94.628] - - - [31744, 31232, 1, 512] - - [526, 94.652] - - - [31232, 31232, 1, 512] - - [526, 94.626] - - - [31488, 30977, 1, 512] - - [526, 94.2] - - - [31232, 30977, 1, 512] - - [525, 94.222] - - - [30976, 30977, 1, 512] - - [526, 94.204] - - - [30976, 512, 1, 512] - - [529, 83.075] - - - [30976, 30720, 1, 512] - - [529, 94.595] - - - [31232, 30720, 1, 512] - - [525, 94.617] - - - [30720, 30720, 1, 512] - - [526, 94.629] - - - [30976, 30465, 1, 512] - - [526, 94.25] - - - [30720, 30465, 1, 512] - - [526, 94.212] - - - [30464, 30465, 1, 512] - - [526, 94.221] - - - [30464, 512, 1, 512] - - [534, 82.019] - - - [30464, 30208, 1, 512] - - [526, 94.606] - - - [30720, 30208, 1, 512] - - [526, 94.608] - - - [30208, 30208, 1, 512] - - [526, 94.62] - - - [30208, 29953, 1, 512] - - [528, 94.216] - - - [30464, 29953, 1, 512] - - [528, 94.193] - - - [29952, 29953, 1, 512] - - [525, 94.23] - - - [29952, 512, 1, 512] - - [528, 81.218] - - - [29952, 29696, 1, 512] - - [525, 94.594] - - - [30208, 29696, 1, 512] - - [526, 94.601] - - - [29696, 29696, 1, 512] - - [526, 94.589] - - - [29696, 29441, 1, 512] - - [526, 94.2] - - - [29952, 29441, 1, 512] - - [530, 94.166] - - - [29440, 29441, 1, 512] - - [526, 94.19] - - - [29440, 512, 1, 512] - - [526, 86.048] - - - [29440, 29184, 1, 512] - - [526, 94.608] - - - [29696, 29184, 1, 512] - - [526, 94.597] - - - [29184, 29184, 1, 512] - - [526, 94.59] - - - [29184, 28929, 1, 512] - - [526, 94.153] - - - [29440, 28929, 1, 512] - - [526, 94.163] - - - [28928, 28929, 1, 512] - - [526, 94.179] - - - [28928, 512, 1, 512] - - [528, 84.889] - - - [28928, 28672, 1, 512] - - [526, 94.602] - - - [29184, 28672, 1, 512] - - [526, 94.593] - - - [28672, 28672, 1, 512] - - [526, 94.615] - - - [28928, 28417, 1, 512] - - [525, 94.179] - - - [28672, 28417, 1, 512] - - [526, 94.155] - - - [28416, 28417, 1, 512] - - [526, 94.132] - - - [28416, 512, 1, 512] - - [532, 83.603] - - - [28416, 28160, 1, 512] - - [526, 94.594] - - - [28672, 28160, 1, 512] - - [526, 94.604] - - - [28160, 28160, 1, 512] - - [526, 94.587] - - - [28160, 27905, 1, 512] - - [526, 94.16] - - - [28416, 27905, 1, 512] - - [526, 94.113] - - - [27904, 27905, 1, 512] - - [526, 94.144] - - - [27904, 512, 1, 512] - - [529, 82.349] - - - [27904, 27648, 1, 512] - - [526, 94.577] - - - [28160, 27648, 1, 512] - - [526, 94.585] - - - [27648, 27648, 1, 512] - - [526, 94.59] - - - [27648, 27393, 1, 512] - - [526, 94.136] - - - [27904, 27393, 1, 512] - - [526, 94.097] - - - [27392, 27393, 1, 512] - - [526, 94.105] - - - [27392, 512, 1, 512] - - [534, 81.24] - - - [27392, 27136, 1, 512] - - [526, 94.545] - - - [27648, 27136, 1, 512] - - [526, 94.548] - - - [27136, 27136, 1, 512] - - [526, 94.538] - - - [27392, 26881, 1, 512] - - [526, 94.137] - - - [27136, 26881, 1, 512] - - [525, 94.097] - - - [26880, 26881, 1, 512] - - [526, 94.158] - - - [26880, 512, 1, 512] - - [534, 80.121] - - - [26880, 26624, 1, 512] - - [526, 94.549] - - - [27136, 26624, 1, 512] - - [526, 94.57] - - - [26624, 26624, 1, 512] - - [526, 94.558] - - - [26624, 26369, 1, 512] - - [526, 94.076] - - - [26880, 26369, 1, 512] - - [526, 94.056] - - - [26368, 26369, 1, 512] - - [525, 94.047] - - - [26368, 512, 1, 512] - - [533, 78.934] - - - [26368, 26112, 1, 512] - - [528, 94.521] - - - [26624, 26112, 1, 512] - - [528, 94.542] - - - [26112, 26112, 1, 512] - - [528, 94.578] - - - [26112, 25857, 1, 512] - - [526, 94.12] - - - [26368, 25857, 1, 512] - - [529, 94.035] - - - [25856, 25857, 1, 512] - - [526, 94.066] - - - [25856, 512, 1, 512] - - [525, 84.441] - - - [25856, 25600, 1, 512] - - [526, 94.532] - - - [26112, 25600, 1, 512] - - [529, 94.571] - - - [25600, 25345, 1, 512] - - [526, 94.06] - - - [25856, 25345, 1, 512] - - [528, 94.097] - - - [25344, 25345, 1, 512] - - [525, 94.07] - - - [25344, 512, 1, 512] - - [532, 82.921] - - - [25344, 25088, 1, 512] - - [528, 94.493] - - - [25600, 25088, 1, 512] - - [526, 94.524] - - - [25088, 25088, 1, 512] - - [526, 94.52] - - - [25088, 24833, 1, 512] - - [526, 93.988] - - - [25344, 24833, 1, 512] - - [526, 93.963] - - - [24832, 24833, 1, 512] - - [526, 94.007] - - - [24832, 512, 1, 512] - - [529, 81.674] - - - [24832, 24576, 1, 512] - - [526, 94.457] - - - [25088, 24576, 1, 512] - - [525, 94.477] - - - [24576, 24576, 1, 512] - - [531, 94.49] - - - [24576, 24321, 1, 512] - - [526, 94.006] - - - [24832, 24321, 1, 512] - - [530, 94.033] - - - [24320, 24321, 1, 512] - - [531, 94.031] - - - [24320, 512, 1, 512] - - [533, 80.137] - - - [24320, 24064, 1, 512] - - [528, 94.478] - - - [24576, 24064, 1, 512] - - [526, 94.507] - - - [24064, 24064, 1, 512] - - [528, 94.456] - - - [24064, 23809, 1, 512] - - [525, 93.987] - - - [24320, 23809, 1, 512] - - [525, 93.97] - - - [23808, 23809, 1, 512] - - [525, 94.02] - - - [23808, 512, 1, 512] - - [530, 78.929] - - - [23808, 23552, 1, 512] - - [528, 94.43] - - - [24064, 23552, 1, 512] - - [528, 94.458] - - - [23552, 23552, 1, 512] - - [528, 94.479] - - - [23552, 23297, 1, 512] - - [526, 94.004] - - - [23808, 23297, 1, 512] - - [526, 93.996] - - - [23296, 23297, 1, 512] - - [526, 93.954] - - - [23296, 512, 1, 512] - - [532, 77.727] - - - [23296, 23040, 1, 512] - - [526, 94.438] - - - [23552, 23040, 1, 512] - - [528, 94.433] - - - [23040, 23040, 1, 512] - - [526, 94.434] - - - [23296, 22785, 1, 512] - - [526, 93.89] - - - [23040, 22785, 1, 512] - - [530, 93.882] - - - [22784, 22785, 1, 512] - - [533, 93.833] - - - [22784, 512, 1, 512] - - [528, 83.584] - - - [22784, 22528, 1, 512] - - [529, 94.424] - - - [23040, 22528, 1, 512] - - [526, 94.393] - - - [22528, 22528, 1, 512] - - [526, 94.41] - - - [22528, 22273, 1, 512] - - [525, 93.932] - - - [22784, 22273, 1, 512] - - [525, 93.927] - - - [22272, 22273, 1, 512] - - [526, 93.819] - - - [22272, 512, 1, 512] - - [534, 82.067] - - - [22272, 22016, 1, 512] - - [525, 94.369] - - - [22528, 22016, 1, 512] - - [529, 94.392] - - - [22016, 22016, 1, 512] - - [525, 94.413] - - - [22016, 21761, 1, 512] - - [531, 93.889] - - - [22272, 21761, 1, 512] - - [530, 93.867] - - - [21760, 21761, 1, 512] - - [528, 93.881] - - - [21760, 512, 1, 512] - - [534, 80.442] - - - [21760, 21504, 1, 512] - - [528, 94.445] - - - [22016, 21504, 1, 512] - - [526, 94.41] - - - [21504, 21504, 1, 512] - - [528, 94.391] - - - [21504, 21249, 1, 512] - - [525, 93.805] - - - [21760, 21249, 1, 512] - - [526, 93.784] - - - [21248, 21249, 1, 512] - - [531, 93.778] - - - [21248, 512, 1, 512] - - [530, 78.782] - - - [21248, 20992, 1, 512] - - [531, 94.35] - - - [21504, 20992, 1, 512] - - [526, 94.375] - - - [20992, 20992, 1, 512] - - [531, 94.353] - - - [20992, 20737, 1, 512] - - [525, 93.849] - - - [21248, 20737, 1, 512] - - [525, 93.845] - - - [20736, 20737, 1, 512] - - [525, 93.822] - - - [20736, 512, 1, 512] - - [526, 77.182] - - - [20736, 20480, 1, 512] - - [526, 94.369] - - - [20992, 20480, 1, 512] - - [525, 94.34] - - - [20480, 20480, 1, 512] - - [526, 94.369] - - - [20736, 20225, 1, 512] - - [526, 93.708] - - - [20480, 20225, 1, 512] - - [526, 93.686] - - - [20224, 20225, 1, 512] - - [535, 93.736] - - - [20224, 512, 1, 512] - - [531, 75.727] - - - [20224, 19968, 1, 512] - - [528, 94.267] - - - [20480, 19968, 1, 512] - - [525, 94.335] - - - [19968, 19968, 1, 512] - - [525, 94.256] - - - [19968, 19713, 1, 512] - - [526, 93.808] - - - [20224, 19713, 1, 512] - - [526, 93.809] - - - [19712, 19713, 1, 512] - - [526, 93.789] - - - [19712, 512, 1, 512] - - [528, 76.114] - - - [19712, 19456, 1, 512] - - [526, 94.238] - - - [19968, 19456, 1, 512] - - [526, 94.263] - - - [19456, 19456, 1, 512] - - [528, 94.26] - - - [19456, 19201, 1, 512] - - [531, 93.767] - - - [19712, 19201, 1, 512] - - [525, 93.738] - - - [19200, 19201, 1, 512] - - [531, 93.755] - - - [19200, 512, 1, 512] - - [531, 80.722] - - - [19200, 18944, 1, 512] - - [528, 94.218] - - - [19456, 18944, 1, 512] - - [525, 94.235] - - - [18944, 18944, 1, 512] - - [526, 94.232] - - - [18944, 18689, 1, 512] - - [535, 93.688] - - - [19200, 18689, 1, 512] - - [531, 93.626] - - - [18688, 18689, 1, 512] - - [528, 93.485] - - - [18688, 512, 1, 512] - - [530, 78.968] - - - [18688, 18432, 1, 512] - - [526, 94.304] - - - [18944, 18432, 1, 512] - - [525, 94.236] - - - [18432, 18432, 1, 512] - - [535, 94.288] - - - [18432, 18177, 1, 512] - - [526, 93.584] - - - [18688, 18177, 1, 512] - - [530, 93.509] - - - [18176, 18177, 1, 512] - - [525, 93.562] - - - [18176, 512, 1, 512] - - [526, 77.065] - - - [18176, 17920, 1, 512] - - [526, 94.222] - - - [18432, 17920, 1, 512] - - [526, 94.194] - - - [17920, 17920, 1, 512] - - [531, 94.289] - - - [18176, 17665, 1, 512] - - [525, 93.533] - - - [17920, 17665, 1, 512] - - [528, 93.452] - - - [17664, 17665, 1, 512] - - [525, 93.641] - - - [17664, 512, 1, 512] - - [529, 75.38] - - - [17664, 17408, 1, 512] - - [526, 94.192] - - - [17920, 17408, 1, 512] - - [529, 94.184] - - - [17408, 17408, 1, 512] - - [526, 94.066] - - - [17664, 17153, 1, 512] - - [526, 93.433] - - - [17408, 17153, 1, 512] - - [525, 93.551] - - - [17152, 17153, 1, 512] - - [526, 93.351] - - - [17152, 512, 1, 512] - - [534, 73.623] - - - [17152, 16896, 1, 512] - - [526, 94.027] - - - [17408, 16896, 1, 512] - - [528, 94.264] - - - [16896, 16896, 1, 512] - - [526, 94.124] - - - [16896, 16641, 1, 512] - - [531, 93.435] - - - [17152, 16641, 1, 512] - - [531, 93.606] - - - [16640, 16641, 1, 512] - - [531, 93.594] - - - [16640, 512, 1, 512] - - [531, 72.219] - - - [16640, 16384, 1, 512] - - [535, 94.213] - - - [16896, 16384, 1, 512] - - [525, 94.04] - - - [16384, 16384, 1, 512] - - [531, 94.033] - - - [16384, 16129, 1, 512] - - [526, 93.307] - - - [16640, 16129, 1, 512] - - [525, 93.517] - - - [16128, 16129, 1, 512] - - [526, 93.448] - - - [16128, 512, 1, 512] - - [534, 71.946] - - - [16128, 15872, 1, 512] - - [525, 94.054] - - - [16384, 15872, 1, 512] - - [528, 94.019] - - - [15872, 15872, 1, 512] - - [531, 94.016] - - - [15872, 15617, 1, 512] - - [530, 93.108] - - - [16128, 15617, 1, 512] - - [531, 93.284] - - - [15616, 15617, 1, 512] - - [526, 93.302] - - - [15616, 512, 1, 512] - - [534, 77.185] - - - [15616, 15360, 1, 512] - - [526, 93.913] - - - [15872, 15360, 1, 512] - - [525, 94.042] - - - [15360, 15360, 1, 512] - - [535, 94.159] - - - [15360, 15105, 1, 512] - - [528, 93.29] - - - [15616, 15105, 1, 512] - - [528, 93.014] - - - [15104, 15105, 1, 512] - - [530, 93.072] - - - [15104, 512, 1, 512] - - [534, 75.22] - - - [15104, 14848, 1, 512] - - [525, 94.01] - - - [15360, 14848, 1, 512] - - [528, 93.841] - - - [14848, 14848, 1, 512] - - [531, 93.994] - - - [14848, 14593, 1, 512] - - [525, 93.277] - - - [15104, 14593, 1, 512] - - [525, 93.373] - - - [14592, 14593, 1, 512] - - [525, 93.13] - - - [14592, 512, 1, 512] - - [529, 72.931] - - - [14592, 14336, 1, 512] - - [531, 94.124] - - - [14848, 14336, 1, 512] - - [526, 93.728] - - - [14336, 14336, 1, 512] - - [535, 93.975] - - - [14336, 14081, 1, 512] - - [530, 93.257] - - - [14592, 14081, 1, 512] - - [533, 93.279] - - - [14080, 14081, 1, 512] - - [530, 93.147] - - - [14080, 512, 1, 512] - - [533, 70.826] - - - [14080, 13824, 1, 512] - - [528, 93.673] - - - [14336, 13824, 1, 512] - - [530, 93.805] - - - [13824, 13824, 1, 512] - - [526, 93.597] - - - [14080, 13569, 1, 512] - - [526, 92.783] - - - [13824, 13569, 1, 512] - - [531, 92.684] - - - [13568, 13569, 1, 512] - - [526, 92.652] - - - [13568, 512, 1, 512] - - [526, 68.748] - - - [13568, 13312, 1, 512] - - [535, 93.975] - - - [13824, 13312, 1, 512] - - [529, 94.008] - - - [13312, 13312, 1, 512] - - [526, 93.936] - - - [13568, 13057, 1, 512] - - [525, 93.266] - - - [13312, 13057, 1, 512] - - [525, 93.191] - - - [13056, 13057, 1, 512] - - [525, 93.176] - - - [13056, 512, 1, 512] - - [528, 77.279] - - - [13056, 12800, 1, 512] - - [526, 93.897] - - - [13312, 12800, 1, 512] - - [530, 93.853] - - - [12800, 12800, 1, 512] - - [528, 93.873] - - - [13056, 12545, 1, 512] - - [526, 93.209] - - - [12800, 12545, 1, 512] - - [534, 93.16] - - - [12544, 12545, 1, 512] - - [526, 93.225] - - - [12544, 512, 1, 512] - - [525, 74.496] - - - [12544, 12288, 1, 512] - - [526, 93.322] - - - [12800, 12288, 1, 512] - - [526, 93.943] - - - [12288, 12288, 1, 512] - - [528, 93.344] - - - [12544, 12033, 1, 512] - - [526, 92.26] - - - [12288, 12033, 1, 512] - - [531, 92.315] - - - [12032, 12033, 1, 512] - - [529, 92.415] - - - [12032, 512, 1, 512] - - [526, 71.762] - - - [12032, 11776, 1, 512] - - [526, 93.632] - - - [12288, 11776, 1, 512] - - [529, 93.508] - - - [11776, 11776, 1, 512] - - [526, 93.724] - - - [11776, 11521, 1, 512] - - [528, 93.042] - - - [12032, 11521, 1, 512] - - [528, 92.913] - - - [11520, 11521, 1, 512] - - [531, 92.337] - - - [11520, 512, 1, 512] - - [529, 69.204] - - - [11520, 11264, 1, 512] - - [528, 93.387] - - - [11776, 11264, 1, 512] - - [528, 93.19] - - - [11264, 11264, 1, 512] - - [526, 93.634] - - - [11264, 11009, 1, 512] - - [531, 92.655] - - - [11520, 11009, 1, 512] - - [528, 92.36] - - - [11008, 11009, 1, 512] - - [531, 91.888] - - - [11008, 512, 1, 512] - - [526, 66.447] - - - [11008, 10752, 1, 512] - - [528, 93.451] - - - [11264, 10752, 1, 512] - - [529, 93.137] - - - [10752, 10752, 1, 512] - - [529, 92.763] - - - [11008, 10497, 1, 512] - - [528, 92.452] - - - [10752, 10497, 1, 512] - - [528, 91.785] - - - [10496, 10497, 1, 512] - - [528, 92.107] - - - [10496, 512, 1, 512] - - [532, 63.85] - - - [10496, 10240, 1, 512] - - [529, 92.743] - - - [10752, 10240, 1, 512] - - [529, 93.381] - - - [10240, 10240, 1, 512] - - [529, 93.136] - - - [10496, 9985, 1, 512] - - [525, 92.048] - - - [10240, 9985, 1, 512] - - [525, 92.428] - - - [9984, 9985, 1, 512] - - [525, 91.845] - - - [9984, 512, 1, 512] - - [526, 61.636] - - - [9984, 9728, 1, 512] - - [530, 93.424] - - - [10240, 9728, 1, 512] - - [526, 92.923] - - - [9728, 9728, 1, 512] - - [529, 92.851] - - - [9728, 9473, 1, 512] - - [526, 91.809] - - - [9984, 9473, 1, 512] - - [528, 91.209] - - - [9472, 9473, 1, 512] - - [531, 91.104] - - - [9472, 512, 1, 512] - - [529, 70.49] - - - [9472, 9216, 1, 512] - - [525, 92.188] - - - [9728, 9216, 1, 512] - - [529, 92.76] - - - [9216, 9216, 1, 512] - - [529, 92.879] - - - [9472, 8961, 1, 512] - - [530, 91.579] - - - [9216, 8961, 1, 512] - - [531, 90.9] - - - [8960, 8961, 1, 512] - - [526, 91.833] - - - [8960, 512, 1, 512] - - [534, 67.08] - - - [8960, 8704, 1, 512] - - [526, 92.49] - - - [9216, 8704, 1, 512] - - [530, 93.052] - - - [8704, 8704, 1, 512] - - [530, 91.892] - - - [8704, 8449, 1, 512] - - [533, 91.459] - - - [8960, 8449, 1, 512] - - [533, 91.796] - - - [8448, 8449, 1, 512] - - [525, 90.829] - - - [8448, 512, 1, 512] - - [533, 63.782] - - - [8448, 8192, 1, 512] - - [527, 91.795] - - - [8704, 8192, 1, 512] - - [525, 92.245] - - - [8192, 8192, 1, 512] - - [525, 91.24] - - - [8192, 7937, 1, 512] - - [533, 90.286] - - - [8448, 7937, 1, 512] - - [534, 90.748] - - - [7936, 7937, 1, 512] - - [533, 89.947] - - - [7936, 512, 1, 512] - - [532, 60.28] - - - [7936, 7680, 1, 512] - - [525, 91.483] - - - [8192, 7680, 1, 512] - - [530, 91.946] - - - [7680, 7680, 1, 512] - - [533, 91.0] - - - [7936, 7425, 1, 512] - - [533, 90.653] - - - [7680, 7425, 1, 512] - - [529, 90.15] - - - [7424, 7425, 1, 512] - - [533, 89.713] - - - [7424, 512, 1, 512] - - [535, 57.085] - - - [7424, 7168, 1, 512] - - [527, 91.8] - - - [7680, 7168, 1, 512] - - [532, 92.101] - - - [7168, 7168, 1, 512] - - [526, 91.486] - - - [7168, 6913, 1, 512] - - [525, 89.149] - - - [7424, 6913, 1, 512] - - [525, 89.476] - - - [6912, 6913, 1, 512] - - [525, 88.786] - - - [6912, 512, 1, 512] - - [529, 53.71] - - - [6912, 6656, 1, 512] - - [531, 90.582] - - - [7168, 6656, 1, 512] - - [530, 90.716] - - - [6656, 6656, 1, 512] - - [528, 90.339] - - - [6912, 6401, 1, 512] - - [530, 90.174] - - - [6656, 6401, 1, 512] - - [530, 90.045] - - - [6400, 6401, 1, 512] - - [530, 89.955] - - - [6400, 512, 1, 512] - - [525, 64.491] - - - [6400, 6144, 1, 512] - - [527, 89.954] - - - [6656, 6144, 1, 512] - - [530, 89.943] - - - [6144, 6144, 1, 512] - - [526, 90.026] - - - [6144, 5889, 1, 512] - - [531, 86.586] - - - [6400, 5889, 1, 512] - - [528, 89.492] - - - [5888, 5889, 1, 512] - - [527, 86.392] - - - [5888, 512, 1, 512] - - [532, 59.38] - - - [5888, 5632, 1, 512] - - [533, 90.191] - - - [6144, 5632, 1, 512] - - [530, 89.78] - - - [5632, 5632, 1, 512] - - [527, 90.379] - - - [5632, 5377, 1, 512] - - [526, 87.09] - - - [5888, 5377, 1, 512] - - [534, 87.541] - - - [5376, 5377, 1, 512] - - [525, 87.33] - - - [5376, 512, 1, 512] - - [534, 54.48] - - - [5376, 5120, 1, 512] - - [527, 87.786] - - - [5632, 5120, 1, 512] - - [528, 87.181] - - - [5120, 5120, 1, 512] - - [525, 88.217] - - - [5120, 4865, 1, 512] - - [535, 84.206] - - - [5376, 4865, 1, 512] - - [528, 87.592] - - - [4864, 4865, 1, 512] - - [528, 84.85] - - - [4864, 512, 1, 512] - - [530, 49.776] - - - [4864, 4608, 1, 512] - - [530, 85.828] - - - [5120, 4608, 1, 512] - - [530, 85.161] - - - [4608, 4608, 1, 512] - - [525, 86.829] - - - [4608, 4353, 1, 512] - - [529, 82.328] - - - [4864, 4353, 1, 512] - - [533, 86.538] - - - [4352, 4353, 1, 512] - - [528, 83.512] - - - [4352, 512, 1, 512] - - [525, 44.941] - - - [4352, 4096, 1, 512] - - [533, 85.25] - - - [4608, 4096, 1, 512] - - [528, 83.871] - - - [4096, 4096, 1, 512] - - [528, 87.059] - - - [4096, 3841, 1, 512] - - [530, 82.392] - - - [4352, 3841, 1, 512] - - [525, 82.131] - - - [3840, 3841, 1, 512] - - [525, 79.6] - - - [3840, 512, 1, 512] - - [530, 40.219] - - - [3840, 3584, 1, 512] - - [532, 80.14] - - - [4096, 3584, 1, 512] - - [530, 84.435] - - - [3584, 3584, 1, 512] - - [526, 82.278] - - - [3840, 3329, 1, 512] - - [527, 81.732] - - - [3584, 3329, 1, 512] - - [527, 78.083] - - - [3328, 3329, 1, 512] - - [527, 79.874] - - - [3328, 512, 1, 512] - - [529, 49.505] - - - [3328, 3072, 1, 512] - - [527, 75.355] - - - [3584, 3072, 1, 512] - - [526, 79.779] - - - [3072, 3072, 1, 512] - - [532, 78.257] - - - [63488, 76800, 1, 512] - - [526, 94.691] - - - [64000, 76800, 1, 512] - - [526, 94.695] - - - [64000, 50177, 1, 512] - - [526, 94.455] - - - [63488, 50177, 1, 512] - - [526, 94.454] - - - [63488, 49665, 1, 512] - - [528, 94.417] - - - [62976, 76800, 1, 512] - - [526, 94.699] - - - [62976, 49153, 1, 512] - - [526, 94.431] - - - [63488, 49153, 1, 512] - - [526, 94.426] - - - [62976, 48641, 1, 512] - - [528, 94.42] - - - [62464, 76800, 1, 512] - - [526, 94.694] - - - [62464, 48129, 1, 512] - - [526, 94.418] - - - [62976, 48129, 1, 512] - - [528, 94.418] - - - [62464, 47617, 1, 512] - - [526, 94.422] - - - [61952, 76800, 1, 512] - - [526, 94.691] - - - [61952, 47105, 1, 512] - - [526, 94.428] - - - [62464, 47105, 1, 512] - - [526, 94.417] - - - [61952, 46593, 1, 512] - - [526, 94.423] - - - [61440, 76800, 1, 512] - - [528, 94.689] - - - [61440, 46081, 1, 512] - - [528, 94.414] - - - [61952, 46081, 1, 512] - - [528, 94.395] - - - [61440, 45569, 1, 512] - - [526, 94.408] - - - [60928, 76800, 1, 512] - - [526, 94.7] - - - [60928, 45057, 1, 512] - - [533, 94.389] - - - [61440, 45057, 1, 512] - - [526, 94.405] - - - [60928, 44545, 1, 512] - - [526, 94.378] - - - [60416, 76800, 1, 512] - - [526, 94.684] - - - [60416, 44033, 1, 512] - - [526, 94.388] - - - [60928, 44033, 1, 512] - - [526, 94.393] - - - [60416, 43521, 1, 512] - - [526, 94.397] - - - [59904, 76800, 1, 512] - - [526, 94.693] - - - [59904, 43009, 1, 512] - - [526, 94.379] - - - [60416, 43009, 1, 512] - - [526, 94.375] - - - [59904, 42497, 1, 512] - - [526, 94.344] - - - [59392, 76800, 1, 512] - - [526, 94.691] - - - [59392, 41985, 1, 512] - - [526, 94.377] - - - [59904, 41985, 1, 512] - - [526, 94.381] - - - [59392, 41473, 1, 512] - - [528, 94.363] - - - [58880, 76800, 1, 512] - - [526, 94.695] - - - [58880, 40961, 1, 512] - - [526, 94.346] - - - [59392, 40961, 1, 512] - - [529, 94.352] - - - [58880, 40449, 1, 512] - - [526, 94.349] - - - [58368, 76800, 1, 512] - - [526, 94.701] - - - [58368, 39937, 1, 512] - - [528, 94.355] - - - [58880, 39937, 1, 512] - - [526, 94.356] - - - [58368, 39425, 1, 512] - - [526, 94.38] - - - [57856, 76800, 1, 512] - - [526, 94.689] - - - [58368, 38913, 1, 512] - - [526, 94.332] - - - [57856, 38913, 1, 512] - - [526, 94.368] - - - [57856, 38401, 1, 512] - - [526, 94.355] - - - [57344, 76800, 1, 512] - - [526, 94.687] - - - [57856, 37889, 1, 512] - - [526, 94.326] - - - [57344, 37889, 1, 512] - - [526, 94.339] - - - [57344, 37377, 1, 512] - - [526, 94.319] - - - [56832, 76800, 1, 512] - - [526, 94.69] - - - [57344, 36865, 1, 512] - - [528, 94.345] - - - [56832, 36865, 1, 512] - - [528, 94.354] - - - [56832, 36353, 1, 512] - - [526, 94.325] - - - [56320, 76800, 1, 512] - - [528, 94.663] - - - [56320, 35841, 1, 512] - - [529, 94.288] - - - [56832, 35841, 1, 512] - - [526, 94.347] - - - [56320, 35329, 1, 512] - - [526, 94.293] - - - [55808, 76800, 1, 512] - - [528, 94.696] - - - [55808, 34817, 1, 512] - - [526, 94.285] - - - [56320, 34817, 1, 512] - - [526, 94.303] - - - [55808, 34305, 1, 512] - - [530, 94.285] - - - [55296, 76800, 1, 512] - - [526, 94.689] - - - [55808, 33793, 1, 512] - - [526, 94.272] - - - [55296, 33793, 1, 512] - - [529, 94.274] - - - [55296, 33281, 1, 512] - - [528, 94.289] - - - [54784, 76800, 1, 512] - - [528, 94.69] - - - [55296, 32769, 1, 512] - - [526, 94.259] - - - [54784, 32769, 1, 512] - - [528, 94.259] - - - [54784, 32257, 1, 512] - - [526, 94.304] - - - [54272, 76800, 1, 512] - - [528, 94.689] - - - [54784, 31745, 1, 512] - - [529, 94.246] - - - [54272, 31745, 1, 512] - - [529, 94.25] - - - [54272, 31233, 1, 512] - - [526, 94.255] - - - [53760, 76800, 1, 512] - - [526, 94.691] - - - [54272, 30721, 1, 512] - - [526, 94.24] - - - [53760, 30721, 1, 512] - - [528, 94.239] - - - [53760, 30209, 1, 512] - - [528, 94.206] - - - [53248, 76800, 1, 512] - - [528, 94.682] - - - [53760, 29697, 1, 512] - - [529, 94.237] - - - [53248, 29697, 1, 512] - - [526, 94.213] - - - [53248, 29185, 1, 512] - - [526, 94.208] - - - [52736, 76800, 1, 512] - - [526, 94.691] - - - [53248, 28673, 1, 512] - - [526, 94.23] - - - [52736, 28673, 1, 512] - - [526, 94.232] - - - [52736, 28161, 1, 512] - - [530, 94.206] - - - [52224, 76800, 1, 512] - - [526, 94.685] - - - [52736, 27649, 1, 512] - - [528, 94.217] - - - [52224, 27649, 1, 512] - - [528, 94.203] - - - [52224, 27137, 1, 512] - - [528, 94.154] - - - [51712, 76800, 1, 512] - - [528, 94.68] - - - [52224, 26625, 1, 512] - - [529, 94.179] - - - [51712, 26625, 1, 512] - - [529, 94.16] - - - [51712, 26113, 1, 512] - - [533, 94.107] - - - [51200, 76800, 1, 512] - - [526, 94.677] - - - [50688, 76800, 1, 512] - - [526, 94.679] - - - [50688, 24577, 1, 512] - - [529, 94.107] - - - [51200, 24577, 1, 512] - - [529, 94.071] - - - [50688, 24065, 1, 512] - - [528, 94.11] - - - [50176, 76800, 1, 512] - - [526, 94.678] - - - [50688, 23553, 1, 512] - - [529, 94.091] - - - [50176, 23553, 1, 512] - - [526, 94.065] - - - [50176, 23041, 1, 512] - - [528, 94.094] - - - [49664, 76800, 1, 512] - - [526, 94.683] - - - [49664, 22529, 1, 512] - - [529, 94.063] - - - [50176, 22529, 1, 512] - - [529, 94.05] - - - [49664, 22017, 1, 512] - - [526, 94.001] - - - [49152, 76800, 1, 512] - - [526, 94.681] - - - [49664, 21505, 1, 512] - - [526, 94.053] - - - [49152, 21505, 1, 512] - - [526, 94.027] - - - [49152, 20993, 1, 512] - - [526, 93.978] - - - [48640, 76800, 1, 512] - - [528, 94.673] - - - [49152, 20481, 1, 512] - - [530, 93.999] - - - [48640, 20481, 1, 512] - - [530, 94.011] - - - [48640, 19969, 1, 512] - - [526, 93.969] - - - [48128, 76800, 1, 512] - - [526, 94.682] - - - [48128, 19457, 1, 512] - - [529, 93.945] - - - [48640, 19457, 1, 512] - - [529, 93.948] - - - [48128, 18945, 1, 512] - - [526, 93.924] - - - [47616, 76800, 1, 512] - - [526, 94.678] - - - [48128, 18433, 1, 512] - - [529, 93.931] - - - [47616, 18433, 1, 512] - - [528, 93.933] - - - [47616, 17921, 1, 512] - - [526, 93.91] - - - [47104, 76800, 1, 512] - - [526, 94.674] - - - [47616, 17409, 1, 512] - - [529, 93.891] - - - [47104, 17409, 1, 512] - - [530, 93.848] - - - [47104, 16897, 1, 512] - - [526, 93.816] - - - [46592, 76800, 1, 512] - - [528, 94.667] - - - [46592, 16385, 1, 512] - - [529, 93.794] - - - [47104, 16385, 1, 512] - - [529, 93.805] - - - [46592, 15873, 1, 512] - - [533, 93.752] - - - [46080, 76800, 1, 512] - - [526, 94.68] - - - [46592, 15361, 1, 512] - - [529, 93.788] - - - [46080, 15361, 1, 512] - - [530, 93.799] - - - [46080, 14849, 1, 512] - - [526, 93.697] - - - [45568, 76800, 1, 512] - - [526, 94.679] - - - [46080, 14337, 1, 512] - - [526, 93.708] - - - [45568, 14337, 1, 512] - - [529, 93.67] - - - [45568, 13825, 1, 512] - - [528, 93.666] - - - [45056, 76800, 1, 512] - - [526, 94.673] - - - [45568, 13313, 1, 512] - - [529, 93.625] - - - [45056, 13313, 1, 512] - - [529, 93.58] - - - [45056, 12801, 1, 512] - - [530, 93.524] - - - [44544, 76800, 1, 512] - - [526, 94.67] - - - [45056, 12289, 1, 512] - - [529, 93.524] - - - [44544, 12289, 1, 512] - - [529, 93.496] - - - [44544, 11777, 1, 512] - - [530, 93.391] - - - [44032, 76800, 1, 512] - - [526, 94.674] - - - [44544, 11265, 1, 512] - - [529, 93.408] - - - [44032, 11265, 1, 512] - - [533, 93.443] - - - [44032, 10753, 1, 512] - - [526, 93.305] - - - [43520, 76800, 1, 512] - - [528, 94.673] - - - [44032, 10241, 1, 512] - - [530, 93.233] - - - [43520, 10241, 1, 512] - - [529, 93.316] - - - [43520, 9729, 1, 512] - - [528, 93.112] - - - [43008, 76800, 1, 512] - - [526, 94.676] - - - [43520, 9217, 1, 512] - - [532, 93.136] - - - [43008, 9217, 1, 512] - - [528, 93.069] - - - [43008, 8705, 1, 512] - - [533, 92.936] - - - [42496, 76800, 1, 512] - - [526, 94.675] - - - [43008, 8193, 1, 512] - - [526, 92.926] - - - [42496, 8193, 1, 512] - - [529, 92.848] - - - [42496, 7681, 1, 512] - - [530, 92.706] - - - [41984, 76800, 1, 512] - - [526, 94.666] - - - [42496, 7169, 1, 512] - - [526, 92.553] - - - [41984, 7169, 1, 512] - - [529, 92.493] - - - [41984, 6657, 1, 512] - - [527, 92.242] - - - [41472, 76800, 1, 512] - - [526, 94.674] - - - [41984, 6145, 1, 512] - - [529, 92.232] - - - [41472, 6145, 1, 512] - - [529, 92.255] - - - [41472, 5633, 1, 512] - - [533, 92.14] - - - [40960, 76800, 1, 512] - - [528, 94.664] - - - [41472, 5121, 1, 512] - - [529, 91.575] - - - [40960, 5121, 1, 512] - - [529, 91.75] - - - [40960, 4609, 1, 512] - - [532, 91.493] - - - [40448, 76800, 1, 512] - - [528, 94.671] - - - [40960, 4097, 1, 512] - - [529, 90.9] - - - [40448, 4097, 1, 512] - - [529, 90.742] - - - [40448, 3585, 1, 512] - - [526, 90.585] - - - [39936, 76800, 1, 512] - - [526, 94.662] - - - [40448, 3073, 1, 512] - - [529, 89.528] - - - [39936, 3073, 1, 512] - - [529, 89.59] - - - [39936, 2561, 1, 512] - - [530, 88.347] - - - [39424, 76800, 1, 512] - - [526, 94.645] - - - [39424, 2049, 1, 512] - - [529, 86.903] - - - [39936, 2049, 1, 512] - - [529, 87.892] - - - [39424, 1537, 1, 512] - - [535, 84.4] - - - [38912, 76800, 1, 512] - - [528, 94.661] - - - [39424, 1025, 1, 512] - - [529, 80.04] - - - [38912, 1025, 1, 512] - - [529, 81.535] - - - [38912, 513, 1, 512] - - [530, 70.685] - - - [38400, 76800, 1, 512] - - [528, 94.655] - - - [89600, 89089, 1, 512] - - [528, 94.385] - - - [89088, 88577, 1, 512] - - [528, 94.403] - - - [88576, 88065, 1, 512] - - [528, 94.405] - - - [88064, 87553, 1, 512] - - [528, 94.404] - - - [87552, 87041, 1, 512] - - [528, 94.39] - - - [87040, 86529, 1, 512] - - [528, 94.391] - - - [86528, 86017, 1, 512] - - [528, 94.385] - - - [86016, 85505, 1, 512] - - [533, 94.406] - - - [85504, 84993, 1, 512] - - [528, 94.399] - - - [84992, 84481, 1, 512] - - [528, 94.388] - - - [84480, 83969, 1, 512] - - [528, 94.382] - - - [83968, 83457, 1, 512] - - [528, 94.377] - - - [83456, 82945, 1, 512] - - [528, 94.413] - - - [82944, 82433, 1, 512] - - [533, 94.381] - - - [82432, 81921, 1, 512] - - [528, 94.38] - - - [81920, 81409, 1, 512] - - [528, 94.394] - - - [81408, 80897, 1, 512] - - [528, 94.375] - - - [80896, 80385, 1, 512] - - [528, 94.357] - - - [80384, 79873, 1, 512] - - [528, 94.38] - - - [79872, 79361, 1, 512] - - [530, 94.346] - - - [79360, 78849, 1, 512] - - [528, 94.378] - - - [78848, 78337, 1, 512] - - [528, 94.381] - - - [78336, 77825, 1, 512] - - [528, 94.36] - - - [77824, 77313, 1, 512] - - [528, 94.366] - - - [77312, 76801, 1, 512] - - [528, 94.379] - - - [76800, 76289, 1, 512] - - [528, 94.364] - - - [76288, 75777, 1, 512] - - [533, 94.354] - - - [75776, 75265, 1, 512] - - [528, 94.364] - - - [75264, 74753, 1, 512] - - [528, 94.378] - - - [74752, 74241, 1, 512] - - [528, 94.351] - - - [74240, 73729, 1, 512] - - [528, 94.38] - - - [73728, 73217, 1, 512] - - [533, 94.362] - - - [73216, 72705, 1, 512] - - [528, 94.34] - - - [72704, 72193, 1, 512] - - [528, 94.365] - - - [72192, 71681, 1, 512] - - [528, 94.338] - - - [71680, 71169, 1, 512] - - [533, 94.376] - - - [71168, 70657, 1, 512] - - [533, 94.341] - - - [70656, 70145, 1, 512] - - [533, 94.336] - - - [70144, 69633, 1, 512] - - [528, 94.334] - - - [69632, 69121, 1, 512] - - [528, 94.35] - - - [69120, 68609, 1, 512] - - [533, 94.344] - - - [68608, 68097, 1, 512] - - [533, 94.351] - - - [68096, 67585, 1, 512] - - [533, 94.346] - - - [67584, 67073, 1, 512] - - [528, 94.329] - - - [67072, 66561, 1, 512] - - [533, 94.306] - - - [66560, 66049, 1, 512] - - [533, 94.305] - - - [66048, 65537, 1, 512] - - [528, 94.321] - - - [65536, 65025, 1, 512] - - [528, 94.328] - - - [65024, 64513, 1, 512] - - [528, 94.332] - - - [64512, 64001, 1, 512] - - [528, 94.319] - - - [64000, 63489, 1, 512] - - [528, 94.302] - - - [63488, 62977, 1, 512] - - [533, 94.337] - - - [62976, 62465, 1, 512] - - [533, 94.328] - - - [62464, 61953, 1, 512] - - [533, 94.339] - - - [61952, 61441, 1, 512] - - [533, 94.295] - - - [61440, 60929, 1, 512] - - [528, 94.284] - - - [60928, 60417, 1, 512] - - [528, 94.309] - - - [60416, 59905, 1, 512] - - [528, 94.319] - - - [59904, 59393, 1, 512] - - [533, 94.277] - - - [59392, 58881, 1, 512] - - [533, 94.275] - - - [58880, 58369, 1, 512] - - [528, 94.303] - - - [58368, 57857, 1, 512] - - [528, 94.293] - - - [57856, 57345, 1, 512] - - [533, 94.298] - - - [57344, 56833, 1, 512] - - [533, 94.261] - - - [56832, 56321, 1, 512] - - [528, 94.282] - - - [56320, 55809, 1, 512] - - [528, 94.286] - - - [55808, 55297, 1, 512] - - [528, 94.29] - - - [55296, 54785, 1, 512] - - [533, 94.267] - - - [54784, 54273, 1, 512] - - [533, 94.229] - - - [54272, 53761, 1, 512] - - [528, 94.262] - - - [53760, 53249, 1, 512] - - [533, 94.275] - - - [53248, 52737, 1, 512] - - [528, 94.24] - - - [52736, 52225, 1, 512] - - [533, 94.251] - - - [52224, 51713, 1, 512] - - [533, 94.258] - - - [51712, 51201, 1, 512] - - [528, 94.242] - - - [51200, 50689, 1, 512] - - [528, 94.25] - - - [50688, 50177, 1, 512] - - [533, 94.225] - - - [50176, 49665, 1, 512] - - [528, 94.234] - - - [49664, 49153, 1, 512] - - [528, 94.228] - - - [49152, 48641, 1, 512] - - [528, 94.213] - - - [48640, 48129, 1, 512] - - [528, 94.214] - - - [48128, 47617, 1, 512] - - [528, 94.194] - - - [47616, 47105, 1, 512] - - [528, 94.224] - - - [47104, 46593, 1, 512] - - [533, 94.21] - - - [46592, 46081, 1, 512] - - [533, 94.2] - - - [46080, 45569, 1, 512] - - [528, 94.182] - - - [45568, 45057, 1, 512] - - [533, 94.22] - - - [45056, 44545, 1, 512] - - [528, 94.169] - - - [44544, 44033, 1, 512] - - [528, 94.14] - - - [44032, 43521, 1, 512] - - [533, 94.168] - - - [43520, 43009, 1, 512] - - [533, 94.152] - - - [43008, 42497, 1, 512] - - [533, 94.156] - - - [42496, 41985, 1, 512] - - [533, 94.1] - - - [41984, 41473, 1, 512] - - [528, 94.121] - - - [41472, 40961, 1, 512] - - [533, 94.096] - - - [40960, 40449, 1, 512] - - [528, 94.081] - - - [40448, 39937, 1, 512] - - [528, 94.091] - - - [39936, 39425, 1, 512] - - [533, 94.082] - - - [39424, 38913, 1, 512] - - [528, 94.035] - - - [38912, 38401, 1, 512] - - [530, 94.023] - - - [38400, 37889, 1, 512] - - [533, 94.088] - - - [37888, 37377, 1, 512] - - [533, 94.073] - - - [37376, 36865, 1, 512] - - [528, 94.086] - - - [36864, 36353, 1, 512] - - [528, 94.014] - - - [36352, 35841, 1, 512] - - [533, 94.032] - - - [35840, 35329, 1, 512] - - [528, 94.02] - - - [35328, 34817, 1, 512] - - [533, 93.985] - - - [34816, 34305, 1, 512] - - [533, 93.962] - - - [34304, 33793, 1, 512] - - [533, 94.035] - - - [33792, 33281, 1, 512] - - [533, 93.969] - - - [33280, 32769, 1, 512] - - [530, 93.912] - - - [32768, 32257, 1, 512] - - [533, 93.944] - - - [32256, 31745, 1, 512] - - [533, 93.939] - - - [31744, 31233, 1, 512] - - [527, 93.912] - - - [31232, 30721, 1, 512] - - [528, 93.895] - - - [30720, 30209, 1, 512] - - [533, 93.874] - - - [30208, 29697, 1, 512] - - [533, 93.857] - - - [29696, 29185, 1, 512] - - [533, 93.82] - - - [29184, 28673, 1, 512] - - [528, 93.834] - - - [28672, 28161, 1, 512] - - [530, 93.794] - - - [28160, 27649, 1, 512] - - [533, 93.767] - - - [27648, 27137, 1, 512] - - [528, 93.713] - - - [27136, 26625, 1, 512] - - [533, 93.722] - - - [26624, 26113, 1, 512] - - [533, 93.71] - - - [26112, 25601, 1, 512] - - [527, 93.649] - - - [25600, 25089, 1, 512] - - [534, 93.648] - - - [25088, 24577, 1, 512] - - [529, 93.596] - - - [24576, 24065, 1, 512] - - [528, 93.588] - - - [24064, 23553, 1, 512] - - [529, 93.598] - - - [23552, 23041, 1, 512] - - [533, 93.471] - - - [23040, 22529, 1, 512] - - [533, 93.583] - - - [22528, 22017, 1, 512] - - [533, 93.458] - - - [22016, 21505, 1, 512] - - [526, 93.446] - - - [21504, 20993, 1, 512] - - [527, 93.374] - - - [20992, 20481, 1, 512] - - [530, 93.405] - - - [20480, 19969, 1, 512] - - [527, 93.367] - - - [19968, 19457, 1, 512] - - [528, 93.233] - - - [19456, 18945, 1, 512] - - [530, 93.202] - - - [18944, 18433, 1, 512] - - [528, 93.201] - - - [18432, 17921, 1, 512] - - [530, 93.115] - - - [17920, 17409, 1, 512] - - [533, 92.994] - - - [17408, 16897, 1, 512] - - [533, 93.041] - - - [16896, 16385, 1, 512] - - [532, 92.921] - - - [16384, 15873, 1, 512] - - [528, 92.91] - - - [15872, 15361, 1, 512] - - [530, 92.787] - - - [15360, 14849, 1, 512] - - [527, 92.693] - - - [14848, 14337, 1, 512] - - [529, 92.662] - - - [14336, 13825, 1, 512] - - [528, 92.681] - - - [13824, 13313, 1, 512] - - [529, 92.402] - - - [13312, 12801, 1, 512] - - [530, 92.384] - - - [12800, 12289, 1, 512] - - [532, 92.292] - - - [12288, 11777, 1, 512] - - [527, 92.046] - - - [11776, 11265, 1, 512] - - [533, 92.001] - - - [11264, 10753, 1, 512] - - [532, 91.91] - - - [10752, 10241, 1, 512] - - [530, 91.883] - - - [10240, 9729, 1, 512] - - [527, 91.138] - - - [9728, 9217, 1, 512] - - [529, 91.205] - - - [9216, 8705, 1, 512] - - [532, 90.541] - - - [8704, 8193, 1, 512] - - [529, 90.392] - - - [8192, 7681, 1, 512] - - [530, 89.827] - - - [7680, 7169, 1, 512] - - [535, 89.697] - - - [7168, 6657, 1, 512] - - [527, 88.368] - - - [6656, 6145, 1, 512] - - [532, 88.868] - - - [6144, 5633, 1, 512] - - [535, 86.592] - - - [5632, 5121, 1, 512] - - [535, 86.576] - - - [5120, 4609, 1, 512] - - [535, 84.11] - - - [4608, 4097, 1, 512] - - [529, 83.015] - - - [4096, 3585, 1, 512] - - [535, 78.76] - - - [3584, 3073, 1, 512] - - [535, 78.362] -- null diff --git a/library/src/blas3/Tensile/Logic/asm_full/aldebaran/aldebaran_Cijk_Ailk_Bjlk_DB_GB.yaml b/library/src/blas3/Tensile/Logic/asm_full/aldebaran/aldebaran_Cijk_Ailk_Bjlk_DB_GB.yaml deleted file mode 100644 index 40387ad..0000000 --- a/library/src/blas3/Tensile/Logic/asm_full/aldebaran/aldebaran_Cijk_Ailk_Bjlk_DB_GB.yaml +++ /dev/null @@ -1,152505 +0,0 @@ -- {MinimumRequiredVersion: 4.8.1} -- aldebaran -- gfx90a -- [Device 0050, Device 0051, Device 0052, Device 0054, Device 0062, Device 7400, Device - 740c] -- AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false -- - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 0 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x064x04_PLR1_TT04_04_WG16_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: 1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 512 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 0 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 1 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 1 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 2 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 4 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 4 - NumGlobalWriteVectorsPerThread: 2 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 4 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 2 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 2] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: 1 - _VectorStore: 1 - _staggerStrideShift: 1 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 16 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsOffsetA: 0 - LdsOffsetB: 512 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 16 - LoopTail: true - LoopUnroll: 16 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 3 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 16 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsOffsetA: 0 - LdsOffsetB: 512 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 4 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 4 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 16 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsOffsetA: 0 - LdsOffsetB: 512 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 5 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 512 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 6 - StaggerU: 32 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 1 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 512 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 4 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 0 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 7 - StaggerU: 32 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 1 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 8 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_SE_FL0_WGM11 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 9 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_SE_FL0_WGM8 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - FractionalLoad: 1 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 10 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_SE_FL1_WGM8 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 11 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x4_SE_FL0_WGM11 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 3 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 12 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG16_16_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 13 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG8_32_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 4 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 14 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SE_1LDSB1_EPS1_IU4_NLCA1_PGR1_SIA2_TT8_32_WG16_16_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 15 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB0_EPS1_IU1_NLCA1_PGR1_SIA3_TT8_32_WG8_32_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 16 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 1 - LSPB: 2 - LVCA: 64 - LVCB: 32 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 17 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_64_WG64_4_1_WGM11 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 18 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 19 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 20 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM10 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 10 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: -1 - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {3: 512} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {0: 128, 1: 128} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: DGEMM_Aldebaran_PKFixedAtomic512Latest - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: Branch - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 1 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 21 - SolutionNameMin: DGEMM_Aldebaran_PKFixedAtomic512Latest - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: -1 - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {3: 512} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {0: 128, 1: 128} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: DGEMM_Aldebaran_PKFixedAtomic512_104 - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: Branch - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 1 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 22 - SolutionNameMin: DGEMM_Aldebaran_PKFixedAtomic512_104 - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 23 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 24 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 25 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 26 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 27 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 28 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 29 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 30 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 31 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 32 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 33 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 34 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 35 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 36 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 37 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 38 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 39 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 40 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 41 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 42 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 43 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 44 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 45 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 46 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 47 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 48 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 49 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 50 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 51 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 52 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 53 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 54 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 55 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 56 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 57 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 58 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 59 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 60 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 61 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 62 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 63 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 64 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 65 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 66 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 67 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 68 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 69 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 70 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 71 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 72 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 73 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 74 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 75 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 76 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 77 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 78 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 79 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 80 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 81 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 82 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 83 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 84 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 85 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 86 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 87 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 88 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 89 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 90 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 91 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 92 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 93 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 94 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 95 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 96 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 97 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 98 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 99 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 100 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 101 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 102 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 103 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 104 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 105 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT8_32_WG16_16_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 106 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 107 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 108 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 109 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 110 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 111 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 112 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 113 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 114 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 115 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 116 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 117 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 118 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 119 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 120 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 121 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 122 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 123 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 124 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 125 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 126 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 127 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 128 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 129 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 130 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 131 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 132 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 133 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 134 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 135 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 136 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS128_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 137 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 138 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 139 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 140 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 141 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 142 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 143 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 144 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 145 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 146 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 147 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 148 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 149 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 150 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 151 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 152 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 153 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 154 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 155 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 156 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 157 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 158 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU32_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 159 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 160 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 161 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 162 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 163 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 164 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 165 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 166 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 167 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 168 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 169 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 170 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 171 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 172 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 173 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 174 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 175 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 176 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 177 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 178 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 179 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 180 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 181 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 182 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 183 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 184 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 185 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 186 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 187 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 188 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 189 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 190 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 191 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 192 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 193 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 194 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 195 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 196 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 197 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 198 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 199 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 200 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 201 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 202 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT4_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 203 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT4_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 204 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 205 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS128_TT4_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 16 - LSCB: 16 - LSPA: 4 - LSPB: 4 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 256 - LdsNumElementsAlignedA: 128 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 256 - LdsOffsetB: 128 - LdsOffsetB_Blk: 384 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 1] - MIWaveTile: [1, 1] - MIWaveTileA: 1 - MIWaveTileB: 1 - MacroTile0: 16 - MacroTile1: 16 - MacroTileA: 16 - MacroTileB: 16 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 4 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 64 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 206 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT16x16x8_MI16x16x4x1_SN_TT1_16_WG16_4_1 - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 16 - SubGroupA: 4 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [1, 16] - ThreadTile0: 4 - ThreadTile1: 1 - ThreadTileA: 4 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 207 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 208 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 209 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 210 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 211 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 212 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 213 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 214 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 215 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 216 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 217 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 218 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 219 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 220 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 221 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 222 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 223 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 224 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 225 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_PLR3_SU0_SUS0_SSO8_TT4_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 226 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_PLR3_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 227 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_PLR5_SU0_SUS0_SSO8_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 1 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 228 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_PLR3_SU0_SUS0_SSO4_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 229 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 8 - LSPB: 2 - LVCA: 32 - LVCB: 128 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 768 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 768 - LdsOffsetB_Blk: 2816 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 230 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x8_MI16x16x4x1_SN_PLR3_SU0_SUS0_SSO4_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 2] - MIWaveTileA: 3 - MIWaveTileB: 2 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 3 - NumLoadsB: 2 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 231 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 32] - ThreadTile0: 12 - ThreadTile1: 2 - ThreadTileA: 12 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 232 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_PLR5_SU32_SUS128_SSO4_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 233 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: 1 - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 234 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 235 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO8_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 236 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 237 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 238 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: 1 - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 239 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 240 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_PLR5_SU32_SUS128_SSO8_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: 1 - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 241 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 242 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU0_SUS0_SSO4_TT2_96_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 243 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW2_NLCA3_NLCB1_PLR5_SU32_SUS128_SSO4_TT3_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 244 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 245 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT2_64_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 246 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT4_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 247 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 248 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 1 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 249 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 250 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 1 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 251 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT2_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 2 - LSPB: 4 - LVCA: 128 - LVCB: 64 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 8 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 252 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 253 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 254 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 255 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 256 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 257 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW2_NLCA3_NLCB1_PLR5_SU0_SUS0_SSO4_TT3_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 258 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU0_SUS0_SSO4_TT2_96_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 259 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW2_NLCA3_NLCB1_PLR5_SU0_SUS0_SSO8_TT3_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 260 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO8_TT4_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 261 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 262 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_32_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 263 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO8_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 264 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 265 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO8_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 266 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT2_64_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 267 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT4_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 268 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 2 - LVCA: 64 - LVCB: 128 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 269 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT4_32_WG16_16_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 16 - LSPB: 16 - LVCA: 16 - LVCB: 16 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2560 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2560 - LdsOffsetB_Blk: 6656 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [5, 3] - MIWaveTileA: 5 - MIWaveTileB: 3 - MacroTile0: 160 - MacroTile1: 96 - MacroTileA: 160 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 60 - NumGlobalWriteVectorsPerThread: 60 - NumLoadsA: 5 - NumLoadsB: 3 - NumLoadsCoalescedA: 5 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 270 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT160x96x16_MI16x16x4x1_SN_GRVW2_NLCA5_NLCB3_PLR5_SU32_SUS128_SSO8_TT5_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [5, 48] - ThreadTile0: 20 - ThreadTile1: 3 - ThreadTileA: 20 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 271 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_32_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 272 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT2_32_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 273 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU32_SUS256_SSO8_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 274 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS256_SSO4_TT2_48_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 275 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO4_TT4_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 276 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT2_32_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 277 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 278 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU32_SUS256_SSO4_TT4_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 279 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO4_TT4_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 280 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS256_SSO4_TT4_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 2 - LSPB: 4 - LVCA: 128 - LVCB: 64 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 8 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 281 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT2_64_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 282 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO4_TT4_48_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2560 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2560 - LdsOffsetB_Blk: 6656 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [5, 3] - MIWaveTileA: 5 - MIWaveTileB: 3 - MacroTile0: 160 - MacroTile1: 96 - MacroTileA: 160 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 60 - NumGlobalWriteVectorsPerThread: 60 - NumLoadsA: 10 - NumLoadsB: 6 - NumLoadsCoalescedA: 5 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 283 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT160x96x16_MI16x16x4x1_SN_GRVW1_NLCA5_NLCB3_PLR5_SU0_SUS0_SSO4_TT5_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [5, 48] - ThreadTile0: 20 - ThreadTile1: 3 - ThreadTileA: 20 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 284 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT4_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 1 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 285 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1280 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1280 - LdsOffsetB_Blk: 3328 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [5, 3] - MIWaveTileA: 5 - MIWaveTileB: 3 - MacroTile0: 160 - MacroTile1: 96 - MacroTileA: 160 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 60 - NumGlobalWriteVectorsPerThread: 60 - NumLoadsA: 5 - NumLoadsB: 3 - NumLoadsCoalescedA: 5 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 286 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT160x96x8_MI16x16x4x1_SN_GRVW1_NLCA5_NLCB3_PLR3_SU32_SUS128_SSO8_TT5_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [5, 48] - ThreadTile0: 20 - ThreadTile1: 3 - ThreadTileA: 20 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 287 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 288 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 8 - LSPB: 2 - LVCA: 32 - LVCB: 128 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 768 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 768 - LdsOffsetB_Blk: 2816 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 289 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x8_MI16x16x4x1_SN_GRVW1_NLCA3_NLCB1_PLR3_SU0_SUS0_SSO4_TT3_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 290 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT4_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 291 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO8_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 2 - LSPB: 4 - LVCA: 128 - LVCB: 64 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 8 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 292 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 293 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 294 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT8_32_WG16_16_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 295 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 296 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT8_32_WG16_16_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 297 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU32_SUS128_SSO8_TT8_32_WG16_16_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 2 - LVCA: 64 - LVCB: 128 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 298 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT2_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 299 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB3_PLR3_SU32_SUS128_SSO8_TT4_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [6, 1] - MIWaveTileA: 6 - MIWaveTileB: 1 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 6 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 300 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x16_MI16x16x4x1_SN_GRVW1_NLCA3_NLCB1_PLR5_SU0_SUS0_SSO4_TT6_16_WG16_16_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [6, 16] - ThreadTile0: 24 - ThreadTile1: 1 - ThreadTileA: 24 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 301 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_64_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 2 - LVCA: 64 - LVCB: 128 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 302 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 303 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 8 - LSPB: 2 - LVCA: 32 - LVCB: 128 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 6 - NumLoadsB: 8 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 304 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW1_NLCA3_NLCB1_PLR5_SU0_SUS0_SSO8_TT3_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 305 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT8_32_WG16_16_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2560 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2560 - LdsOffsetB_Blk: 6656 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [5, 3] - MIWaveTileA: 5 - MIWaveTileB: 3 - MacroTile0: 160 - MacroTile1: 96 - MacroTileA: 160 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 60 - NumGlobalWriteVectorsPerThread: 60 - NumLoadsA: 10 - NumLoadsB: 6 - NumLoadsCoalescedA: 5 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 306 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT160x96x16_MI16x16x4x1_SN_GRVW1_NLCA5_NLCB3_PLR5_SU32_SUS256_SSO4_TT5_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [5, 48] - ThreadTile0: 20 - ThreadTile1: 3 - ThreadTileA: 20 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 307 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT8_32_WG16_16_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 308 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_32_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 309 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT4_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 310 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 311 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS256_SSO4_TT4_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2560 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2560 - LdsOffsetB_Blk: 6656 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [5, 3] - MIWaveTileA: 5 - MIWaveTileB: 3 - MacroTile0: 160 - MacroTile1: 96 - MacroTileA: 160 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 60 - NumGlobalWriteVectorsPerThread: 60 - NumLoadsA: 10 - NumLoadsB: 6 - NumLoadsCoalescedA: 5 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 312 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT160x96x16_MI16x16x4x1_SN_GRVW1_NLCA5_NLCB3_PLR5_SU32_SUS128_SSO8_TT5_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [5, 48] - ThreadTile0: 20 - ThreadTile1: 3 - ThreadTileA: 20 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 2 - LVCA: 64 - LVCB: 128 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 313 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT4_32_WG16_16_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 6 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 314 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO8_TT2_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 315 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 2 - LSPB: 4 - LVCA: 128 - LVCB: 64 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 8 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 316 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO8_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [1, 6] - MIWaveTileA: 1 - MIWaveTileB: 6 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 6 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 317 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB3_PLR5_SU32_SUS256_SSO4_TT1_96_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [1, 96] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 2 - LVCA: 64 - LVCB: 128 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 318 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU32_SUS128_SSO4_TT2_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 319 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT8_32_WG16_16_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 320 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT8_32_WG16_16_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 321 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT8_32_WG16_16_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 8 - NumLoadsB: 6 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 322 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO4_TT2_96_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 323 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT2_32_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 8 - NumLoadsB: 6 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 324 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO4_TT4_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 8 - LSPB: 2 - LVCA: 32 - LVCB: 128 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 6 - NumLoadsB: 8 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 325 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW1_NLCA3_NLCB1_PLR5_SU32_SUS128_SSO4_TT3_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 326 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_32_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 327 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO8_TT2_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 328 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT2_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 329 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO8_TT2_32_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 330 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT2_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 331 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 332 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 333 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_48_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 334 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_48_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 335 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS256_SVW2_TT2_48_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 336 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 337 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 338 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 339 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 340 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 341 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 342 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 343 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 344 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 345 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_48_VW2_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 346 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 347 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 348 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 349 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU32_SUS256_SVW2_TT2_48_VW2_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 350 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 351 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 352 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 353 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 354 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 355 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 356 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 357 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 358 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 359 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 360 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 361 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 362 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 363 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 364 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 365 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 366 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 367 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 368 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 369 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 370 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 371 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 372 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 373 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 374 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 375 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 376 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 377 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 378 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 379 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 380 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 381 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 382 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 383 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 384 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 385 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 386 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 387 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 388 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 389 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 390 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 391 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 392 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 393 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 394 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 395 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 396 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 397 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 398 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 399 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 400 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 401 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 402 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 403 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 404 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 405 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_48_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 406 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 407 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 408 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_64_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 409 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 410 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 411 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 412 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 2] - MIWaveTileA: 3 - MIWaveTileB: 2 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 3 - NumLoadsB: 2 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 413 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU0_SUS0_SVW1_TT3_32_VW1_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 32] - ThreadTile0: 12 - ThreadTile1: 2 - ThreadTileA: 12 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 2] - MIWaveTileA: 3 - MIWaveTileB: 2 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 3 - NumLoadsB: 2 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 414 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU32_SUS256_SVW1_TT3_32_VW1_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 32] - ThreadTile0: 12 - ThreadTile1: 2 - ThreadTileA: 12 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 415 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 8 - LSPB: 2 - LVCA: 32 - LVCB: 128 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 768 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 768 - LdsOffsetB_Blk: 2816 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 416 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x8_MI16x16x4x1_SN_AMAS0_GRVW1_NEPBS2_NLCA3_NLCB1_PLR3_SU32_SUS256_SVW1_TT3_64_VW1_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 417 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU32_SUS256_SVW1_TT3_64_VW1_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 418 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 419 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 420 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU0_SUS0_SVW1_TT3_64_VW1_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 421 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 422 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 423 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU32_SUS256_SVW1_TT3_64_VW1_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 424 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 425 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 426 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 427 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 428 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 429 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 430 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 431 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 432 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 433 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 434 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 435 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 436 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 437 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB4_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 438 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB4_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 439 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB4_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 440 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 441 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 442 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 443 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 444 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 445 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 446 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 447 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 448 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 449 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 450 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 451 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 452 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 453 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_32_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 454 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 455 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS256_SVW2_TT2_48_VW2_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 456 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 457 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 458 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 1 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 459 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 460 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 461 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 462 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 463 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 464 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 465 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 466 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 467 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 468 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 469 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 470 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 471 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 472 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 473 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 474 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 475 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 476 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 477 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 478 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 479 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 480 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 481 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 482 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 483 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 484 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 485 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 486 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB2_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 487 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 488 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 489 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_ - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StorePriorityOpt: 0 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 -- [2, 3, 0, 1] -- - - [38144, 38144, 1, 256] - - [23, 74.893] - - - [29568, 128, 1, 384] - - [35, 47.913] - - - [30848, 128, 1, 256] - - [24, 41.422] - - - [25728, 128, 1, 384] - - [41, 57.278] - - - [32256, 32256, 1, 256] - - [25, 75.584] - - - [7680, 7680, 1, 256] - - [58, 73.456] - - - [41984, 41984, 1, 256] - - [26, 75.197] - - - [40448, 40448, 1, 256] - - [25, 74.937] - - - [25728, 128, 1, 256] - - [27, 49.099] - - - [64, 64, 1, 64] - - [206, 0.117] - - - [15104, 15104, 1, 256] - - [28, 75.487] - - - [17280, 17280, 1, 384] - - [28, 90.896] - - - [34688, 128, 1, 384] - - [24, 54.573] - - - [27392, 27392, 1, 256] - - [29, 74.926] - - - [6528, 128, 1, 256] - - [109, 37.791] - - - [35328, 35328, 1, 256] - - [52, 75.386] - - - [18432, 18432, 1, 384] - - [30, 89.766] - - - [31232, 31232, 1, 256] - - [49, 75.543] - - - [7808, 128, 1, 256] - - [110, 35.357] - - - [38400, 38400, 1, 384] - - [26, 90.583] - - - [16128, 16128, 1, 256] - - [38, 75.578] - - - [9472, 9472, 1, 256] - - [28, 74.227] - - - [21888, 21888, 1, 384] - - [29, 89.215] - - - [38656, 38656, 1, 256] - - [31, 74.685] - - - [20224, 20224, 1, 256] - - [38, 75.749] - - - [8960, 8960, 1, 256] - - [32, 73.31] - - - [29952, 29952, 1, 384] - - [28, 90.917] - - - [36864, 36864, 1, 384] - - [33, 87.926] - - - [33408, 33408, 1, 384] - - [30, 90.847] - - - [20608, 128, 1, 384] - - [34, 48.729] - - - [23424, 23424, 1, 384] - - [28, 91.11] - - - [4864, 4864, 1, 256] - - [35, 68.048] - - - [21504, 21504, 1, 384] - - [42, 89.214] - - - [25600, 25600, 1, 256] - - [30, 75.967] - - - [40960, 40960, 1, 256] - - [55, 66.623] - - - [19200, 19200, 1, 384] - - [30, 90.96] - - - [64, 1, 1, 64] - - [206, 0.002] - - - [25088, 25088, 1, 256] - - [43, 75.905] - - - [41728, 41728, 1, 256] - - [29, 73.787] - - - [35840, 35840, 1, 256] - - [38, 75.487] - - - [34560, 34560, 1, 256] - - [28, 75.1] - - - [26368, 26368, 1, 256] - - [26, 75.32] - - - [5888, 5888, 1, 256] - - [36, 70.595] - - - [28032, 28032, 1, 384] - - [28, 91.053] - - - [42496, 42496, 1, 256] - - [29, 75.137] - - - [27008, 128, 1, 256] - - [34, 50.615] - - - [38400, 38400, 1, 256] - - [26, 75.384] - - - [11008, 11008, 1, 256] - - [86, 73.892] - - - [32000, 32000, 1, 256] - - [28, 75.005] - - - [37248, 37248, 1, 384] - - [26, 90.692] - - - [10496, 10496, 1, 256] - - [26, 74.439] - - - [16640, 16640, 1, 256] - - [26, 75.73] - - - [24960, 24960, 1, 384] - - [26, 91.209] - - - [18688, 18688, 1, 256] - - [28, 75.593] - - - [22272, 22272, 1, 384] - - [37, 91.067] - - - [15488, 128, 1, 256] - - [60, 32.909] - - - [28416, 28416, 1, 384] - - [26, 90.867] - - - [3840, 3840, 1, 256] - - [35, 64.173] - - - [19968, 19968, 1, 384] - - [26, 90.738] - - - [43776, 43776, 1, 256] - - [29, 74.28] - - - [35072, 35072, 1, 256] - - [26, 75.071] - - - [20736, 20736, 1, 256] - - [28, 75.611] - - - [7168, 7168, 1, 256] - - [38, 72.544] - - - [18432, 18432, 1, 256] - - [28, 76.049] - - - [38016, 38016, 1, 384] - - [26, 91.017] - - - [35328, 35328, 1, 384] - - [26, 90.526] - - - [38784, 38784, 1, 384] - - [26, 90.979] - - - [26112, 26112, 1, 384] - - [30, 91.076] - - - [27264, 27264, 1, 384] - - [30, 90.776] - - - [44928, 44928, 1, 384] - - [28, 90.696] - - - [41088, 128, 1, 384] - - [37, 60.328] - - - [42368, 128, 1, 256] - - [24, 44.691] - - - [10752, 10752, 1, 256] - - [28, 75.284] - - - [9088, 128, 1, 384] - - [110, 46.368] - - - [17152, 17152, 1, 256] - - [48, 75.562] - - - [44928, 128, 1, 384] - - [37, 54.276] - - - [7808, 128, 1, 384] - - [111, 40.491] - - - [29184, 29184, 1, 256] - - [38, 75.558] - - - [11776, 11776, 1, 256] - - [26, 75.132] - - - [1, 64, 1, 64] - - [206, 0.002] - - - [27136, 27136, 1, 256] - - [43, 75.911] - - - [33408, 128, 1, 256] - - [36, 44.177] - - - [33792, 33792, 1, 384] - - [42, 89.232] - - - [43520, 43520, 1, 256] - - [39, 75.193] - - - [14592, 14592, 1, 384] - - [38, 90.309] - - - [41472, 41472, 1, 256] - - [26, 74.99] - - - [14080, 14080, 1, 256] - - [58, 74.446] - - - [34688, 128, 1, 256] - - [40, 45.463] - - - [16896, 16896, 1, 256] - - [38, 76.106] - - - [15744, 15744, 1, 384] - - [26, 90.68] - - - [28416, 28416, 1, 256] - - [26, 74.869] - - - [23808, 23808, 1, 256] - - [28, 75.555] - - - [27648, 27648, 1, 256] - - [26, 75.552] - - - [1152, 3072, 1, 384] - - [41, 63.809] - - - [21888, 128, 1, 256] - - [36, 43.504] - - - [34816, 34816, 1, 256] - - [30, 75.387] - - - [43776, 43776, 1, 384] - - [42, 89.916] - - - [36096, 36096, 1, 256] - - [42, 74.203] - - - [24320, 24320, 1, 256] - - [25, 75.666] - - - [12544, 12544, 1, 256] - - [28, 74.955] - - - [29184, 29184, 1, 384] - - [30, 90.781] - - - [29568, 29568, 1, 384] - - [25, 90.297] - - - [12928, 128, 1, 384] - - [112, 54.014] - - - [36480, 36480, 1, 384] - - [26, 90.923] - - - [30720, 30720, 1, 256] - - [30, 75.552] - - - [25728, 25728, 1, 384] - - [59, 90.933] - - - [34048, 34048, 1, 256] - - [43, 74.525] - - - [12928, 128, 1, 256] - - [113, 45.988] - - - [9728, 9728, 1, 256] - - [30, 74.669] - - - [128, 128, 1, 256] - - [114, 1.074] - - - [33024, 33024, 1, 256] - - [44, 75.213] - - - [15488, 128, 1, 384] - - [58, 37.868] - - - [39808, 128, 1, 384] - - [45, 59.233] - - - [18176, 18176, 1, 256] - - [28, 75.705] - - - [21504, 21504, 1, 256] - - [28, 76.075] - - - [16384, 16384, 1, 256] - - [55, 62.756] - - - [27008, 128, 1, 384] - - [46, 59.129] - - - [27904, 27904, 1, 256] - - [39, 75.163] - - - [24448, 128, 1, 384] - - [47, 55.216] - - - [35968, 128, 1, 384] - - [54, 55.639] - - - [37632, 37632, 1, 256] - - [30, 74.955] - - - [14848, 14848, 1, 256] - - [29, 75.624] - - - [23552, 23552, 1, 256] - - [30, 76.111] - - - [4608, 4608, 1, 50000] - - [54, 97.942] - - - [13056, 13056, 1, 256] - - [30, 75.313] - - - [38528, 128, 1, 256] - - [59, 48.177] - - - [19584, 19584, 1, 384] - - [38, 91.095] - - - [16768, 128, 1, 384] - - [58, 40.639] - - - [22784, 22784, 1, 256] - - [48, 75.492] - - - [44160, 44160, 1, 384] - - [30, 90.838] - - - [28160, 28160, 1, 256] - - [49, 75.558] - - - [14592, 14592, 1, 256] - - [69, 74.291] - - - [20992, 20992, 1, 256] - - [38, 76.172] - - - [41216, 41216, 1, 256] - - [26, 74.854] - - - [21760, 21760, 1, 256] - - [28, 75.734] - - - [25344, 25344, 1, 256] - - [42, 74.966] - - - [4608, 4608, 1, 256] - - [50, 69.948] - - - [2560, 2048, 1, 256] - - [50, 56.06] - - - [30464, 30464, 1, 256] - - [51, 74.371] - - - [19200, 19200, 1, 256] - - [38, 75.675] - - - [22272, 22272, 1, 256] - - [52, 75.594] - - - [29952, 29952, 1, 256] - - [48, 75.275] - - - [20480, 20480, 1, 256] - - [30, 75.812] - - - [17408, 17408, 1, 256] - - [26, 75.599] - - - [32768, 32768, 1, 256] - - [53, 57.634] - - - [18816, 18816, 1, 384] - - [59, 90.869] - - - [34944, 34944, 1, 384] - - [38, 90.977] - - - [18048, 18048, 1, 384] - - [59, 90.958] - - - [34560, 34560, 1, 384] - - [38, 90.826] - - - [9088, 128, 1, 256] - - [115, 40.41] - - - [24576, 24576, 1, 256] - - [55, 70.192] - - - [32128, 128, 1, 384] - - [50, 51.321] - - - [8448, 8448, 1, 256] - - [58, 73.715] - - - [42752, 42752, 1, 256] - - [28, 74.735] - - - [5376, 5376, 1, 256] - - [50, 69.366] - - - [18048, 128, 1, 256] - - [40, 37.6] - - - [3584, 3584, 1, 256] - - [36, 62.742] - - - [37120, 37120, 1, 256] - - [28, 74.921] - - - [39936, 39936, 1, 384] - - [62, 89.019] - - - [20736, 20736, 1, 384] - - [54, 91.075] - - - [35584, 35584, 1, 256] - - [38, 74.893] - - - [26112, 26112, 1, 256] - - [49, 76.0] - - - [16896, 16896, 1, 384] - - [38, 90.639] - - - [40704, 40704, 1, 384] - - [26, 90.834] - - - [33280, 33280, 1, 256] - - [25, 75.635] - - - [5632, 5632, 1, 256] - - [35, 71.658] - - - [19456, 19456, 1, 256] - - [28, 75.886] - - - [22016, 22016, 1, 256] - - [31, 76.148] - - - [14208, 128, 1, 256] - - [34, 30.277] - - - [13568, 13568, 1, 256] - - [26, 75.309] - - - [30848, 128, 1, 384] - - [61, 49.625] - - - [1408, 128, 1, 384] - - [116, 13.439] - - - [5760, 5760, 1, 5760] - - [240, 98.181] - - - [39936, 39936, 1, 256] - - [26, 75.103] - - - [1920, 3072, 1, 384] - - [36, 62.558] - - - [9984, 9984, 1, 256] - - [36, 74.421] - - - [2816, 2048, 1, 256] - - [36, 51.272] - - - [23168, 128, 1, 256] - - [48, 45.196] - - - [19968, 19968, 1, 256] - - [29, 76.018] - - - [44800, 44800, 1, 256] - - [26, 74.449] - - - [14976, 14976, 1, 384] - - [45, 90.548] - - - [35712, 35712, 1, 384] - - [26, 90.953] - - - [43008, 43008, 1, 384] - - [55, 89.03] - - - [41088, 41088, 1, 384] - - [42, 90.052] - - - [16128, 16128, 1, 384] - - [26, 90.362] - - - [5120, 5120, 1, 256] - - [36, 70.401] - - - [25856, 25856, 1, 256] - - [30, 75.426] - - - [12288, 12288, 1, 256] - - [28, 75.012] - - - [6400, 6400, 1, 256] - - [56, 72.272] - - - [2688, 128, 1, 256] - - [117, 20.246] - - - [11648, 128, 1, 256] - - [118, 42.668] - - - [43264, 43264, 1, 256] - - [26, 74.81] - - - [19712, 19712, 1, 256] - - [33, 75.0] - - - [34176, 34176, 1, 384] - - [28, 90.755] - - - [31104, 31104, 1, 384] - - [28, 90.907] - - - [36608, 36608, 1, 256] - - [39, 74.827] - - - [39808, 128, 1, 256] - - [41, 49.04] - - - [13824, 13824, 1, 384] - - [30, 90.095] - - - [42624, 42624, 1, 384] - - [29, 87.986] - - - [21120, 21120, 1, 384] - - [38, 91.055] - - - [23296, 23296, 1, 256] - - [26, 75.634] - - - [42240, 42240, 1, 256] - - [28, 74.739] - - - [33408, 128, 1, 384] - - [36, 52.984] - - - [43648, 128, 1, 256] - - [41, 45.265] - - - [19328, 128, 1, 384] - - [57, 46.245] - - - [33792, 33792, 1, 256] - - [26, 75.516] - - - [31488, 31488, 1, 256] - - [38, 74.936] - - - [768, 3072, 1, 384] - - [58, 44.769] - - - [6144, 6144, 1, 256] - - [38, 71.929] - - - [20352, 20352, 1, 384] - - [28, 91.117] - - - [23168, 128, 1, 384] - - [40, 53.248] - - - [33536, 33536, 1, 256] - - [51, 75.178] - - - [32640, 32640, 1, 384] - - [30, 90.087] - - - [1536, 3072, 1, 384] - - [36, 63.449] - - - [19328, 128, 1, 256] - - [41, 40.04] - - - [2688, 3072, 1, 384] - - [50, 72.101] - - - [24192, 24192, 1, 384] - - [28, 91.241] - - - [6912, 6912, 1, 256] - - [35, 72.526] - - - [15360, 15360, 1, 256] - - [26, 76.063] - - - [18944, 18944, 1, 256] - - [25, 76.049] - - - [37376, 37376, 1, 256] - - [42, 75.399] - - - [31488, 31488, 1, 384] - - [30, 90.749] - - - [26880, 26880, 1, 256] - - [26, 75.31] - - - [44928, 128, 1, 128] - - [41, 31.959] - - - [24448, 128, 1, 256] - - [58, 46.959] - - - [31872, 31872, 1, 384] - - [30, 90.846] - - - [1408, 128, 1, 256] - - [119, 11.358] - - - [38528, 128, 1, 384] - - [59, 58.149] - - - [15616, 15616, 1, 256] - - [26, 75.505] - - - [39552, 39552, 1, 384] - - [26, 90.918] - - - [4352, 4352, 1, 256] - - [35, 67.359] - - - [28288, 128, 1, 384] - - [35, 46.221] - - - [10368, 128, 1, 256] - - [110, 43.982] - - - [32128, 128, 1, 256] - - [41, 42.898] - - - [4608, 4608, 1, 4608] - - [234, 97.048] - - - [8704, 8704, 1, 256] - - [30, 74.167] - - - [17664, 17664, 1, 256] - - [31, 74.999] - - - [24576, 24576, 1, 384] - - [33, 83.07] - - - [37248, 128, 1, 384] - - [56, 56.876] - - - [34304, 34304, 1, 256] - - [44, 75.525] - - - [42368, 128, 1, 384] - - [28, 52.092] - - - [17664, 17664, 1, 384] - - [28, 90.703] - - - [12800, 12800, 1, 256] - - [28, 75.345] - - - [26624, 26624, 1, 256] - - [38, 75.83] - - - [36864, 36864, 1, 256] - - [30, 74.612] - - - [40704, 40704, 1, 256] - - [28, 74.86] - - - [12032, 12032, 1, 256] - - [28, 74.709] - - - [33024, 33024, 1, 384] - - [26, 90.647] - - - [28800, 28800, 1, 384] - - [54, 91.023] - - - [22656, 22656, 1, 384] - - [26, 91.179] - - - [41472, 41472, 1, 384] - - [28, 90.471] - - - [39680, 39680, 1, 256] - - [30, 74.613] - - - [44032, 44032, 1, 256] - - [38, 75.129] - - - [43392, 43392, 1, 384] - - [26, 90.699] - - - [42240, 42240, 1, 384] - - [28, 90.742] - - - [38912, 38912, 1, 256] - - [30, 75.302] - - - [23040, 23040, 1, 384] - - [28, 91.024] - - - [13312, 13312, 1, 256] - - [26, 75.812] - - - [128, 128, 1, 384] - - [120, 1.211] - - - [39168, 39168, 1, 256] - - [42, 74.682] - - - [25344, 25344, 1, 384] - - [28, 91.121] - - - [5248, 128, 1, 256] - - [110, 31.75] - - - [30208, 30208, 1, 256] - - [43, 75.621] - - - [40192, 40192, 1, 256] - - [30, 74.915] - - - [15872, 15872, 1, 256] - - [26, 76.103] - - - [44544, 44544, 1, 256] - - [23, 74.966] - - - [11520, 11520, 1, 256] - - [30, 74.674] - - - [15360, 15360, 1, 384] - - [48, 89.355] - - - [23040, 23040, 1, 256] - - [28, 76.194] - - - [26496, 26496, 1, 384] - - [30, 91.077] - - - [11264, 11264, 1, 256] - - [26, 75.369] - - - [18048, 128, 1, 384] - - [60, 43.408] - - - [30976, 30976, 1, 256] - - [29, 74.406] - - - [11648, 128, 1, 384] - - [121, 50.278] - - - [2304, 3072, 1, 384] - - [36, 73.922] - - - [28928, 28928, 1, 256] - - [31, 74.911] - - - [43008, 43008, 1, 256] - - [26, 74.922] - - - [29440, 29440, 1, 256] - - [31, 75.408] - - - [36352, 36352, 1, 256] - - [42, 75.406] - - - [32256, 32256, 1, 384] - - [38, 90.64] - - - [23808, 23808, 1, 384] - - [54, 91.162] - - - [37248, 128, 1, 256] - - [34, 47.064] - - - [1, 1, 1, 64] - - [0, 0.0] - - - [37888, 37888, 1, 256] - - [30, 75.212] - - - [35968, 128, 1, 256] - - [60, 46.038] - - - [13824, 13824, 1, 256] - - [52, 75.675] - - - [39168, 39168, 1, 384] - - [26, 90.701] - - - [37632, 37632, 1, 384] - - [26, 90.859] - - - [29568, 128, 1, 256] - - [61, 39.885] - - - [14336, 14336, 1, 256] - - [30, 75.668] - - - [28288, 128, 1, 256] - - [61, 38.51] - - - [16512, 16512, 1, 384] - - [55, 89.482] - - - [30720, 30720, 1, 384] - - [33, 89.477] - - - [21248, 21248, 1, 256] - - [38, 75.578] - - - [29696, 29696, 1, 256] - - [38, 75.577] - - - [384, 3072, 1, 384] - - [110, 48.474] - - - [28672, 28672, 1, 256] - - [28, 75.173] - - - [32512, 32512, 1, 256] - - [25, 75.093] - - - [9216, 9216, 1, 256] - - [28, 74.101] - - - [6656, 6656, 1, 256] - - [35, 72.68] - - - [30336, 30336, 1, 384] - - [26, 90.946] - - - [20608, 128, 1, 256] - - [48, 41.527] - - - [7936, 7936, 1, 256] - - [36, 73.5] - - - [41856, 41856, 1, 384] - - [30, 90.931] - - - [44288, 44288, 1, 256] - - [38, 74.601] - - - [7744, 7744, 1, 7744] - - [26, 97.287] - - - [7424, 7424, 1, 256] - - [50, 73.14] - - - [39424, 39424, 1, 256] - - [26, 75.347] - - - [43648, 128, 1, 384] - - [37, 52.576] - - - [14208, 14208, 1, 384] - - [38, 89.943] - - - [36096, 36096, 1, 384] - - [29, 89.977] - - - [44544, 44544, 1, 384] - - [38, 90.389] - - - [22528, 22528, 1, 256] - - [30, 75.715] - - - [4096, 4096, 1, 256] - - [36, 66.031] - - - [31744, 31744, 1, 256] - - [38, 75.52] - - - [3968, 128, 1, 384] - - [109, 28.344] - - - [17920, 17920, 1, 256] - - [30, 76.073] - - - [5248, 128, 1, 384] - - [109, 36.66] - - - [26880, 26880, 1, 384] - - [38, 91.155] - - - [8192, 8192, 1, 256] - - [26, 72.824] - - - [3968, 128, 1, 256] - - [111, 24.407] - - - [41088, 128, 1, 256] - - [60, 49.94] - - - [21888, 128, 1, 384] - - [48, 50.833] - - - [16768, 128, 1, 256] - - [60, 35.074] - - - [24064, 24064, 1, 256] - - [25, 76.017] - - - [44928, 128, 1, 256] - - [28, 46.495] - - - [27648, 27648, 1, 384] - - [62, 89.139] - - - [24832, 24832, 1, 256] - - [25, 75.662] - - - [10240, 10240, 1, 256] - - [28, 75.096] - - - [40320, 40320, 1, 384] - - [30, 90.989] - - - [18432, 2688, 1, 384] - - [28, 85.949] - - - [43008, 2304, 1, 384] - - [30, 89.185] - - - [3840, 3072, 1, 384] - - [56, 76.792] - - - [33408, 1920, 1, 384] - - [30, 87.863] - - - [33792, 2688, 1, 384] - - [38, 87.542] - - - [8064, 2688, 1, 384] - - [38, 79.14] - - - [33408, 2304, 1, 384] - - [30, 88.628] - - - [31872, 1536, 1, 384] - - [26, 85.612] - - - [41088, 1920, 1, 384] - - [62, 86.524] - - - [41088, 2304, 1, 384] - - [75, 87.329] - - - [5376, 1536, 1, 384] - - [35, 71.681] - - - [16128, 1536, 1, 384] - - [30, 83.019] - - - [36480, 2688, 1, 384] - - [37, 88.977] - - - [15360, 768, 1, 384] - - [61, 75.527] - - - [42624, 768, 1, 384] - - [54, 82.371] - - - [4992, 1536, 1, 384] - - [36, 67.772] - - - [29952, 1536, 1, 384] - - [26, 86.156] - - - [10752, 2688, 1, 384] - - [37, 82.264] - - - [42240, 2688, 1, 384] - - [54, 89.835] - - - [36096, 1536, 1, 384] - - [25, 85.555] - - - [26496, 1536, 1, 384] - - [26, 85.909] - - - [42624, 2688, 1, 384] - - [28, 87.217] - - - [17664, 2688, 1, 384] - - [37, 86.621] - - - [37248, 1536, 1, 384] - - [38, 87.151] - - - [16896, 2304, 1, 384] - - [38, 86.299] - - - [22272, 1920, 1, 384] - - [26, 87.094] - - - [26880, 2688, 1, 384] - - [37, 87.772] - - - [384, 768, 1, 384] - - [111, 21.71] - - - [16896, 1920, 1, 384] - - [38, 86.53] - - - [32640, 2304, 1, 384] - - [26, 85.785] - - - [5760, 2304, 1, 384] - - [50, 76.76] - - - [11904, 2304, 1, 384] - - [28, 82.809] - - - [24576, 2304, 1, 384] - - [38, 85.039] - - - [33024, 1536, 1, 384] - - [38, 84.384] - - - [36096, 2304, 1, 384] - - [25, 86.916] - - - [20352, 2688, 1, 384] - - [45, 87.058] - - - [14592, 2304, 1, 384] - - [38, 85.738] - - - [16128, 1920, 1, 384] - - [28, 82.946] - - - [16512, 1920, 1, 384] - - [42, 79.688] - - - [35712, 1920, 1, 384] - - [38, 87.275] - - - [9216, 2688, 1, 384] - - [28, 83.651] - - - [23808, 2688, 1, 384] - - [54, 88.104] - - - [18048, 768, 1, 384] - - [45, 78.673] - - - [14592, 2688, 1, 384] - - [26, 86.006] - - - [14208, 1920, 1, 384] - - [30, 82.418] - - - [14976, 2688, 1, 384] - - [37, 85.749] - - - [17280, 2304, 1, 384] - - [30, 84.863] - - - [11520, 2304, 1, 384] - - [26, 84.796] - - - [18432, 768, 1, 384] - - [45, 79.512] - - - [4608, 768, 1, 384] - - [36, 65.306] - - - [34944, 1920, 1, 384] - - [38, 87.524] - - - [13824, 2688, 1, 384] - - [54, 86.081] - - - [39936, 2304, 1, 384] - - [26, 88.199] - - - [7680, 2688, 1, 384] - - [28, 81.415] - - - [19968, 2304, 1, 384] - - [28, 86.824] - - - [27648, 2688, 1, 384] - - [38, 86.702] - - - [4224, 768, 1, 384] - - [35, 59.788] - - - [24192, 1920, 1, 384] - - [28, 87.591] - - - [32640, 1920, 1, 384] - - [26, 84.928] - - - [34176, 2688, 1, 384] - - [57, 88.958] - - - [35328, 1536, 1, 384] - - [28, 85.698] - - - [8832, 2688, 1, 384] - - [54, 81.421] - - - [18048, 1920, 1, 384] - - [28, 84.439] - - - [31488, 768, 1, 384] - - [54, 82.286] - - - [21504, 2304, 1, 384] - - [26, 86.898] - - - [11136, 2688, 1, 384] - - [54, 84.703] - - - [768, 1152, 1, 384] - - [110, 46.95] - - - [29184, 2688, 1, 384] - - [45, 88.494] - - - [4608, 2688, 1, 384] - - [50, 79.851] - - - [21504, 2688, 1, 384] - - [28, 85.527] - - - [34176, 768, 1, 384] - - [54, 83.561] - - - [23808, 1536, 1, 384] - - [28, 84.573] - - - [43392, 1536, 1, 384] - - [26, 87.946] - - - [13824, 768, 1, 384] - - [37, 77.721] - - - [38016, 1536, 1, 384] - - [30, 86.692] - - - [20736, 2688, 1, 384] - - [59, 88.463] - - - [15744, 1536, 1, 384] - - [26, 82.237] - - - [16512, 1536, 1, 384] - - [39, 75.676] - - - [3072, 2304, 1, 384] - - [50, 73.318] - - - [5760, 2688, 1, 384] - - [447, 81.265] - - - [38400, 2304, 1, 384] - - [26, 88.2] - - - [15360, 2688, 1, 384] - - [28, 86.326] - - - [29952, 2688, 1, 384] - - [59, 88.86] - - - [43008, 2688, 1, 384] - - [30, 87.686] - - - [13440, 1920, 1, 384] - - [28, 82.655] - - - [6528, 2688, 1, 384] - - [26, 81.588] - - - [2304, 1536, 1, 384] - - [50, 64.822] - - - [40320, 1536, 1, 384] - - [28, 86.967] - - - [13440, 1536, 1, 384] - - [28, 81.341] - - - [40320, 2688, 1, 384] - - [54, 88.955] - - - [30336, 2304, 1, 384] - - [37, 88.619] - - - [24192, 2688, 1, 384] - - [45, 87.204] - - - [35328, 768, 1, 384] - - [54, 81.961] - - - [23040, 768, 1, 384] - - [59, 81.55] - - - [29952, 2304, 1, 384] - - [28, 87.87] - - - [33024, 1920, 1, 384] - - [38, 86.647] - - - [14976, 768, 1, 384] - - [59, 74.424] - - - [42624, 1920, 1, 384] - - [38, 87.484] - - - [32640, 2688, 1, 384] - - [29, 85.433] - - - [11520, 1536, 1, 384] - - [30, 82.185] - - - [6912, 768, 1, 384] - - [61, 69.191] - - - [39552, 1920, 1, 384] - - [28, 87.96] - - - [32256, 1920, 1, 384] - - [28, 87.539] - - - [10752, 1536, 1, 384] - - [26, 77.35] - - - [24576, 2688, 1, 384] - - [38, 84.613] - - - [12672, 2688, 1, 384] - - [54, 86.586] - - - [10752, 1920, 1, 384] - - [28, 81.453] - - - [40704, 1536, 1, 384] - - [26, 87.483] - - - [32256, 768, 1, 384] - - [54, 83.799] - - - [18816, 2688, 1, 384] - - [45, 85.896] - - - [11520, 2688, 1, 384] - - [37, 83.45] - - - [35712, 2688, 1, 384] - - [54, 88.852] - - - [29952, 1920, 1, 384] - - [26, 88.372] - - - [26880, 1920, 1, 384] - - [26, 87.563] - - - [33408, 2688, 1, 384] - - [59, 89.067] - - - [35328, 2688, 1, 384] - - [45, 88.976] - - - [21120, 2688, 1, 384] - - [63, 86.9] - - - [19584, 1920, 1, 384] - - [28, 86.96] - - - [17664, 1536, 1, 384] - - [30, 81.463] - - - [36864, 768, 1, 384] - - [45, 83.653] - - - [14592, 1536, 1, 384] - - [26, 81.728] - - - [11136, 2304, 1, 384] - - [26, 82.333] - - - [9600, 2688, 1, 384] - - [30, 82.609] - - - [9216, 2304, 1, 384] - - [30, 83.156] - - - [21120, 768, 1, 384] - - [37, 81.735] - - - [4992, 2688, 1, 384] - - [447, 81.907] - - - [41472, 768, 1, 384] - - [45, 85.184] - - - [37632, 1536, 1, 384] - - [28, 86.027] - - - [38784, 2304, 1, 384] - - [38, 88.806] - - - [8448, 2688, 1, 384] - - [45, 82.716] - - - [36864, 2304, 1, 384] - - [38, 87.729] - - - [40704, 1920, 1, 384] - - [30, 88.34] - - - [39552, 2688, 1, 384] - - [54, 89.664] - - - [26112, 768, 1, 384] - - [45, 79.562] - - - [29184, 1536, 1, 384] - - [26, 86.987] - - - [32640, 1536, 1, 384] - - [30, 83.298] - - - [5376, 2688, 1, 384] - - [383, 81.218] - - - [13056, 768, 1, 384] - - [54, 73.998] - - - [13824, 2304, 1, 384] - - [28, 85.541] - - - [16896, 768, 1, 384] - - [37, 74.655] - - - [30336, 1920, 1, 384] - - [30, 87.334] - - - [27264, 2304, 1, 384] - - [30, 88.195] - - - [7680, 1536, 1, 384] - - [26, 76.276] - - - [30720, 2688, 1, 384] - - [28, 87.806] - - - [36096, 2688, 1, 384] - - [29, 87.086] - - - [5760, 1920, 1, 384] - - [36, 72.496] - - - [42240, 1536, 1, 384] - - [28, 87.779] - - - [8448, 1920, 1, 384] - - [35, 82.584] - - - [32256, 1536, 1, 384] - - [38, 86.303] - - - [44160, 2304, 1, 384] - - [59, 89.108] - - - [30336, 2688, 1, 384] - - [45, 88.332] - - - [6144, 2688, 1, 384] - - [447, 80.619] - - - [39168, 1536, 1, 384] - - [26, 86.962] - - - [11904, 1920, 1, 384] - - [59, 83.359] - - - [8064, 1536, 1, 384] - - [56, 78.707] - - - [21120, 1920, 1, 384] - - [28, 86.0] - - - [22656, 2304, 1, 384] - - [45, 88.383] - - - [19968, 2688, 1, 384] - - [45, 87.953] - - - [10752, 768, 1, 384] - - [64, 71.538] - - - [18432, 2304, 1, 384] - - [30, 86.351] - - - [14976, 1920, 1, 384] - - [59, 86.02] - - - [33024, 2688, 1, 384] - - [59, 87.96] - - - [1536, 768, 1, 384] - - [345, 47.037] - - - [33024, 2304, 1, 384] - - [45, 87.324] - - - [14208, 2688, 1, 384] - - [45, 84.865] - - - [38016, 2304, 1, 384] - - [54, 88.859] - - - [16896, 2688, 1, 384] - - [37, 85.692] - - - [31104, 768, 1, 384] - - [54, 81.555] - - - [41472, 2304, 1, 384] - - [26, 88.447] - - - [23424, 2688, 1, 384] - - [30, 87.378] - - - [26496, 2688, 1, 384] - - [37, 88.317] - - - [16512, 2304, 1, 384] - - [29, 80.998] - - - [11520, 1920, 1, 384] - - [26, 81.321] - - - [39552, 768, 1, 384] - - [54, 85.633] - - - [6144, 2304, 1, 384] - - [35, 80.665] - - - [14208, 2304, 1, 384] - - [28, 83.895] - - - [19584, 2304, 1, 384] - - [30, 85.331] - - - [36480, 768, 1, 384] - - [59, 83.786] - - - [15744, 2688, 1, 384] - - [54, 86.233] - - - [34560, 1536, 1, 384] - - [30, 86.554] - - - [8448, 2304, 1, 384] - - [26, 82.987] - - - [26112, 2688, 1, 384] - - [54, 88.934] - - - [39936, 768, 1, 384] - - [37, 82.561] - - - [19200, 1920, 1, 384] - - [28, 85.398] - - - [38400, 768, 1, 384] - - [54, 83.34] - - - [8448, 1536, 1, 384] - - [35, 74.895] - - - [13824, 1536, 1, 384] - - [26, 82.987] - - - [9600, 768, 1, 384] - - [24, 64.683] - - - [10368, 768, 1, 384] - - [64, 69.246] - - - [20736, 1536, 1, 384] - - [26, 85.105] - - - [28800, 768, 1, 384] - - [37, 80.331] - - - [10368, 1536, 1, 384] - - [28, 81.171] - - - [21888, 1536, 1, 384] - - [28, 82.833] - - - [38784, 2688, 1, 384] - - [37, 89.453] - - - [27648, 2304, 1, 384] - - [30, 87.437] - - - [11136, 1920, 1, 384] - - [28, 83.769] - - - [37248, 768, 1, 384] - - [45, 85.091] - - - [23040, 2688, 1, 384] - - [37, 87.36] - - - [37632, 1920, 1, 384] - - [26, 87.723] - - - [7680, 768, 1, 384] - - [24, 62.177] - - - [38016, 1920, 1, 384] - - [26, 88.34] - - - [35712, 2304, 1, 384] - - [45, 88.776] - - - [37248, 2688, 1, 384] - - [45, 88.92] - - - [29568, 1920, 1, 384] - - [52, 86.846] - - - [38400, 2688, 1, 384] - - [54, 88.725] - - - [25728, 768, 1, 384] - - [59, 83.582] - - - [8832, 1920, 1, 384] - - [54, 79.336] - - - [43776, 1920, 1, 384] - - [31, 86.47] - - - [15744, 768, 1, 384] - - [54, 77.357] - - - [27264, 1920, 1, 384] - - [30, 85.812] - - - [33792, 2304, 1, 384] - - [28, 87.915] - - - [8832, 2304, 1, 384] - - [28, 80.583] - - - [39168, 2688, 1, 384] - - [37, 89.136] - - - [35328, 1920, 1, 384] - - [28, 88.337] - - - [35328, 2304, 1, 384] - - [28, 88.11] - - - [29184, 768, 1, 384] - - [54, 81.9] - - - [18048, 2688, 1, 384] - - [59, 87.839] - - - [32256, 2688, 1, 384] - - [37, 88.127] - - - [18816, 1536, 1, 384] - - [28, 81.82] - - - [13056, 1536, 1, 384] - - [26, 78.981] - - - [34944, 1536, 1, 384] - - [28, 87.087] - - - [38400, 1920, 1, 384] - - [38, 88.679] - - - [15360, 2304, 1, 384] - - [26, 85.764] - - - [27264, 2688, 1, 384] - - [28, 87.083] - - - [11136, 1536, 1, 384] - - [28, 79.702] - - - [30720, 2304, 1, 384] - - [38, 87.779] - - - [24960, 2688, 1, 384] - - [59, 87.696] - - - [13824, 1920, 1, 384] - - [26, 84.703] - - - [17280, 2688, 1, 384] - - [45, 87.383] - - - [31872, 768, 1, 384] - - [37, 83.012] - - - [11904, 2688, 1, 384] - - [37, 85.826] - - - [7296, 768, 1, 384] - - [64, 58.889] - - - [19200, 1536, 1, 384] - - [30, 83.176] - - - [12288, 768, 1, 384] - - [37, 69.334] - - - [33792, 768, 1, 384] - - [54, 82.575] - - - [21888, 2688, 1, 384] - - [65, 83.957] - - - [2688, 1920, 1, 384] - - [36, 67.714] - - - [19968, 768, 1, 384] - - [45, 78.744] - - - [12288, 2688, 1, 384] - - [28, 83.987] - - - [12288, 2304, 1, 384] - - [26, 84.682] - - - [28416, 768, 1, 384] - - [45, 80.118] - - - [34560, 768, 1, 384] - - [54, 84.329] - - - [39936, 2688, 1, 384] - - [28, 88.07] - - - [8064, 1920, 1, 384] - - [38, 79.126] - - - [26880, 1536, 1, 384] - - [38, 86.713] - - - [28032, 2688, 1, 384] - - [37, 89.022] - - - [41472, 2688, 1, 384] - - [59, 89.01] - - - [29568, 2688, 1, 384] - - [54, 87.448] - - - [31104, 2688, 1, 384] - - [54, 88.587] - - - [5376, 1920, 1, 384] - - [50, 76.125] - - - [41856, 2688, 1, 384] - - [37, 89.376] - - - [9984, 768, 1, 384] - - [61, 67.304] - - - [3456, 2688, 1, 384] - - [36, 70.02] - - - [43392, 2688, 1, 384] - - [45, 89.587] - - - [36480, 1920, 1, 384] - - [45, 88.451] - - - [29568, 1536, 1, 384] - - [48, 84.24] - - - [36864, 2688, 1, 384] - - [38, 87.725] - - - [12672, 768, 1, 384] - - [24, 72.052] - - - [24064, 3072, 1, 256] - - [32, 72.89] - - - [256, 512, 1, 256] - - [123, 8.398] - - - [40960, 27648, 1, 256] - - [23, 67.36] - - - [31744, 3072, 1, 256] - - [30, 73.477] - - - [13056, 1792, 1, 256] - - [56, 68.961] - - - [35328, 22785, 1, 256] - - [74, 73.723] - - - [28160, 15872, 1, 256] - - [66, 75.611] - - - [39168, 1792, 1, 256] - - [60, 71.707] - - - [23808, 11265, 1, 256] - - [33, 71.85] - - - [16640, 4353, 1, 256] - - [74, 69.432] - - - [38912, 26624, 1, 256] - - [38, 75.182] - - - [6912, 3585, 1, 256] - - [30, 65.964] - - - [32768, 1792, 1, 256] - - [23, 58.928] - - - [30976, 18688, 1, 256] - - [29, 74.426] - - - [512, 2048, 1, 256] - - [122, 37.096] - - - [15872, 3584, 1, 256] - - [32, 72.985] - - - [6400, 1792, 1, 256] - - [36, 62.034] - - - [39680, 27393, 1, 256] - - [55, 72.507] - - - [36864, 24577, 1, 256] - - [33, 71.805] - - - [26112, 1536, 1, 256] - - [50, 70.393] - - - [26368, 1536, 1, 256] - - [50, 69.857] - - - [16896, 4353, 1, 256] - - [31, 71.01] - - - [14336, 1793, 1, 256] - - [26, 64.85] - - - [3840, 3072, 1, 256] - - [35, 63.197] - - - [2560, 3072, 1, 256] - - [36, 56.708] - - - [6656, 1536, 1, 256] - - [50, 61.995] - - - [27136, 1792, 1, 256] - - [35, 71.697] - - - [43776, 3072, 1, 256] - - [29, 71.088] - - - [23296, 1792, 1, 256] - - [35, 70.456] - - - [11264, 7937, 1, 256] - - [33, 72.469] - - - [768, 3072, 1, 256] - - [58, 38.337] - - - [6912, 3841, 1, 256] - - [35, 66.587] - - - [40960, 769, 1, 256] - - [30, 55.161] - - - [40448, 9216, 1, 256] - - [43, 74.661] - - - [7680, 4353, 1, 256] - - [26, 67.882] - - - [23296, 3072, 1, 256] - - [26, 72.423] - - - [7936, 4609, 1, 256] - - [58, 69.169] - - - [20736, 8448, 1, 256] - - [30, 74.596] - - - [768, 1024, 1, 256] - - [124, 36.135] - - - [38656, 3072, 1, 256] - - [31, 72.725] - - - [28160, 1792, 1, 256] - - [32, 71.494] - - - [13824, 3072, 1, 256] - - [28, 71.6] - - - [42752, 1792, 1, 256] - - [30, 72.151] - - - [35584, 23041, 1, 256] - - [29, 73.114] - - - [13056, 3072, 1, 256] - - [28, 71.139] - - - [37888, 768, 1, 256] - - [45, 68.261] - - - [19456, 3072, 1, 256] - - [28, 72.97] - - - [15872, 9216, 1, 256] - - [38, 74.823] - - - [30976, 1792, 1, 256] - - [58, 71.133] - - - [26368, 14081, 1, 256] - - [55, 72.876] - - - [35328, 23041, 1, 256] - - [44, 73.686] - - - [27648, 15105, 1, 256] - - [26, 74.161] - - - [25856, 13568, 1, 256] - - [38, 75.047] - - - [23296, 9216, 1, 256] - - [30, 74.259] - - - [2048, 1024, 1, 256] - - [60, 34.871] - - - [12032, 1792, 1, 256] - - [50, 69.388] - - - [11520, 1536, 1, 256] - - [56, 67.434] - - - [16128, 768, 1, 256] - - [61, 62.842] - - - [15360, 3072, 1, 256] - - [28, 72.025] - - - [38912, 26369, 1, 256] - - [55, 73.922] - - - [25344, 13056, 1, 256] - - [42, 74.792] - - - [39168, 26880, 1, 256] - - [39, 74.625] - - - [39424, 768, 1, 256] - - [64, 68.846] - - - [10496, 1792, 1, 256] - - [36, 66.402] - - - [28672, 3072, 1, 256] - - [26, 73.07] - - - [27392, 768, 1, 256] - - [40, 63.787] - - - [39680, 768, 1, 256] - - [45, 68.165] - - - [11520, 8193, 1, 256] - - [55, 70.412] - - - [17408, 4865, 1, 256] - - [26, 71.47] - - - [14080, 1537, 1, 256] - - [28, 63.305] - - - [29184, 768, 1, 256] - - [40, 67.284] - - - [19200, 6913, 1, 256] - - [32, 71.548] - - - [33536, 9216, 1, 256] - - [42, 74.284] - - - [5632, 3072, 1, 256] - - [56, 66.804] - - - [32768, 20480, 1, 256] - - [67, 58.823] - - - [29440, 9216, 1, 256] - - [48, 74.429] - - - [40960, 1792, 1, 256] - - [39, 66.694] - - - [10240, 3072, 1, 256] - - [35, 70.816] - - - [20992, 1792, 1, 256] - - [28, 71.07] - - - [42240, 9216, 1, 256] - - [29, 74.271] - - - [19200, 6912, 1, 256] - - [28, 74.327] - - - [27392, 1792, 1, 256] - - [87, 68.339] - - - [42496, 1536, 1, 256] - - [38, 71.485] - - - [29440, 16897, 1, 256] - - [29, 73.235] - - - [20480, 8192, 1, 256] - - [26, 74.921] - - - [11264, 8193, 1, 256] - - [30, 70.854] - - - [26880, 14337, 1, 256] - - [55, 72.433] - - - [28928, 16641, 1, 256] - - [74, 73.123] - - - [15360, 2817, 1, 256] - - [28, 67.942] - - - [44288, 1536, 1, 256] - - [30, 70.471] - - - [7936, 1536, 1, 256] - - [35, 63.735] - - - [18176, 5633, 1, 256] - - [26, 71.129] - - - [8448, 3072, 1, 256] - - [50, 68.909] - - - [17920, 5632, 1, 256] - - [25, 74.331] - - - [1792, 2048, 1, 256] - - [36, 40.334] - - - [39936, 3072, 1, 256] - - [38, 73.824] - - - [20480, 3072, 1, 256] - - [30, 72.732] - - - [24832, 1792, 1, 256] - - [68, 71.196] - - - [37376, 25088, 1, 256] - - [38, 75.229] - - - [7168, 4097, 1, 256] - - [26, 67.186] - - - [21504, 768, 1, 256] - - [54, 64.581] - - - [13312, 3072, 1, 256] - - [28, 71.791] - - - [40960, 1025, 1, 256] - - [30, 57.587] - - - [12032, 1536, 1, 256] - - [35, 65.445] - - - [9216, 768, 1, 256] - - [61, 59.444] - - - [44288, 27648, 1, 256] - - [29, 74.594] - - - [32512, 1792, 1, 256] - - [35, 71.363] - - - [23808, 11520, 1, 256] - - [26, 75.256] - - - [25600, 13057, 1, 256] - - [55, 73.924] - - - [40448, 1792, 1, 256] - - [30, 72.271] - - - [25088, 12800, 1, 256] - - [48, 75.683] - - - [22784, 10496, 1, 256] - - [68, 74.874] - - - [38400, 26113, 1, 256] - - [29, 73.709] - - - [9728, 3072, 1, 256] - - [28, 70.812] - - - [20736, 1792, 1, 256] - - [56, 70.596] - - - [7680, 3072, 1, 256] - - [36, 67.931] - - - [5376, 2305, 1, 256] - - [50, 58.991] - - - [12800, 3072, 1, 256] - - [32, 71.387] - - - [43520, 3584, 1, 256] - - [38, 74.091] - - - [12288, 3072, 1, 256] - - [28, 71.431] - - - [12800, 1536, 1, 256] - - [68, 67.084] - - - [21504, 8961, 1, 256] - - [26, 73.17] - - - [39680, 9216, 1, 256] - - [28, 74.129] - - - [3584, 513, 1, 256] - - [34, 30.521] - - - [1280, 3072, 1, 256] - - [41, 43.215] - - - [13056, 9216, 1, 256] - - [38, 74.371] - - - [22016, 768, 1, 256] - - [59, 64.586] - - - [33024, 1536, 1, 256] - - [56, 69.126] - - - [26880, 9216, 1, 256] - - [28, 74.278] - - - [44032, 27648, 1, 256] - - [28, 74.877] - - - [7680, 768, 1, 256] - - [41, 51.917] - - - [32000, 19712, 1, 256] - - [38, 75.0] - - - [26880, 14593, 1, 256] - - [55, 72.979] - - - [24064, 9216, 1, 256] - - [44, 74.935] - - - [39424, 26881, 1, 256] - - [55, 73.642] - - - [27392, 3072, 1, 256] - - [44, 70.269] - - - [10752, 1792, 1, 256] - - [56, 67.663] - - - [8960, 5633, 1, 256] - - [30, 70.165] - - - [34560, 3072, 1, 256] - - [26, 72.869] - - - [23808, 9216, 1, 256] - - [26, 74.572] - - - [29696, 17153, 1, 256] - - [30, 74.326] - - - [11776, 1536, 1, 256] - - [35, 64.33] - - - [13568, 1536, 1, 256] - - [35, 67.459] - - - [30208, 9216, 1, 256] - - [43, 75.008] - - - [36608, 1536, 1, 256] - - [56, 70.529] - - - [12800, 513, 1, 256] - - [38, 48.009] - - - [7680, 1792, 1, 256] - - [36, 65.222] - - - [42496, 2305, 1, 256] - - [38, 68.926] - - - [37376, 1536, 1, 256] - - [30, 70.914] - - - [20224, 1792, 1, 256] - - [35, 70.026] - - - [43520, 1536, 1, 256] - - [26, 71.431] - - - [26368, 768, 1, 256] - - [41, 66.092] - - - [18176, 3072, 1, 256] - - [28, 72.178] - - - [24320, 12033, 1, 256] - - [74, 73.218] - - - [17408, 9216, 1, 256] - - [28, 74.837] - - - [36352, 1792, 1, 256] - - [30, 72.37] - - - [20992, 8705, 1, 256] - - [33, 72.643] - - - [19712, 7424, 1, 256] - - [60, 73.379] - - - [38144, 768, 1, 256] - - [24, 67.922] - - - [10752, 1536, 1, 256] - - [30, 64.306] - - - [4096, 3072, 1, 256] - - [52, 64.649] - - - [29696, 17409, 1, 256] - - [55, 72.74] - - - [10240, 6913, 1, 256] - - [33, 71.858] - - - [18944, 1536, 1, 256] - - [50, 68.276] - - - [38656, 26113, 1, 256] - - [42, 72.926] - - - [37376, 25089, 1, 256] - - [42, 73.77] - - - [38400, 1536, 1, 256] - - [26, 70.995] - - - [8448, 1792, 1, 256] - - [56, 64.45] - - - [13056, 769, 1, 256] - - [45, 54.241] - - - [24320, 11777, 1, 256] - - [74, 73.035] - - - [17664, 9216, 1, 256] - - [25, 73.709] - - - [8192, 4865, 1, 256] - - [38, 69.019] - - - [17920, 1792, 1, 256] - - [60, 70.638] - - - [32000, 19713, 1, 256] - - [55, 73.259] - - - [8960, 768, 1, 256] - - [61, 58.995] - - - [31232, 3072, 1, 256] - - [68, 73.089] - - - [12544, 257, 1, 256] - - [37, 35.771] - - - [43776, 3585, 1, 256] - - [29, 68.553] - - - [11008, 1792, 1, 256] - - [69, 65.563] - - - [29696, 17408, 1, 256] - - [28, 75.352] - - - [34560, 22272, 1, 256] - - [26, 74.996] - - - [256, 2048, 1, 256] - - [115, 25.406] - - - [32768, 20481, 1, 256] - - [67, 55.844] - - - [14336, 3072, 1, 256] - - [56, 71.953] - - - [19456, 7168, 1, 256] - - [28, 74.12] - - - [13312, 9216, 1, 256] - - [38, 74.894] - - - [22272, 768, 1, 256] - - [61, 65.541] - - - [24064, 1792, 1, 256] - - [32, 71.538] - - - [16896, 1792, 1, 256] - - [56, 70.328] - - - [27904, 15616, 1, 256] - - [28, 75.078] - - - [37888, 3072, 1, 256] - - [26, 73.795] - - - [13056, 513, 1, 256] - - [37, 48.579] - - - [36608, 24065, 1, 256] - - [29, 72.88] - - - [40704, 3072, 1, 256] - - [28, 73.142] - - - [28928, 16640, 1, 256] - - [48, 75.132] - - - [24576, 12288, 1, 256] - - [55, 69.898] - - - [17152, 3072, 1, 256] - - [50, 71.874] - - - [17152, 4864, 1, 256] - - [36, 73.546] - - - [42496, 9216, 1, 256] - - [29, 74.809] - - - [32256, 768, 1, 256] - - [40, 68.635] - - - [4352, 1792, 1, 256] - - [35, 55.99] - - - [5632, 768, 1, 256] - - [40, 46.604] - - - [40704, 513, 1, 256] - - [59, 56.078] - - - [19712, 768, 1, 256] - - [70, 61.993] - - - [33536, 20993, 1, 256] - - [74, 73.277] - - - [2816, 3072, 1, 256] - - [56, 61.168] - - - [3584, 3072, 1, 256] - - [26, 59.974] - - - [4608, 1537, 1, 256] - - [36, 51.08] - - - [44032, 9216, 1, 256] - - [26, 74.553] - - - [33792, 21249, 1, 256] - - [33, 74.025] - - - [32512, 20225, 1, 256] - - [74, 73.399] - - - [38656, 9216, 1, 256] - - [42, 74.324] - - - [17664, 5377, 1, 256] - - [31, 70.624] - - - [19456, 7169, 1, 256] - - [33, 71.093] - - - [8448, 5121, 1, 256] - - [30, 68.835] - - - [29440, 17152, 1, 256] - - [31, 75.486] - - - [40448, 513, 1, 256] - - [37, 56.249] - - - [41472, 1792, 1, 256] - - [30, 72.408] - - - [17920, 3072, 1, 256] - - [58, 72.264] - - - [35072, 9216, 1, 256] - - [26, 74.214] - - - [34816, 22273, 1, 256] - - [55, 74.005] - - - [35072, 22785, 1, 256] - - [33, 73.107] - - - [39168, 9216, 1, 256] - - [29, 74.256] - - - [42752, 2817, 1, 256] - - [55, 69.038] - - - [11776, 3072, 1, 256] - - [26, 70.075] - - - [24832, 12289, 1, 256] - - [44, 72.414] - - - [24576, 12033, 1, 256] - - [33, 68.22] - - - [6400, 1536, 1, 256] - - [35, 60.956] - - - [32512, 3072, 1, 256] - - [58, 72.723] - - - [30976, 3072, 1, 256] - - [32, 71.129] - - - [22016, 9473, 1, 256] - - [25, 73.241] - - - [19968, 1792, 1, 256] - - [56, 71.033] - - - [29440, 3072, 1, 256] - - [31, 73.013] - - - [43776, 3840, 1, 256] - - [29, 72.616] - - - [41472, 768, 1, 256] - - [54, 68.878] - - - [8192, 1792, 1, 256] - - [36, 63.092] - - - [35840, 3072, 1, 256] - - [28, 73.788] - - - [8704, 3072, 1, 256] - - [35, 70.142] - - - [9728, 1792, 1, 256] - - [50, 67.491] - - - [22272, 9729, 1, 256] - - [71, 72.315] - - - [32768, 3072, 1, 256] - - [23, 59.221] - - - [3072, 2048, 1, 256] - - [36, 54.959] - - - [36864, 24576, 1, 256] - - [26, 74.304] - - - [9984, 1536, 1, 256] - - [56, 64.234] - - - [12032, 8961, 1, 256] - - [48, 71.392] - - - [38400, 25857, 1, 256] - - [29, 73.786] - - - [20224, 7937, 1, 256] - - [55, 72.001] - - - [34304, 21761, 1, 256] - - [44, 73.859] - - - [30720, 18432, 1, 256] - - [28, 75.598] - - - [31744, 9216, 1, 256] - - [28, 74.725] - - - [27136, 14848, 1, 256] - - [48, 75.662] - - - [34048, 9216, 1, 256] - - [44, 73.879] - - - [3584, 257, 1, 256] - - [115, 33.197] - - - [18688, 6145, 1, 256] - - [33, 70.227] - - - [36096, 768, 1, 256] - - [40, 66.163] - - - [36608, 9216, 1, 256] - - [29, 74.237] - - - [35584, 9216, 1, 256] - - [29, 74.453] - - - [29952, 17664, 1, 256] - - [25, 75.301] - - - [34816, 1792, 1, 256] - - [30, 72.664] - - - [24064, 11776, 1, 256] - - [25, 75.569] - - - [40448, 3072, 1, 256] - - [52, 73.246] - - - [18688, 6401, 1, 256] - - [55, 70.993] - - - [20480, 1536, 1, 256] - - [28, 69.236] - - - [18432, 3072, 1, 256] - - [26, 72.829] - - - [20224, 768, 1, 256] - - [61, 64.775] - - - [25344, 768, 1, 256] - - [70, 66.478] - - - [36608, 24320, 1, 256] - - [39, 74.884] - - - [34816, 9216, 1, 256] - - [26, 74.783] - - - [41216, 27648, 1, 256] - - [29, 74.634] - - - [30464, 9216, 1, 256] - - [42, 73.654] - - - [7424, 3072, 1, 256] - - [36, 69.039] - - - [20480, 1792, 1, 256] - - [50, 70.652] - - - [41984, 1793, 1, 256] - - [30, 67.812] - - - [18688, 1792, 1, 256] - - [50, 70.3] - - - [13824, 1792, 1, 256] - - [36, 70.014] - - - [38144, 3072, 1, 256] - - [38, 73.144] - - - [33280, 3072, 1, 256] - - [52, 73.418] - - - [35584, 23296, 1, 256] - - [51, 74.79] - - - [43520, 768, 1, 256] - - [37, 69.457] - - - [40704, 1536, 1, 256] - - [30, 71.038] - - - [29696, 3072, 1, 256] - - [38, 73.642] - - - [32256, 19969, 1, 256] - - [74, 73.669] - - - [40960, 9216, 1, 256] - - [39, 67.099] - - - [37632, 9216, 1, 256] - - [30, 73.983] - - - [42240, 2305, 1, 256] - - [33, 68.068] - - - [17920, 5377, 1, 256] - - [25, 71.454] - - - [27904, 9216, 1, 256] - - [42, 74.405] - - - [34304, 22016, 1, 256] - - [43, 75.492] - - - [11776, 8705, 1, 256] - - [52, 72.055] - - - [22272, 1536, 1, 256] - - [32, 69.158] - - - [25856, 9216, 1, 256] - - [26, 74.358] - - - [19712, 3072, 1, 256] - - [52, 70.534] - - - [41472, 9216, 1, 256] - - [43, 74.605] - - - [42496, 27648, 1, 256] - - [29, 75.01] - - - [44288, 4352, 1, 256] - - [39, 73.828] - - - [42496, 2561, 1, 256] - - [26, 68.824] - - - [9984, 6657, 1, 256] - - [30, 71.101] - - - [43008, 3073, 1, 256] - - [55, 69.489] - - - [36352, 24065, 1, 256] - - [42, 73.734] - - - [24832, 3072, 1, 256] - - [58, 72.622] - - - [29184, 16641, 1, 256] - - [33, 73.633] - - - [1024, 2048, 1, 256] - - [73, 34.714] - - - [42240, 27648, 1, 256] - - [29, 74.587] - - - [9984, 1792, 1, 256] - - [56, 67.945] - - - [44288, 3072, 1, 256] - - [55, 72.705] - - - [11008, 768, 1, 256] - - [73, 56.84] - - - [28672, 16129, 1, 256] - - [55, 73.603] - - - [17920, 9216, 1, 256] - - [75, 74.862] - - - [25088, 12801, 1, 256] - - [44, 73.417] - - - [19712, 9216, 1, 256] - - [42, 74.04] - - - [31744, 19457, 1, 256] - - [26, 72.863] - - - [36864, 1792, 1, 256] - - [28, 72.242] - - - [42496, 1792, 1, 256] - - [26, 72.873] - - - [39936, 9216, 1, 256] - - [26, 74.655] - - - [8960, 1792, 1, 256] - - [35, 66.553] - - - [17664, 5121, 1, 256] - - [52, 69.096] - - - [38144, 25601, 1, 256] - - [29, 72.464] - - - [27136, 14849, 1, 256] - - [44, 73.926] - - - [31744, 19456, 1, 256] - - [28, 75.314] - - - [33024, 3072, 1, 256] - - [25, 72.339] - - - [37888, 9216, 1, 256] - - [28, 74.723] - - - [6912, 1792, 1, 256] - - [36, 64.897] - - - [42240, 2049, 1, 256] - - [59, 66.103] - - - [34048, 3072, 1, 256] - - [32, 72.477] - - - [37120, 9216, 1, 256] - - [42, 74.336] - - - [14080, 9216, 1, 256] - - [52, 73.043] - - - [38400, 1792, 1, 256] - - [38, 72.333] - - - [43776, 9216, 1, 256] - - [42, 74.086] - - - [14336, 2049, 1, 256] - - [30, 63.371] - - - [37120, 24577, 1, 256] - - [42, 72.438] - - - [30976, 18433, 1, 256] - - [29, 72.375] - - - [37632, 3072, 1, 256] - - [26, 72.923] - - - [34560, 1792, 1, 256] - - [56, 71.797] - - - [5120, 3072, 1, 256] - - [36, 66.228] - - - [21760, 9217, 1, 256] - - [33, 71.598] - - - [24064, 11521, 1, 256] - - [31, 73.393] - - - [7936, 3072, 1, 256] - - [50, 68.835] - - - [21760, 9472, 1, 256] - - [30, 75.121] - - - [9216, 6145, 1, 256] - - [28, 69.56] - - - [8192, 1536, 1, 256] - - [28, 63.981] - - - [39936, 27648, 1, 256] - - [26, 74.967] - - - [21248, 9216, 1, 256] - - [28, 74.493] - - - [5376, 2049, 1, 256] - - [28, 56.9] - - - [35072, 22529, 1, 256] - - [55, 72.351] - - - [13312, 769, 1, 256] - - [59, 55.124] - - - [35840, 9216, 1, 256] - - [26, 74.674] - - - [39424, 27136, 1, 256] - - [39, 75.313] - - - [26368, 9216, 1, 256] - - [30, 74.3] - - - [34048, 21505, 1, 256] - - [29, 72.296] - - - [26112, 1792, 1, 256] - - [60, 71.895] - - - [23296, 768, 1, 256] - - [64, 67.076] - - - [43264, 27648, 1, 256] - - [28, 74.283] - - - [18432, 9216, 1, 256] - - [38, 75.027] - - - [38912, 3072, 1, 256] - - [28, 73.845] - - - [30464, 17921, 1, 256] - - [42, 72.452] - - - [37376, 9216, 1, 256] - - [42, 74.853] - - - [256, 3072, 1, 256] - - [124, 36.28] - - - [9472, 3072, 1, 256] - - [30, 69.062] - - - [35840, 23552, 1, 256] - - [30, 75.421] - - - [8960, 3072, 1, 256] - - [35, 68.93] - - - [34816, 3072, 1, 256] - - [26, 73.722] - - - [11008, 3072, 1, 256] - - [69, 69.116] - - - [36864, 1536, 1, 256] - - [28, 70.739] - - - [23552, 9216, 1, 256] - - [26, 74.931] - - - [31232, 18945, 1, 256] - - [42, 73.789] - - - [27136, 9216, 1, 256] - - [44, 74.972] - - - [19968, 7681, 1, 256] - - [42, 72.249] - - - [31488, 18945, 1, 256] - - [33, 73.064] - - - [33280, 1792, 1, 256] - - [32, 72.275] - - - [14592, 3072, 1, 256] - - [60, 70.802] - - - [30976, 18689, 1, 256] - - [42, 72.645] - - - [4096, 769, 1, 256] - - [46, 35.775] - - - [31488, 3072, 1, 256] - - [38, 72.752] - - - [33024, 1792, 1, 256] - - [60, 70.613] - - - [11520, 8449, 1, 256] - - [31, 71.344] - - - [44544, 4353, 1, 256] - - [33, 71.136] - - - [18176, 5889, 1, 256] - - [26, 70.944] - - - [5632, 2305, 1, 256] - - [35, 60.921] - - - [39936, 27393, 1, 256] - - [55, 73.828] - - - [10240, 7169, 1, 256] - - [38, 70.424] - - - [39168, 26625, 1, 256] - - [29, 72.658] - - - [10752, 7681, 1, 256] - - [28, 71.73] - - - [13824, 1536, 1, 256] - - [28, 68.41] - - - [14336, 9216, 1, 256] - - [38, 74.678] - - - [37632, 25345, 1, 256] - - [55, 73.074] - - - [35840, 23553, 1, 256] - - [33, 73.019] - - - [23552, 3072, 1, 256] - - [26, 73.223] - - - [19712, 7169, 1, 256] - - [55, 69.862] - - - [5888, 2561, 1, 256] - - [35, 63.656] - - - [27136, 768, 1, 256] - - [40, 67.459] - - - [22272, 1792, 1, 256] - - [35, 70.322] - - - [15616, 1536, 1, 256] - - [30, 67.46] - - - [3840, 769, 1, 256] - - [35, 46.289] - - - [42240, 2304, 1, 256] - - [38, 72.628] - - - [24576, 3072, 1, 256] - - [23, 68.65] - - - [27136, 1536, 1, 256] - - [32, 70.093] - - - [25344, 12801, 1, 256] - - [29, 72.186] - - - [32512, 20224, 1, 256] - - [48, 75.143] - - - [17664, 3072, 1, 256] - - [58, 71.451] - - - [28160, 15873, 1, 256] - - [33, 73.603] - - - [40960, 3072, 1, 256] - - [39, 66.819] - - - [14592, 9216, 1, 256] - - [72, 72.836] - - - [22784, 10497, 1, 256] - - [42, 72.571] - - - [22272, 3072, 1, 256] - - [58, 72.388] - - - [39680, 27137, 1, 256] - - [33, 72.405] - - - [20992, 8704, 1, 256] - - [38, 75.323] - - - [24320, 1536, 1, 256] - - [58, 69.964] - - - [7936, 4865, 1, 256] - - [58, 69.997] - - - [17664, 5376, 1, 256] - - [58, 73.382] - - - [37888, 25345, 1, 256] - - [55, 73.968] - - - [23296, 10753, 1, 256] - - [33, 72.528] - - - [28416, 15873, 1, 256] - - [71, 73.027] - - - [27648, 15361, 1, 256] - - [33, 72.759] - - - [39424, 1536, 1, 256] - - [32, 70.892] - - - [15104, 2817, 1, 256] - - [30, 68.216] - - - [19456, 9216, 1, 256] - - [26, 74.881] - - - [24064, 11777, 1, 256] - - [44, 73.397] - - - [40448, 1536, 1, 256] - - [31, 71.11] - - - [512, 3072, 1, 256] - - [118, 44.68] - - - [38912, 9216, 1, 256] - - [30, 74.717] - - - [19456, 6913, 1, 256] - - [28, 72.86] - - - [29440, 1792, 1, 256] - - [56, 71.532] - - - [41984, 9216, 1, 256] - - [30, 74.61] - - - [14080, 1793, 1, 256] - - [28, 65.232] - - - [20992, 8449, 1, 256] - - [33, 72.683] - - - [17920, 768, 1, 256] - - [41, 64.388] - - - [10496, 7169, 1, 256] - - [55, 69.631] - - - [40704, 27648, 1, 256] - - [28, 74.406] - - - [13568, 1025, 1, 256] - - [41, 59.023] - - - [38144, 9216, 1, 256] - - [42, 74.19] - - - [27392, 15104, 1, 256] - - [44, 74.726] - - - [2304, 3072, 1, 256] - - [50, 60.894] - - - [9472, 6401, 1, 256] - - [26, 70.416] - - - [39424, 1792, 1, 256] - - [60, 72.285] - - - [41728, 768, 1, 256] - - [73, 67.967] - - - [11264, 3072, 1, 256] - - [30, 70.634] - - - [25344, 3072, 1, 256] - - [58, 72.1] - - - [24576, 1792, 1, 256] - - [26, 67.901] - - - [27392, 14849, 1, 256] - - [42, 72.861] - - - [14848, 2561, 1, 256] - - [28, 66.512] - - - [28160, 3072, 1, 256] - - [28, 73.026] - - - [23552, 11009, 1, 256] - - [55, 73.699] - - - [11776, 8449, 1, 256] - - [55, 72.092] - - - [16640, 1792, 1, 256] - - [56, 67.869] - - - [24576, 12289, 1, 256] - - [33, 67.331] - - - [38656, 26369, 1, 256] - - [29, 72.925] - - - [13824, 9216, 1, 256] - - [62, 74.326] - - - [28928, 1792, 1, 256] - - [28, 71.293] - - - [27904, 15361, 1, 256] - - [42, 72.692] - - - [3840, 1792, 1, 256] - - [35, 60.099] - - - [14848, 3072, 1, 256] - - [38, 71.769] - - - [27904, 1536, 1, 256] - - [68, 70.041] - - - [34816, 1536, 1, 256] - - [28, 71.375] - - - [14592, 2305, 1, 256] - - [32, 66.345] - - - [22528, 9985, 1, 256] - - [30, 73.686] - - - [26368, 13825, 1, 256] - - [74, 72.918] - - - [4096, 1792, 1, 256] - - [35, 53.676] - - - [30720, 18177, 1, 256] - - [55, 74.079] - - - [37120, 24833, 1, 256] - - [42, 73.124] - - - [24320, 3072, 1, 256] - - [26, 72.794] - - - [2560, 1536, 1, 256] - - [56, 43.756] - - - [44032, 4097, 1, 256] - - [55, 70.115] - - - [44544, 27648, 1, 256] - - [29, 74.734] - - - [34048, 21761, 1, 256] - - [29, 72.608] - - - [24064, 1536, 1, 256] - - [50, 69.843] - - - [24832, 12545, 1, 256] - - [74, 73.308] - - - [44032, 3841, 1, 256] - - [26, 70.954] - - - [40448, 257, 1, 256] - - [37, 45.308] - - - [26624, 14337, 1, 256] - - [33, 72.796] - - - [8192, 5121, 1, 256] - - [26, 68.016] - - - [42240, 1536, 1, 256] - - [28, 70.765] - - - [5888, 2817, 1, 256] - - [56, 63.618] - - - [6144, 1792, 1, 256] - - [26, 60.128] - - - [16384, 1792, 1, 256] - - [23, 60.792] - - - [35584, 23297, 1, 256] - - [33, 73.151] - - - [36352, 24064, 1, 256] - - [29, 75.214] - - - [23040, 1536, 1, 256] - - [68, 69.287] - - - [8704, 1536, 1, 256] - - [56, 63.632] - - - [18432, 6145, 1, 256] - - [38, 70.988] - - - [12032, 3072, 1, 256] - - [28, 70.991] - - - [39168, 3072, 1, 256] - - [31, 72.818] - - - [28160, 1536, 1, 256] - - [55, 70.098] - - - [41728, 27648, 1, 256] - - [42, 73.797] - - - [28416, 1792, 1, 256] - - [32, 70.953] - - - [24320, 12032, 1, 256] - - [25, 75.433] - - - [28928, 16385, 1, 256] - - [55, 72.397] - - - [34816, 22528, 1, 256] - - [26, 75.054] - - - [26368, 1792, 1, 256] - - [56, 71.252] - - - [25856, 13569, 1, 256] - - [29, 72.951] - - - [25600, 13312, 1, 256] - - [28, 75.771] - - - [31232, 18689, 1, 256] - - [29, 73.883] - - - [20736, 9216, 1, 256] - - [38, 74.492] - - - [34304, 9216, 1, 256] - - [43, 75.021] - - - [43264, 3073, 1, 256] - - [55, 69.053] - - - [8704, 5633, 1, 256] - - [38, 70.581] - - - [4864, 1793, 1, 256] - - [35, 53.98] - - - [41984, 3072, 1, 256] - - [30, 73.928] - - - [20992, 3072, 1, 256] - - [26, 72.977] - - - [9728, 6401, 1, 256] - - [30, 70.891] - - - [16640, 4097, 1, 256] - - [55, 67.999] - - - [38400, 9216, 1, 256] - - [29, 74.802] - - - [38656, 1536, 1, 256] - - [32, 70.607] - - - [1536, 3072, 1, 256] - - [36, 50.841] - - - [12544, 1792, 1, 256] - - [35, 68.266] - - - [37632, 1792, 1, 256] - - [68, 71.968] - - - [17152, 4609, 1, 256] - - [28, 70.191] - - - [18944, 6656, 1, 256] - - [31, 74.667] - - - [34560, 22017, 1, 256] - - [33, 73.029] - - - [23296, 11008, 1, 256] - - [28, 74.913] - - - [14848, 768, 1, 256] - - [41, 60.74] - - - [38656, 1792, 1, 256] - - [32, 71.937] - - - [8448, 5377, 1, 256] - - [32, 70.346] - - - [29952, 17665, 1, 256] - - [74, 73.321] - - - [33792, 21504, 1, 256] - - [28, 75.406] - - - [24576, 1536, 1, 256] - - [39, 66.688] - - - [37376, 1792, 1, 256] - - [26, 72.422] - - - [42752, 768, 1, 256] - - [37, 68.776] - - - [4096, 1025, 1, 256] - - [36, 45.482] - - - [35840, 768, 1, 256] - - [61, 68.381] - - - [19200, 3072, 1, 256] - - [50, 71.676] - - - [33536, 1792, 1, 256] - - [35, 71.782] - - - [36864, 9216, 1, 256] - - [23, 74.017] - - - [38656, 26368, 1, 256] - - [29, 74.56] - - - [44288, 9216, 1, 256] - - [42, 74.399] - - - [44288, 4097, 1, 256] - - [55, 69.82] - - - [26112, 3072, 1, 256] - - [28, 73.247] - - - [512, 768, 1, 256] - - [109, 22.789] - - - [36096, 3072, 1, 256] - - [25, 71.493] - - - [4864, 1537, 1, 256] - - [56, 53.44] - - - [31232, 18944, 1, 256] - - [31, 75.689] - - - [20224, 7681, 1, 256] - - [29, 71.553] - - - [26112, 9216, 1, 256] - - [42, 74.995] - - - [21504, 3072, 1, 256] - - [26, 73.129] - - - [12544, 3072, 1, 256] - - [36, 71.024] - - - [32256, 19713, 1, 256] - - [71, 73.799] - - - [40704, 1792, 1, 256] - - [58, 72.126] - - - [18176, 5888, 1, 256] - - [35, 73.909] - - - [33792, 9216, 1, 256] - - [28, 74.743] - - - [26624, 14336, 1, 256] - - [38, 75.385] - - - [38912, 1792, 1, 256] - - [30, 72.82] - - - [7936, 1792, 1, 256] - - [35, 66.373] - - - [28672, 16385, 1, 256] - - [55, 72.361] - - - [18944, 3072, 1, 256] - - [60, 72.445] - - - [33280, 20993, 1, 256] - - [44, 73.865] - - - [37120, 24832, 1, 256] - - [28, 74.866] - - - [43520, 1792, 1, 256] - - [28, 72.876] - - - [16896, 4609, 1, 256] - - [31, 70.699] - - - [41472, 1536, 1, 256] - - [38, 71.132] - - - [39936, 768, 1, 256] - - [41, 68.67] - - - [23296, 11009, 1, 256] - - [55, 72.525] - - - [26624, 9216, 1, 256] - - [38, 74.827] - - - [29184, 9216, 1, 256] - - [75, 74.75] - - - [36352, 9216, 1, 256] - - [29, 74.823] - - - [37632, 25344, 1, 256] - - [30, 74.856] - - - [37888, 25600, 1, 256] - - [28, 75.356] - - - [16640, 9216, 1, 256] - - [42, 74.324] - - - [44544, 9216, 1, 256] - - [29, 74.614] - - - [14080, 1792, 1, 256] - - [32, 68.979] - - - [33536, 21249, 1, 256] - - [71, 73.389] - - - [34048, 21760, 1, 256] - - [65, 74.55] - - - [9984, 768, 1, 256] - - [24, 54.574] - - - [40192, 1536, 1, 256] - - [26, 71.09] - - - [41728, 3072, 1, 256] - - [31, 71.64] - - - [35328, 9216, 1, 256] - - [44, 74.937] - - - [32512, 768, 1, 256] - - [40, 67.213] - - - [14592, 2049, 1, 256] - - [40, 63.564] - - - [14848, 9216, 1, 256] - - [42, 74.684] - - - [23808, 3072, 1, 256] - - [26, 72.789] - - - [13568, 9216, 1, 256] - - [30, 74.107] - - - [42496, 2560, 1, 256] - - [26, 73.708] - - - [42752, 3072, 1, 256] - - [26, 73.224] - - - [39680, 27392, 1, 256] - - [38, 74.582] - - - [14592, 1792, 1, 256] - - [58, 68.208] - - - [25600, 13313, 1, 256] - - [33, 72.762] - - - [26624, 1792, 1, 256] - - [28, 72.142] - - - [20480, 8193, 1, 256] - - [55, 71.705] - - - [36096, 23808, 1, 256] - - [42, 74.247] - - - [15104, 2561, 1, 256] - - [41, 67.093] - - - [43520, 3072, 1, 256] - - [28, 73.788] - - - [1280, 2048, 1, 256] - - [58, 42.367] - - - [43008, 1792, 1, 256] - - [26, 72.967] - - - [18688, 3072, 1, 256] - - [26, 72.245] - - - [35328, 23040, 1, 256] - - [52, 75.432] - - - [18944, 6401, 1, 256] - - [74, 71.873] - - - [16128, 3585, 1, 256] - - [26, 69.134] - - - [29952, 1536, 1, 256] - - [38, 70.233] - - - [17408, 5121, 1, 256] - - [55, 70.042] - - - [36608, 1792, 1, 256] - - [35, 71.769] - - - [13056, 768, 1, 256] - - [54, 61.047] - - - [26112, 13824, 1, 256] - - [66, 75.824] - - - [43520, 3585, 1, 256] - - [55, 70.545] - - - [40704, 9216, 1, 256] - - [38, 74.131] - - - [27904, 15617, 1, 256] - - [43, 73.165] - - - [21248, 3072, 1, 256] - - [50, 72.225] - - - [38912, 1536, 1, 256] - - [30, 71.532] - - - [28672, 1792, 1, 256] - - [28, 71.832] - - - [18432, 1792, 1, 256] - - [26, 70.635] - - - [29952, 9216, 1, 256] - - [48, 74.232] - - - [4352, 1025, 1, 256] - - [24, 47.919] - - - [34304, 22017, 1, 256] - - [44, 73.821] - - - [28160, 15617, 1, 256] - - [55, 73.595] - - - [19968, 9216, 1, 256] - - [29, 74.953] - - - [7424, 4353, 1, 256] - - [35, 68.112] - - - [19200, 1792, 1, 256] - - [60, 69.134] - - - [27648, 15360, 1, 256] - - [30, 75.75] - - - [23040, 10497, 1, 256] - - [71, 73.426] - - - [21248, 8961, 1, 256] - - [71, 72.223] - - - [32256, 1792, 1, 256] - - [58, 72.227] - - - [26112, 13569, 1, 256] - - [44, 73.707] - - - [12288, 8961, 1, 256] - - [38, 72.875] - - - [6656, 3585, 1, 256] - - [60, 66.739] - - - [19968, 7425, 1, 256] - - [55, 72.526] - - - [9472, 768, 1, 256] - - [54, 52.718] - - - [33792, 3072, 1, 256] - - [30, 73.595] - - - [15616, 3072, 1, 256] - - [38, 71.795] - - - [8704, 5377, 1, 256] - - [28, 70.194] - - - [11520, 3072, 1, 256] - - [68, 70.214] - - - [25856, 1536, 1, 256] - - [36, 69.702] - - - [28416, 768, 1, 256] - - [41, 66.259] - - - [32256, 3072, 1, 256] - - [26, 73.426] - - - [20736, 1536, 1, 256] - - [56, 69.294] - - - [22784, 10241, 1, 256] - - [42, 71.799] - - - [36608, 24321, 1, 256] - - [42, 72.971] - - - [36096, 9216, 1, 256] - - [29, 73.759] - - - [10752, 768, 1, 256] - - [41, 58.055] - - - [38400, 26112, 1, 256] - - [28, 75.376] - - - [9216, 5889, 1, 256] - - [56, 70.705] - - - [41472, 27648, 1, 256] - - [44, 74.643] - - - [38144, 25856, 1, 256] - - [38, 74.767] - - - [15360, 3073, 1, 256] - - [28, 67.924] - - - [29184, 16896, 1, 256] - - [26, 75.487] - - - [16128, 1792, 1, 256] - - [35, 68.836] - - - [32768, 20225, 1, 256] - - [96, 57.927] - - - [23040, 10752, 1, 256] - - [48, 75.65] - - - [15872, 3585, 1, 256] - - [26, 69.965] - - - [11008, 7681, 1, 256] - - [76, 70.59] - - - [15360, 9216, 1, 256] - - [38, 74.876] - - - [28416, 16128, 1, 256] - - [52, 74.891] - - - [30208, 1792, 1, 256] - - [30, 71.803] - - - [41728, 1792, 1, 256] - - [65, 71.002] - - - [32256, 19968, 1, 256] - - [52, 75.446] - - - [18944, 1792, 1, 256] - - [60, 70.622] - - - [41728, 1793, 1, 256] - - [26, 66.538] - - - [31488, 19201, 1, 256] - - [33, 73.037] - - - [40192, 257, 1, 256] - - [26, 44.954] - - - [42752, 27648, 1, 256] - - [30, 74.278] - - - [40704, 768, 1, 256] - - [34, 68.398] - - - [25088, 12545, 1, 256] - - [25, 73.619] - - - [24576, 9216, 1, 256] - - [39, 70.049] - - - [33024, 20737, 1, 256] - - [74, 73.586] - - - [29696, 9216, 1, 256] - - [28, 74.863] - - - [31232, 1536, 1, 256] - - [26, 70.863] - - - [30208, 17920, 1, 256] - - [25, 75.773] - - - [44544, 4609, 1, 256] - - [55, 71.087] - - - [22016, 9728, 1, 256] - - [38, 75.247] - - - [30208, 17921, 1, 256] - - [43, 73.75] - - - [19200, 6657, 1, 256] - - [71, 71.173] - - - [22016, 9729, 1, 256] - - [25, 73.212] - - - [18176, 768, 1, 256] - - [24, 64.992] - - - [29184, 1792, 1, 256] - - [60, 71.58] - - - [12288, 1792, 1, 256] - - [36, 67.603] - - - [22528, 1536, 1, 256] - - [35, 69.808] - - - [14848, 2305, 1, 256] - - [26, 66.318] - - - [41216, 1025, 1, 256] - - [70, 63.189] - - - [8192, 3072, 1, 256] - - [28, 69.015] - - - [5888, 1792, 1, 256] - - [35, 64.218] - - - [21760, 3072, 1, 256] - - [36, 72.375] - - - [22272, 9985, 1, 256] - - [74, 72.608] - - - [29184, 1536, 1, 256] - - [32, 70.393] - - - [22016, 3072, 1, 256] - - [48, 72.804] - - - [30720, 9216, 1, 256] - - [28, 74.894] - - - [39680, 1792, 1, 256] - - [30, 72.047] - - - [9728, 1536, 1, 256] - - [36, 63.991] - - - [34560, 9216, 1, 256] - - [38, 74.255] - - - [12032, 8705, 1, 256] - - [38, 71.355] - - - [10752, 7425, 1, 256] - - [52, 71.821] - - - [18688, 1536, 1, 256] - - [56, 69.385] - - - [16128, 3840, 1, 256] - - [56, 72.828] - - - [38656, 768, 1, 256] - - [34, 68.285] - - - [21248, 1792, 1, 256] - - [35, 70.279] - - - [36352, 3072, 1, 256] - - [28, 73.493] - - - [19968, 7680, 1, 256] - - [30, 75.125] - - - [3840, 513, 1, 256] - - [35, 32.748] - - - [38400, 3072, 1, 256] - - [38, 73.324] - - - [5376, 768, 1, 256] - - [50, 44.727] - - - [20224, 9216, 1, 256] - - [38, 74.585] - - - [17408, 5120, 1, 256] - - [28, 74.321] - - - [28928, 9216, 1, 256] - - [26, 74.117] - - - [35072, 1792, 1, 256] - - [30, 71.863] - - - [31488, 19200, 1, 256] - - [28, 75.016] - - - [11008, 7937, 1, 256] - - [72, 70.797] - - - [21248, 8705, 1, 256] - - [74, 72.196] - - - [13568, 3072, 1, 256] - - [35, 71.084] - - - [34560, 22273, 1, 256] - - [33, 73.128] - - - [34048, 768, 1, 256] - - [41, 67.567] - - - [40448, 27648, 1, 256] - - [43, 74.661] - - - [28416, 16129, 1, 256] - - [55, 73.072] - - - [34816, 22529, 1, 256] - - [33, 72.704] - - - [22528, 3072, 1, 256] - - [28, 73.326] - - - [27136, 14593, 1, 256] - - [71, 73.833] - - - [35584, 3072, 1, 256] - - [30, 72.72] - - - [43008, 3072, 1, 256] - - [28, 73.831] - - - [30464, 1792, 1, 256] - - [32, 70.013] - - - [16384, 4097, 1, 256] - - [28, 59.138] - - - [20992, 9216, 1, 256] - - [42, 75.069] - - - [31488, 1792, 1, 256] - - [35, 71.727] - - - [31488, 9216, 1, 256] - - [28, 74.117] - - - [22272, 9984, 1, 256] - - [58, 75.013] - - - [41728, 1537, 1, 256] - - [28, 65.171] - - - [26880, 1792, 1, 256] - - [28, 71.327] - - - [30464, 768, 1, 256] - - [34, 66.682] - - - [2816, 1792, 1, 256] - - [36, 54.473] - - - [41472, 1537, 1, 256] - - [38, 66.289] - - - [43008, 27648, 1, 256] - - [26, 74.892] - - - [39424, 27137, 1, 256] - - [42, 73.628] - - - [24320, 1792, 1, 256] - - [32, 70.912] - - - [32000, 3072, 1, 256] - - [26, 72.768] - - - [12800, 1792, 1, 256] - - [32, 68.573] - - - [15872, 3072, 1, 256] - - [26, 72.001] - - - [15872, 1792, 1, 256] - - [36, 70.342] - - - [10496, 7425, 1, 256] - - [55, 71.106] - - - [16896, 4608, 1, 256] - - [26, 73.401] - - - [9984, 6913, 1, 256] - - [26, 71.091] - - - [21248, 8960, 1, 256] - - [28, 74.939] - - - [14336, 1792, 1, 256] - - [50, 69.041] - - - [24832, 12544, 1, 256] - - [32, 75.359] - - - [30464, 18176, 1, 256] - - [29, 74.481] - - - [31744, 19201, 1, 256] - - [28, 74.065] - - - [1792, 768, 1, 256] - - [125, 40.595] - - - [1536, 2048, 1, 256] - - [58, 49.311] - - - [40192, 3072, 1, 256] - - [38, 73.329] - - - [42240, 3072, 1, 256] - - [30, 73.139] - - - [32256, 9216, 1, 256] - - [43, 74.881] - - - [41984, 2049, 1, 256] - - [75, 66.57] - - - [6656, 1792, 1, 256] - - [60, 63.224] - - - [13824, 1537, 1, 256] - - [28, 63.771] - - - [20736, 3072, 1, 256] - - [28, 72.45] - - - [36096, 23809, 1, 256] - - [42, 72.71] - - - [41728, 9216, 1, 256] - - [29, 73.108] - - - [25600, 768, 1, 256] - - [61, 67.051] - - - [37632, 768, 1, 256] - - [24, 68.035] - - - [25600, 9216, 1, 256] - - [30, 75.002] - - - [19968, 3072, 1, 256] - - [30, 72.482] - - - [15616, 9216, 1, 256] - - [30, 74.411] - - - [29184, 16897, 1, 256] - - [33, 73.625] - - - [7168, 3841, 1, 256] - - [30, 67.57] - - - [40704, 769, 1, 256] - - [34, 60.707] - - - [6144, 3073, 1, 256] - - [38, 64.657] - - - [34304, 1792, 1, 256] - - [60, 72.252] - - - [18688, 6400, 1, 256] - - [38, 74.512] - - - [20992, 1536, 1, 256] - - [28, 69.992] - - - [21760, 768, 1, 256] - - [61, 64.205] - - - [43264, 3072, 1, 256] - - [26, 73.263] - - - [21760, 9216, 1, 256] - - [26, 74.502] - - - [11264, 768, 1, 256] - - [54, 60.064] - - - [42496, 3072, 1, 256] - - [38, 73.716] - - - [30208, 17665, 1, 256] - - [74, 73.922] - - - [27392, 15105, 1, 256] - - [29, 72.88] - - - [29952, 17409, 1, 256] - - [74, 72.467] - - - [44032, 3072, 1, 256] - - [38, 73.868] - - - [41216, 9216, 1, 256] - - [42, 74.596] - - - [8448, 1536, 1, 256] - - [50, 61.889] - - - [36352, 768, 1, 256] - - [64, 68.459] - - - [23552, 768, 1, 256] - - [59, 64.869] - - - [7168, 3072, 1, 256] - - [36, 67.869] - - - [44288, 4353, 1, 256] - - [55, 70.362] - - - [36608, 768, 1, 256] - - [64, 68.316] - - - [15616, 3073, 1, 256] - - [28, 68.049] - - - [37376, 24833, 1, 256] - - [29, 73.814] - - - [38144, 25857, 1, 256] - - [29, 72.949] - - - [26880, 14592, 1, 256] - - [30, 75.253] - - - [6144, 2817, 1, 256] - - [38, 62.577] - - - [23808, 768, 1, 256] - - [64, 64.822] - - - [39168, 26881, 1, 256] - - [42, 72.928] - - - [5120, 1793, 1, 256] - - [50, 55.962] - - - [32512, 19969, 1, 256] - - [44, 73.225] - - - [43008, 2817, 1, 256] - - [26, 69.9] - - - [26112, 13825, 1, 256] - - [44, 73.821] - - - [33536, 3072, 1, 256] - - [52, 72.881] - - - [9728, 6657, 1, 256] - - [28, 71.388] - - - [2048, 3072, 1, 256] - - [50, 55.21] - - - [24832, 9216, 1, 256] - - [30, 74.486] - - - [5632, 2561, 1, 256] - - [28, 61.714] - - - [33280, 20992, 1, 256] - - [52, 75.582] - - - [20224, 7936, 1, 256] - - [28, 74.62] - - - [28672, 16384, 1, 256] - - [28, 74.944] - - - [28416, 9216, 1, 256] - - [33, 73.936] - - - [7936, 768, 1, 256] - - [28, 52.95] - - - [23552, 11265, 1, 256] - - [26, 72.46] - - - [25088, 3072, 1, 256] - - [30, 72.818] - - - [32000, 19457, 1, 256] - - [33, 72.725] - - - [44800, 3072, 1, 256] - - [26, 72.719] - - - [37120, 1792, 1, 256] - - [28, 71.806] - - - [30464, 18177, 1, 256] - - [29, 72.706] - - - [44544, 4608, 1, 256] - - [29, 73.899] - - - [7168, 768, 1, 256] - - [28, 49.311] - - - [18944, 9216, 1, 256] - - [43, 74.773] - - - [33280, 20737, 1, 256] - - [71, 73.951] - - - [25856, 3072, 1, 256] - - [50, 72.641] - - - [27648, 9216, 1, 256] - - [38, 74.837] - - - [5120, 2049, 1, 256] - - [77, 55.943] - - - [28160, 9216, 1, 256] - - [49, 74.804] - - - [37632, 25089, 1, 256] - - [33, 73.058] - - - [22016, 1792, 1, 256] - - [30, 71.088] - - - [16384, 9216, 1, 256] - - [23, 63.2] - - - [21504, 9217, 1, 256] - - [33, 72.143] - - - [20480, 7937, 1, 256] - - [33, 72.788] - - - [33536, 21248, 1, 256] - - [52, 75.086] - - - [12800, 768, 1, 256] - - [41, 60.244] - - - [28672, 9216, 1, 256] - - [38, 74.277] - - - [32000, 9216, 1, 256] - - [28, 74.163] - - - [44544, 3072, 1, 256] - - [26, 73.355] - - - [5376, 3072, 1, 256] - - [26, 64.581] - - - [35840, 23297, 1, 256] - - [33, 74.027] - - - [23808, 11521, 1, 256] - - [33, 72.524] - - - [13312, 1025, 1, 256] - - [37, 58.916] - - - [18176, 9216, 1, 256] - - [38, 74.483] - - - [17920, 5633, 1, 256] - - [60, 71.536] - - - [27648, 3072, 1, 256] - - [26, 73.527] - - - [1024, 3072, 1, 256] - - [73, 49.177] - - - [22016, 9216, 1, 256] - - [42, 75.131] - - - [21760, 9473, 1, 256] - - [33, 72.475] - - - [6144, 1536, 1, 256] - - [35, 58.455] - - - [16896, 1536, 1, 256] - - [56, 68.548] - - - [19968, 768, 1, 256] - - [41, 64.87] - - - [23552, 11264, 1, 256] - - [38, 75.562] - - - [27904, 3072, 1, 256] - - [60, 72.351] - - - [19712, 7425, 1, 256] - - [74, 71.126] - - - [26624, 14081, 1, 256] - - [38, 73.889] - - - [3328, 257, 1, 256] - - [111, 31.019] - - - [24320, 9216, 1, 256] - - [48, 74.48] - - - [14080, 3072, 1, 256] - - [60, 70.439] - - - [17408, 3072, 1, 256] - - [38, 72.646] - - - [21504, 9216, 1, 256] - - [28, 74.988] - - - [14848, 2560, 1, 256] - - [35, 71.053] - - - [34304, 3072, 1, 256] - - [48, 73.279] - - - [15104, 9216, 1, 256] - - [26, 74.312] - - - [17152, 4865, 1, 256] - - [26, 70.61] - - - [38912, 26625, 1, 256] - - [55, 72.727] - - - [41216, 1792, 1, 256] - - [26, 71.942] - - - [39424, 3072, 1, 256] - - [30, 73.417] - - - [30720, 18433, 1, 256] - - [33, 72.945] - - - [18944, 6657, 1, 256] - - [25, 72.217] - - - [5632, 1792, 1, 256] - - [35, 61.719] - - - [18176, 1792, 1, 256] - - [56, 69.577] - - - [31232, 9216, 1, 256] - - [29, 74.854] - - - [42752, 2561, 1, 256] - - [33, 68.13] - - - [18688, 9216, 1, 256] - - [38, 74.553] - - - [43776, 1792, 1, 256] - - [42, 69.792] - - - [10240, 1792, 1, 256] - - [36, 66.547] - - - [33792, 21505, 1, 256] - - [33, 72.863] - - - [25856, 13313, 1, 256] - - [29, 72.38] - - - [29952, 3072, 1, 256] - - [31, 72.875] - - - [5888, 768, 1, 256] - - [56, 48.723] - - - [20480, 9216, 1, 256] - - [38, 74.621] - - - [17152, 1792, 1, 256] - - [35, 69.112] - - - [23040, 10753, 1, 256] - - [74, 73.355] - - - [8960, 5889, 1, 256] - - [58, 70.343] - - - [16640, 4352, 1, 256] - - [35, 72.785] - - - [30464, 3072, 1, 256] - - [58, 71.363] - - - [16128, 9216, 1, 256] - - [26, 74.406] - - - [25344, 13057, 1, 256] - - [42, 72.612] - - - [39424, 9216, 1, 256] - - [42, 74.936] - - - [25600, 3072, 1, 256] - - [30, 73.346] - - - [28416, 3072, 1, 256] - - [48, 72.357] - - - [12800, 257, 1, 256] - - [61, 36.642] - - - [43264, 1792, 1, 256] - - [26, 72.181] - - - [20736, 8193, 1, 256] - - [55, 71.591] - - - [30976, 9216, 1, 256] - - [42, 73.289] - - - [40192, 27648, 1, 256] - - [26, 74.477] - - - [31232, 1792, 1, 256] - - [36, 71.89] - - - [36352, 23809, 1, 256] - - [29, 73.868] - - - [9984, 3072, 1, 256] - - [26, 69.262] - - - [11776, 1792, 1, 256] - - [58, 68.461] - - - [37120, 1536, 1, 256] - - [50, 70.444] - - - [14592, 2304, 1, 256] - - [60, 69.777] - - - [7424, 768, 1, 256] - - [61, 50.368] - - - [10240, 1536, 1, 256] - - [50, 65.762] - - - [27392, 9216, 1, 256] - - [42, 74.083] - - - [15104, 3072, 1, 256] - - [35, 71.582] - - - [19200, 9216, 1, 256] - - [29, 74.336] - - - [36096, 23553, 1, 256] - - [29, 72.256] - - - [16128, 3841, 1, 256] - - [28, 69.422] - - - [18432, 5889, 1, 256] - - [26, 72.062] - - - [43776, 3841, 1, 256] - - [29, 69.142] - - - [22528, 10241, 1, 256] - - [55, 72.313] - - - [20224, 3072, 1, 256] - - [30, 72.17] - - - [39680, 3072, 1, 256] - - [38, 72.868] - - - [20736, 8449, 1, 256] - - [33, 72.238] - - - [30720, 1792, 1, 256] - - [56, 72.468] - - - [36864, 24321, 1, 256] - - [33, 73.196] - - - [22784, 1536, 1, 256] - - [35, 69.341] - - - [7424, 4097, 1, 256] - - [26, 66.537] - - - [7680, 4609, 1, 256] - - [26, 68.021] - - - [12032, 768, 1, 256] - - [64, 57.752] - - - [1792, 3072, 1, 256] - - [56, 49.169] - - - [6400, 3073, 1, 256] - - [26, 63.363] - - - [29440, 17153, 1, 256] - - [74, 73.313] - - - [8704, 1792, 1, 256] - - [28, 65.844] - - - [30720, 3072, 1, 256] - - [30, 73.577] - - - [16384, 3841, 1, 256] - - [39, 58.468] - - - [40192, 9216, 1, 256] - - [30, 74.318] - - - [23040, 1792, 1, 256] - - [50, 71.001] - - - [37888, 25601, 1, 256] - - [55, 72.878] - - - [26368, 14080, 1, 256] - - [26, 75.251] - - - [30208, 3072, 1, 256] - - [30, 72.997] - - - [33024, 20736, 1, 256] - - [44, 75.079] - - - [35072, 22784, 1, 256] - - [26, 74.903] - - - [9472, 6145, 1, 256] - - [28, 69.301] - - - [22784, 1792, 1, 256] - - [35, 70.849] - - - [768, 2048, 1, 256] - - [113, 44.352] - - - [1024, 1280, 1, 256] - - [126, 43.314] - - - [41984, 27648, 1, 256] - - [38, 74.925] - - - [33024, 20481, 1, 256] - - [43, 72.827] - - - [33280, 1536, 1, 256] - - [50, 71.098] - - - [9216, 3072, 1, 256] - - [35, 70.425] - - - [22528, 1792, 1, 256] - - [35, 71.567] - - - [25088, 768, 1, 256] - - [24, 66.789] - - - [13825, 128, 1, 128] - - [127, 26.713] - - - [20609, 128, 1, 256] - - [207, 33.58] - - - [6017, 128, 1, 256] - - [128, 30.524] - - - [2305, 128, 1, 128] - - [129, 10.976] - - - [15745, 128, 1, 256] - - [207, 27.345] - - - [8833, 128, 1, 128] - - [130, 23.452] - - - [641, 128, 1, 128] - - [131, 3.197] - - - [9217, 128, 1, 128] - - [132, 23.704] - - - [15361, 128, 1, 256] - - [220, 26.647] - - - [22913, 128, 1, 256] - - [208, 36.177] - - - [2177, 128, 1, 128] - - [119, 10.367] - - - [19073, 128, 1, 256] - - [215, 31.601] - - - [28289, 128, 1, 128] - - [207, 20.694] - - - [13057, 128, 1, 256] - - [121, 37.574] - - - [1793, 128, 1, 128] - - [133, 8.538] - - - [16769, 128, 1, 128] - - [216, 18.474] - - - [23681, 128, 1, 256] - - [209, 36.993] - - - [14593, 128, 1, 256] - - [212, 25.525] - - - [24449, 128, 1, 128] - - [208, 24.486] - - - [4609, 128, 1, 256] - - [134, 24.923] - - - [10625, 128, 1, 128] - - [135, 23.134] - - - [12545, 128, 1, 256] - - [136, 36.242] - - - [5633, 128, 1, 128] - - [137, 18.479] - - - [641, 128, 1, 256] - - [134, 4.731] - - - [18305, 128, 1, 256] - - [208, 30.812] - - - [23297, 128, 1, 256] - - [219, 36.822] - - - [21377, 128, 1, 256] - - [209, 34.302] - - - [9601, 128, 1, 128] - - [138, 24.866] - - - [13697, 128, 1, 256] - - [139, 39.415] - - - [23681, 128, 1, 128] - - [211, 23.717] - - - [24833, 128, 1, 256] - - [210, 38.387] - - - [25985, 128, 1, 128] - - [216, 25.37] - - - [9601, 128, 1, 256] - - [140, 35.612] - - - [17153, 128, 1, 128] - - [211, 18.729] - - - [9985, 128, 1, 128] - - [141, 25.327] - - - [23297, 128, 1, 128] - - [216, 23.719] - - - [19073, 128, 1, 128] - - [212, 20.373] - - - [2689, 128, 1, 256] - - [119, 18.463] - - - [4993, 128, 1, 128] - - [119, 19.814] - - - [6913, 128, 1, 256] - - [142, 33.681] - - - [6785, 128, 1, 128] - - [133, 22.258] - - - [27905, 128, 1, 128] - - [207, 26.646] - - - [7169, 128, 1, 256] - - [143, 27.705] - - - [11905, 128, 1, 256] - - [138, 35.504] - - - [1409, 128, 1, 128] - - [144, 7.027] - - - [12673, 128, 1, 128] - - [140, 26.276] - - - [27521, 128, 1, 256] - - [213, 41.204] - - - [1409, 128, 1, 256] - - [133, 10.4] - - - [25217, 128, 1, 128] - - [216, 24.917] - - - [7297, 128, 1, 128] - - [138, 20.023] - - - [14081, 128, 1, 128] - - [207, 15.796] - - - [22913, 128, 1, 128] - - [212, 23.168] - - - [10753, 128, 1, 256] - - [145, 32.662] - - - [7937, 128, 1, 128] - - [137, 21.46] - - - [11393, 128, 1, 128] - - [146, 24.445] - - - [26369, 128, 1, 128] - - [207, 25.951] - - - [12161, 128, 1, 256] - - [138, 36.415] - - - [8449, 128, 1, 128] - - [130, 22.514] - - - [22145, 128, 1, 256] - - [214, 35.265] - - - [20225, 128, 1, 256] - - [214, 33.174] - - - [10241, 128, 1, 256] - - [147, 35.572] - - - [6913, 128, 1, 128] - - [133, 21.806] - - - [4993, 128, 1, 256] - - [148, 26.999] - - - [6401, 128, 1, 256] - - [149, 32.584] - - - [13057, 128, 1, 128] - - [150, 26.404] - - - [2945, 128, 1, 128] - - [151, 13.846] - - - [3713, 128, 1, 256] - - [119, 20.452] - - - [10753, 128, 1, 128] - - [152, 23.073] - - - [14849, 128, 1, 256] - - [208, 26.003] - - - [3841, 128, 1, 128] - - [119, 15.578] - - - [28289, 128, 1, 256] - - [215, 32.474] - - - [12929, 128, 1, 128] - - [153, 26.806] - - - [14081, 128, 1, 256] - - [211, 24.426] - - - [14977, 128, 1, 256] - - [208, 26.165] - - - [12545, 128, 1, 128] - - [150, 25.865] - - - [16129, 128, 1, 256] - - [208, 27.59] - - - [11777, 128, 1, 256] - - [145, 34.359] - - - [11777, 128, 1, 128] - - [154, 24.695] - - - [17537, 128, 1, 256] - - [207, 29.553] - - - [5377, 128, 1, 128] - - [155, 17.639] - - - [8065, 128, 1, 256] - - [140, 31.167] - - - [6145, 128, 1, 128] - - [149, 19.981] - - - [20993, 128, 1, 128] - - [215, 21.979] - - - [15617, 128, 1, 128] - - [207, 17.518] - - - [5633, 128, 1, 256] - - [156, 28.675] - - - [4865, 128, 1, 128] - - [134, 19.203] - - - [385, 128, 1, 256] - - [129, 2.856] - - - [3841, 128, 1, 256] - - [148, 21.236] - - - [8833, 128, 1, 256] - - [153, 33.694] - - - [4225, 128, 1, 128] - - [129, 17.135] - - - [11009, 128, 1, 256] - - [153, 33.235] - - - [385, 128, 1, 128] - - [129, 1.973] - - - [9473, 128, 1, 256] - - [153, 35.137] - - - [5761, 128, 1, 128] - - [133, 19.155] - - - [11905, 128, 1, 128] - - [150, 25.324] - - - [4097, 128, 1, 256] - - [157, 22.568] - - - [25217, 128, 1, 256] - - [209, 38.617] - - - [9089, 128, 1, 256] - - [158, 34.055] - - - [10369, 128, 1, 256] - - [153, 37.425] - - - [14209, 128, 1, 256] - - [214, 24.794] - - - [6401, 128, 1, 128] - - [159, 20.814] - - - [27137, 128, 1, 256] - - [209, 41.004] - - - [16385, 128, 1, 256] - - [211, 28.591] - - - [24833, 128, 1, 128] - - [208, 24.538] - - - [18689, 128, 1, 128] - - [220, 19.906] - - - [7553, 128, 1, 256] - - [156, 29.265] - - - [8321, 128, 1, 128] - - [152, 22.334] - - - [15361, 128, 1, 128] - - [207, 17.337] - - - [1153, 128, 1, 128] - - [133, 5.75] - - - [1025, 128, 1, 128] - - [129, 5.182] - - - [19841, 128, 1, 256] - - [217, 32.4] - - - [15233, 128, 1, 128] - - [207, 17.088] - - - [21761, 128, 1, 256] - - [208, 34.842] - - - [17153, 128, 1, 256] - - [212, 29.478] - - - [15617, 128, 1, 256] - - [215, 27.123] - - - [4865, 128, 1, 256] - - [133, 26.404] - - - [14209, 128, 1, 128] - - [216, 16.135] - - - [19457, 128, 1, 256] - - [207, 32.31] - - - [9857, 128, 1, 256] - - [160, 36.107] - - - [11521, 128, 1, 128] - - [149, 24.366] - - - [8449, 128, 1, 256] - - [141, 32.146] - - - [4097, 128, 1, 128] - - [119, 16.708] - - - [28673, 128, 1, 256] - - [212, 31.898] - - - [12161, 128, 1, 128] - - [152, 25.868] - - - [1921, 128, 1, 256] - - [129, 13.314] - - - [9985, 128, 1, 256] - - [141, 36.306] - - - [7937, 128, 1, 256] - - [152, 30.434] - - - [9857, 128, 1, 128] - - [132, 25.528] - - - [13825, 128, 1, 256] - - [139, 39.476] - - - [9089, 128, 1, 128] - - [132, 24.22] - - - [6785, 128, 1, 256] - - [142, 33.387] - - - [5249, 128, 1, 256] - - [152, 27.094] - - - [7681, 128, 1, 256] - - [140, 29.376] - - - [3329, 128, 1, 128] - - [129, 15.075] - - - [14465, 128, 1, 128] - - [216, 16.375] - - - [11137, 128, 1, 256] - - [152, 34.039] - - - [1153, 128, 1, 256] - - [133, 8.596] - - - [16001, 128, 1, 128] - - [211, 17.707] - - - [26753, 128, 1, 128] - - [207, 25.881] - - - [13697, 128, 1, 128] - - [150, 27.251] - - - [3073, 128, 1, 128] - - [151, 14.356] - - - [22529, 128, 1, 256] - - [208, 35.8] - - - [18689, 128, 1, 256] - - [221, 30.86] - - - [257, 128, 1, 128] - - [144, 1.282] - - - [15233, 128, 1, 256] - - [212, 26.581] - - - [27521, 128, 1, 128] - - [215, 26.45] - - - [16385, 128, 1, 128] - - [212, 18.492] - - - [4481, 128, 1, 256] - - [161, 24.683] - - - [6017, 128, 1, 128] - - [162, 19.916] - - - [7297, 128, 1, 256] - - [128, 28.273] - - - [7553, 128, 1, 128] - - [163, 20.572] - - - [21761, 128, 1, 128] - - [215, 22.216] - - - [11393, 128, 1, 256] - - [138, 34.324] - - - [11521, 128, 1, 256] - - [164, 34.428] - - - [12929, 128, 1, 256] - - [160, 37.424] - - - [20225, 128, 1, 128] - - [211, 21.175] - - - [13313, 128, 1, 128] - - [165, 25.928] - - - [2561, 128, 1, 128] - - [166, 12.04] - - - [1537, 128, 1, 128] - - [134, 7.718] - - - [24449, 128, 1, 256] - - [224, 37.636] - - - [12289, 128, 1, 256] - - [145, 35.432] - - - [4225, 128, 1, 256] - - [119, 23.101] - - - [26369, 128, 1, 256] - - [218, 40.172] - - - [17921, 128, 1, 256] - - [216, 30.409] - - - [2945, 128, 1, 256] - - [119, 20.127] - - - [24065, 128, 1, 128] - - [208, 24.069] - - - [6529, 128, 1, 128] - - [134, 21.514] - - - [6145, 128, 1, 256] - - [152, 30.96] - - - [25985, 128, 1, 256] - - [214, 39.546] - - - [8705, 128, 1, 256] - - [141, 32.533] - - - [384, 128, 1, 256] - - [167, 3.132] - - - [25601, 128, 1, 256] - - [219, 38.882] - - - [28673, 128, 1, 128] - - [208, 20.508] - - - [20609, 128, 1, 128] - - [211, 21.547] - - - [19457, 128, 1, 128] - - [220, 20.546] - - - [16769, 128, 1, 256] - - [214, 28.553] - - - [12673, 128, 1, 256] - - [112, 36.9] - - - [8321, 128, 1, 256] - - [164, 31.823] - - - [5249, 128, 1, 128] - - [133, 17.854] - - - [16129, 128, 1, 128] - - [207, 18.011] - - - [13441, 128, 1, 256] - - [113, 38.454] - - - [5377, 128, 1, 256] - - [164, 27.466] - - - [21377, 128, 1, 128] - - [208, 21.885] - - - [14465, 128, 1, 256] - - [215, 25.211] - - - [11137, 128, 1, 128] - - [165, 24.107] - - - [7681, 128, 1, 128] - - [132, 20.692] - - - [7169, 128, 1, 128] - - [128, 19.671] - - - [22145, 128, 1, 128] - - [207, 22.957] - - - [11009, 128, 1, 128] - - [168, 23.76] - - - [20993, 128, 1, 256] - - [215, 34.018] - - - [13313, 128, 1, 256] - - [169, 37.292] - - - [25601, 128, 1, 128] - - [216, 24.669] - - - [4609, 128, 1, 128] - - [133, 18.29] - - - [5761, 128, 1, 256] - - [149, 29.529] - - - [17921, 128, 1, 128] - - [207, 19.773] - - - [2689, 128, 1, 128] - - [144, 12.405] - - - [8705, 128, 1, 128] - - [132, 22.785] - - - [10241, 128, 1, 128] - - [154, 25.624] - - - [14977, 128, 1, 128] - - [208, 16.955] - - - [18305, 128, 1, 128] - - [216, 19.753] - - - [3457, 128, 1, 128] - - [129, 15.372] - - - [24065, 128, 1, 256] - - [208, 37.356] - - - [12289, 128, 1, 128] - - [150, 25.127] - - - [14593, 128, 1, 128] - - [207, 16.546] - - - [2177, 128, 1, 256] - - [170, 15.018] - - - [4481, 128, 1, 128] - - [134, 18.173] - - - [8065, 128, 1, 128] - - [128, 21.886] - - - [3457, 128, 1, 256] - - [129, 22.482] - - - [6529, 128, 1, 256] - - [138, 32.895] - - - [26753, 128, 1, 256] - - [222, 40.673] - - - [17537, 128, 1, 128] - - [215, 19.148] - - - [22529, 128, 1, 128] - - [208, 22.748] - - - [10625, 128, 1, 256] - - [164, 32.474] - - - [14849, 128, 1, 128] - - [207, 17.019] - - - [9217, 128, 1, 256] - - [154, 33.105] - - - [19841, 128, 1, 128] - - [207, 21.072] - - - [15745, 128, 1, 128] - - [215, 17.609] - - - [13441, 128, 1, 128] - - [153, 27.181] - - - [3713, 128, 1, 128] - - [119, 15.058] - - - [27137, 128, 1, 128] - - [221, 26.286] - - - [16001, 128, 1, 256] - - [216, 27.499] - - - [10369, 128, 1, 128] - - [132, 26.301] - - - [1921, 128, 1, 128] - - [119, 8.975] - - - [9473, 128, 1, 128] - - [141, 24.62] - - - [27905, 128, 1, 256] - - [223, 41.737] - - - [30976, 1024, 1, 128] - - [28, 42.513] - - - [42240, 26369, 1, 128] - - [23, 45.705] - - - [33024, 17025, 1, 128] - - [28, 46.355] - - - [39168, 512, 1, 128] - - [50, 41.06] - - - [30848, 1024, 1, 128] - - [61, 42.967] - - - [41728, 8192, 1, 128] - - [65, 45.865] - - - [39552, 23553, 1, 128] - - [23, 45.133] - - - [35072, 512, 1, 128] - - [59, 41.862] - - - [29952, 14081, 1, 128] - - [26, 45.934] - - - [33280, 2048, 1, 128] - - [40, 44.567] - - - [40320, 128, 1, 128] - - [35, 32.824] - - - [35456, 1024, 1, 128] - - [37, 42.971] - - - [36096, 1024, 1, 128] - - [60, 40.444] - - - [36992, 20993, 1, 128] - - [23, 45.407] - - - [36096, 20097, 1, 128] - - [65, 44.703] - - - [31488, 15489, 1, 128] - - [23, 45.828] - - - [39552, 23681, 1, 128] - - [23, 45.361] - - - [36864, 128, 1, 128] - - [54, 30.903] - - - [40320, 4096, 1, 128] - - [28, 45.636] - - - [35200, 2048, 1, 128] - - [54, 44.094] - - - [29824, 2048, 1, 128] - - [68, 43.127] - - - [34688, 2048, 1, 128] - - [41, 44.433] - - - [42752, 26753, 1, 128] - - [23, 45.655] - - - [34304, 4096, 1, 128] - - [68, 46.126] - - - [36480, 20481, 1, 128] - - [51, 45.313] - - - [33408, 128, 1, 128] - - [41, 28.689] - - - [38784, 4096, 1, 128] - - [35, 45.248] - - - [43264, 27393, 1, 128] - - [39, 45.698] - - - [34560, 128, 1, 128] - - [35, 29.604] - - - [30336, 4096, 1, 128] - - [68, 45.255] - - - [29056, 2048, 1, 128] - - [36, 43.507] - - - [34816, 512, 1, 128] - - [24, 42.118] - - - [38272, 2048, 1, 128] - - [97, 36.638] - - - [39808, 23937, 1, 128] - - [65, 43.271] - - - [30848, 512, 1, 128] - - [35, 40.458] - - - [40448, 512, 1, 128] - - [61, 41.853] - - - [40448, 24577, 1, 128] - - [65, 45.618] - - - [44544, 28545, 1, 128] - - [51, 45.737] - - - [30208, 14209, 1, 128] - - [31, 46.282] - - - [34688, 18689, 1, 128] - - [30, 45.523] - - - [31360, 512, 1, 128] - - [30, 40.429] - - - [38912, 512, 1, 128] - - [35, 41.394] - - - [39680, 1024, 1, 128] - - [35, 42.937] - - - [34048, 1024, 1, 128] - - [35, 43.28] - - - [39552, 4096, 1, 128] - - [39, 45.488] - - - [40320, 24321, 1, 128] - - [23, 45.408] - - - [40832, 24833, 1, 128] - - [39, 45.035] - - - [36736, 1024, 1, 128] - - [36, 42.751] - - - [44672, 1024, 1, 128] - - [61, 43.451] - - - [32000, 128, 1, 128] - - [41, 27.968] - - - [40704, 4096, 1, 128] - - [28, 45.772] - - - [38144, 1024, 1, 128] - - [30, 43.025] - - - [30720, 14849, 1, 128] - - [39, 46.757] - - - [38144, 8192, 1, 128] - - [38, 46.834] - - - [30208, 1024, 1, 128] - - [36, 42.874] - - - [43136, 1024, 1, 128] - - [35, 42.913] - - - [38528, 1024, 1, 128] - - [37, 42.892] - - - [43264, 2048, 1, 128] - - [59, 44.133] - - - [38400, 22529, 1, 128] - - [65, 45.897] - - - [37120, 128, 1, 128] - - [32, 31.012] - - - [32256, 128, 1, 128] - - [59, 28.12] - - - [29952, 13953, 1, 128] - - [38, 45.983] - - - [34560, 8192, 1, 128] - - [28, 46.641] - - - [37504, 21505, 1, 128] - - [52, 44.546] - - - [33536, 128, 1, 128] - - [50, 28.894] - - - [41856, 2048, 1, 128] - - [61, 44.478] - - - [32896, 4096, 1, 128] - - [75, 43.557] - - - [41856, 8192, 1, 128] - - [38, 46.566] - - - [29440, 4096, 1, 128] - - [39, 45.779] - - - [33664, 8192, 1, 128] - - [48, 46.232] - - - [36992, 512, 1, 128] - - [73, 39.86] - - - [33280, 512, 1, 128] - - [32, 40.592] - - - [41728, 128, 1, 128] - - [60, 33.563] - - - [31744, 128, 1, 128] - - [35, 27.614] - - - [31360, 1024, 1, 128] - - [61, 42.98] - - - [29952, 8192, 1, 128] - - [52, 46.599] - - - [38016, 2048, 1, 128] - - [64, 44.223] - - - [34176, 8192, 1, 128] - - [48, 46.412] - - - [30464, 512, 1, 128] - - [61, 39.585] - - - [41984, 2048, 1, 128] - - [62, 44.722] - - - [40448, 4096, 1, 128] - - [52, 46.116] - - - [33920, 4096, 1, 128] - - [28, 45.581] - - - [41088, 8192, 1, 128] - - [62, 45.778] - - - [39808, 8192, 1, 128] - - [23, 44.228] - - - [40832, 4096, 1, 128] - - [26, 45.606] - - - [30592, 2048, 1, 128] - - [61, 43.833] - - - [36352, 1024, 1, 128] - - [41, 43.102] - - - [30336, 2048, 1, 128] - - [41, 43.833] - - - [30976, 512, 1, 128] - - [41, 40.197] - - - [42368, 1024, 1, 128] - - [41, 43.084] - - - [29056, 1024, 1, 128] - - [41, 42.348] - - - [38784, 22913, 1, 128] - - [23, 45.544] - - - [28928, 512, 1, 128] - - [35, 39.211] - - - [40576, 512, 1, 128] - - [28, 41.519] - - - [34816, 4096, 1, 128] - - [28, 46.696] - - - [41600, 2048, 1, 128] - - [23, 43.878] - - - [29696, 8192, 1, 128] - - [26, 47.304] - - - [41856, 4096, 1, 128] - - [28, 45.762] - - - [35584, 2048, 1, 128] - - [64, 44.122] - - - [30848, 14849, 1, 128] - - [38, 45.765] - - - [33280, 17281, 1, 128] - - [48, 46.514] - - - [43776, 2048, 1, 128] - - [61, 43.542] - - - [42112, 8192, 1, 128] - - [26, 46.462] - - - [37376, 128, 1, 128] - - [64, 31.226] - - - [41600, 4096, 1, 128] - - [39, 45.603] - - - [36224, 20353, 1, 128] - - [39, 45.672] - - - [29952, 1024, 1, 128] - - [36, 42.84] - - - [34176, 1024, 1, 128] - - [35, 42.871] - - - [31744, 512, 1, 128] - - [24, 40.693] - - - [42624, 8192, 1, 128] - - [28, 41.712] - - - [41216, 128, 1, 128] - - [45, 33.152] - - - [42624, 26753, 1, 128] - - [78, 40.481] - - - [32512, 2048, 1, 128] - - [45, 44.312] - - - [40064, 4096, 1, 128] - - [23, 45.152] - - - [32640, 4096, 1, 128] - - [28, 45.254] - - - [42112, 26241, 1, 128] - - [23, 45.551] - - - [32256, 512, 1, 128] - - [58, 40.905] - - - [40960, 1024, 1, 128] - - [28, 40.673] - - - [35968, 128, 1, 128] - - [41, 30.387] - - - [32384, 8192, 1, 128] - - [30, 46.529] - - - [42880, 512, 1, 128] - - [56, 40.788] - - - [33024, 8192, 1, 128] - - [30, 47.007] - - - [43904, 1024, 1, 128] - - [41, 42.98] - - - [33664, 17665, 1, 128] - - [39, 45.702] - - - [41856, 512, 1, 128] - - [54, 42.279] - - - [40704, 128, 1, 128] - - [34, 33.254] - - - [33408, 17537, 1, 128] - - [39, 45.966] - - - [37120, 512, 1, 128] - - [61, 40.735] - - - [41216, 25345, 1, 128] - - [65, 45.806] - - - [39680, 8192, 1, 128] - - [38, 46.567] - - - [40192, 24193, 1, 128] - - [39, 45.525] - - - [33024, 17153, 1, 128] - - [23, 46.343] - - - [38272, 1024, 1, 128] - - [31, 38.221] - - - [35328, 1024, 1, 128] - - [36, 43.01] - - - [31104, 8192, 1, 128] - - [28, 46.434] - - - [40320, 8192, 1, 128] - - [26, 46.483] - - - [29312, 2048, 1, 128] - - [59, 43.896] - - - [36608, 20737, 1, 128] - - [23, 45.83] - - - [42240, 4096, 1, 128] - - [26, 45.827] - - - [43520, 2048, 1, 128] - - [34, 44.598] - - - [29056, 512, 1, 128] - - [61, 39.186] - - - [35328, 19329, 1, 128] - - [51, 46.24] - - - [30464, 128, 1, 128] - - [24, 26.971] - - - [29696, 13697, 1, 128] - - [28, 46.562] - - - [43904, 28033, 1, 128] - - [28, 45.375] - - - [35584, 19713, 1, 128] - - [23, 45.7] - - - [41088, 4096, 1, 128] - - [37, 45.105] - - - [42368, 2048, 1, 128] - - [59, 44.078] - - - [36736, 128, 1, 128] - - [36, 30.789] - - - [30336, 8192, 1, 128] - - [36, 46.227] - - - [43008, 128, 1, 128] - - [35, 31.442] - - - [37120, 1024, 1, 128] - - [41, 42.986] - - - [31104, 2048, 1, 128] - - [61, 44.22] - - - [33152, 4096, 1, 128] - - [35, 45.818] - - - [43392, 27521, 1, 128] - - [39, 45.358] - - - [37248, 21249, 1, 128] - - [39, 45.316] - - - [33920, 17921, 1, 128] - - [39, 45.594] - - - [39680, 4096, 1, 128] - - [26, 45.776] - - - [43264, 512, 1, 128] - - [56, 40.916] - - - [35712, 8192, 1, 128] - - [26, 46.366] - - - [31616, 2048, 1, 128] - - [39, 43.609] - - - [35328, 512, 1, 128] - - [35, 40.342] - - - [43136, 27265, 1, 128] - - [23, 45.393] - - - [30208, 128, 1, 128] - - [61, 26.654] - - - [40320, 24449, 1, 128] - - [39, 45.352] - - - [44288, 2048, 1, 128] - - [37, 44.467] - - - [35072, 1024, 1, 128] - - [28, 43.152] - - - [30464, 14465, 1, 128] - - [75, 43.511] - - - [44160, 8192, 1, 128] - - [26, 46.21] - - - [33792, 17793, 1, 128] - - [30, 46.448] - - - [37632, 1024, 1, 128] - - [35, 43.181] - - - [35968, 2048, 1, 128] - - [61, 44.208] - - - [38400, 8192, 1, 128] - - [31, 46.925] - - - [32512, 4096, 1, 128] - - [28, 45.842] - - - [32512, 16641, 1, 128] - - [52, 46.246] - - - [39424, 128, 1, 128] - - [58, 32.351] - - - [30976, 8192, 1, 128] - - [77, 45.168] - - - [35968, 20097, 1, 128] - - [26, 45.717] - - - [38656, 512, 1, 128] - - [59, 40.966] - - - [34944, 18945, 1, 128] - - [28, 45.624] - - - [33664, 17793, 1, 128] - - [23, 45.754] - - - [38656, 22657, 1, 128] - - [51, 45.798] - - - [34944, 1024, 1, 128] - - [35, 43.102] - - - [31872, 16001, 1, 128] - - [39, 45.704] - - - [43392, 8192, 1, 128] - - [38, 46.144] - - - [38016, 512, 1, 128] - - [41, 40.69] - - - [29440, 8192, 1, 128] - - [23, 46.686] - - - [35200, 1024, 1, 128] - - [41, 42.748] - - - [34304, 18433, 1, 128] - - [51, 46.162] - - - [44672, 28801, 1, 128] - - [39, 45.545] - - - [29184, 4096, 1, 128] - - [58, 45.73] - - - [33408, 8192, 1, 128] - - [28, 46.532] - - - [39040, 128, 1, 128] - - [40, 31.994] - - - [39680, 23681, 1, 128] - - [39, 45.518] - - - [38144, 4096, 1, 128] - - [26, 45.883] - - - [42368, 26497, 1, 128] - - [52, 45.346] - - - [42368, 4096, 1, 128] - - [68, 45.365] - - - [31872, 128, 1, 128] - - [64, 27.758] - - - [41984, 512, 1, 128] - - [45, 42.307] - - - [39296, 2048, 1, 128] - - [39, 43.934] - - - [33920, 2048, 1, 128] - - [37, 44.095] - - - [36736, 20865, 1, 128] - - [23, 45.681] - - - [34432, 8192, 1, 128] - - [31, 44.91] - - - [30848, 14977, 1, 128] - - [38, 45.811] - - - [31744, 15873, 1, 128] - - [30, 46.488] - - - [42880, 27009, 1, 128] - - [65, 44.895] - - - [42240, 26241, 1, 128] - - [39, 45.662] - - - [38400, 4096, 1, 128] - - [38, 45.966] - - - [42624, 26625, 1, 128] - - [28, 40.471] - - - [35072, 4096, 1, 128] - - [38, 45.875] - - - [40576, 4096, 1, 128] - - [26, 45.622] - - - [39296, 8192, 1, 128] - - [30, 46.525] - - - [42624, 512, 1, 128] - - [62, 40.003] - - - [32768, 8192, 1, 128] - - [38, 35.565] - - - [36864, 1024, 1, 128] - - [36, 42.756] - - - [43392, 128, 1, 128] - - [60, 31.053] - - - [41344, 2048, 1, 128] - - [41, 44.187] - - - [35584, 4096, 1, 128] - - [28, 45.526] - - - [40064, 2048, 1, 128] - - [23, 42.994] - - - [40576, 24705, 1, 128] - - [39, 45.497] - - - [39808, 1024, 1, 128] - - [28, 42.27] - - - [36992, 1024, 1, 128] - - [73, 42.438] - - - [42496, 1024, 1, 128] - - [64, 43.847] - - - [43904, 128, 1, 128] - - [48, 31.389] - - - [31232, 512, 1, 128] - - [50, 40.907] - - - [42112, 128, 1, 128] - - [41, 33.659] - - - [37376, 2048, 1, 128] - - [59, 44.486] - - - [38016, 128, 1, 128] - - [61, 31.37] - - - [42368, 8192, 1, 128] - - [31, 46.148] - - - [43392, 512, 1, 128] - - [24, 41.032] - - - [41984, 1024, 1, 128] - - [59, 43.616] - - - [42240, 2048, 1, 128] - - [54, 44.379] - - - [29952, 128, 1, 128] - - [24, 26.55] - - - [36608, 8192, 1, 128] - - [39, 46.612] - - - [32512, 16513, 1, 128] - - [65, 46.298] - - - [29568, 512, 1, 128] - - [32, 38.598] - - - [34304, 1024, 1, 128] - - [58, 43.493] - - - [41984, 4096, 1, 128] - - [26, 46.4] - - - [30464, 4096, 1, 128] - - [77, 43.433] - - - [41216, 2048, 1, 128] - - [37, 44.506] - - - [36480, 20609, 1, 128] - - [39, 45.594] - - - [44800, 4096, 1, 128] - - [30, 45.493] - - - [36864, 512, 1, 128] - - [24, 40.566] - - - [39680, 2048, 1, 128] - - [30, 43.919] - - - [43648, 4096, 1, 128] - - [26, 45.327] - - - [33664, 128, 1, 128] - - [24, 29.078] - - - [41600, 512, 1, 128] - - [75, 42.222] - - - [43776, 1024, 1, 128] - - [64, 41.689] - - - [37632, 512, 1, 128] - - [36, 40.728] - - - [44160, 128, 1, 128] - - [40, 31.541] - - - [37248, 8192, 1, 128] - - [38, 46.313] - - - [34816, 18817, 1, 128] - - [30, 46.801] - - - [38528, 22529, 1, 128] - - [39, 45.314] - - - [40192, 24321, 1, 128] - - [39, 45.492] - - - [40832, 128, 1, 128] - - [60, 32.914] - - - [29312, 8192, 1, 128] - - [28, 46.211] - - - [43776, 27777, 1, 128] - - [62, 43.983] - - - [37632, 21633, 1, 128] - - [39, 45.804] - - - [33792, 4096, 1, 128] - - [26, 46.375] - - - [35968, 1024, 1, 128] - - [24, 43.0] - - - [37888, 512, 1, 128] - - [35, 41.05] - - - [35968, 512, 1, 128] - - [56, 40.217] - - - [30592, 1024, 1, 128] - - [54, 42.842] - - - [38400, 512, 1, 128] - - [40, 40.674] - - - [43264, 1024, 1, 128] - - [26, 43.316] - - - [38528, 4096, 1, 128] - - [50, 45.377] - - - [28928, 1024, 1, 128] - - [36, 42.433] - - - [33152, 1024, 1, 128] - - [24, 42.341] - - - [41344, 1024, 1, 128] - - [41, 43.141] - - - [30848, 8192, 1, 128] - - [38, 46.564] - - - [41344, 4096, 1, 128] - - [30, 45.428] - - - [38912, 2048, 1, 128] - - [30, 44.728] - - - [38272, 128, 1, 128] - - [85, 31.539] - - - [31488, 4096, 1, 128] - - [26, 45.665] - - - [44416, 4096, 1, 128] - - [26, 45.725] - - - [39552, 2048, 1, 128] - - [41, 44.397] - - - [37760, 1024, 1, 128] - - [41, 43.235] - - - [34304, 18305, 1, 128] - - [51, 46.325] - - - [44544, 28673, 1, 128] - - [65, 45.502] - - - [44416, 8192, 1, 128] - - [26, 46.505] - - - [38144, 512, 1, 128] - - [28, 40.804] - - - [30208, 14337, 1, 128] - - [25, 46.144] - - - [38144, 2048, 1, 128] - - [38, 44.289] - - - [40448, 128, 1, 128] - - [79, 33.044] - - - [42240, 8192, 1, 128] - - [26, 46.544] - - - [39424, 2048, 1, 128] - - [37, 44.584] - - - [41088, 512, 1, 128] - - [58, 39.121] - - - [36224, 2048, 1, 128] - - [61, 43.974] - - - [31744, 4096, 1, 128] - - [28, 46.397] - - - [44160, 512, 1, 128] - - [50, 40.882] - - - [32000, 1024, 1, 128] - - [35, 42.737] - - - [42752, 1024, 1, 128] - - [41, 43.363] - - - [42496, 2048, 1, 128] - - [40, 44.786] - - - [32640, 2048, 1, 128] - - [26, 43.065] - - - [42752, 26881, 1, 128] - - [39, 45.625] - - - [32256, 8192, 1, 128] - - [52, 47.12] - - - [44800, 512, 1, 128] - - [38, 41.246] - - - [34816, 128, 1, 128] - - [64, 29.795] - - - [38272, 8192, 1, 128] - - [62, 42.399] - - - [44800, 28929, 1, 128] - - [23, 45.322] - - - [37120, 8192, 1, 128] - - [23, 46.547] - - - [43776, 512, 1, 128] - - [34, 37.791] - - - [43008, 1024, 1, 128] - - [36, 43.999] - - - [34432, 18561, 1, 128] - - [75, 44.536] - - - [36736, 4096, 1, 128] - - [28, 45.56] - - - [36224, 512, 1, 128] - - [41, 40.29] - - - [32768, 512, 1, 128] - - [37, 36.086] - - - [30592, 128, 1, 128] - - [41, 26.897] - - - [43008, 27137, 1, 128] - - [23, 46.575] - - - [34048, 18177, 1, 128] - - [65, 45.732] - - - [43136, 2048, 1, 128] - - [38, 43.944] - - - [29184, 13313, 1, 128] - - [23, 45.766] - - - [40064, 24193, 1, 128] - - [65, 44.519] - - - [40960, 128, 1, 128] - - [54, 32.703] - - - [29184, 2048, 1, 128] - - [40, 44.175] - - - [37248, 128, 1, 128] - - [35, 31.154] - - - [35328, 128, 1, 128] - - [35, 30.192] - - - [43264, 128, 1, 128] - - [58, 31.142] - - - [29952, 4096, 1, 128] - - [58, 45.58] - - - [36736, 20737, 1, 128] - - [39, 45.629] - - - [34176, 4096, 1, 128] - - [32, 45.382] - - - [32768, 1024, 1, 128] - - [26, 36.303] - - - [44160, 4096, 1, 128] - - [26, 45.544] - - - [31104, 1024, 1, 128] - - [50, 43.24] - - - [33792, 512, 1, 128] - - [61, 41.448] - - - [41216, 25217, 1, 128] - - [65, 45.833] - - - [31872, 1024, 1, 128] - - [26, 42.624] - - - [38528, 8192, 1, 128] - - [25, 46.201] - - - [44672, 4096, 1, 128] - - [30, 45.698] - - - [32512, 1024, 1, 128] - - [24, 42.329] - - - [39168, 8192, 1, 128] - - [65, 46.6] - - - [31360, 15361, 1, 128] - - [39, 45.526] - - - [38016, 22145, 1, 128] - - [39, 45.578] - - - [35712, 128, 1, 128] - - [35, 30.005] - - - [30208, 4096, 1, 128] - - [60, 45.872] - - - [33920, 128, 1, 128] - - [36, 29.163] - - - [30336, 128, 1, 128] - - [41, 26.666] - - - [42368, 128, 1, 128] - - [27, 30.858] - - - [38912, 4096, 1, 128] - - [38, 46.69] - - - [34176, 512, 1, 128] - - [26, 41.65] - - - [42752, 8192, 1, 128] - - [28, 46.612] - - - [31488, 1024, 1, 128] - - [35, 43.165] - - - [36608, 1024, 1, 128] - - [61, 42.897] - - - [41856, 128, 1, 128] - - [41, 33.455] - - - [29312, 13441, 1, 128] - - [26, 45.652] - - - [43520, 128, 1, 128] - - [37, 31.296] - - - [31616, 8192, 1, 128] - - [26, 45.923] - - - [40448, 2048, 1, 128] - - [34, 44.652] - - - [35328, 2048, 1, 128] - - [45, 44.595] - - - [36864, 20865, 1, 128] - - [39, 46.566] - - - [32000, 2048, 1, 128] - - [28, 44.031] - - - [34176, 18177, 1, 128] - - [65, 45.62] - - - [37504, 128, 1, 128] - - [41, 31.255] - - - [33792, 1024, 1, 128] - - [36, 43.346] - - - [31872, 8192, 1, 128] - - [26, 46.308] - - - [40704, 512, 1, 128] - - [35, 41.912] - - - [37632, 128, 1, 128] - - [35, 31.221] - - - [32640, 1024, 1, 128] - - [26, 41.437] - - - [44544, 8192, 1, 128] - - [23, 46.449] - - - [39424, 8192, 1, 128] - - [26, 46.96] - - - [39296, 512, 1, 128] - - [35, 41.084] - - - [35840, 128, 1, 128] - - [28, 30.25] - - - [39168, 1024, 1, 128] - - [50, 43.49] - - - [35712, 19841, 1, 128] - - [39, 45.634] - - - [29568, 13569, 1, 128] - - [23, 45.271] - - - [34944, 4096, 1, 128] - - [28, 45.826] - - - [32768, 2048, 1, 128] - - [30, 34.701] - - - [39296, 128, 1, 128] - - [28, 32.103] - - - [29568, 4096, 1, 128] - - [62, 44.664] - - - [39040, 1024, 1, 128] - - [59, 43.154] - - - [37376, 1024, 1, 128] - - [35, 43.367] - - - [33536, 2048, 1, 128] - - [26, 44.356] - - - [31488, 8192, 1, 128] - - [28, 46.506] - - - [37888, 1024, 1, 128] - - [41, 43.644] - - - [41472, 4096, 1, 128] - - [31, 45.992] - - - [30592, 512, 1, 128] - - [35, 40.13] - - - [34560, 18561, 1, 128] - - [30, 45.867] - - - [29184, 512, 1, 128] - - [35, 39.349] - - - [32256, 16257, 1, 128] - - [25, 46.612] - - - [43392, 27393, 1, 128] - - [39, 45.405] - - - [29312, 4096, 1, 128] - - [36, 45.362] - - - [43648, 2048, 1, 128] - - [30, 43.517] - - - [44288, 1024, 1, 128] - - [61, 43.32] - - - [35456, 128, 1, 128] - - [68, 30.098] - - - [44160, 28289, 1, 128] - - [38, 45.17] - - - [40320, 1024, 1, 128] - - [41, 43.195] - - - [37888, 22017, 1, 128] - - [39, 46.388] - - - [29696, 512, 1, 128] - - [35, 39.605] - - - [35840, 2048, 1, 128] - - [45, 44.646] - - - [37504, 2048, 1, 128] - - [64, 43.735] - - - [41728, 4096, 1, 128] - - [68, 45.048] - - - [42752, 4096, 1, 128] - - [38, 45.734] - - - [29824, 4096, 1, 128] - - [62, 44.859] - - - [44800, 1024, 1, 128] - - [36, 43.392] - - - [30592, 4096, 1, 128] - - [26, 45.582] - - - [43904, 4096, 1, 128] - - [28, 45.454] - - - [39552, 8192, 1, 128] - - [39, 46.1] - - - [37632, 2048, 1, 128] - - [61, 44.242] - - - [29312, 128, 1, 128] - - [24, 25.614] - - - [30080, 512, 1, 128] - - [40, 39.423] - - - [33664, 2048, 1, 128] - - [54, 44.161] - - - [43520, 27521, 1, 128] - - [65, 46.018] - - - [36224, 128, 1, 128] - - [35, 30.533] - - - [28928, 12929, 1, 128] - - [38, 45.873] - - - [29440, 1024, 1, 128] - - [36, 42.53] - - - [35840, 19969, 1, 128] - - [39, 46.357] - - - [42880, 4096, 1, 128] - - [58, 44.933] - - - [42496, 8192, 1, 128] - - [26, 46.936] - - - [39936, 24065, 1, 128] - - [39, 46.243] - - - [33408, 1024, 1, 128] - - [26, 42.705] - - - [32256, 2048, 1, 128] - - [37, 44.555] - - - [35712, 19713, 1, 128] - - [39, 45.682] - - - [40192, 4096, 1, 128] - - [38, 45.816] - - - [32000, 16129, 1, 128] - - [28, 46.012] - - - [44032, 512, 1, 128] - - [24, 41.391] - - - [35584, 128, 1, 128] - - [41, 30.103] - - - [35584, 8192, 1, 128] - - [30, 46.454] - - - [37888, 21889, 1, 128] - - [39, 46.38] - - - [37504, 1024, 1, 128] - - [60, 42.416] - - - [33664, 512, 1, 128] - - [61, 41.523] - - - [32384, 1024, 1, 128] - - [95, 42.482] - - - [38400, 1024, 1, 128] - - [73, 43.355] - - - [35200, 128, 1, 128] - - [64, 29.841] - - - [43648, 1024, 1, 128] - - [36, 42.242] - - - [36608, 128, 1, 128] - - [24, 30.717] - - - [32768, 128, 1, 128] - - [41, 28.982] - - - [28928, 4096, 1, 128] - - [30, 45.55] - - - [35200, 19329, 1, 128] - - [23, 45.615] - - - [41216, 8192, 1, 128] - - [26, 46.646] - - - [36864, 8192, 1, 128] - - [30, 47.323] - - - [40064, 128, 1, 128] - - [58, 32.473] - - - [42624, 1024, 1, 128] - - [38, 41.156] - - - [34688, 128, 1, 128] - - [35, 29.748] - - - [43648, 27777, 1, 128] - - [23, 44.988] - - - [37888, 8192, 1, 128] - - [38, 47.162] - - - [41472, 25601, 1, 128] - - [65, 45.782] - - - [38272, 512, 1, 128] - - [28, 37.579] - - - [35456, 4096, 1, 128] - - [50, 45.661] - - - [42496, 26625, 1, 128] - - [51, 45.874] - - - [43136, 4096, 1, 128] - - [38, 45.517] - - - [44800, 8192, 1, 128] - - [30, 46.255] - - - [36480, 8192, 1, 128] - - [30, 46.305] - - - [37504, 4096, 1, 128] - - [64, 44.983] - - - [39040, 8192, 1, 128] - - [28, 46.156] - - - [31104, 512, 1, 128] - - [24, 40.757] - - - [34176, 2048, 1, 128] - - [41, 44.049] - - - [31616, 512, 1, 128] - - [64, 40.723] - - - [35456, 2048, 1, 128] - - [24, 44.265] - - - [43136, 8192, 1, 128] - - [38, 46.264] - - - [33024, 128, 1, 128] - - [61, 28.36] - - - [38656, 4096, 1, 128] - - [52, 45.653] - - - [33408, 17409, 1, 128] - - [39, 45.715] - - - [39424, 1024, 1, 128] - - [40, 43.586] - - - [29312, 13313, 1, 128] - - [28, 45.515] - - - [35840, 4096, 1, 128] - - [30, 46.364] - - - [42496, 512, 1, 128] - - [36, 41.174] - - - [37632, 8192, 1, 128] - - [38, 46.625] - - - [41088, 2048, 1, 128] - - [59, 43.811] - - - [38528, 512, 1, 128] - - [26, 40.582] - - - [35072, 2048, 1, 128] - - [54, 44.333] - - - [31104, 4096, 1, 128] - - [39, 45.51] - - - [33280, 4096, 1, 128] - - [68, 46.23] - - - [43904, 8192, 1, 128] - - [38, 46.33] - - - [34816, 8192, 1, 128] - - [28, 47.493] - - - [38016, 1024, 1, 128] - - [56, 43.185] - - - [33152, 128, 1, 128] - - [60, 28.403] - - - [42496, 128, 1, 128] - - [43, 31.099] - - - [40832, 24961, 1, 128] - - [39, 45.111] - - - [41728, 1024, 1, 128] - - [68, 42.868] - - - [41472, 25473, 1, 128] - - [51, 46.035] - - - [34560, 2048, 1, 128] - - [64, 44.545] - - - [31616, 15617, 1, 128] - - [23, 45.318] - - - [33664, 4096, 1, 128] - - [39, 45.445] - - - [35328, 8192, 1, 128] - - [31, 47.021] - - - [39808, 4096, 1, 128] - - [38, 43.349] - - - [37248, 512, 1, 128] - - [41, 40.754] - - - [31360, 4096, 1, 128] - - [28, 45.616] - - - [41344, 8192, 1, 128] - - [38, 46.237] - - - [32000, 512, 1, 128] - - [50, 40.442] - - - [35968, 19969, 1, 128] - - [28, 45.615] - - - [30080, 14081, 1, 128] - - [48, 43.41] - - - [35840, 8192, 1, 128] - - [38, 47.132] - - - [44672, 2048, 1, 128] - - [64, 44.597] - - - [31872, 2048, 1, 128] - - [38, 43.727] - - - [42496, 4096, 1, 128] - - [23, 46.1] - - - [43776, 128, 1, 128] - - [59, 31.207] - - - [40704, 2048, 1, 128] - - [45, 44.366] - - - [34432, 128, 1, 128] - - [36, 29.398] - - - [44544, 2048, 1, 128] - - [70, 44.537] - - - [32384, 16385, 1, 128] - - [26, 45.9] - - - [43776, 27905, 1, 128] - - [62, 44.016] - - - [44032, 4096, 1, 128] - - [38, 46.373] - - - [36480, 512, 1, 128] - - [40, 40.173] - - - [44160, 1024, 1, 128] - - [64, 43.268] - - - [41216, 4096, 1, 128] - - [38, 45.814] - - - [44032, 2048, 1, 128] - - [59, 44.796] - - - [33152, 2048, 1, 128] - - [64, 44.512] - - - [41984, 25985, 1, 128] - - [39, 46.352] - - - [39552, 512, 1, 128] - - [36, 41.308] - - - [41344, 25473, 1, 128] - - [30, 45.432] - - - [40960, 4096, 1, 128] - - [23, 41.231] - - - [32640, 128, 1, 128] - - [61, 28.095] - - - [35968, 4096, 1, 128] - - [36, 45.611] - - - [33536, 4096, 1, 128] - - [30, 45.86] - - - [30976, 15105, 1, 128] - - [77, 44.642] - - - [35072, 8192, 1, 128] - - [30, 46.675] - - - [39424, 23425, 1, 128] - - [65, 45.929] - - - [43520, 1024, 1, 128] - - [34, 43.744] - - - [44288, 28417, 1, 128] - - [51, 45.373] - - - [30848, 128, 1, 128] - - [41, 27.219] - - - [35712, 512, 1, 128] - - [28, 40.09] - - - [44160, 2048, 1, 128] - - [54, 44.248] - - - [34048, 8192, 1, 128] - - [65, 46.442] - - - [40448, 24449, 1, 128] - - [51, 46.019] - - - [39168, 23297, 1, 128] - - [51, 45.842] - - - [32128, 1024, 1, 128] - - [61, 42.82] - - - [36864, 20993, 1, 128] - - [39, 46.547] - - - [40064, 1024, 1, 128] - - [38, 42.512] - - - [38784, 8192, 1, 128] - - [38, 46.238] - - - [37248, 2048, 1, 128] - - [38, 43.739] - - - [34560, 4096, 1, 128] - - [28, 45.799] - - - [39040, 23041, 1, 128] - - [23, 45.415] - - - [36480, 1024, 1, 128] - - [61, 42.697] - - - [39040, 2048, 1, 128] - - [54, 44.221] - - - [39808, 23809, 1, 128] - - [65, 43.154] - - - [36992, 4096, 1, 128] - - [38, 45.424] - - - [32768, 16897, 1, 128] - - [38, 34.596] - - - [30976, 2048, 1, 128] - - [32, 42.926] - - - [32640, 16769, 1, 128] - - [30, 45.395] - - - [29824, 13953, 1, 128] - - [65, 45.133] - - - [29184, 128, 1, 128] - - [41, 25.776] - - - [30720, 8192, 1, 128] - - [30, 47.523] - - - [30848, 2048, 1, 128] - - [41, 43.898] - - - [38016, 4096, 1, 128] - - [23, 45.516] - - - [35456, 8192, 1, 128] - - [28, 46.552] - - - [36992, 21121, 1, 128] - - [23, 45.411] - - - [36736, 2048, 1, 128] - - [54, 44.039] - - - [37888, 128, 1, 128] - - [61, 31.44] - - - [39808, 2048, 1, 128] - - [28, 42.312] - - - [41856, 25985, 1, 128] - - [23, 45.662] - - - [34688, 4096, 1, 128] - - [50, 45.486] - - - [38784, 1024, 1, 128] - - [41, 42.968] - - - [40960, 25089, 1, 128] - - [23, 40.357] - - - [32000, 4096, 1, 128] - - [26, 45.811] - - - [41600, 25601, 1, 128] - - [23, 45.352] - - - [37504, 512, 1, 128] - - [58, 40.171] - - - [32128, 16129, 1, 128] - - [30, 45.863] - - - [37248, 21377, 1, 128] - - [39, 45.342] - - - [35840, 512, 1, 128] - - [50, 40.542] - - - [36096, 128, 1, 128] - - [37, 30.362] - - - [32512, 8192, 1, 128] - - [26, 46.813] - - - [36736, 8192, 1, 128] - - [30, 46.469] - - - [42880, 1024, 1, 128] - - [40, 43.034] - - - [44288, 8192, 1, 128] - - [28, 46.275] - - - [36224, 1024, 1, 128] - - [41, 42.95] - - - [41344, 25345, 1, 128] - - [30, 45.426] - - - [32384, 512, 1, 128] - - [36, 39.622] - - - [38272, 4096, 1, 128] - - [62, 41.461] - - - [37120, 2048, 1, 128] - - [24, 44.324] - - - [33152, 8192, 1, 128] - - [30, 46.687] - - - [36096, 4096, 1, 128] - - [77, 44.072] - - - [34560, 18689, 1, 128] - - [39, 45.867] - - - [36864, 4096, 1, 128] - - [26, 46.428] - - - [34944, 512, 1, 128] - - [38, 41.997] - - - [37760, 128, 1, 128] - - [35, 31.327] - - - [31616, 128, 1, 128] - - [40, 27.502] - - - [36224, 4096, 1, 128] - - [38, 45.663] - - - [40576, 24577, 1, 128] - - [23, 45.04] - - - [34688, 1024, 1, 128] - - [64, 43.274] - - - [40192, 1024, 1, 128] - - [24, 43.302] - - - [44672, 512, 1, 128] - - [28, 41.103] - - - [33664, 1024, 1, 128] - - [35, 43.004] - - - [39424, 512, 1, 128] - - [58, 41.0] - - - [44416, 1024, 1, 128] - - [41, 43.311] - - - [33408, 2048, 1, 128] - - [41, 44.337] - - - [43648, 8192, 1, 128] - - [38, 46.058] - - - [43520, 27649, 1, 128] - - [65, 45.797] - - - [40448, 1024, 1, 128] - - [64, 43.246] - - - [33152, 17153, 1, 128] - - [38, 46.124] - - - [33024, 512, 1, 128] - - [41, 40.032] - - - [39680, 128, 1, 128] - - [58, 32.482] - - - [29696, 4096, 1, 128] - - [38, 46.382] - - - [42112, 2048, 1, 128] - - [28, 44.11] - - - [38016, 8192, 1, 128] - - [26, 46.349] - - - [30464, 8192, 1, 128] - - [79, 44.92] - - - [43648, 128, 1, 128] - - [54, 31.115] - - - [32896, 16897, 1, 128] - - [75, 44.774] - - - [43008, 8192, 1, 128] - - [28, 47.387] - - - [34304, 512, 1, 128] - - [70, 41.954] - - - [38528, 128, 1, 128] - - [41, 31.686] - - - [41216, 1024, 1, 128] - - [73, 43.046] - - - [38272, 22401, 1, 128] - - [75, 40.701] - - - [34048, 4096, 1, 128] - - [36, 45.503] - - - [30720, 512, 1, 128] - - [61, 40.496] - - - [41728, 512, 1, 128] - - [30, 42.02] - - - [43136, 512, 1, 128] - - [61, 40.856] - - - [41088, 1024, 1, 128] - - [85, 42.006] - - - [33536, 1024, 1, 128] - - [28, 43.074] - - - [41088, 25089, 1, 128] - - [75, 44.401] - - - [36352, 20353, 1, 128] - - [23, 46.128] - - - [29184, 1024, 1, 128] - - [60, 42.608] - - - [44800, 128, 1, 128] - - [60, 32.154] - - - [41600, 8192, 1, 128] - - [23, 46.389] - - - [44416, 28545, 1, 128] - - [23, 45.558] - - - [34048, 512, 1, 128] - - [35, 41.814] - - - [32128, 16257, 1, 128] - - [38, 45.982] - - - [44288, 4096, 1, 128] - - [30, 45.46] - - - [34432, 18433, 1, 128] - - [75, 44.145] - - - [41856, 25857, 1, 128] - - [23, 45.643] - - - [32128, 2048, 1, 128] - - [59, 44.016] - - - [34688, 512, 1, 128] - - [64, 41.886] - - - [39936, 4096, 1, 128] - - [38, 46.381] - - - [38656, 1024, 1, 128] - - [38, 42.922] - - - [37760, 512, 1, 128] - - [35, 40.533] - - - [30336, 512, 1, 128] - - [36, 40.215] - - - [38016, 22017, 1, 128] - - [39, 45.593] - - - [44544, 4096, 1, 128] - - [51, 45.793] - - - [38912, 8192, 1, 128] - - [28, 47.448] - - - [39936, 128, 1, 128] - - [61, 32.476] - - - [36480, 2048, 1, 128] - - [41, 44.212] - - - [35200, 4096, 1, 128] - - [26, 45.526] - - - [30976, 14977, 1, 128] - - [75, 44.608] - - - [31104, 15105, 1, 128] - - [30, 45.786] - - - [40832, 1024, 1, 128] - - [38, 42.57] - - - [32384, 16513, 1, 128] - - [26, 46.085] - - - [43392, 4096, 1, 128] - - [39, 45.466] - - - [32768, 4096, 1, 128] - - [26, 35.663] - - - [38272, 22273, 1, 128] - - [79, 40.665] - - - [32128, 512, 1, 128] - - [35, 40.604] - - - [32896, 2048, 1, 128] - - [29, 41.278] - - - [37376, 21505, 1, 128] - - [38, 45.849] - - - [41856, 1024, 1, 128] - - [64, 43.584] - - - [33536, 8192, 1, 128] - - [65, 46.748] - - - [29568, 1024, 1, 128] - - [59, 41.624] - - - [44032, 28033, 1, 128] - - [39, 46.336] - - - [33280, 8192, 1, 128] - - [52, 47.149] - - - [39296, 4096, 1, 128] - - [26, 45.716] - - - [30592, 14593, 1, 128] - - [30, 45.761] - - - [37504, 8192, 1, 128] - - [62, 45.639] - - - [30336, 14465, 1, 128] - - [30, 45.698] - - - [29952, 2048, 1, 128] - - [24, 44.232] - - - [40832, 512, 1, 128] - - [45, 40.792] - - - [44672, 28673, 1, 128] - - [23, 45.372] - - - [30080, 4096, 1, 128] - - [79, 43.362] - - - [37888, 2048, 1, 128] - - [62, 44.739] - - - [37632, 21761, 1, 128] - - [23, 45.775] - - - [29824, 8192, 1, 128] - - [75, 45.814] - - - [35328, 19457, 1, 128] - - [65, 46.076] - - - [37376, 4096, 1, 128] - - [36, 45.946] - - - [33792, 17921, 1, 128] - - [38, 46.337] - - - [34304, 8192, 1, 128] - - [48, 47.014] - - - [42752, 512, 1, 128] - - [41, 40.648] - - - [36992, 2048, 1, 128] - - [40, 43.993] - - - [39168, 4096, 1, 128] - - [65, 45.615] - - - [31360, 15489, 1, 128] - - [38, 45.835] - - - [43520, 8192, 1, 128] - - [26, 46.897] - - - [30080, 2048, 1, 128] - - [74, 41.788] - - - [30720, 4096, 1, 128] - - [26, 46.713] - - - [34176, 128, 1, 128] - - [64, 29.275] - - - [32768, 16769, 1, 128] - - [38, 34.684] - - - [35072, 128, 1, 128] - - [36, 29.938] - - - [35712, 4096, 1, 128] - - [39, 45.542] - - - [36480, 4096, 1, 128] - - [38, 45.435] - - - [39424, 4096, 1, 128] - - [28, 45.949] - - - [38400, 128, 1, 128] - - [26, 31.61] - - - [34432, 2048, 1, 128] - - [60, 42.941] - - - [41344, 512, 1, 128] - - [50, 41.676] - - - [35200, 512, 1, 128] - - [24, 41.728] - - - [39936, 8192, 1, 128] - - [30, 47.064] - - - [31488, 128, 1, 128] - - [36, 27.718] - - - [43008, 512, 1, 128] - - [61, 41.471] - - - [33024, 4096, 1, 128] - - [26, 46.064] - - - [36608, 512, 1, 128] - - [50, 40.617] - - - [37376, 8192, 1, 128] - - [38, 46.86] - - - [29824, 13825, 1, 128] - - [51, 45.018] - - - [36352, 2048, 1, 128] - - [41, 44.472] - - - [30336, 1024, 1, 128] - - [36, 42.605] - - - [44416, 28417, 1, 128] - - [39, 45.544] - - - [38144, 22273, 1, 128] - - [39, 45.787] - - - [28928, 2048, 1, 128] - - [61, 43.673] - - - [29568, 13697, 1, 128] - - [39, 45.321] - - - [43136, 27137, 1, 128] - - [23, 45.344] - - - [42112, 4096, 1, 128] - - [38, 45.686] - - - [40960, 512, 1, 128] - - [45, 40.455] - - - [35584, 1024, 1, 128] - - [64, 43.056] - - - [31232, 15361, 1, 128] - - [26, 45.928] - - - [40960, 8192, 1, 128] - - [26, 41.571] - - - [31232, 1024, 1, 128] - - [35, 43.531] - - - [29312, 512, 1, 128] - - [41, 39.334] - - - [44416, 512, 1, 128] - - [36, 41.196] - - - [42240, 512, 1, 128] - - [54, 41.921] - - - [31232, 8192, 1, 128] - - [28, 46.787] - - - [35072, 19201, 1, 128] - - [38, 45.897] - - - [29568, 128, 1, 128] - - [36, 26.027] - - - [33792, 2048, 1, 128] - - [54, 44.646] - - - [35712, 2048, 1, 128] - - [64, 44.276] - - - [40576, 128, 1, 128] - - [64, 32.708] - - - [40704, 1024, 1, 128] - - [35, 43.238] - - - [29824, 1024, 1, 128] - - [32, 42.016] - - - [33536, 17665, 1, 128] - - [65, 46.186] - - - [43008, 27009, 1, 128] - - [23, 46.557] - - - [34304, 2048, 1, 128] - - [73, 44.427] - - - [37120, 21249, 1, 128] - - [30, 45.742] - - - [41600, 1024, 1, 128] - - [26, 43.256] - - - [33024, 1024, 1, 128] - - [35, 42.574] - - - [42368, 512, 1, 128] - - [58, 40.466] - - - [30592, 14721, 1, 128] - - [39, 45.745] - - - [29696, 2048, 1, 128] - - [45, 44.525] - - - [31232, 128, 1, 128] - - [41, 27.421] - - - [38784, 22785, 1, 128] - - [23, 45.55] - - - [32896, 1024, 1, 128] - - [98, 37.593] - - - [32128, 128, 1, 128] - - [36, 27.849] - - - [35968, 8192, 1, 128] - - [38, 46.444] - - - [38400, 2048, 1, 128] - - [57, 44.604] - - - [36864, 2048, 1, 128] - - [28, 44.373] - - - [31616, 4096, 1, 128] - - [65, 45.071] - - - [34688, 18817, 1, 128] - - [23, 45.573] - - - [42624, 4096, 1, 128] - - [30, 41.862] - - - [29312, 1024, 1, 128] - - [35, 42.221] - - - [37760, 2048, 1, 128] - - [59, 44.239] - - - [39808, 512, 1, 128] - - [45, 40.895] - - - [41472, 128, 1, 128] - - [41, 33.401] - - - [32128, 4096, 1, 128] - - [38, 45.8] - - - [43520, 4096, 1, 128] - - [28, 46.119] - - - [41472, 512, 1, 128] - - [26, 42.267] - - - [38912, 22913, 1, 128] - - [39, 46.708] - - - [30464, 1024, 1, 128] - - [28, 42.124] - - - [33280, 128, 1, 128] - - [41, 28.747] - - - [31872, 15873, 1, 128] - - [39, 45.676] - - - [36352, 4096, 1, 128] - - [26, 45.942] - - - [30720, 2048, 1, 128] - - [38, 44.652] - - - [33792, 128, 1, 128] - - [64, 29.019] - - - [36096, 8192, 1, 128] - - [60, 45.087] - - - [38784, 128, 1, 128] - - [24, 32.111] - - - [30208, 2048, 1, 128] - - [40, 44.35] - - - [34432, 4096, 1, 128] - - [51, 44.362] - - - [42880, 128, 1, 128] - - [32, 30.95] - - - [31616, 15745, 1, 128] - - [23, 45.341] - - - [40960, 2048, 1, 128] - - [23, 39.906] - - - [41344, 128, 1, 128] - - [24, 33.298] - - - [41728, 25857, 1, 128] - - [65, 45.145] - - - [32896, 512, 1, 128] - - [72, 32.523] - - - [41728, 2048, 1, 128] - - [60, 43.251] - - - [42368, 26369, 1, 128] - - [48, 45.288] - - - [30720, 14721, 1, 128] - - [30, 46.809] - - - [37376, 512, 1, 128] - - [41, 41.061] - - - [35456, 19457, 1, 128] - - [28, 45.548] - - - [29184, 13185, 1, 128] - - [23, 45.997] - - - [34944, 128, 1, 128] - - [73, 29.692] - - - [36608, 20609, 1, 128] - - [28, 45.82] - - - [35584, 19585, 1, 128] - - [39, 45.685] - - - [42880, 8192, 1, 128] - - [48, 45.728] - - - [39936, 1024, 1, 128] - - [64, 43.517] - - - [34944, 19073, 1, 128] - - [30, 45.75] - - - [32512, 128, 1, 128] - - [64, 28.309] - - - [40064, 512, 1, 128] - - [28, 41.616] - - - [30464, 2048, 1, 128] - - [52, 42.134] - - - [30592, 8192, 1, 128] - - [28, 46.473] - - - [39040, 512, 1, 128] - - [35, 40.947] - - - [41088, 128, 1, 128] - - [80, 33.019] - - - [29824, 128, 1, 128] - - [61, 26.31] - - - [32384, 128, 1, 128] - - [40, 28.038] - - - [41728, 25729, 1, 128] - - [51, 45.1] - - - [30976, 4096, 1, 128] - - [52, 44.37] - - - [42624, 128, 1, 128] - - [80, 30.795] - - - [42112, 512, 1, 128] - - [30, 41.873] - - - [38784, 2048, 1, 128] - - [61, 44.143] - - - [35200, 8192, 1, 128] - - [28, 46.357] - - - [30976, 128, 1, 128] - - [36, 26.972] - - - [32640, 16641, 1, 128] - - [28, 45.367] - - - [41984, 8192, 1, 128] - - [28, 47.088] - - - [30080, 128, 1, 128] - - [41, 26.605] - - - [35584, 512, 1, 128] - - [70, 39.6] - - - [44800, 2048, 1, 128] - - [39, 44.034] - - - [34048, 128, 1, 128] - - [41, 29.335] - - - [35712, 1024, 1, 128] - - [50, 42.79] - - - [43136, 128, 1, 128] - - [34, 31.165] - - - [33280, 1024, 1, 128] - - [32, 42.848] - - - [34816, 18945, 1, 128] - - [23, 46.764] - - - [40704, 8192, 1, 128] - - [38, 46.688] - - - [34304, 128, 1, 128] - - [45, 29.487] - - - [39936, 512, 1, 128] - - [35, 41.614] - - - [36096, 2048, 1, 128] - - [79, 42.932] - - - [40832, 8192, 1, 128] - - [30, 46.421] - - - [37760, 4096, 1, 128] - - [38, 45.537] - - - [36736, 512, 1, 128] - - [56, 40.645] - - - [31744, 8192, 1, 128] - - [28, 47.209] - - - [33920, 1024, 1, 128] - - [64, 42.96] - - - [39808, 128, 1, 128] - - [34, 32.336] - - - [36608, 2048, 1, 128] - - [61, 44.236] - - - [30464, 14593, 1, 128] - - [79, 42.867] - - - [35200, 19201, 1, 128] - - [39, 45.538] - - - [41472, 1024, 1, 128] - - [28, 43.428] - - - [30720, 128, 1, 128] - - [41, 26.818] - - - [41600, 128, 1, 128] - - [60, 33.352] - - - [38144, 22145, 1, 128] - - [23, 45.754] - - - [37120, 4096, 1, 128] - - [39, 45.686] - - - [40704, 24705, 1, 128] - - [23, 45.747] - - - [41088, 25217, 1, 128] - - [39, 44.43] - - - [43776, 8192, 1, 128] - - [62, 45.549] - - - [38912, 1024, 1, 128] - - [35, 43.79] - - - [43008, 2048, 1, 128] - - [26, 44.887] - - - [42496, 26497, 1, 128] - - [65, 46.082] - - - [33536, 512, 1, 128] - - [56, 41.487] - - - [43520, 512, 1, 128] - - [36, 41.357] - - - [39040, 23169, 1, 128] - - [23, 45.446] - - - [29568, 2048, 1, 128] - - [41, 43.692] - - - [44672, 8192, 1, 128] - - [28, 46.483] - - - [29824, 512, 1, 128] - - [61, 39.696] - - - [34944, 2048, 1, 128] - - [41, 44.314] - - - [33408, 4096, 1, 128] - - [26, 45.857] - - - [41600, 25729, 1, 128] - - [39, 45.529] - - - [40832, 2048, 1, 128] - - [28, 44.139] - - - [38912, 128, 1, 128] - - [41, 32.145] - - - [34048, 2048, 1, 128] - - [23, 43.336] - - - [43904, 2048, 1, 128] - - [54, 44.341] - - - [39296, 23297, 1, 128] - - [23, 45.524] - - - [31232, 4096, 1, 128] - - [38, 45.845] - - - [35840, 1024, 1, 128] - - [24, 43.597] - - - [28928, 128, 1, 128] - - [41, 25.434] - - - [42752, 2048, 1, 128] - - [39, 44.089] - - - [44032, 1024, 1, 128] - - [24, 43.694] - - - [29440, 13569, 1, 128] - - [26, 45.99] - - - [35456, 19585, 1, 128] - - [30, 45.703] - - - [35840, 19841, 1, 128] - - [38, 46.395] - - - [31360, 128, 1, 128] - - [37, 27.409] - - - [40192, 2048, 1, 128] - - [61, 44.606] - - - [33920, 8192, 1, 128] - - [28, 46.43] - - - [43648, 512, 1, 128] - - [26, 40.198] - - - [30080, 14209, 1, 128] - - [46, 43.283] - - - [39680, 23809, 1, 128] - - [23, 45.422] - - - [32512, 512, 1, 128] - - [73, 39.476] - - - [34816, 2048, 1, 128] - - [26, 44.568] - - - [43392, 1024, 1, 128] - - [56, 42.75] - - - [39040, 4096, 1, 128] - - [28, 45.454] - - - [43264, 4096, 1, 128] - - [30, 45.807] - - - [44416, 2048, 1, 128] - - [37, 44.36] - - - [31488, 512, 1, 128] - - [38, 40.673] - - - [31616, 1024, 1, 128] - - [28, 43.131] - - - [44032, 8192, 1, 128] - - [28, 47.089] - - - [39424, 23553, 1, 128] - - [39, 45.76] - - - [31360, 8192, 1, 128] - - [28, 46.446] - - - [42752, 128, 1, 128] - - [24, 31.193] - - - [40192, 512, 1, 128] - - [28, 41.544] - - - [36096, 20225, 1, 128] - - [48, 44.722] - - - [41984, 26113, 1, 128] - - [23, 46.343] - - - [39936, 2048, 1, 128] - - [59, 44.806] - - - [42880, 2048, 1, 128] - - [34, 43.617] - - - [29440, 128, 1, 128] - - [35, 26.096] - - - [40192, 128, 1, 128] - - [36, 32.619] - - - [36608, 4096, 1, 128] - - [38, 45.643] - - - [37760, 21761, 1, 128] - - [23, 45.57] - - - [44160, 28161, 1, 128] - - [30, 45.034] - - - [44288, 512, 1, 128] - - [50, 40.758] - - - [29056, 13185, 1, 128] - - [30, 45.782] - - - [43904, 512, 1, 128] - - [56, 40.392] - - - [29696, 128, 1, 128] - - [36, 26.197] - - - [36224, 8192, 1, 128] - - [28, 46.49] - - - [33024, 2048, 1, 128] - - [41, 44.345] - - - [44032, 28161, 1, 128] - - [39, 46.258] - - - [44032, 128, 1, 128] - - [38, 31.413] - - - [38784, 512, 1, 128] - - [73, 40.788] - - - [29056, 8192, 1, 128] - - [38, 46.411] - - - [33920, 18049, 1, 128] - - [39, 45.703] - - - [34816, 1024, 1, 128] - - [36, 43.646] - - - [29056, 128, 1, 128] - - [41, 25.546] - - - [39552, 1024, 1, 128] - - [24, 43.178] - - - [36992, 8192, 1, 128] - - [38, 46.344] - - - [44544, 1024, 1, 128] - - [41, 43.492] - - - [43904, 27905, 1, 128] - - [26, 45.391] - - - [29440, 512, 1, 128] - - [36, 39.452] - - - [29568, 8192, 1, 128] - - [30, 46.022] - - - [41472, 2048, 1, 128] - - [73, 44.493] - - - [29184, 8192, 1, 128] - - [36, 46.616] - - - [33408, 512, 1, 128] - - [61, 40.96] - - - [38656, 22785, 1, 128] - - [65, 45.782] - - - [31744, 15745, 1, 128] - - [38, 46.536] - - - [38656, 2048, 1, 128] - - [64, 44.176] - - - [30080, 8192, 1, 128] - - [75, 44.644] - - - [44672, 128, 1, 128] - - [26, 31.601] - - - [40704, 24833, 1, 128] - - [51, 45.754] - - - [33792, 8192, 1, 128] - - [28, 47.181] - - - [33920, 512, 1, 128] - - [36, 41.639] - - - [40576, 1024, 1, 128] - - [54, 43.164] - - - [36224, 20225, 1, 128] - - [39, 45.701] - - - [34432, 1024, 1, 128] - - [28, 42.711] - - - [31488, 15617, 1, 128] - - [28, 45.941] - - - [40576, 2048, 1, 128] - - [54, 44.334] - - - [30208, 512, 1, 128] - - [26, 40.117] - - - [36480, 128, 1, 128] - - [61, 30.685] - - - [37504, 21633, 1, 128] - - [31, 44.759] - - - [32896, 17025, 1, 128] - - [75, 44.944] - - - [39168, 2048, 1, 128] - - [24, 44.037] - - - [29440, 2048, 1, 128] - - [56, 43.835] - - - [29440, 13441, 1, 128] - - [28, 45.952] - - - [32640, 8192, 1, 128] - - [38, 46.225] - - - [35072, 19073, 1, 128] - - [39, 45.943] - - - [33152, 512, 1, 128] - - [41, 40.22] - - - [40576, 8192, 1, 128] - - [30, 46.372] - - - [34944, 8192, 1, 128] - - [28, 46.479] - - - [38656, 128, 1, 128] - - [90, 31.82] - - - [33536, 17537, 1, 128] - - [51, 46.122] - - - [29952, 512, 1, 128] - - [35, 39.867] - - - [31488, 2048, 1, 128] - - [41, 44.242] - - - [31872, 4096, 1, 128] - - [39, 45.435] - - - [31232, 15233, 1, 128] - - [51, 46.235] - - - [38912, 23041, 1, 128] - - [23, 46.582] - - - [31232, 2048, 1, 128] - - [70, 44.357] - - - [40448, 8192, 1, 128] - - [48, 46.84] - - - [36352, 128, 1, 128] - - [37, 30.612] - - - [43776, 4096, 1, 128] - - [62, 44.709] - - - [32000, 8192, 1, 128] - - [30, 46.687] - - - [37760, 8192, 1, 128] - - [26, 46.459] - - - [30080, 1024, 1, 128] - - [32, 41.65] - - - [44544, 128, 1, 128] - - [58, 32.088] - - - [29696, 1024, 1, 128] - - [81, 42.71] - - - [32640, 512, 1, 128] - - [26, 39.477] - - - [44416, 128, 1, 128] - - [58, 31.571] - - - [41216, 512, 1, 128] - - [60, 41.238] - - - [31872, 512, 1, 128] - - [41, 40.288] - - - [34432, 512, 1, 128] - - [28, 41.713] - - - [34560, 1024, 1, 128] - - [41, 43.417] - - - [42240, 128, 1, 128] - - [73, 33.682] - - - [44288, 28289, 1, 128] - - [23, 45.366] - - - [30336, 14337, 1, 128] - - [48, 45.421] - - - [32384, 2048, 1, 128] - - [64, 44.101] - - - [38400, 22401, 1, 128] - - [65, 46.128] - - - [39296, 1024, 1, 128] - - [35, 43.051] - - - [28928, 8192, 1, 128] - - [28, 46.561] - - - [40320, 2048, 1, 128] - - [64, 44.54] - - - [31104, 15233, 1, 128] - - [26, 45.866] - - - [39680, 512, 1, 128] - - [45, 41.26] - - - [34048, 18049, 1, 128] - - [65, 45.789] - - - [30720, 1024, 1, 128] - - [36, 43.411] - - - [42880, 26881, 1, 128] - - [65, 44.892] - - - [32896, 8192, 1, 128] - - [75, 45.437] - - - [43264, 8192, 1, 128] - - [38, 46.531] - - - [37632, 4096, 1, 128] - - [23, 45.643] - - - [32256, 4096, 1, 128] - - [58, 46.088] - - - [37248, 4096, 1, 128] - - [28, 45.415] - - - [33280, 17409, 1, 128] - - [65, 46.304] - - - [36096, 512, 1, 128] - - [45, 39.045] - - - [37120, 21121, 1, 128] - - [28, 45.711] - - - [32896, 128, 1, 128] - - [24, 27.925] - - - [36352, 20481, 1, 128] - - [39, 45.734] - - - [43392, 2048, 1, 128] - - [38, 43.776] - - - [36352, 512, 1, 128] - - [61, 40.948] - - - [29056, 13057, 1, 128] - - [23, 45.668] - - - [29056, 4096, 1, 128] - - [56, 45.545] - - - [37888, 4096, 1, 128] - - [38, 46.418] - - - [40320, 512, 1, 128] - - [35, 41.566] - - - [39168, 128, 1, 128] - - [35, 32.284] - - - [41472, 8192, 1, 128] - - [52, 46.812] - - - [34560, 512, 1, 128] - - [61, 42.049] - - - [34176, 18305, 1, 128] - - [52, 45.623] - - - [34688, 8192, 1, 128] - - [30, 46.367] - - - [29696, 13825, 1, 128] - - [38, 46.514] - - - [33152, 17281, 1, 128] - - [38, 46.072] - - - [30208, 8192, 1, 128] - - [52, 46.932] - - - [43648, 27649, 1, 128] - - [39, 44.631] - - - [31360, 2048, 1, 128] - - [61, 43.773] - - - [41984, 128, 1, 128] - - [26, 33.557] - - - [38528, 2048, 1, 128] - - [24, 43.961] - - - [32256, 16385, 1, 128] - - [31, 46.375] - - - [42240, 1024, 1, 128] - - [61, 43.243] - - - [32000, 16001, 1, 128] - - [23, 46.01] - - - [37248, 1024, 1, 128] - - [28, 42.966] - - - [32256, 1024, 1, 128] - - [26, 43.039] - - - [39296, 23425, 1, 128] - - [23, 45.506] - - - [43008, 4096, 1, 128] - - [26, 46.736] - - - [31104, 128, 1, 128] - - [61, 27.18] - - - [38656, 8192, 1, 128] - - [65, 46.561] - - - [44288, 128, 1, 128] - - [52, 31.718] - - - [38528, 22657, 1, 128] - - [23, 45.514] - - - [39552, 128, 1, 128] - - [24, 32.199] - - - [37376, 21377, 1, 128] - - [39, 46.05] - - - [28928, 13057, 1, 128] - - [28, 45.96] - - - [43264, 27265, 1, 128] - - [23, 45.734] - - - [35328, 4096, 1, 128] - - [65, 46.012] - - - [30848, 4096, 1, 128] - - [36, 45.587] - - - [44800, 28801, 1, 128] - - [39, 45.329] - - - [35456, 512, 1, 128] - - [81, 39.962] - - - [40960, 24961, 1, 128] - - [23, 40.336] - - - [39936, 23937, 1, 128] - - [23, 46.264] - - - [31744, 1024, 1, 128] - - [45, 43.136] - - - [32128, 8192, 1, 128] - - [23, 46.465] - - - [42112, 26113, 1, 128] - - [23, 45.455] - - - [31744, 2048, 1, 128] - - [37, 44.56] - - - [42112, 1024, 1, 128] - - [24, 43.21] - - - [40064, 8192, 1, 128] - - [23, 45.948] - - - [38144, 128, 1, 128] - - [41, 31.54] - - - [42624, 2048, 1, 128] - - [28, 40.989] - - - [36992, 128, 1, 128] - - [36, 30.794] - - - [40192, 8192, 1, 128] - - [30, 46.619] - - - [40064, 24065, 1, 128] - - [51, 44.533] - - - [37760, 21889, 1, 128] - - [23, 45.545] - - - [36352, 8192, 1, 128] - - [26, 46.787] - - - [44544, 512, 1, 128] - - [58, 41.173] - - - [32384, 4096, 1, 128] - - [28, 45.721] - - - [39168, 23169, 1, 128] - - [51, 45.761] - - - [1408, 897, 1, 128] - - [124, 29.971] - - - [16512, 512, 1, 128] - - [99, 30.672] - - - [20480, 12673, 1, 128] - - [30, 47.23] - - - [20992, 512, 1, 128] - - [28, 41.196] - - - [9344, 512, 1, 128] - - [35, 32.285] - - - [18048, 2048, 1, 128] - - [26, 42.348] - - - [20352, 12673, 1, 128] - - [30, 45.916] - - - [640, 128, 1, 128] - - [129, 3.662] - - - [28160, 512, 1, 128] - - [68, 41.669] - - - [20608, 4096, 1, 128] - - [28, 44.987] - - - [19328, 1024, 1, 128] - - [35, 41.669] - - - [26496, 4096, 1, 128] - - [68, 44.965] - - - [10624, 512, 1, 128] - - [50, 32.382] - - - [20352, 1024, 1, 128] - - [30, 42.206] - - - [10240, 6529, 1, 128] - - [38, 45.357] - - - [22144, 14465, 1, 128] - - [28, 45.898] - - - [13184, 2048, 1, 128] - - [41, 41.458] - - - [14720, 6913, 1, 128] - - [26, 44.874] - - - [21248, 512, 1, 128] - - [56, 36.614] - - - [10496, 128, 1, 128] - - [115, 31.751] - - - [13056, 5377, 1, 128] - - [26, 44.678] - - - [10880, 128, 1, 128] - - [115, 29.524] - - - [18688, 512, 1, 128] - - [35, 37.935] - - - [22656, 4096, 1, 128] - - [56, 45.081] - - - [15232, 1024, 1, 128] - - [38, 40.493] - - - [20224, 4096, 1, 128] - - [36, 45.205] - - - [6016, 2305, 1, 128] - - [30, 37.786] - - - [13184, 4096, 1, 128] - - [37, 43.952] - - - [256, 129, 1, 128] - - [171, 1.465] - - - [11264, 7553, 1, 128] - - [38, 45.632] - - - [18176, 128, 1, 128] - - [34, 26.582] - - - [15872, 8193, 1, 128] - - [38, 45.629] - - - [26112, 4096, 1, 128] - - [58, 45.885] - - - [22784, 2048, 1, 128] - - [54, 43.471] - - - [10880, 7297, 1, 128] - - [38, 44.655] - - - [14720, 2048, 1, 128] - - [24, 42.96] - - - [9216, 5633, 1, 128] - - [28, 44.514] - - - [23040, 15233, 1, 128] - - [38, 46.577] - - - [8832, 5121, 1, 128] - - [30, 43.392] - - - [18816, 1024, 1, 128] - - [32, 40.938] - - - [128, 129, 1, 128] - - [120, 0.738] - - - [15488, 512, 1, 128] - - [36, 35.007] - - - [18176, 1024, 1, 128] - - [30, 40.995] - - - [16128, 8449, 1, 128] - - [28, 45.56] - - - [16000, 2048, 1, 128] - - [36, 42.288] - - - [24960, 9089, 1, 128] - - [38, 45.317] - - - [14336, 1024, 1, 128] - - [30, 39.739] - - - [25472, 8192, 1, 128] - - [26, 46.584] - - - [23040, 128, 1, 128] - - [24, 31.621] - - - [9472, 512, 1, 128] - - [28, 33.193] - - - [19072, 128, 1, 128] - - [58, 27.345] - - - [10624, 6913, 1, 128] - - [38, 44.484] - - - [7808, 1024, 1, 128] - - [30, 35.397] - - - [27008, 11137, 1, 128] - - [51, 45.459] - - - [21504, 4096, 1, 128] - - [30, 46.162] - - - [7936, 1024, 1, 128] - - [56, 36.263] - - - [12928, 5121, 1, 128] - - [54, 43.389] - - - [26240, 8192, 1, 128] - - [30, 42.362] - - - [18304, 2048, 1, 128] - - [64, 43.223] - - - [24576, 1024, 1, 128] - - [28, 41.641] - - - [10624, 128, 1, 128] - - [110, 28.619] - - - [24576, 128, 1, 128] - - [35, 33.175] - - - [25600, 9601, 1, 128] - - [39, 46.181] - - - [5248, 128, 1, 128] - - [166, 22.135] - - - [24448, 4096, 1, 128] - - [36, 45.432] - - - [19328, 128, 1, 128] - - [35, 27.717] - - - [24064, 512, 1, 128] - - [61, 39.773] - - - [11136, 512, 1, 128] - - [37, 33.263] - - - [14592, 1024, 1, 128] - - [30, 39.377] - - - [12544, 4737, 1, 128] - - [28, 44.272] - - - [17280, 128, 1, 128] - - [64, 25.266] - - - [25344, 8192, 1, 128] - - [65, 46.021] - - - [4608, 512, 1, 128] - - [40, 26.791] - - - [4608, 128, 1, 128] - - [117, 20.246] - - - [21760, 512, 1, 128] - - [35, 37.072] - - - [7936, 128, 1, 128] - - [109, 26.151] - - - [11008, 7425, 1, 128] - - [86, 44.734] - - - [13824, 2048, 1, 128] - - [41, 43.654] - - - [18048, 512, 1, 128] - - [28, 37.199] - - - [19584, 11905, 1, 128] - - [26, 45.771] - - - [22656, 512, 1, 128] - - [41, 38.172] - - - [4608, 3073, 1, 128] - - [36, 38.869] - - - [5504, 128, 1, 128] - - [124, 22.823] - - - [4864, 1024, 1, 128] - - [28, 33.684] - - - [17664, 1024, 1, 128] - - [28, 40.529] - - - [18176, 2048, 1, 128] - - [41, 43.601] - - - [2048, 1537, 1, 128] - - [60, 33.885] - - - [22528, 128, 1, 128] - - [24, 30.976] - - - [21760, 13953, 1, 128] - - [23, 46.306] - - - [7040, 128, 1, 128] - - [111, 27.349] - - - [3328, 1665, 1, 128] - - [28, 32.65] - - - [768, 512, 1, 128] - - [120, 16.196] - - - [21504, 13697, 1, 128] - - [23, 46.635] - - - [18560, 10881, 1, 128] - - [30, 45.473] - - - [2560, 128, 1, 128] - - [117, 13.997] - - - [15616, 1024, 1, 128] - - [50, 40.824] - - - [19456, 4096, 1, 128] - - [28, 45.916] - - - [25600, 2048, 1, 128] - - [29, 44.245] - - - [2304, 128, 1, 128] - - [151, 12.504] - - - [1664, 1025, 1, 128] - - [35, 19.797] - - - [23168, 15361, 1, 128] - - [26, 45.518] - - - [9856, 128, 1, 128] - - [109, 30.957] - - - [13312, 2048, 1, 128] - - [26, 42.713] - - - [19200, 512, 1, 128] - - [36, 38.611] - - - [19200, 2048, 1, 128] - - [24, 43.454] - - - [23168, 2048, 1, 128] - - [54, 43.985] - - - [18688, 128, 1, 128] - - [45, 27.163] - - - [13568, 1024, 1, 128] - - [28, 41.514] - - - [17792, 9985, 1, 128] - - [51, 45.306] - - - [20608, 1024, 1, 128] - - [48, 42.014] - - - [11648, 8065, 1, 128] - - [38, 44.991] - - - [1280, 128, 1, 128] - - [114, 7.324] - - - [16256, 4096, 1, 128] - - [30, 44.813] - - - [17024, 1024, 1, 128] - - [26, 41.744] - - - [19456, 128, 1, 128] - - [35, 28.119] - - - [20736, 512, 1, 128] - - [54, 40.399] - - - [14464, 6785, 1, 128] - - [26, 44.794] - - - [20736, 13057, 1, 128] - - [38, 46.071] - - - [8704, 2048, 1, 128] - - [28, 42.448] - - - [640, 512, 1, 128] - - [114, 13.792] - - - [768, 129, 1, 128] - - [171, 4.395] - - - [27776, 1024, 1, 128] - - [50, 42.75] - - - [19200, 11521, 1, 128] - - [28, 45.806] - - - [6400, 2048, 1, 128] - - [28, 40.695] - - - [14976, 7297, 1, 128] - - [28, 44.964] - - - [7040, 2048, 1, 128] - - [28, 42.133] - - - [25984, 128, 1, 128] - - [35, 34.207] - - - [13696, 128, 1, 128] - - [115, 33.364] - - - [2688, 1153, 1, 128] - - [81, 32.625] - - - [15232, 2048, 1, 128] - - [30, 42.804] - - - [11776, 128, 1, 128] - - [172, 29.869] - - - [3328, 512, 1, 128] - - [124, 32.644] - - - [11648, 7937, 1, 128] - - [26, 44.948] - - - [19456, 2048, 1, 128] - - [24, 43.9] - - - [11008, 128, 1, 128] - - [110, 29.228] - - - [9984, 6401, 1, 128] - - [28, 44.572] - - - [25856, 9857, 1, 128] - - [38, 45.618] - - - [4224, 512, 1, 128] - - [58, 24.753] - - - [13568, 5761, 1, 128] - - [26, 44.662] - - - [5632, 2049, 1, 128] - - [60, 37.024] - - - [8832, 2048, 1, 128] - - [26, 40.482] - - - [5632, 3969, 1, 128] - - [60, 41.798] - - - [25856, 2048, 1, 128] - - [41, 43.873] - - - [25472, 2048, 1, 128] - - [24, 43.855] - - - [20736, 12929, 1, 128] - - [28, 46.039] - - - [14592, 128, 1, 128] - - [34, 22.261] - - - [1792, 512, 1, 128] - - [110, 24.27] - - - [14208, 2048, 1, 128] - - [41, 42.396] - - - [15360, 7681, 1, 128] - - [38, 45.649] - - - [5760, 2048, 1, 128] - - [38, 38.982] - - - [6400, 512, 1, 128] - - [36, 34.368] - - - [5248, 3713, 1, 128] - - [35, 40.632] - - - [16768, 1024, 1, 128] - - [54, 40.445] - - - [10752, 512, 1, 128] - - [61, 32.745] - - - [26624, 2048, 1, 128] - - [39, 44.32] - - - [384, 128, 1, 128] - - [117, 2.232] - - - [27392, 8192, 1, 128] - - [75, 45.808] - - - [24448, 512, 1, 128] - - [36, 38.266] - - - [11136, 7553, 1, 128] - - [30, 44.898] - - - [17024, 9345, 1, 128] - - [28, 45.64] - - - [16000, 8193, 1, 128] - - [28, 45.091] - - - [5888, 2048, 1, 128] - - [37, 39.591] - - - [18304, 10497, 1, 128] - - [26, 45.496] - - - [3968, 128, 1, 128] - - [114, 17.751] - - - [14336, 6529, 1, 128] - - [30, 45.676] - - - [19840, 128, 1, 128] - - [58, 28.341] - - - [25600, 8192, 1, 128] - - [30, 47.314] - - - [18688, 11009, 1, 128] - - [26, 45.921] - - - [7680, 1024, 1, 128] - - [26, 35.722] - - - [7168, 128, 1, 128] - - [173, 24.049] - - - [1664, 512, 1, 128] - - [109, 26.701] - - - [12544, 1024, 1, 128] - - [28, 39.443] - - - [6528, 2048, 1, 128] - - [30, 40.762] - - - [19072, 4096, 1, 128] - - [35, 45.091] - - - [2048, 512, 1, 128] - - [174, 27.115] - - - [13568, 5889, 1, 128] - - [28, 44.448] - - - [23680, 16001, 1, 128] - - [65, 45.198] - - - [26112, 10113, 1, 128] - - [51, 46.101] - - - [15872, 128, 1, 128] - - [58, 23.922] - - - [16384, 512, 1, 128] - - [30, 33.327] - - - [9856, 6273, 1, 128] - - [38, 42.549] - - - [26368, 1024, 1, 128] - - [61, 42.303] - - - [16256, 2048, 1, 128] - - [28, 42.32] - - - [3968, 2305, 1, 128] - - [35, 36.793] - - - [28672, 8192, 1, 128] - - [26, 47.54] - - - [10368, 1024, 1, 128] - - [35, 40.775] - - - [11008, 1024, 1, 128] - - [69, 35.775] - - - [11776, 4097, 1, 128] - - [32, 43.237] - - - [26496, 2048, 1, 128] - - [40, 43.274] - - - [17792, 4096, 1, 128] - - [58, 45.005] - - - [2304, 512, 1, 128] - - [110, 29.196] - - - [9216, 2048, 1, 128] - - [28, 41.282] - - - [12416, 512, 1, 128] - - [81, 36.201] - - - [18048, 128, 1, 128] - - [60, 26.342] - - - [21888, 14209, 1, 128] - - [77, 41.982] - - - [9344, 5761, 1, 128] - - [26, 44.358] - - - [19712, 2048, 1, 128] - - [42, 41.116] - - - [12288, 1024, 1, 128] - - [26, 39.436] - - - [3584, 1921, 1, 128] - - [50, 33.069] - - - [22784, 128, 1, 128] - - [59, 31.335] - - - [26880, 128, 1, 128] - - [64, 35.57] - - - [17408, 1024, 1, 128] - - [28, 41.997] - - - [15488, 4096, 1, 128] - - [26, 44.919] - - - [13312, 5633, 1, 128] - - [30, 45.186] - - - [22016, 14337, 1, 128] - - [31, 46.472] - - - [19328, 2048, 1, 128] - - [64, 43.728] - - - [25600, 128, 1, 128] - - [61, 33.998] - - - [22784, 15105, 1, 128] - - [65, 46.187] - - - [5376, 3713, 1, 128] - - [35, 41.376] - - - [14208, 512, 1, 128] - - [54, 32.868] - - - [12928, 4096, 1, 128] - - [35, 44.633] - - - [768, 257, 1, 128] - - [120, 8.688] - - - [27776, 11777, 1, 128] - - [38, 45.416] - - - [12032, 1024, 1, 128] - - [35, 40.109] - - - [14208, 4096, 1, 128] - - [26, 44.998] - - - [19840, 12161, 1, 128] - - [38, 45.78] - - - [17536, 512, 1, 128] - - [54, 38.659] - - - [19840, 4096, 1, 128] - - [35, 45.206] - - - [26624, 512, 1, 128] - - [35, 40.688] - - - [27136, 11137, 1, 128] - - [31, 46.29] - - - [11008, 512, 1, 128] - - [35, 32.65] - - - [1024, 513, 1, 128] - - [167, 18.031] - - - [15744, 512, 1, 128] - - [54, 35.825] - - - [22016, 128, 1, 128] - - [82, 30.621] - - - [9344, 1024, 1, 128] - - [28, 38.357] - - - [28544, 1024, 1, 128] - - [30, 42.309] - - - [13440, 5633, 1, 128] - - [51, 43.699] - - - [21632, 13825, 1, 128] - - [39, 45.844] - - - [24064, 4096, 1, 128] - - [31, 45.78] - - - [24192, 512, 1, 128] - - [24, 39.572] - - - [22912, 15233, 1, 128] - - [28, 46.056] - - - [20864, 13185, 1, 128] - - [30, 45.554] - - - [8064, 4353, 1, 128] - - [28, 42.586] - - - [8704, 5121, 1, 128] - - [38, 43.442] - - - [19840, 1024, 1, 128] - - [26, 41.559] - - - [15616, 128, 1, 128] - - [41, 23.393] - - - [21632, 512, 1, 128] - - [36, 36.974] - - - [13440, 512, 1, 128] - - [35, 38.697] - - - [23936, 128, 1, 128] - - [35, 32.37] - - - [8960, 5377, 1, 128] - - [28, 43.666] - - - [27008, 512, 1, 128] - - [35, 40.97] - - - [13440, 5761, 1, 128] - - [58, 43.687] - - - [3072, 512, 1, 128] - - [175, 30.642] - - - [4096, 1024, 1, 128] - - [58, 29.651] - - - [7296, 3585, 1, 128] - - [36, 41.586] - - - [12416, 4737, 1, 128] - - [32, 43.615] - - - [6912, 512, 1, 128] - - [35, 36.398] - - - [11136, 2048, 1, 128] - - [41, 41.832] - - - [18176, 10369, 1, 128] - - [26, 45.848] - - - [14976, 4096, 1, 128] - - [26, 44.963] - - - [19712, 4096, 1, 128] - - [41, 43.38] - - - [8064, 1024, 1, 128] - - [26, 34.668] - - - [9600, 128, 1, 128] - - [124, 30.281] - - - [26240, 1024, 1, 128] - - [28, 41.079] - - - [5248, 3585, 1, 128] - - [30, 40.924] - - - [16768, 2048, 1, 128] - - [41, 42.748] - - - [13184, 128, 1, 128] - - [115, 32.655] - - - [19328, 11521, 1, 128] - - [38, 45.479] - - - [4864, 512, 1, 128] - - [40, 27.847] - - - [3584, 2049, 1, 128] - - [28, 33.674] - - - [18560, 128, 1, 128] - - [35, 26.93] - - - [27392, 11393, 1, 128] - - [26, 44.508] - - - [27520, 512, 1, 128] - - [73, 40.386] - - - [18176, 4096, 1, 128] - - [38, 45.274] - - - [7808, 4225, 1, 128] - - [50, 42.956] - - - [15232, 128, 1, 128] - - [73, 22.86] - - - [25728, 1024, 1, 128] - - [35, 42.049] - - - [23936, 512, 1, 128] - - [35, 39.341] - - - [23424, 2048, 1, 128] - - [33, 42.136] - - - [28032, 12161, 1, 128] - - [26, 45.603] - - - [27136, 512, 1, 128] - - [32, 41.025] - - - [14336, 6657, 1, 128] - - [23, 45.789] - - - [15616, 4096, 1, 128] - - [38, 45.338] - - - [3328, 1793, 1, 128] - - [26, 34.56] - - - [28416, 512, 1, 128] - - [77, 38.809] - - - [16384, 8705, 1, 128] - - [30, 37.424] - - - [3200, 1537, 1, 128] - - [38, 32.96] - - - [26368, 128, 1, 128] - - [64, 35.08] - - - [16000, 512, 1, 128] - - [70, 34.729] - - - [25216, 9345, 1, 128] - - [28, 45.479] - - - [28288, 4096, 1, 128] - - [23, 45.457] - - - [24832, 512, 1, 128] - - [70, 38.5] - - - [18048, 10369, 1, 128] - - [25, 44.334] - - - [20480, 4096, 1, 128] - - [35, 46.324] - - - [17792, 10113, 1, 128] - - [25, 45.432] - - - [13312, 5505, 1, 128] - - [30, 45.025] - - - [17024, 2048, 1, 128] - - [45, 43.476] - - - [20608, 12929, 1, 128] - - [30, 45.702] - - - [16896, 4096, 1, 128] - - [58, 45.64] - - - [27776, 2048, 1, 128] - - [64, 43.876] - - - [6912, 3201, 1, 128] - - [50, 41.76] - - - [15744, 2048, 1, 128] - - [61, 43.428] - - - [24448, 128, 1, 128] - - [41, 32.828] - - - [2688, 128, 1, 128] - - [117, 14.172] - - - [7808, 2048, 1, 128] - - [28, 40.69] - - - [1408, 512, 1, 128] - - [124, 23.224] - - - [12032, 512, 1, 128] - - [35, 35.58] - - - [26752, 512, 1, 128] - - [35, 40.792] - - - [16128, 8321, 1, 128] - - [30, 45.623] - - - [25856, 128, 1, 128] - - [60, 34.214] - - - [24064, 8192, 1, 128] - - [58, 47.018] - - - [28160, 4096, 1, 128] - - [58, 45.418] - - - [13312, 128, 1, 128] - - [115, 31.595] - - - [10112, 6401, 1, 128] - - [32, 43.844] - - - [16384, 4096, 1, 128] - - [38, 38.021] - - - [16512, 2048, 1, 128] - - [83, 39.086] - - - [27520, 11521, 1, 128] - - [28, 45.627] - - - [8192, 4481, 1, 128] - - [36, 42.962] - - - [16768, 512, 1, 128] - - [28, 35.966] - - - [6144, 128, 1, 128] - - [109, 25.055] - - - [13568, 512, 1, 128] - - [35, 38.952] - - - [9344, 5633, 1, 128] - - [35, 44.287] - - - [13440, 4096, 1, 128] - - [26, 44.134] - - - [2176, 1665, 1, 128] - - [26, 25.457] - - - [28288, 128, 1, 128] - - [41, 25.262] - - - [11776, 4096, 1, 128] - - [60, 44.562] - - - [17280, 512, 1, 128] - - [36, 38.266] - - - [5504, 3841, 1, 128] - - [100, 36.733] - - - [14848, 7041, 1, 128] - - [26, 45.163] - - - [3584, 128, 1, 128] - - [114, 16.958] - - - [26880, 8192, 1, 128] - - [38, 46.482] - - - [2944, 1409, 1, 128] - - [36, 28.732] - - - [26368, 10369, 1, 128] - - [38, 45.792] - - - [21888, 512, 1, 128] - - [59, 35.508] - - - [15872, 2048, 1, 128] - - [45, 43.266] - - - [20224, 512, 1, 128] - - [28, 40.45] - - - [24320, 8449, 1, 128] - - [38, 45.716] - - - [5632, 1024, 1, 128] - - [28, 34.129] - - - [17152, 9473, 1, 128] - - [28, 45.713] - - - [4096, 128, 1, 128] - - [167, 18.547] - - - [8832, 128, 1, 128] - - [172, 28.344] - - - [2048, 1409, 1, 128] - - [58, 31.533] - - - [28160, 12289, 1, 128] - - [52, 45.864] - - - [9088, 5505, 1, 128] - - [38, 43.99] - - - [19200, 1024, 1, 128] - - [28, 41.453] - - - [18048, 4096, 1, 128] - - [38, 44.168] - - - [12928, 512, 1, 128] - - [45, 37.398] - - - [20864, 4096, 1, 128] - - [56, 45.018] - - - [27008, 2048, 1, 128] - - [41, 43.353] - - - [16640, 128, 1, 128] - - [73, 24.927] - - - [24960, 8192, 1, 128] - - [28, 46.265] - - - [24320, 1024, 1, 128] - - [38, 42.944] - - - [23552, 15873, 1, 128] - - [26, 46.69] - - - [26240, 4096, 1, 128] - - [28, 42.717] - - - [24320, 128, 1, 128] - - [58, 32.949] - - - [26240, 128, 1, 128] - - [24, 34.73] - - - [3200, 1665, 1, 128] - - [28, 31.363] - - - [11776, 2048, 1, 128] - - [38, 42.593] - - - [6144, 512, 1, 128] - - [35, 33.481] - - - [24960, 128, 1, 128] - - [24, 33.334] - - - [23424, 128, 1, 128] - - [73, 31.852] - - - [11776, 8065, 1, 128] - - [68, 45.272] - - - [19072, 11265, 1, 128] - - [26, 45.403] - - - [8192, 4609, 1, 128] - - [35, 43.552] - - - [21888, 4096, 1, 128] - - [75, 40.382] - - - [14976, 2048, 1, 128] - - [45, 42.838] - - - [23680, 4096, 1, 128] - - [51, 44.576] - - - [14080, 1024, 1, 128] - - [28, 42.357] - - - [19968, 4096, 1, 128] - - [56, 45.484] - - - [8704, 128, 1, 128] - - [111, 28.178] - - - [23424, 15745, 1, 128] - - [30, 43.83] - - - [8320, 2048, 1, 128] - - [37, 40.584] - - - [6144, 2433, 1, 128] - - [30, 39.949] - - - [19200, 11393, 1, 128] - - [38, 45.778] - - - [28416, 128, 1, 128] - - [64, 25.438] - - - [14080, 2048, 1, 128] - - [28, 42.359] - - - [12544, 4096, 1, 128] - - [26, 45.013] - - - [17024, 128, 1, 128] - - [48, 24.991] - - - [23936, 16257, 1, 128] - - [28, 46.159] - - - [12288, 128, 1, 128] - - [124, 30.85] - - - [28800, 1024, 1, 128] - - [68, 41.985] - - - [13824, 6017, 1, 128] - - [38, 45.055] - - - [23040, 2048, 1, 128] - - [37, 43.938] - - - [9984, 6273, 1, 128] - - [28, 44.695] - - - [23680, 512, 1, 128] - - [35, 39.204] - - - [7936, 4353, 1, 128] - - [26, 43.529] - - - [24192, 2048, 1, 128] - - [37, 44.024] - - - [8448, 512, 1, 128] - - [32, 29.71] - - - [5760, 2177, 1, 128] - - [36, 38.6] - - - [22656, 14977, 1, 128] - - [23, 45.731] - - - [17024, 4096, 1, 128] - - [30, 45.364] - - - [24960, 8961, 1, 128] - - [26, 45.169] - - - [5888, 1024, 1, 128] - - [41, 34.886] - - - [9344, 2048, 1, 128] - - [28, 41.138] - - - [11520, 1024, 1, 128] - - [28, 38.952] - - - [17024, 9217, 1, 128] - - [30, 45.351] - - - [10368, 6657, 1, 128] - - [28, 44.392] - - - [21632, 2048, 1, 128] - - [54, 43.672] - - - [26880, 2048, 1, 128] - - [54, 43.869] - - - [20736, 4096, 1, 128] - - [28, 45.481] - - - [26624, 8192, 1, 128] - - [26, 47.587] - - - [26752, 2048, 1, 128] - - [41, 43.698] - - - [24192, 8321, 1, 128] - - [38, 45.195] - - - [4736, 1024, 1, 128] - - [28, 33.114] - - - [27648, 8192, 1, 128] - - [28, 47.306] - - - [27392, 11521, 1, 128] - - [75, 44.51] - - - [27776, 4096, 1, 128] - - [36, 45.284] - - - [28672, 12801, 1, 128] - - [26, 46.648] - - - [13056, 512, 1, 128] - - [64, 37.842] - - - [25088, 2048, 1, 128] - - [70, 43.926] - - - [17408, 9601, 1, 128] - - [38, 46.144] - - - [5120, 3585, 1, 128] - - [35, 40.938] - - - [13824, 512, 1, 128] - - [35, 39.335] - - - [8576, 1024, 1, 128] - - [84, 38.069] - - - [16768, 4096, 1, 128] - - [36, 45.208] - - - [25728, 9729, 1, 128] - - [28, 45.32] - - - [27392, 512, 1, 128] - - [70, 38.691] - - - [13824, 128, 1, 128] - - [115, 32.393] - - - [27264, 1024, 1, 128] - - [36, 41.537] - - - [22272, 14465, 1, 128] - - [31, 46.223] - - - [19840, 2048, 1, 128] - - [54, 43.568] - - - [18176, 10497, 1, 128] - - [26, 45.819] - - - [4992, 3329, 1, 128] - - [50, 41.312] - - - [14976, 7169, 1, 128] - - [30, 44.407] - - - [10112, 512, 1, 128] - - [59, 34.534] - - - [24704, 128, 1, 128] - - [28, 33.286] - - - [16896, 128, 1, 128] - - [70, 25.26] - - - [10880, 7169, 1, 128] - - [28, 44.091] - - - [9600, 512, 1, 128] - - [50, 33.131] - - - [22528, 1024, 1, 128] - - [28, 42.125] - - - [27008, 128, 1, 128] - - [35, 35.055] - - - [4480, 2945, 1, 128] - - [35, 40.291] - - - [15872, 8065, 1, 128] - - [26, 45.84] - - - [28672, 128, 1, 128] - - [41, 25.507] - - - [9344, 128, 1, 128] - - [111, 29.6] - - - [15360, 2048, 1, 128] - - [54, 43.411] - - - [11392, 512, 1, 128] - - [35, 33.918] - - - [9216, 128, 1, 128] - - [173, 28.702] - - - [8192, 2048, 1, 128] - - [38, 40.844] - - - [14464, 1024, 1, 128] - - [26, 39.175] - - - [4096, 2433, 1, 128] - - [60, 39.249] - - - [6528, 2945, 1, 128] - - [38, 40.713] - - - [12672, 512, 1, 128] - - [79, 36.955] - - - [26624, 128, 1, 128] - - [24, 35.168] - - - [19712, 1024, 1, 128] - - [50, 40.481] - - - [4480, 2817, 1, 128] - - [36, 39.015] - - - [13440, 2048, 1, 128] - - [56, 42.115] - - - [256, 257, 1, 128] - - [167, 2.987] - - - [16000, 128, 1, 128] - - [58, 23.867] - - - [7552, 3969, 1, 128] - - [50, 42.069] - - - [12416, 2048, 1, 128] - - [28, 41.554] - - - [18432, 512, 1, 128] - - [50, 37.578] - - - [14464, 512, 1, 128] - - [35, 33.638] - - - [1280, 769, 1, 128] - - [124, 25.116] - - - [14976, 512, 1, 128] - - [37, 34.617] - - - [28032, 4096, 1, 128] - - [50, 45.397] - - - [27904, 128, 1, 128] - - [35, 36.155] - - - [20224, 12545, 1, 128] - - [28, 45.854] - - - [15872, 4096, 1, 128] - - [38, 45.739] - - - [3456, 1793, 1, 128] - - [35, 35.322] - - - [14336, 128, 1, 128] - - [32, 21.74] - - - [21248, 2048, 1, 128] - - [61, 43.649] - - - [23040, 1024, 1, 128] - - [32, 41.885] - - - [15232, 7425, 1, 128] - - [23, 44.631] - - - [14592, 512, 1, 128] - - [35, 33.889] - - - [22912, 15105, 1, 128] - - [30, 45.856] - - - [22528, 2048, 1, 128] - - [28, 44.374] - - - [3072, 1024, 1, 128] - - [58, 32.874] - - - [17536, 4096, 1, 128] - - [26, 45.101] - - - [384, 257, 1, 128] - - [123, 4.378] - - - [14464, 6657, 1, 128] - - [30, 44.75] - - - [20096, 1024, 1, 128] - - [28, 41.66] - - - [26880, 4096, 1, 128] - - [26, 45.686] - - - [18816, 2048, 1, 128] - - [59, 43.333] - - - [17152, 512, 1, 128] - - [26, 38.061] - - - [18432, 4096, 1, 128] - - [28, 46.42] - - - [10368, 2048, 1, 128] - - [41, 42.664] - - - [1408, 769, 1, 128] - - [115, 27.266] - - - [7168, 2048, 1, 128] - - [56, 39.766] - - - [17664, 128, 1, 128] - - [60, 26.138] - - - [1152, 513, 1, 128] - - [116, 19.814] - - - [7296, 3713, 1, 128] - - [26, 42.233] - - - [24064, 2048, 1, 128] - - [73, 43.879] - - - [8576, 2048, 1, 128] - - [41, 42.014] - - - [23168, 15489, 1, 128] - - [30, 45.702] - - - [14848, 7169, 1, 128] - - [28, 44.814] - - - [2432, 512, 1, 128] - - [176, 30.686] - - - [19712, 12033, 1, 128] - - [51, 44.208] - - - [25856, 4096, 1, 128] - - [26, 45.649] - - - [17152, 9345, 1, 128] - - [31, 45.62] - - - [3712, 128, 1, 128] - - [166, 16.913] - - - [22272, 128, 1, 128] - - [60, 30.86] - - - [25600, 9729, 1, 128] - - [28, 46.199] - - - [6016, 2433, 1, 128] - - [35, 39.116] - - - [12928, 128, 1, 128] - - [115, 32.348] - - - [25088, 8192, 1, 128] - - [65, 46.92] - - - [7040, 1024, 1, 128] - - [36, 39.056] - - - [4736, 3201, 1, 128] - - [58, 39.823] - - - [16000, 1024, 1, 128] - - [38, 39.786] - - - [1920, 512, 1, 128] - - [124, 25.535] - - - [8192, 1024, 1, 128] - - [35, 36.996] - - - [8448, 4865, 1, 128] - - [28, 43.535] - - - [11136, 7425, 1, 128] - - [38, 45.079] - - - [23296, 4096, 1, 128] - - [28, 45.463] - - - [27904, 2048, 1, 128] - - [59, 43.926] - - - [23552, 4096, 1, 128] - - [38, 46.138] - - - [24960, 2048, 1, 128] - - [59, 44.025] - - - [2816, 128, 1, 128] - - [129, 14.953] - - - [7424, 3841, 1, 128] - - [38, 42.525] - - - [20480, 128, 1, 128] - - [24, 29.193] - - - [18816, 11137, 1, 128] - - [26, 45.576] - - - [26496, 128, 1, 128] - - [35, 34.688] - - - [16896, 9217, 1, 128] - - [26, 45.874] - - - [23296, 512, 1, 128] - - [36, 38.895] - - - [8064, 2048, 1, 128] - - [26, 40.104] - - - [19968, 128, 1, 128] - - [58, 28.741] - - - [8320, 4737, 1, 128] - - [76, 42.111] - - - [27648, 1024, 1, 128] - - [36, 43.172] - - - [3712, 512, 1, 128] - - [36, 22.02] - - - [256, 128, 1, 128] - - [167, 1.5] - - - [3072, 1537, 1, 128] - - [36, 31.708] - - - [5504, 1024, 1, 128] - - [28, 32.134] - - - [20992, 2048, 1, 128] - - [64, 44.058] - - - [20480, 1024, 1, 128] - - [35, 41.912] - - - [20864, 128, 1, 128] - - [73, 29.177] - - - [28544, 12545, 1, 128] - - [38, 45.641] - - - [1152, 512, 1, 128] - - [167, 20.49] - - - [24320, 8321, 1, 128] - - [28, 45.636] - - - [2688, 512, 1, 128] - - [126, 29.178] - - - [27904, 8192, 1, 128] - - [51, 46.305] - - - [3840, 2177, 1, 128] - - [38, 36.982] - - - [25344, 128, 1, 128] - - [70, 33.786] - - - [13184, 512, 1, 128] - - [41, 38.064] - - - [7680, 512, 1, 128] - - [28, 27.702] - - - [11904, 2048, 1, 128] - - [45, 42.903] - - - [12544, 512, 1, 128] - - [37, 36.617] - - - [8448, 4737, 1, 128] - - [28, 43.198] - - - [28544, 128, 1, 128] - - [64, 25.216] - - - [21760, 14081, 1, 128] - - [28, 46.293] - - - [12800, 128, 1, 128] - - [122, 31.388] - - - [17664, 4096, 1, 128] - - [35, 45.009] - - - [2432, 1793, 1, 128] - - [50, 29.769] - - - [16384, 8577, 1, 128] - - [30, 37.574] - - - [28544, 512, 1, 128] - - [60, 38.922] - - - [28032, 12033, 1, 128] - - [38, 45.692] - - - [4864, 3329, 1, 128] - - [30, 40.718] - - - [12928, 5249, 1, 128] - - [38, 44.052] - - - [4736, 512, 1, 128] - - [34, 27.214] - - - [27264, 2048, 1, 128] - - [38, 42.91] - - - [19840, 12033, 1, 128] - - [28, 45.797] - - - [19584, 4096, 1, 128] - - [28, 45.273] - - - [21376, 4096, 1, 128] - - [38, 45.323] - - - [20352, 4096, 1, 128] - - [56, 45.354] - - - [6400, 2689, 1, 128] - - [36, 41.995] - - - [24704, 8192, 1, 128] - - [62, 45.87] - - - [22528, 14849, 1, 128] - - [28, 46.927] - - - [18304, 512, 1, 128] - - [36, 37.603] - - - [6656, 1024, 1, 128] - - [35, 38.433] - - - [13568, 4096, 1, 128] - - [30, 45.258] - - - [6016, 512, 1, 128] - - [85, 32.844] - - - [17664, 2048, 1, 128] - - [30, 42.904] - - - [17408, 512, 1, 128] - - [61, 38.114] - - - [24960, 4096, 1, 128] - - [38, 45.456] - - - [20608, 12801, 1, 128] - - [38, 45.592] - - - [27648, 11649, 1, 128] - - [26, 46.394] - - - [5760, 128, 1, 128] - - [110, 23.233] - - - [17792, 512, 1, 128] - - [41, 36.774] - - - [17664, 512, 1, 128] - - [37, 37.061] - - - [19968, 12161, 1, 128] - - [26, 46.15] - - - [19840, 512, 1, 128] - - [38, 39.521] - - - [12032, 4353, 1, 128] - - [28, 43.908] - - - [25984, 512, 1, 128] - - [79, 40.063] - - - [27648, 4096, 1, 128] - - [38, 46.345] - - - [10752, 7041, 1, 128] - - [26, 45.241] - - - [28544, 2048, 1, 128] - - [45, 43.938] - - - [7680, 2048, 1, 128] - - [38, 41.027] - - - [13184, 5377, 1, 128] - - [32, 43.357] - - - [6784, 3201, 1, 128] - - [36, 40.973] - - - [16384, 2048, 1, 128] - - [26, 36.708] - - - [22656, 1024, 1, 128] - - [28, 41.7] - - - [12800, 512, 1, 128] - - [41, 37.247] - - - [23936, 1024, 1, 128] - - [28, 42.686] - - - [15360, 1024, 1, 128] - - [28, 40.76] - - - [15488, 2048, 1, 128] - - [61, 43.276] - - - [11392, 1024, 1, 128] - - [28, 38.453] - - - [15744, 1024, 1, 128] - - [35, 41.287] - - - [9856, 2048, 1, 128] - - [26, 41.072] - - - [5888, 2305, 1, 128] - - [56, 40.7] - - - [10496, 512, 1, 128] - - [85, 35.51] - - - [1664, 1153, 1, 128] - - [60, 22.184] - - - [3456, 1024, 1, 128] - - [36, 36.005] - - - [20992, 13313, 1, 128] - - [30, 46.058] - - - [11904, 4096, 1, 128] - - [28, 44.407] - - - [13056, 1024, 1, 128] - - [28, 40.26] - - - [12800, 2048, 1, 128] - - [70, 42.43] - - - [12160, 512, 1, 128] - - [54, 36.03] - - - [5760, 2049, 1, 128] - - [24, 38.035] - - - [11392, 128, 1, 128] - - [115, 30.138] - - - [5632, 128, 1, 128] - - [110, 22.716] - - - [11520, 2048, 1, 128] - - [38, 41.898] - - - [11648, 2048, 1, 128] - - [54, 42.355] - - - [28544, 8192, 1, 128] - - [38, 46.393] - - - [22912, 1024, 1, 128] - - [28, 42.084] - - - [10752, 7169, 1, 128] - - [26, 45.01] - - - [8320, 128, 1, 128] - - [110, 26.935] - - - [23808, 1024, 1, 128] - - [56, 42.655] - - - [25984, 8192, 1, 128] - - [25, 46.384] - - - [22656, 2048, 1, 128] - - [59, 43.771] - - - [7296, 1024, 1, 128] - - [26, 33.942] - - - [28032, 512, 1, 128] - - [30, 41.815] - - - [22400, 2048, 1, 128] - - [45, 43.599] - - - [22144, 512, 1, 128] - - [24, 37.331] - - - [13312, 4096, 1, 128] - - [26, 45.594] - - - [10240, 2048, 1, 128] - - [38, 42.314] - - - [12672, 128, 1, 128] - - [115, 32.477] - - - [10752, 2048, 1, 128] - - [61, 42.028] - - - [1152, 128, 1, 128] - - [120, 6.643] - - - [13696, 5889, 1, 128] - - [50, 42.592] - - - [9216, 1024, 1, 128] - - [36, 37.865] - - - [17152, 128, 1, 128] - - [41, 25.431] - - - [24320, 2048, 1, 128] - - [41, 44.178] - - - [16512, 8705, 1, 128] - - [28, 44.284] - - - [3072, 1409, 1, 128] - - [30, 29.52] - - - [1024, 128, 1, 128] - - [117, 5.905] - - - [22400, 14593, 1, 128] - - [23, 45.652] - - - [4096, 512, 1, 128] - - [34, 24.897] - - - [4992, 128, 1, 128] - - [114, 21.548] - - - [9472, 5889, 1, 128] - - [38, 44.292] - - - [9472, 5761, 1, 128] - - [26, 43.936] - - - [27136, 1024, 1, 128] - - [32, 43.113] - - - [6528, 1024, 1, 128] - - [36, 37.555] - - - [25472, 1024, 1, 128] - - [41, 42.1] - - - [5120, 512, 1, 128] - - [77, 29.255] - - - [5504, 512, 1, 128] - - [35, 29.719] - - - [21120, 13441, 1, 128] - - [51, 44.906] - - - [4352, 128, 1, 128] - - [114, 19.236] - - - [8832, 5249, 1, 128] - - [26, 43.648] - - - [1536, 1025, 1, 128] - - [173, 29.099] - - - [11520, 512, 1, 128] - - [85, 34.41] - - - [5632, 2048, 1, 128] - - [25, 38.174] - - - [7424, 128, 1, 128] - - [111, 24.684] - - - [18432, 128, 1, 128] - - [59, 27.171] - - - [12672, 2048, 1, 128] - - [30, 42.183] - - - [14208, 128, 1, 128] - - [41, 21.502] - - - [15360, 7553, 1, 128] - - [28, 45.889] - - - [26496, 1024, 1, 128] - - [36, 42.264] - - - [27136, 128, 1, 128] - - [77, 35.338] - - - [12032, 2048, 1, 128] - - [59, 42.76] - - - [11648, 1024, 1, 128] - - [38, 39.161] - - - [11776, 512, 1, 128] - - [35, 35.139] - - - [1024, 512, 1, 128] - - [133, 18.435] - - - [11264, 7681, 1, 128] - - [28, 45.447] - - - [19456, 11777, 1, 128] - - [38, 46.512] - - - [14080, 4096, 1, 128] - - [38, 43.534] - - - [7040, 3329, 1, 128] - - [36, 41.787] - - - [27392, 4096, 1, 128] - - [77, 44.661] - - - [14720, 7041, 1, 128] - - [30, 44.828] - - - [19584, 1024, 1, 128] - - [37, 41.301] - - - [21376, 13569, 1, 128] - - [30, 45.75] - - - [20480, 12801, 1, 128] - - [30, 47.065] - - - [21248, 128, 1, 128] - - [41, 29.72] - - - [9728, 1024, 1, 128] - - [35, 39.467] - - - [18688, 10881, 1, 128] - - [30, 45.714] - - - [21120, 13313, 1, 128] - - [52, 44.673] - - - [20096, 2048, 1, 128] - - [41, 43.167] - - - [16640, 4096, 1, 128] - - [26, 45.585] - - - [28160, 12161, 1, 128] - - [51, 45.981] - - - [640, 129, 1, 128] - - [167, 3.662] - - - [28672, 512, 1, 128] - - [38, 39.061] - - - [12416, 4096, 1, 128] - - [26, 44.607] - - - [25344, 9473, 1, 128] - - [58, 44.899] - - - [18304, 1024, 1, 128] - - [24, 40.763] - - - [25600, 4096, 1, 128] - - [38, 46.394] - - - [22272, 512, 1, 128] - - [61, 37.987] - - - [21504, 13825, 1, 128] - - [39, 46.611] - - - [4736, 128, 1, 128] - - [117, 20.684] - - - [26496, 10625, 1, 128] - - [65, 45.111] - - - [7040, 512, 1, 128] - - [36, 36.099] - - - [14336, 4096, 1, 128] - - [28, 46.045] - - - [9216, 512, 1, 128] - - [41, 32.251] - - - [1280, 641, 1, 128] - - [109, 24.901] - - - [16768, 8961, 1, 128] - - [30, 45.569] - - - [18944, 11137, 1, 128] - - [51, 46.171] - - - [21504, 2048, 1, 128] - - [62, 44.054] - - - [21888, 1024, 1, 128] - - [35, 37.4] - - - [11264, 512, 1, 128] - - [28, 34.059] - - - [27776, 8192, 1, 128] - - [26, 46.27] - - - [10368, 6785, 1, 128] - - [36, 44.491] - - - [18432, 10753, 1, 128] - - [38, 46.711] - - - [19968, 2048, 1, 128] - - [64, 43.687] - - - [16640, 512, 1, 128] - - [85, 35.332] - - - [24576, 8577, 1, 128] - - [39, 42.526] - - - [28672, 2048, 1, 128] - - [30, 43.883] - - - [11136, 128, 1, 128] - - [109, 29.674] - - - [12288, 4609, 1, 128] - - [26, 44.544] - - - [14848, 1024, 1, 128] - - [28, 40.215] - - - [14848, 128, 1, 128] - - [36, 22.334] - - - [7424, 1024, 1, 128] - - [26, 34.422] - - - [2560, 1024, 1, 128] - - [60, 29.199] - - - [6400, 128, 1, 128] - - [124, 25.674] - - - [15488, 7809, 1, 128] - - [30, 45.064] - - - [17920, 2048, 1, 128] - - [70, 43.624] - - - [5760, 512, 1, 128] - - [36, 31.739] - - - [16640, 1024, 1, 128] - - [28, 41.308] - - - [28160, 2048, 1, 128] - - [24, 44.07] - - - [5504, 3969, 1, 128] - - [69, 37.291] - - - [11776, 1024, 1, 128] - - [28, 39.659] - - - [18816, 128, 1, 128] - - [41, 27.194] - - - [27904, 12033, 1, 128] - - [51, 45.586] - - - [11520, 7937, 1, 128] - - [28, 44.746] - - - [18944, 11265, 1, 128] - - [65, 46.055] - - - [5376, 1024, 1, 128] - - [26, 32.116] - - - [12032, 4225, 1, 128] - - [26, 43.658] - - - [5376, 128, 1, 128] - - [109, 22.045] - - - [9856, 1024, 1, 128] - - [26, 39.055] - - - [26752, 10881, 1, 128] - - [28, 45.393] - - - [20352, 128, 1, 128] - - [60, 28.794] - - - [14464, 128, 1, 128] - - [58, 21.619] - - - [1024, 385, 1, 128] - - [123, 15.36] - - - [3840, 128, 1, 128] - - [116, 17.389] - - - [24192, 128, 1, 128] - - [61, 32.783] - - - [28544, 12673, 1, 128] - - [28, 45.789] - - - [1664, 128, 1, 128] - - [116, 9.305] - - - [26752, 8192, 1, 128] - - [23, 46.339] - - - [16896, 1024, 1, 128] - - [58, 41.586] - - - [9728, 128, 1, 128] - - [110, 30.297] - - - [11264, 2048, 1, 128] - - [30, 42.067] - - - [11392, 2048, 1, 128] - - [28, 41.914] - - - [20224, 2048, 1, 128] - - [41, 43.332] - - - [26880, 1024, 1, 128] - - [59, 42.478] - - - [15104, 512, 1, 128] - - [26, 34.75] - - - [26368, 2048, 1, 128] - - [26, 43.697] - - - [6784, 3073, 1, 128] - - [56, 40.31] - - - [23168, 128, 1, 128] - - [24, 31.797] - - - [8448, 1024, 1, 128] - - [68, 37.138] - - - [16896, 9089, 1, 128] - - [25, 46.11] - - - [17536, 128, 1, 128] - - [35, 25.85] - - - [22912, 512, 1, 128] - - [37, 38.232] - - - [28032, 128, 1, 128] - - [24, 36.13] - - - [19584, 512, 1, 128] - - [59, 39.162] - - - [27136, 11265, 1, 128] - - [25, 46.115] - - - [4992, 512, 1, 128] - - [34, 28.409] - - - [8448, 128, 1, 128] - - [172, 27.47] - - - [27648, 128, 1, 128] - - [24, 35.822] - - - [16640, 2048, 1, 128] - - [37, 42.998] - - - [26752, 10753, 1, 128] - - [39, 45.477] - - - [2944, 1281, 1, 128] - - [36, 26.563] - - - [5376, 3841, 1, 128] - - [28, 42.384] - - - [10496, 6913, 1, 128] - - [30, 44.723] - - - [17024, 512, 1, 128] - - [36, 37.559] - - - [11008, 7297, 1, 128] - - [101, 44.623] - - - [14080, 128, 1, 128] - - [61, 25.612] - - - [5888, 512, 1, 128] - - [35, 32.264] - - - [19200, 128, 1, 128] - - [35, 27.968] - - - [14208, 6529, 1, 128] - - [28, 44.534] - - - [22912, 4096, 1, 128] - - [35, 45.339] - - - [14336, 2048, 1, 128] - - [38, 43.096] - - - [17792, 128, 1, 128] - - [34, 26.123] - - - [22656, 14849, 1, 128] - - [23, 45.629] - - - [19712, 512, 1, 128] - - [28, 39.479] - - - [5248, 1024, 1, 128] - - [26, 35.108] - - - [3712, 2049, 1, 128] - - [28, 34.365] - - - [24448, 8449, 1, 128] - - [28, 45.255] - - - [8192, 512, 1, 128] - - [50, 28.879] - - - [25472, 4096, 1, 128] - - [26, 45.651] - - - [25088, 512, 1, 128] - - [28, 39.464] - - - [23168, 1024, 1, 128] - - [36, 42.044] - - - [24320, 8192, 1, 128] - - [58, 46.734] - - - [24192, 8192, 1, 128] - - [26, 46.474] - - - [2176, 512, 1, 128] - - [122, 28.055] - - - [4992, 3457, 1, 128] - - [36, 41.978] - - - [896, 257, 1, 128] - - [177, 9.692] - - - [28288, 1024, 1, 128] - - [38, 42.197] - - - [20864, 1024, 1, 128] - - [38, 42.728] - - - [18432, 2048, 1, 128] - - [26, 43.921] - - - [17280, 9601, 1, 128] - - [30, 45.382] - - - [18944, 4096, 1, 128] - - [32, 45.601] - - - [13440, 128, 1, 128] - - [173, 32.958] - - - [7424, 2048, 1, 128] - - [28, 40.427] - - - [768, 128, 1, 128] - - [117, 4.429] - - - [16128, 512, 1, 128] - - [58, 34.593] - - - [28288, 12289, 1, 128] - - [30, 45.397] - - - [23552, 128, 1, 128] - - [54, 31.909] - - - [24832, 8192, 1, 128] - - [60, 46.933] - - - [10240, 1024, 1, 128] - - [26, 40.434] - - - [8960, 2048, 1, 128] - - [50, 40.69] - - - [17664, 9985, 1, 128] - - [30, 45.395] - - - [25088, 4096, 1, 128] - - [68, 45.937] - - - [7552, 2048, 1, 128] - - [61, 40.591] - - - [15104, 7297, 1, 128] - - [26, 45.225] - - - [7168, 1024, 1, 128] - - [28, 33.982] - - - [26112, 8192, 1, 128] - - [58, 47.055] - - - [24192, 1024, 1, 128] - - [28, 42.824] - - - [22912, 2048, 1, 128] - - [36, 43.139] - - - [10368, 512, 1, 128] - - [37, 35.36] - - - [22528, 4096, 1, 128] - - [30, 46.425] - - - [6528, 128, 1, 128] - - [110, 26.046] - - - [26752, 4096, 1, 128] - - [39, 45.179] - - - [2816, 512, 1, 128] - - [109, 30.344] - - - [22016, 14209, 1, 128] - - [25, 46.54] - - - [8832, 1024, 1, 128] - - [81, 36.792] - - - [16384, 128, 1, 128] - - [61, 24.795] - - - [5120, 1024, 1, 128] - - [26, 34.644] - - - [24832, 8833, 1, 128] - - [51, 45.868] - - - [11520, 128, 1, 128] - - [122, 30.26] - - - [24960, 512, 1, 128] - - [26, 38.841] - - - [27520, 2048, 1, 128] - - [64, 43.992] - - - [22272, 14593, 1, 128] - - [51, 45.954] - - - [2048, 128, 1, 128] - - [117, 11.281] - - - [2176, 1537, 1, 128] - - [36, 25.126] - - - [10496, 1024, 1, 128] - - [26, 41.279] - - - [12160, 4353, 1, 128] - - [28, 43.944] - - - [6144, 1024, 1, 128] - - [26, 35.581] - - - [26752, 1024, 1, 128] - - [41, 42.239] - - - [17280, 4096, 1, 128] - - [35, 45.135] - - - [16896, 512, 1, 128] - - [73, 36.62] - - - [4480, 128, 1, 128] - - [117, 19.801] - - - [18944, 128, 1, 128] - - [57, 27.644] - - - [9600, 2048, 1, 128] - - [28, 41.598] - - - [19456, 1024, 1, 128] - - [38, 41.595] - - - [9984, 2048, 1, 128] - - [26, 42.255] - - - [25216, 9217, 1, 128] - - [23, 45.153] - - - [19968, 1024, 1, 128] - - [30, 42.029] - - - [13952, 2048, 1, 128] - - [64, 43.274] - - - [10496, 2048, 1, 128] - - [30, 42.812] - - - [12672, 1024, 1, 128] - - [28, 39.816] - - - [19072, 11393, 1, 128] - - [39, 45.652] - - - [11008, 2048, 1, 128] - - [76, 40.313] - - - [27520, 11649, 1, 128] - - [28, 45.664] - - - [10880, 512, 1, 128] - - [35, 32.199] - - - [14592, 6785, 1, 128] - - [60, 44.037] - - - [7424, 512, 1, 128] - - [28, 26.773] - - - [13056, 5249, 1, 128] - - [38, 44.423] - - - [23296, 15489, 1, 128] - - [26, 46.155] - - - [28416, 8192, 1, 128] - - [28, 46.394] - - - [11392, 7681, 1, 128] - - [28, 44.48] - - - [18048, 1024, 1, 128] - - [26, 40.637] - - - [15616, 7809, 1, 128] - - [30, 45.364] - - - [128, 128, 1, 128] - - [114, 0.75] - - - [24704, 512, 1, 128] - - [60, 36.765] - - - [7680, 4097, 1, 128] - - [28, 42.92] - - - [16640, 8961, 1, 128] - - [26, 45.941] - - - [18944, 1024, 1, 128] - - [50, 41.631] - - - [12928, 2048, 1, 128] - - [41, 42.458] - - - [22272, 2048, 1, 128] - - [54, 43.542] - - - [27904, 11905, 1, 128] - - [52, 45.645] - - - [26240, 2048, 1, 128] - - [30, 41.678] - - - [9728, 6017, 1, 128] - - [35, 44.661] - - - [20736, 1024, 1, 128] - - [38, 42.428] - - - [3456, 1921, 1, 128] - - [36, 37.335] - - - [8064, 512, 1, 128] - - [36, 27.566] - - - [4224, 1024, 1, 128] - - [28, 29.888] - - - [25984, 10113, 1, 128] - - [31, 45.563] - - - [13696, 6017, 1, 128] - - [86, 42.514] - - - [27520, 8192, 1, 128] - - [38, 46.441] - - - [18944, 512, 1, 128] - - [35, 38.516] - - - [6272, 128, 1, 128] - - [111, 25.437] - - - [27264, 4096, 1, 128] - - [50, 45.21] - - - [1792, 1153, 1, 128] - - [52, 23.607] - - - [17536, 9729, 1, 128] - - [38, 45.307] - - - [13184, 5505, 1, 128] - - [32, 43.612] - - - [2944, 128, 1, 128] - - [116, 15.411] - - - [25344, 512, 1, 128] - - [26, 39.459] - - - [23040, 15361, 1, 128] - - [38, 46.101] - - - [8704, 512, 1, 128] - - [41, 30.861] - - - [20864, 13057, 1, 128] - - [23, 45.465] - - - [19328, 4096, 1, 128] - - [38, 45.15] - - - [28288, 8192, 1, 128] - - [39, 46.469] - - - [10112, 1024, 1, 128] - - [28, 40.387] - - - [17536, 2048, 1, 128] - - [64, 43.799] - - - [7552, 128, 1, 128] - - [109, 24.997] - - - [15616, 7937, 1, 128] - - [30, 45.393] - - - [23040, 512, 1, 128] - - [70, 38.562] - - - [25984, 2048, 1, 128] - - [40, 43.611] - - - [14720, 128, 1, 128] - - [60, 22.186] - - - [23424, 1024, 1, 128] - - [26, 41.891] - - - [1920, 1281, 1, 128] - - [35, 27.133] - - - [27136, 2048, 1, 128] - - [45, 44.323] - - - [28800, 8192, 1, 128] - - [30, 46.396] - - - [15488, 128, 1, 128] - - [36, 23.15] - - - [28800, 12929, 1, 128] - - [38, 45.572] - - - [21888, 14081, 1, 128] - - [77, 41.962] - - - [25600, 1024, 1, 128] - - [35, 42.273] - - - [21632, 1024, 1, 128] - - [41, 41.233] - - - [24448, 1024, 1, 128] - - [50, 41.88] - - - [4352, 2689, 1, 128] - - [36, 38.62] - - - [20480, 512, 1, 128] - - [28, 39.934] - - - [7296, 128, 1, 128] - - [115, 24.703] - - - [4992, 1024, 1, 128] - - [26, 34.368] - - - [27264, 11393, 1, 128] - - [30, 45.662] - - - [26752, 128, 1, 128] - - [61, 35.086] - - - [24960, 1024, 1, 128] - - [41, 41.824] - - - [21504, 512, 1, 128] - - [24, 37.142] - - - [6272, 2561, 1, 128] - - [28, 40.26] - - - [25088, 9089, 1, 128] - - [48, 45.869] - - - [20864, 512, 1, 128] - - [36, 40.683] - - - [4224, 2561, 1, 128] - - [32, 36.881] - - - [15744, 8065, 1, 128] - - [26, 45.33] - - - [21632, 128, 1, 128] - - [58, 30.024] - - - [15104, 4096, 1, 128] - - [26, 45.379] - - - [20352, 512, 1, 128] - - [64, 40.397] - - - [25472, 9601, 1, 128] - - [30, 45.504] - - - [27904, 512, 1, 128] - - [28, 41.498] - - - [19968, 512, 1, 128] - - [35, 40.046] - - - [5760, 1024, 1, 128] - - [36, 34.613] - - - [28416, 12545, 1, 128] - - [38, 45.589] - - - [16512, 8833, 1, 128] - - [28, 44.373] - - - [6016, 128, 1, 128] - - [122, 24.399] - - - [13056, 4096, 1, 128] - - [50, 45.349] - - - [19968, 12289, 1, 128] - - [28, 45.801] - - - [7424, 3713, 1, 128] - - [35, 43.114] - - - [28800, 128, 1, 128] - - [41, 25.256] - - - [512, 512, 1, 128] - - [116, 10.955] - - - [24832, 2048, 1, 128] - - [40, 43.926] - - - [20736, 128, 1, 128] - - [59, 29.22] - - - [26368, 512, 1, 128] - - [77, 40.028] - - - [26496, 8192, 1, 128] - - [58, 45.86] - - - [13824, 4096, 1, 128] - - [35, 45.2] - - - [27264, 128, 1, 128] - - [41, 35.195] - - - [21760, 1024, 1, 128] - - [36, 41.232] - - - [2432, 1921, 1, 128] - - [36, 31.632] - - - [27136, 8192, 1, 128] - - [48, 47.092] - - - [6784, 2048, 1, 128] - - [38, 41.505] - - - [11264, 128, 1, 128] - - [178, 28.769] - - - [7552, 512, 1, 128] - - [59, 27.24] - - - [19328, 11649, 1, 128] - - [28, 45.711] - - - [17152, 2048, 1, 128] - - [61, 43.822] - - - [23808, 16129, 1, 128] - - [28, 46.207] - - - [20224, 12417, 1, 128] - - [23, 45.968] - - - [27904, 1024, 1, 128] - - [56, 42.724] - - - [3456, 512, 1, 128] - - [109, 33.676] - - - [13312, 512, 1, 128] - - [79, 37.959] - - - [26368, 4096, 1, 128] - - [26, 45.563] - - - [23296, 15617, 1, 128] - - [30, 46.156] - - - [26112, 10241, 1, 128] - - [25, 45.807] - - - [26240, 512, 1, 128] - - [35, 39.854] - - - [4352, 1024, 1, 128] - - [28, 30.751] - - - [10624, 2048, 1, 128] - - [50, 41.133] - - - [23808, 16001, 1, 128] - - [26, 46.117] - - - [17536, 9857, 1, 128] - - [26, 45.598] - - - [23936, 4096, 1, 128] - - [30, 45.442] - - - [1408, 128, 1, 128] - - [167, 7.933] - - - [14848, 512, 1, 128] - - [36, 34.702] - - - [8704, 4993, 1, 128] - - [28, 44.036] - - - [15104, 2048, 1, 128] - - [61, 43.131] - - - [2560, 512, 1, 128] - - [122, 30.724] - - - [27264, 8192, 1, 128] - - [30, 46.344] - - - [23808, 4096, 1, 128] - - [39, 45.544] - - - [14080, 6273, 1, 128] - - [36, 43.219] - - - [10112, 6529, 1, 128] - - [36, 43.982] - - - [27648, 512, 1, 128] - - [41, 41.305] - - - [20992, 128, 1, 128] - - [41, 29.53] - - - [15104, 128, 1, 128] - - [58, 22.807] - - - [7808, 128, 1, 128] - - [124, 25.729] - - - [3584, 1024, 1, 128] - - [56, 26.305] - - - [15232, 512, 1, 128] - - [26, 34.983] - - - [21376, 13697, 1, 128] - - [28, 45.838] - - - [11392, 7809, 1, 128] - - [55, 44.52] - - - [11904, 1024, 1, 128] - - [26, 39.736] - - - [28800, 2048, 1, 128] - - [37, 43.549] - - - [8960, 512, 1, 128] - - [56, 31.586] - - - [19456, 11649, 1, 128] - - [30, 46.451] - - - [11904, 128, 1, 128] - - [115, 31.381] - - - [18560, 512, 1, 128] - - [28, 37.97] - - - [6656, 128, 1, 128] - - [111, 26.414] - - - [17792, 2048, 1, 128] - - [54, 43.056] - - - [21632, 4096, 1, 128] - - [38, 45.217] - - - [25728, 4096, 1, 128] - - [50, 45.444] - - - [18048, 10241, 1, 128] - - [52, 44.043] - - - [1792, 1281, 1, 128] - - [58, 25.763] - - - [512, 385, 1, 128] - - [116, 8.547] - - - [26112, 512, 1, 128] - - [24, 40.26] - - - [16128, 1024, 1, 128] - - [24, 39.928] - - - [4480, 1024, 1, 128] - - [28, 31.324] - - - [14720, 4096, 1, 128] - - [30, 44.948] - - - [23552, 2048, 1, 128] - - [45, 44.203] - - - [22528, 512, 1, 128] - - [45, 38.262] - - - [22912, 128, 1, 128] - - [64, 31.394] - - - [25344, 1024, 1, 128] - - [35, 42.112] - - - [24064, 16257, 1, 128] - - [51, 46.571] - - - [9088, 5377, 1, 128] - - [30, 44.163] - - - [27776, 128, 1, 128] - - [61, 35.926] - - - [15616, 512, 1, 128] - - [77, 35.235] - - - [13568, 128, 1, 128] - - [173, 32.943] - - - [15488, 7681, 1, 128] - - [26, 44.725] - - - [20096, 512, 1, 128] - - [41, 39.969] - - - [24832, 4096, 1, 128] - - [28, 45.784] - - - [28800, 4096, 1, 128] - - [35, 45.168] - - - [11904, 4225, 1, 128] - - [38, 43.69] - - - [3968, 1024, 1, 128] - - [35, 28.48] - - - [6400, 2817, 1, 128] - - [28, 40.365] - - - [24576, 4096, 1, 128] - - [30, 42.846] - - - [9088, 128, 1, 128] - - [115, 29.292] - - - [17152, 4096, 1, 128] - - [36, 45.392] - - - [22528, 14721, 1, 128] - - [38, 46.947] - - - [27392, 2048, 1, 128] - - [87, 42.968] - - - [8832, 512, 1, 128] - - [37, 30.98] - - - [8960, 5249, 1, 128] - - [60, 43.23] - - - [3200, 1024, 1, 128] - - [81, 33.876] - - - [4736, 3073, 1, 128] - - [28, 39.057] - - - [28032, 2048, 1, 128] - - [41, 44.151] - - - [14592, 2048, 1, 128] - - [36, 41.79] - - - [13440, 1024, 1, 128] - - [28, 41.263] - - - [14464, 2048, 1, 128] - - [64, 42.686] - - - [6912, 2048, 1, 128] - - [28, 42.332] - - - [19584, 2048, 1, 128] - - [41, 43.263] - - - [17920, 128, 1, 128] - - [58, 26.677] - - - [19584, 11777, 1, 128] - - [38, 45.594] - - - [23936, 16129, 1, 128] - - [30, 46.1] - - - [10496, 6785, 1, 128] - - [28, 44.614] - - - [27648, 2048, 1, 128] - - [61, 44.374] - - - [23808, 128, 1, 128] - - [61, 32.621] - - - [20864, 2048, 1, 128] - - [30, 43.381] - - - [9088, 512, 1, 128] - - [81, 31.659] - - - [3584, 512, 1, 128] - - [40, 22.053] - - - [8576, 4993, 1, 128] - - [26, 43.512] - - - [3328, 1024, 1, 128] - - [80, 34.549] - - - [20608, 2048, 1, 128] - - [62, 43.039] - - - [23552, 15745, 1, 128] - - [38, 46.73] - - - [23424, 15617, 1, 128] - - [46, 43.609] - - - [21120, 512, 1, 128] - - [26, 41.201] - - - [6656, 512, 1, 128] - - [61, 35.556] - - - [12544, 128, 1, 128] - - [115, 31.709] - - - [24448, 8577, 1, 128] - - [38, 45.437] - - - [9984, 512, 1, 128] - - [37, 34.496] - - - [18304, 4096, 1, 128] - - [36, 45.075] - - - [17920, 512, 1, 128] - - [40, 37.359] - - - [12160, 4096, 1, 128] - - [36, 45.148] - - - [3968, 2433, 1, 128] - - [35, 38.283] - - - [27008, 4096, 1, 128] - - [65, 45.118] - - - [22272, 1024, 1, 128] - - [38, 41.907] - - - [14336, 512, 1, 128] - - [30, 33.551] - - - [18560, 10753, 1, 128] - - [38, 45.476] - - - [6272, 2048, 1, 128] - - [35, 39.859] - - - [12800, 1024, 1, 128] - - [30, 40.056] - - - [9600, 5889, 1, 128] - - [30, 43.887] - - - [13056, 128, 1, 128] - - [173, 32.667] - - - [7296, 2048, 1, 128] - - [36, 39.711] - - - [21376, 512, 1, 128] - - [41, 36.706] - - - [11904, 512, 1, 128] - - [35, 35.236] - - - [6400, 1024, 1, 128] - - [28, 37.431] - - - [27008, 1024, 1, 128] - - [36, 42.529] - - - [22400, 14721, 1, 128] - - [23, 45.543] - - - [6272, 1024, 1, 128] - - [81, 36.799] - - - [17408, 128, 1, 128] - - [41, 25.811] - - - [26624, 10625, 1, 128] - - [30, 46.721] - - - [22400, 1024, 1, 128] - - [50, 41.793] - - - [18304, 10625, 1, 128] - - [38, 45.509] - - - [15872, 1024, 1, 128] - - [58, 40.799] - - - [21120, 128, 1, 128] - - [62, 29.43] - - - [22784, 4096, 1, 128] - - [58, 45.163] - - - [25728, 9857, 1, 128] - - [30, 45.466] - - - [16256, 1024, 1, 128] - - [26, 39.961] - - - [18560, 4096, 1, 128] - - [38, 45.201] - - - [7936, 4225, 1, 128] - - [30, 43.039] - - - [7680, 3969, 1, 128] - - [36, 42.949] - - - [9472, 2048, 1, 128] - - [41, 41.446] - - - [28160, 128, 1, 128] - - [79, 35.981] - - - [18816, 512, 1, 128] - - [37, 38.09] - - - [9856, 512, 1, 128] - - [50, 33.895] - - - [17664, 9857, 1, 128] - - [39, 45.257] - - - [27392, 128, 1, 128] - - [59, 35.553] - - - [24448, 2048, 1, 128] - - [28, 43.225] - - - [7808, 512, 1, 128] - - [28, 27.757] - - - [13952, 512, 1, 128] - - [26, 39.585] - - - [24576, 512, 1, 128] - - [26, 39.435] - - - [27520, 128, 1, 128] - - [60, 35.463] - - - [26496, 512, 1, 128] - - [61, 40.326] - - - [8576, 512, 1, 128] - - [41, 30.226] - - - [11648, 512, 1, 128] - - [61, 34.583] - - - [17408, 2048, 1, 128] - - [61, 43.526] - - - [17920, 10241, 1, 128] - - [48, 45.751] - - - [16384, 1024, 1, 128] - - [28, 37.324] - - - [6016, 2048, 1, 128] - - [28, 40.14] - - - [9728, 512, 1, 128] - - [50, 33.763] - - - [19712, 128, 1, 128] - - [35, 28.104] - - - [26112, 1024, 1, 128] - - [61, 42.492] - - - [16768, 128, 1, 128] - - [40, 24.867] - - - [8960, 1024, 1, 128] - - [26, 37.168] - - - [6784, 128, 1, 128] - - [124, 26.778] - - - [12800, 4993, 1, 128] - - [60, 44.444] - - - [6144, 2561, 1, 128] - - [36, 39.951] - - - [26880, 10881, 1, 128] - - [39, 45.591] - - - [12928, 1024, 1, 128] - - [28, 40.278] - - - [7040, 3457, 1, 128] - - [50, 42.449] - - - [15744, 4096, 1, 128] - - [30, 45.477] - - - [20096, 4096, 1, 128] - - [56, 44.898] - - - [21760, 128, 1, 128] - - [58, 30.145] - - - [7936, 2048, 1, 128] - - [38, 40.79] - - - [24448, 8192, 1, 128] - - [30, 46.582] - - - [21120, 2048, 1, 128] - - [70, 42.844] - - - [12160, 1024, 1, 128] - - [30, 40.129] - - - [7168, 3457, 1, 128] - - [36, 41.781] - - - [15232, 7553, 1, 128] - - [39, 44.833] - - - [26624, 1024, 1, 128] - - [36, 42.803] - - - [25344, 2048, 1, 128] - - [28, 43.168] - - - [12544, 4865, 1, 128] - - [30, 44.396] - - - [21120, 4096, 1, 128] - - [60, 44.469] - - - [20224, 128, 1, 128] - - [37, 28.834] - - - [14592, 4096, 1, 128] - - [32, 43.855] - - - [16256, 8577, 1, 128] - - [30, 45.045] - - - [24192, 4096, 1, 128] - - [30, 45.486] - - - [21248, 1024, 1, 128] - - [28, 41.174] - - - [25216, 1024, 1, 128] - - [26, 42.089] - - - [5888, 2177, 1, 128] - - [30, 39.11] - - - [21504, 1024, 1, 128] - - [28, 41.566] - - - [17536, 1024, 1, 128] - - [30, 42.672] - - - [9728, 2048, 1, 128] - - [41, 41.96] - - - [13952, 6273, 1, 128] - - [30, 44.621] - - - [28800, 512, 1, 128] - - [73, 38.717] - - - [2304, 1793, 1, 128] - - [28, 28.752] - - - [12416, 128, 1, 128] - - [111, 31.601] - - - [20224, 1024, 1, 128] - - [36, 42.303] - - - [22144, 128, 1, 128] - - [35, 30.683] - - - [22784, 1024, 1, 128] - - [30, 42.032] - - - [27136, 4096, 1, 128] - - [50, 46.044] - - - [27264, 512, 1, 128] - - [28, 39.551] - - - [26240, 10241, 1, 128] - - [26, 41.87] - - - [27904, 4096, 1, 128] - - [65, 45.283] - - - [21504, 128, 1, 128] - - [35, 30.135] - - - [3712, 2177, 1, 128] - - [36, 35.942] - - - [18432, 1024, 1, 128] - - [38, 41.126] - - - [28672, 4096, 1, 128] - - [28, 46.523] - - - [25344, 4096, 1, 128] - - [68, 45.027] - - - [26880, 512, 1, 128] - - [64, 40.713] - - - [21888, 2048, 1, 128] - - [88, 36.567] - - - [1792, 128, 1, 128] - - [114, 10.02] - - - [6016, 1024, 1, 128] - - [50, 35.431] - - - [15104, 7425, 1, 128] - - [28, 45.481] - - - [22016, 2048, 1, 128] - - [73, 43.988] - - - [13952, 4096, 1, 128] - - [30, 44.891] - - - [20992, 4096, 1, 128] - - [50, 45.563] - - - [8064, 4481, 1, 128] - - [38, 42.758] - - - [12672, 4096, 1, 128] - - [28, 44.756] - - - [20096, 12289, 1, 128] - - [26, 45.138] - - - [14848, 2048, 1, 128] - - [26, 43.185] - - - [23168, 512, 1, 128] - - [41, 38.602] - - - [7680, 128, 1, 128] - - [109, 25.42] - - - [13312, 1024, 1, 128] - - [38, 40.722] - - - [10624, 1024, 1, 128] - - [38, 36.621] - - - [3840, 512, 1, 128] - - [28, 22.912] - - - [22144, 14337, 1, 128] - - [38, 45.549] - - - [3200, 128, 1, 128] - - [177, 16.517] - - - [25472, 9473, 1, 128] - - [38, 45.643] - - - [16768, 9089, 1, 128] - - [26, 45.453] - - - [12288, 2048, 1, 128] - - [28, 42.3] - - - [20608, 512, 1, 128] - - [77, 40.424] - - - [2816, 1024, 1, 128] - - [58, 31.092] - - - [7552, 1024, 1, 128] - - [28, 34.968] - - - [5120, 3457, 1, 128] - - [30, 40.119] - - - [25216, 2048, 1, 128] - - [24, 43.233] - - - [12672, 4865, 1, 128] - - [28, 43.916] - - - [10880, 2048, 1, 128] - - [50, 40.607] - - - [18176, 512, 1, 128] - - [30, 37.436] - - - [8320, 4609, 1, 128] - - [28, 42.483] - - - [16000, 4096, 1, 128] - - [36, 44.862] - - - [22144, 2048, 1, 128] - - [61, 43.669] - - - [22784, 512, 1, 128] - - [56, 38.3] - - - [4096, 2561, 1, 128] - - [60, 36.95] - - - [24576, 2048, 1, 128] - - [28, 41.044] - - - [26624, 4096, 1, 128] - - [26, 46.684] - - - [18560, 2048, 1, 128] - - [41, 43.454] - - - [19584, 128, 1, 128] - - [54, 28.304] - - - [23936, 2048, 1, 128] - - [41, 43.931] - - - [23552, 512, 1, 128] - - [35, 39.286] - - - [12032, 4096, 1, 128] - - [50, 44.644] - - - [3840, 2305, 1, 128] - - [35, 36.194] - - - [25088, 128, 1, 128] - - [24, 33.681] - - - [16640, 8833, 1, 128] - - [30, 46.213] - - - [896, 128, 1, 128] - - [114, 5.167] - - - [17280, 2048, 1, 128] - - [37, 43.519] - - - [16896, 2048, 1, 128] - - [24, 43.271] - - - [22656, 128, 1, 128] - - [24, 31.094] - - - [25728, 8192, 1, 128] - - [30, 46.455] - - - [16128, 128, 1, 128] - - [73, 24.353] - - - [3840, 1024, 1, 128] - - [28, 27.77] - - - [2944, 512, 1, 128] - - [124, 31.266] - - - [24064, 1024, 1, 128] - - [26, 42.816] - - - [896, 385, 1, 128] - - [179, 13.814] - - - [8064, 128, 1, 128] - - [115, 26.572] - - - [12416, 1024, 1, 128] - - [68, 39.02] - - - [20608, 128, 1, 128] - - [41, 28.874] - - - [2944, 1024, 1, 128] - - [60, 31.974] - - - [6656, 2048, 1, 128] - - [28, 40.68] - - - [24064, 128, 1, 128] - - [59, 32.609] - - - [15744, 7937, 1, 128] - - [26, 45.436] - - - [2688, 1024, 1, 128] - - [48, 30.021] - - - [24192, 8193, 1, 128] - - [30, 44.945] - - - [24320, 4096, 1, 128] - - [38, 45.703] - - - [24576, 8705, 1, 128] - - [23, 42.522] - - - [13824, 1024, 1, 128] - - [26, 41.992] - - - [27776, 512, 1, 128] - - [41, 41.391] - - - [10240, 128, 1, 128] - - [122, 30.977] - - - [26240, 10369, 1, 128] - - [38, 41.875] - - - [16512, 4096, 1, 128] - - [38, 42.737] - - - [9856, 6145, 1, 128] - - [26, 42.117] - - - [27392, 1024, 1, 128] - - [41, 40.737] - - - [14976, 1024, 1, 128] - - [28, 40.146] - - - [1280, 512, 1, 128] - - [167, 22.362] - - - [6528, 2817, 1, 128] - - [28, 40.738] - - - [12288, 512, 1, 128] - - [54, 35.624] - - - [5248, 512, 1, 128] - - [79, 29.412] - - - [28544, 4096, 1, 128] - - [56, 45.444] - - - [21248, 13569, 1, 128] - - [52, 45.898] - - - [26112, 2048, 1, 128] - - [61, 44.634] - - - [14208, 6401, 1, 128] - - [28, 44.543] - - - [13952, 128, 1, 128] - - [115, 33.545] - - - [2304, 1665, 1, 128] - - [35, 26.796] - - - [6912, 1024, 1, 128] - - [30, 39.297] - - - [28672, 1024, 1, 128] - - [50, 42.181] - - - [14592, 6913, 1, 128] - - [58, 43.881] - - - [24704, 1024, 1, 128] - - [64, 39.923] - - - [22400, 512, 1, 128] - - [28, 37.762] - - - [23424, 4096, 1, 128] - - [36, 43.442] - - - [24832, 128, 1, 128] - - [41, 33.52] - - - [23680, 2048, 1, 128] - - [38, 42.883] - - - [25984, 9985, 1, 128] - - [31, 45.49] - - - [15360, 512, 1, 128] - - [61, 35.422] - - - [21376, 2048, 1, 128] - - [37, 43.564] - - - [16128, 2048, 1, 128] - - [64, 43.058] - - - [15872, 512, 1, 128] - - [41, 36.255] - - - [3072, 128, 1, 128] - - [117, 16.313] - - - [27520, 4096, 1, 128] - - [28, 45.446] - - - [25216, 4096, 1, 128] - - [36, 45.281] - - - [28672, 12673, 1, 128] - - [38, 46.81] - - - [28288, 512, 1, 128] - - [35, 38.591] - - - [22400, 4096, 1, 128] - - [39, 45.103] - - - [25344, 9345, 1, 128] - - [48, 45.119] - - - [9984, 128, 1, 128] - - [115, 30.964] - - - [28416, 1024, 1, 128] - - [38, 42.268] - - - [27008, 8192, 1, 128] - - [51, 46.31] - - - [13184, 1024, 1, 128] - - [28, 40.381] - - - [10240, 512, 1, 128] - - [50, 34.804] - - - [3456, 128, 1, 128] - - [167, 17.354] - - - [16000, 8321, 1, 128] - - [30, 45.088] - - - [27520, 1024, 1, 128] - - [59, 42.634] - - - [25088, 1024, 1, 128] - - [26, 42.278] - - - [6784, 512, 1, 128] - - [61, 35.653] - - - [18432, 10625, 1, 128] - - [38, 46.727] - - - [16128, 4096, 1, 128] - - [28, 45.42] - - - [26880, 11009, 1, 128] - - [30, 45.732] - - - [28800, 12801, 1, 128] - - [30, 45.526] - - - [12288, 4096, 1, 128] - - [28, 45.625] - - - [20096, 12417, 1, 128] - - [30, 45.622] - - - [1920, 128, 1, 128] - - [117, 10.498] - - - [13056, 2048, 1, 128] - - [38, 42.641] - - - [384, 385, 1, 128] - - [180, 6.362] - - - [9088, 1024, 1, 128] - - [24, 37.203] - - - [6784, 1024, 1, 128] - - [28, 39.104] - - - [21760, 4096, 1, 128] - - [35, 45.548] - - - [27008, 11009, 1, 128] - - [51, 45.472] - - - [14208, 1024, 1, 128] - - [26, 39.018] - - - [25600, 512, 1, 128] - - [29, 39.615] - - - [23680, 1024, 1, 128] - - [26, 42.082] - - - [28160, 8192, 1, 128] - - [52, 46.763] - - - [22016, 4096, 1, 128] - - [31, 45.788] - - - [18688, 4096, 1, 128] - - [26, 45.475] - - - [10752, 1024, 1, 128] - - [28, 37.471] - - - [2432, 128, 1, 128] - - [181, 13.103] - - - [7296, 512, 1, 128] - - [37, 26.543] - - - [19200, 4096, 1, 128] - - [28, 45.06] - - - [4608, 2945, 1, 128] - - [35, 41.341] - - - [18816, 11009, 1, 128] - - [28, 45.335] - - - [9600, 1024, 1, 128] - - [81, 38.743] - - - [7168, 512, 1, 128] - - [61, 26.137] - - - [11904, 4097, 1, 128] - - [38, 43.378] - - - [17920, 1024, 1, 128] - - [26, 40.839] - - - [11520, 7809, 1, 128] - - [50, 44.699] - - - [22784, 14977, 1, 128] - - [48, 46.074] - - - [13696, 1024, 1, 128] - - [26, 41.476] - - - [15104, 1024, 1, 128] - - [38, 40.371] - - - [25216, 512, 1, 128] - - [36, 39.351] - - - [5376, 512, 1, 128] - - [58, 29.296] - - - [17408, 4096, 1, 128] - - [26, 45.895] - - - [25728, 512, 1, 128] - - [37, 39.372] - - - [896, 512, 1, 128] - - [119, 16.743] - - - [6912, 3329, 1, 128] - - [56, 41.631] - - - [22016, 512, 1, 128] - - [58, 37.306] - - - [22144, 4096, 1, 128] - - [56, 45.107] - - - [10368, 128, 1, 128] - - [109, 31.493] - - - [23296, 2048, 1, 128] - - [64, 43.486] - - - [17920, 10113, 1, 128] - - [65, 46.032] - - - [14848, 4096, 1, 128] - - [38, 45.374] - - - [26112, 128, 1, 128] - - [24, 34.81] - - - [28032, 8192, 1, 128] - - [38, 46.536] - - - [20096, 128, 1, 128] - - [34, 28.324] - - - [15360, 4096, 1, 128] - - [26, 45.83] - - - [3328, 128, 1, 128] - - [119, 17.059] - - - [25472, 512, 1, 128] - - [41, 39.729] - - - [18304, 128, 1, 128] - - [61, 26.658] - - - [20352, 12545, 1, 128] - - [30, 45.911] - - - [26624, 10753, 1, 128] - - [38, 46.719] - - - [20480, 2048, 1, 128] - - [26, 43.517] - - - [26496, 10497, 1, 128] - - [48, 45.036] - - - [22400, 128, 1, 128] - - [41, 30.98] - - - [9216, 5505, 1, 128] - - [23, 44.431] - - - [24064, 8193, 1, 128] - - [31, 45.758] - - - [4224, 128, 1, 128] - - [119, 18.782] - - - [6656, 3073, 1, 128] - - [38, 41.59] - - - [10880, 1024, 1, 128] - - [26, 36.012] - - - [23808, 512, 1, 128] - - [61, 39.24] - - - [15488, 1024, 1, 128] - - [50, 40.525] - - - [24704, 8705, 1, 128] - - [26, 44.615] - - - [12416, 4609, 1, 128] - - [58, 43.37] - - - [3712, 1024, 1, 128] - - [36, 26.643] - - - [25856, 8192, 1, 128] - - [39, 46.514] - - - [8320, 1024, 1, 128] - - [60, 35.307] - - - [16256, 512, 1, 128] - - [41, 34.925] - - - [18944, 2048, 1, 128] - - [59, 43.781] - - - [23168, 4096, 1, 128] - - [28, 45.243] - - - [15616, 2048, 1, 128] - - [38, 43.257] - - - [24320, 512, 1, 128] - - [58, 39.869] - - - [2688, 1025, 1, 128] - - [58, 29.542] - - - [12800, 5121, 1, 128] - - [48, 43.607] - - - [5120, 128, 1, 128] - - [117, 21.595] - - - [4352, 512, 1, 128] - - [35, 25.253] - - - [24576, 8192, 1, 128] - - [39, 43.71] - - - [8320, 512, 1, 128] - - [36, 27.956] - - - [12160, 4481, 1, 128] - - [30, 43.502] - - - [2560, 1025, 1, 128] - - [60, 28.887] - - - [19072, 1024, 1, 128] - - [36, 41.458] - - - [2816, 1153, 1, 128] - - [38, 33.809] - - - [6912, 128, 1, 128] - - [115, 27.43] - - - [9088, 2048, 1, 128] - - [38, 40.832] - - - [26368, 8192, 1, 128] - - [38, 46.613] - - - [17408, 9729, 1, 128] - - [38, 46.24] - - - [18816, 4096, 1, 128] - - [38, 45.154] - - - [4480, 512, 1, 128] - - [68, 25.95] - - - [11648, 128, 1, 128] - - [124, 30.816] - - - [1536, 897, 1, 128] - - [173, 27.683] - - - [11136, 1024, 1, 128] - - [61, 37.354] - - - [8704, 1024, 1, 128] - - [54, 38.492] - - - [19072, 2048, 1, 128] - - [24, 43.381] - - - [25856, 1024, 1, 128] - - [41, 42.188] - - - [7552, 3841, 1, 128] - - [30, 42.352] - - - [23296, 128, 1, 128] - - [61, 31.743] - - - [23424, 512, 1, 128] - - [54, 38.672] - - - [26368, 10497, 1, 128] - - [28, 45.865] - - - [18560, 1024, 1, 128] - - [28, 40.909] - - - [8192, 128, 1, 128] - - [110, 26.405] - - - [27776, 11905, 1, 128] - - [38, 45.554] - - - [18688, 1024, 1, 128] - - [61, 41.268] - - - [21248, 4096, 1, 128] - - [36, 45.353] - - - [16256, 8449, 1, 128] - - [30, 44.905] - - - [1920, 1409, 1, 128] - - [28, 28.905] - - - [24704, 4096, 1, 128] - - [62, 44.774] - - - [13824, 6145, 1, 128] - - [28, 44.647] - - - [6528, 512, 1, 128] - - [36, 34.74] - - - [21376, 128, 1, 128] - - [24, 29.786] - - - [11264, 1024, 1, 128] - - [26, 38.262] - - - [4352, 2817, 1, 128] - - [36, 38.359] - - - [22272, 4096, 1, 128] - - [23, 45.315] - - - [27264, 11265, 1, 128] - - [26, 45.335] - - - [28160, 1024, 1, 128] - - [32, 43.105] - - - [16256, 128, 1, 128] - - [73, 24.011] - - - [18688, 2048, 1, 128] - - [41, 43.547] - - - [9600, 6017, 1, 128] - - [56, 44.038] - - - [23552, 1024, 1, 128] - - [26, 42.519] - - - [8576, 128, 1, 128] - - [115, 27.764] - - - [20992, 13185, 1, 128] - - [30, 46.298] - - - [20992, 1024, 1, 128] - - [38, 42.981] - - - [14720, 512, 1, 128] - - [37, 34.106] - - - [28032, 1024, 1, 128] - - [26, 43.043] - - - [20352, 2048, 1, 128] - - [61, 43.425] - - - [15360, 128, 1, 128] - - [41, 23.052] - - - [8448, 2048, 1, 128] - - [30, 41.013] - - - [6272, 2689, 1, 128] - - [36, 41.19] - - - [7808, 4097, 1, 128] - - [26, 42.053] - - - [25472, 128, 1, 128] - - [61, 33.95] - - - [12288, 4481, 1, 128] - - [28, 44.377] - - - [28416, 4096, 1, 128] - - [35, 45.681] - - - [2176, 128, 1, 128] - - [117, 11.724] - - - [21760, 2048, 1, 128] - - [30, 43.441] - - - [21376, 1024, 1, 128] - - [28, 40.864] - - - [13696, 2048, 1, 128] - - [30, 41.689] - - - [28288, 12417, 1, 128] - - [26, 45.754] - - - [5632, 512, 1, 128] - - [60, 31.392] - - - [22016, 1024, 1, 128] - - [60, 41.391] - - - [25216, 128, 1, 128] - - [61, 33.676] - - - [25216, 8192, 1, 128] - - [23, 46.249] - - - [12032, 128, 1, 128] - - [173, 30.73] - - - [6144, 2048, 1, 128] - - [30, 39.869] - - - [23680, 128, 1, 128] - - [28, 32.26] - - - [15744, 128, 1, 128] - - [24, 23.537] - - - [3968, 512, 1, 128] - - [60, 23.346] - - - [16512, 1024, 1, 128] - - [76, 34.182] - - - [1536, 128, 1, 128] - - [120, 8.857] - - - [25984, 4096, 1, 128] - - [52, 45.164] - - - [19456, 512, 1, 128] - - [35, 39.313] - - - [9984, 1024, 1, 128] - - [28, 39.768] - - - [14080, 6401, 1, 128] - - [60, 43.172] - - - [20736, 2048, 1, 128] - - [64, 43.746] - - - [4224, 2689, 1, 128] - - [32, 37.664] - - - [13696, 512, 1, 128] - - [36, 38.896] - - - [17280, 1024, 1, 128] - - [36, 42.398] - - - [10752, 128, 1, 128] - - [115, 29.285] - - - [1536, 512, 1, 128] - - [124, 24.918] - - - [25728, 2048, 1, 128] - - [38, 43.634] - - - [9472, 128, 1, 128] - - [111, 29.878] - - - [7168, 3585, 1, 128] - - [28, 41.883] - - - [14720, 1024, 1, 128] - - [28, 39.831] - - - [25728, 128, 1, 128] - - [59, 34.113] - - - [14976, 128, 1, 128] - - [61, 22.567] - - - [24832, 1024, 1, 128] - - [41, 41.954] - - - [14080, 512, 1, 128] - - [35, 39.158] - - - [17152, 1024, 1, 128] - - [38, 42.312] - - - [19072, 512, 1, 128] - - [68, 38.216] - - - [21120, 1024, 1, 128] - - [28, 42.256] - - - [4864, 128, 1, 128] - - [114, 20.996] - - - [7936, 512, 1, 128] - - [28, 28.619] - - - [21248, 13441, 1, 128] - - [30, 45.979] - - - [12160, 2048, 1, 128] - - [64, 42.781] - - - [19712, 11905, 1, 128] - - [65, 44.366] - - - [23296, 1024, 1, 128] - - [36, 42.225] - - - [24832, 8961, 1, 128] - - [52, 45.757] - - - [13568, 2048, 1, 128] - - [64, 43.078] - - - [13696, 4096, 1, 128] - - [50, 43.197] - - - [5888, 128, 1, 128] - - [109, 24.145] - - - [10112, 2048, 1, 128] - - [38, 42.288] - - - [21632, 13953, 1, 128] - - [23, 45.908] - - - [19328, 512, 1, 128] - - [41, 38.887] - - - [6272, 512, 1, 128] - - [64, 33.934] - - - [4864, 3201, 1, 128] - - [50, 41.106] - - - [15232, 4096, 1, 128] - - [28, 44.986] - - - [23040, 4096, 1, 128] - - [36, 45.59] - - - [2816, 1281, 1, 128] - - [26, 26.313] - - - [8960, 128, 1, 128] - - [111, 28.755] - - - [9472, 1024, 1, 128] - - [50, 38.481] - - - [27648, 11777, 1, 128] - - [38, 46.371] - - - [28416, 2048, 1, 128] - - [30, 43.52] - - - [13952, 6145, 1, 128] - - [35, 44.191] - - - [13952, 1024, 1, 128] - - [28, 41.942] - - - [12544, 2048, 1, 128] - - [28, 42.423] - - - [10624, 7041, 1, 128] - - [30, 44.381] - - - [24704, 2048, 1, 128] - - [29, 43.324] - - - [17280, 9473, 1, 128] - - [39, 45.407] - - - [25088, 9217, 1, 128] - - [48, 45.765] - - - [10240, 6657, 1, 128] - - [30, 45.524] - - - [12800, 4096, 1, 128] - - [68, 44.802] - - - [17792, 1024, 1, 128] - - [30, 40.407] - - - [12160, 128, 1, 128] - - [173, 31.828] - - - [16512, 128, 1, 128] - - [64, 24.389] - - - [25856, 512, 1, 128] - - [36, 39.671] - - - [8576, 4865, 1, 128] - - [26, 43.84] - - - [25984, 1024, 1, 128] - - [34, 41.903] - - - [512, 128, 1, 128] - - [179, 2.93] - - - [10112, 128, 1, 128] - - [111, 31.099] - - - [28288, 2048, 1, 128] - - [24, 44.215] - - - [1152, 641, 1, 128] - - [110, 23.269] - - - [17920, 4096, 1, 128] - - [58, 45.539] - - - [2560, 1921, 1, 128] - - [36, 33.5] - - - [24704, 8833, 1, 128] - - [30, 44.724] - - - [3200, 512, 1, 128] - - [124, 32.245] - - - [6656, 2945, 1, 128] - - [58, 41.037] - - - [12672, 4993, 1, 128] - - [28, 43.969] - - - [4608, 1024, 1, 128] - - [28, 32.411] - - - [25856, 9985, 1, 128] - - [23, 45.586] - - - [23808, 2048, 1, 128] - - [61, 43.971] - - - [9728, 6145, 1, 128] - - [56, 44.553] - - - [28416, 12417, 1, 128] - - [30, 45.628] - - - [14464, 4096, 1, 128] - - [30, 44.862] - - - [21888, 128, 1, 128] - - [40, 30.215] - - - [23680, 15873, 1, 128] - - [51, 45.366] - - - [22144, 1024, 1, 128] - - [50, 41.574] - - - [17664, 512, 1, 256] - - [61, 56.371] - - - [25600, 1024, 1, 256] - - [54, 69.238] - - - [28928, 512, 1, 256] - - [64, 62.502] - - - [15104, 512, 1, 256] - - [24, 55.153] - - - [38912, 1024, 1, 256] - - [24, 70.324] - - - [34304, 8192, 1, 256] - - [43, 74.829] - - - [23552, 1024, 1, 256] - - [64, 68.047] - - - [39424, 23552, 1, 256] - - [42, 75.106] - - - [9472, 1024, 1, 256] - - [41, 59.523] - - - [28928, 13056, 1, 256] - - [52, 74.869] - - - [42496, 1024, 1, 256] - - [70, 70.355] - - - [18432, 1024, 1, 256] - - [41, 66.708] - - - [40192, 24320, 1, 256] - - [28, 74.909] - - - [33280, 17152, 1, 256] - - [25, 75.787] - - - [27904, 512, 1, 256] - - [28, 65.331] - - - [39680, 8192, 1, 256] - - [26, 74.356] - - - [28160, 8192, 1, 256] - - [30, 74.637] - - - [25088, 8192, 1, 256] - - [25, 74.701] - - - [23040, 15360, 1, 256] - - [26, 75.671] - - - [19712, 11776, 1, 256] - - [42, 74.182] - - - [43520, 27648, 1, 256] - - [42, 74.947] - - - [44544, 4096, 1, 256] - - [42, 73.451] - - - [20224, 4096, 1, 256] - - [28, 72.535] - - - [31744, 4096, 1, 256] - - [29, 73.158] - - - [33024, 16896, 1, 256] - - [43, 75.252] - - - [32768, 8192, 1, 256] - - [89, 58.919] - - - [42752, 4096, 1, 256] - - [28, 72.964] - - - [19968, 512, 1, 256] - - [59, 61.638] - - - [10496, 512, 1, 256] - - [36, 55.906] - - - [36864, 4096, 1, 256] - - [55, 72.634] - - - [12288, 1024, 1, 256] - - [38, 62.574] - - - [22784, 14848, 1, 256] - - [29, 75.154] - - - [17152, 9472, 1, 256] - - [32, 74.859] - - - [31488, 1024, 1, 256] - - [61, 69.495] - - - [25344, 1024, 1, 256] - - [70, 67.149] - - - [33536, 512, 1, 256] - - [41, 65.285] - - - [28672, 8192, 1, 256] - - [30, 74.718] - - - [15104, 7168, 1, 256] - - [28, 73.585] - - - [38144, 22272, 1, 256] - - [39, 74.833] - - - [25344, 4096, 1, 256] - - [44, 72.024] - - - [6400, 2560, 1, 256] - - [35, 64.98] - - - [21248, 13568, 1, 256] - - [38, 75.179] - - - [2304, 1536, 1, 256] - - [50, 53.512] - - - [20992, 512, 1, 256] - - [28, 63.931] - - - [3072, 1024, 1, 256] - - [36, 49.378] - - - [36864, 20736, 1, 256] - - [26, 75.147] - - - [39936, 24064, 1, 256] - - [38, 75.501] - - - [2816, 512, 1, 256] - - [182, 42.312] - - - [37888, 512, 1, 256] - - [24, 66.052] - - - [39680, 1024, 1, 256] - - [38, 69.234] - - - [35584, 19712, 1, 256] - - [23, 74.82] - - - [25600, 9728, 1, 256] - - [26, 75.44] - - - [2816, 1024, 1, 256] - - [60, 46.28] - - - [13056, 1024, 1, 256] - - [61, 62.574] - - - [39680, 4096, 1, 256] - - [38, 72.728] - - - [4864, 3072, 1, 256] - - [56, 64.205] - - - [27648, 11776, 1, 256] - - [30, 75.73] - - - [13056, 4096, 1, 256] - - [38, 72.06] - - - [4096, 2304, 1, 256] - - [35, 59.057] - - - [34048, 1024, 1, 256] - - [40, 68.468] - - - [6400, 512, 1, 256] - - [41, 51.306] - - - [15872, 4096, 1, 256] - - [38, 72.281] - - - [29440, 1024, 1, 256] - - [61, 68.859] - - - [7424, 512, 1, 256] - - [64, 41.766] - - - [19200, 4096, 1, 256] - - [38, 72.099] - - - [37376, 21504, 1, 256] - - [42, 75.215] - - - [37888, 1024, 1, 256] - - [59, 70.228] - - - [40704, 24832, 1, 256] - - [28, 74.84] - - - [26112, 1024, 1, 256] - - [70, 69.213] - - - [25088, 8960, 1, 256] - - [60, 75.437] - - - [27136, 512, 1, 256] - - [28, 64.388] - - - [4608, 512, 1, 256] - - [28, 38.941] - - - [31232, 8192, 1, 256] - - [42, 74.803] - - - [33024, 512, 1, 256] - - [36, 60.923] - - - [27648, 512, 1, 256] - - [85, 65.08] - - - [28928, 4096, 1, 256] - - [28, 72.464] - - - [44544, 2048, 1, 256] - - [42, 72.039] - - - [43776, 27648, 1, 256] - - [42, 74.437] - - - [19456, 4096, 1, 256] - - [26, 72.834] - - - [33536, 17664, 1, 256] - - [25, 75.192] - - - [35328, 4096, 1, 256] - - [44, 73.657] - - - [13312, 5376, 1, 256] - - [35, 74.004] - - - [32768, 1024, 1, 256] - - [39, 59.341] - - - [39168, 4096, 1, 256] - - [42, 73.18] - - - [15616, 7936, 1, 256] - - [38, 74.429] - - - [41472, 25600, 1, 256] - - [39, 74.75] - - - [14592, 4096, 1, 256] - - [32, 70.479] - - - [37632, 21760, 1, 256] - - [26, 74.813] - - - [37376, 21248, 1, 256] - - [38, 75.33] - - - [14336, 6656, 1, 256] - - [30, 74.692] - - - [36608, 20480, 1, 256] - - [42, 74.602] - - - [32256, 16384, 1, 256] - - [44, 75.223] - - - [44544, 28416, 1, 256] - - [39, 75.158] - - - [26112, 512, 1, 256] - - [36, 62.957] - - - [41216, 25344, 1, 256] - - [28, 74.919] - - - [16640, 512, 1, 256] - - [102, 56.008] - - - [30464, 14336, 1, 256] - - [42, 73.987] - - - [13312, 4096, 1, 256] - - [30, 72.101] - - - [22528, 1024, 1, 256] - - [41, 68.222] - - - [5632, 1024, 1, 256] - - [27, 50.714] - - - [27392, 1024, 1, 256] - - [52, 64.891] - - - [27648, 8192, 1, 256] - - [26, 75.078] - - - [26368, 1024, 1, 256] - - [24, 68.193] - - - [43776, 4096, 1, 256] - - [29, 73.189] - - - [23552, 15872, 1, 256] - - [26, 76.241] - - - [26624, 10496, 1, 256] - - [28, 75.778] - - - [27392, 8192, 1, 256] - - [29, 74.062] - - - [17408, 9728, 1, 256] - - [26, 75.456] - - - [16896, 9216, 1, 256] - - [28, 74.885] - - - [26880, 11008, 1, 256] - - [28, 74.818] - - - [31488, 512, 1, 256] - - [26, 64.98] - - - [14336, 6400, 1, 256] - - [50, 74.846] - - - [17152, 512, 1, 256] - - [64, 60.79] - - - [7168, 512, 1, 256] - - [50, 41.029] - - - [41984, 26112, 1, 256] - - [23, 75.545] - - - [11776, 512, 1, 256] - - [41, 52.738] - - - [16128, 8448, 1, 256] - - [35, 74.56] - - - [11520, 1024, 1, 256] - - [26, 61.695] - - - [27904, 1024, 1, 256] - - [40, 68.722] - - - [37888, 8192, 1, 256] - - [30, 74.991] - - - [20480, 12544, 1, 256] - - [23, 75.828] - - - [23552, 15616, 1, 256] - - [30, 76.268] - - - [21504, 13824, 1, 256] - - [28, 76.158] - - - [27136, 11008, 1, 256] - - [25, 75.587] - - - [32000, 512, 1, 256] - - [35, 63.572] - - - [26624, 1024, 1, 256] - - [41, 68.378] - - - [34816, 8192, 1, 256] - - [38, 75.041] - - - [23040, 512, 1, 256] - - [70, 61.25] - - - [36608, 1024, 1, 256] - - [41, 69.423] - - - [43264, 8192, 1, 256] - - [30, 74.358] - - - [30208, 14336, 1, 256] - - [43, 75.063] - - - [43520, 512, 1, 256] - - [60, 66.566] - - - [32256, 4096, 1, 256] - - [44, 73.355] - - - [33792, 17664, 1, 256] - - [28, 75.868] - - - [10752, 6912, 1, 256] - - [28, 74.308] - - - [29696, 8192, 1, 256] - - [30, 74.986] - - - [41472, 512, 1, 256] - - [26, 66.983] - - - [44544, 8192, 1, 256] - - [28, 74.517] - - - [41472, 8192, 1, 256] - - [44, 74.446] - - - [38656, 4096, 1, 256] - - [42, 73.31] - - - [44800, 512, 1, 256] - - [30, 66.471] - - - [37376, 4096, 1, 256] - - [42, 73.81] - - - [19200, 1024, 1, 256] - - [73, 66.532] - - - [39680, 23552, 1, 256] - - [38, 74.227] - - - [30976, 8192, 1, 256] - - [42, 73.668] - - - [25856, 1024, 1, 256] - - [41, 68.429] - - - [22016, 14336, 1, 256] - - [38, 75.309] - - - [17152, 9216, 1, 256] - - [26, 74.262] - - - [18432, 10752, 1, 256] - - [26, 75.858] - - - [5376, 1024, 1, 256] - - [41, 48.52] - - - [21760, 13824, 1, 256] - - [30, 75.487] - - - [15360, 512, 1, 256] - - [35, 55.596] - - - [2560, 512, 1, 256] - - [126, 43.314] - - - [36096, 8192, 1, 256] - - [29, 73.872] - - - [42752, 26624, 1, 256] - - [28, 74.446] - - - [35584, 19456, 1, 256] - - [42, 74.674] - - - [6144, 2304, 1, 256] - - [56, 66.961] - - - [42240, 1024, 1, 256] - - [45, 69.828] - - - [26880, 4096, 1, 256] - - [26, 72.673] - - - [28160, 12032, 1, 256] - - [28, 75.503] - - - [18688, 10752, 1, 256] - - [28, 75.22] - - - [43520, 8192, 1, 256] - - [30, 74.662] - - - [8192, 4352, 1, 256] - - [38, 70.832] - - - [6912, 3072, 1, 256] - - [35, 68.847] - - - [31744, 15616, 1, 256] - - [38, 75.832] - - - [36352, 20224, 1, 256] - - [42, 75.358] - - - [41216, 25088, 1, 256] - - [23, 74.747] - - - [37632, 1024, 1, 256] - - [24, 69.481] - - - [18944, 512, 1, 256] - - [26, 59.145] - - - [15616, 1024, 1, 256] - - [61, 65.832] - - - [44288, 512, 1, 256] - - [32, 66.064] - - - [24832, 8704, 1, 256] - - [52, 74.982] - - - [21504, 13568, 1, 256] - - [38, 76.086] - - - [18176, 10496, 1, 256] - - [28, 75.104] - - - [21248, 1024, 1, 256] - - [61, 66.339] - - - [16384, 1024, 1, 256] - - [55, 58.261] - - - [25600, 8192, 1, 256] - - [38, 75.08] - - - [28672, 12544, 1, 256] - - [23, 75.492] - - - [16128, 1024, 1, 256] - - [36, 62.905] - - - [22272, 14592, 1, 256] - - [48, 75.412] - - - [1280, 512, 1, 256] - - [122, 30.85] - - - [36864, 20992, 1, 256] - - [38, 75.008] - - - [3584, 1792, 1, 256] - - [56, 56.36] - - - [35072, 19200, 1, 256] - - [26, 74.991] - - - [32000, 4096, 1, 256] - - [55, 72.71] - - - [28416, 1024, 1, 256] - - [64, 67.534] - - - [20480, 12800, 1, 256] - - [26, 75.76] - - - [21760, 4096, 1, 256] - - [92, 72.584] - - - [44288, 8192, 1, 256] - - [42, 74.353] - - - [33280, 4096, 1, 256] - - [29, 73.587] - - - [32512, 1024, 1, 256] - - [34, 68.481] - - - [38400, 22528, 1, 256] - - [29, 75.187] - - - [40448, 1024, 1, 256] - - [61, 70.029] - - - [5120, 512, 1, 256] - - [36, 42.717] - - - [29952, 8192, 1, 256] - - [52, 74.456] - - - [40448, 24576, 1, 256] - - [44, 74.509] - - - [29696, 4096, 1, 256] - - [26, 73.234] - - - [21504, 1024, 1, 256] - - [59, 67.191] - - - [19968, 1024, 1, 256] - - [61, 66.898] - - - [16896, 512, 1, 256] - - [61, 59.514] - - - [33536, 17408, 1, 256] - - [29, 74.787] - - - [19712, 512, 1, 256] - - [28, 60.709] - - - [16384, 8704, 1, 256] - - [39, 63.818] - - - [29952, 13824, 1, 256] - - [31, 75.112] - - - [14592, 6656, 1, 256] - - [68, 72.415] - - - [36864, 1024, 1, 256] - - [38, 68.423] - - - [31744, 15872, 1, 256] - - [28, 75.897] - - - [24832, 8960, 1, 256] - - [32, 75.119] - - - [23808, 1024, 1, 256] - - [41, 68.233] - - - [19200, 11264, 1, 256] - - [29, 74.727] - - - [23296, 15360, 1, 256] - - [30, 75.187] - - - [34304, 18432, 1, 256] - - [43, 75.048] - - - [22016, 1024, 1, 256] - - [64, 67.571] - - - [40704, 4096, 1, 256] - - [29, 73.066] - - - [25600, 4096, 1, 256] - - [29, 73.152] - - - [3328, 1024, 1, 256] - - [36, 52.353] - - - [30464, 8192, 1, 256] - - [42, 73.421] - - - [39424, 8192, 1, 256] - - [38, 74.587] - - - [23808, 15872, 1, 256] - - [38, 75.524] - - - [8960, 1024, 1, 256] - - [54, 56.85] - - - [44032, 4096, 1, 256] - - [38, 73.194] - - - [35584, 8192, 1, 256] - - [43, 74.268] - - - [29184, 8192, 1, 256] - - [31, 74.597] - - - [13824, 1024, 1, 256] - - [45, 65.854] - - - [36608, 8192, 1, 256] - - [26, 74.225] - - - [30976, 512, 1, 256] - - [24, 64.406] - - - [33024, 4096, 1, 256] - - [43, 73.199] - - - [11776, 7936, 1, 256] - - [28, 74.374] - - - [23808, 16128, 1, 256] - - [28, 75.37] - - - [22272, 14336, 1, 256] - - [42, 74.944] - - - [27392, 11520, 1, 256] - - [43, 74.289] - - - [30464, 4096, 1, 256] - - [42, 71.791] - - - [20992, 13312, 1, 256] - - [29, 75.507] - - - [44800, 1024, 1, 256] - - [30, 69.404] - - - [32512, 4096, 1, 256] - - [90, 72.937] - - - [23296, 15616, 1, 256] - - [28, 75.408] - - - [9216, 1024, 1, 256] - - [37, 58.517] - - - [20224, 12544, 1, 256] - - [26, 75.301] - - - [32256, 1024, 1, 256] - - [61, 69.566] - - - [38400, 512, 1, 256] - - [73, 65.71] - - - [29952, 1024, 1, 256] - - [61, 68.339] - - - [36352, 512, 1, 256] - - [50, 65.269] - - - [41728, 25600, 1, 256] - - [42, 73.877] - - - [32000, 1024, 1, 256] - - [24, 68.824] - - - [38144, 22016, 1, 256] - - [23, 74.81] - - - [27136, 11264, 1, 256] - - [48, 75.389] - - - [34048, 18176, 1, 256] - - [66, 74.655] - - - [22016, 14080, 1, 256] - - [60, 76.113] - - - [19712, 12032, 1, 256] - - [42, 74.357] - - - [23552, 4096, 1, 256] - - [42, 72.963] - - - [15872, 1024, 1, 256] - - [41, 64.076] - - - [37120, 512, 1, 256] - - [50, 65.473] - - - [9984, 1024, 1, 256] - - [41, 61.703] - - - [32512, 8192, 1, 256] - - [25, 74.44] - - - [15360, 4096, 1, 256] - - [28, 72.39] - - - [13056, 512, 1, 256] - - [35, 57.597] - - - [44032, 8192, 1, 256] - - [38, 74.818] - - - [24576, 8192, 1, 256] - - [55, 70.1] - - - [36352, 8192, 1, 256] - - [29, 74.844] - - - [26368, 8192, 1, 256] - - [38, 74.449] - - - [20480, 1024, 1, 256] - - [28, 65.317] - - - [35072, 8192, 1, 256] - - [38, 74.34] - - - [32000, 15872, 1, 256] - - [28, 75.055] - - - [40704, 24576, 1, 256] - - [29, 74.2] - - - [15104, 7424, 1, 256] - - [30, 74.433] - - - [25856, 4096, 1, 256] - - [29, 72.553] - - - [14848, 512, 1, 256] - - [26, 54.42] - - - [39424, 4096, 1, 256] - - [66, 73.507] - - - [24832, 512, 1, 256] - - [24, 60.301] - - - [44288, 28416, 1, 256] - - [23, 74.674] - - - [12544, 4608, 1, 256] - - [36, 72.008] - - - [12800, 4864, 1, 256] - - [60, 73.386] - - - [29440, 512, 1, 256] - - [64, 63.112] - - - [40192, 24064, 1, 256] - - [38, 74.898] - - - [18176, 4096, 1, 256] - - [30, 72.486] - - - [40960, 8192, 1, 256] - - [55, 67.262] - - - [42240, 512, 1, 256] - - [37, 67.25] - - - [9728, 512, 1, 256] - - [50, 52.767] - - - [14848, 7168, 1, 256] - - [26, 73.661] - - - [44800, 28672, 1, 256] - - [55, 74.049] - - - [15616, 7680, 1, 256] - - [30, 74.594] - - - [33280, 17408, 1, 256] - - [43, 75.374] - - - [42752, 1024, 1, 256] - - [37, 69.863] - - - [35328, 8192, 1, 256] - - [48, 74.689] - - - [36352, 1024, 1, 256] - - [79, 69.823] - - - [35840, 1024, 1, 256] - - [75, 70.114] - - - [41472, 4096, 1, 256] - - [29, 73.656] - - - [3584, 1024, 1, 256] - - [61, 41.069] - - - [22528, 14592, 1, 256] - - [26, 76.245] - - - [44032, 512, 1, 256] - - [45, 66.675] - - - [30720, 1024, 1, 256] - - [59, 69.647] - - - [39680, 512, 1, 256] - - [59, 65.345] - - - [22272, 1024, 1, 256] - - [41, 67.222] - - - [42240, 26368, 1, 256] - - [30, 74.63] - - - [10240, 6400, 1, 256] - - [56, 74.368] - - - [30976, 14848, 1, 256] - - [29, 74.145] - - - [41728, 25856, 1, 256] - - [42, 74.256] - - - [28928, 12800, 1, 256] - - [48, 74.82] - - - [21760, 14080, 1, 256] - - [26, 75.486] - - - [5888, 1024, 1, 256] - - [61, 52.899] - - - [24576, 8704, 1, 256] - - [39, 70.872] - - - [38912, 4096, 1, 256] - - [26, 73.118] - - - [15360, 1024, 1, 256] - - [64, 65.031] - - - [18688, 512, 1, 256] - - [35, 58.254] - - - [27392, 512, 1, 256] - - [58, 62.874] - - - [22784, 512, 1, 256] - - [41, 61.037] - - - [40448, 4096, 1, 256] - - [62, 73.641] - - - [19200, 512, 1, 256] - - [36, 59.411] - - - [26368, 10496, 1, 256] - - [28, 74.947] - - - [25088, 9216, 1, 256] - - [43, 74.874] - - - [33536, 1024, 1, 256] - - [24, 68.643] - - - [25600, 9472, 1, 256] - - [28, 75.887] - - - [13824, 4096, 1, 256] - - [28, 71.984] - - - [5632, 3840, 1, 256] - - [35, 70.207] - - - [9216, 5376, 1, 256] - - [50, 72.915] - - - [8960, 5120, 1, 256] - - [32, 71.62] - - - [19456, 512, 1, 256] - - [28, 60.468] - - - [24576, 4096, 1, 256] - - [33, 68.369] - - - [27392, 11264, 1, 256] - - [42, 74.537] - - - [35072, 4096, 1, 256] - - [42, 72.993] - - - [44288, 4096, 1, 256] - - [42, 73.28] - - - [40448, 8192, 1, 256] - - [52, 74.465] - - - [33280, 512, 1, 256] - - [61, 64.893] - - - [22272, 4096, 1, 256] - - [42, 72.461] - - - [35584, 512, 1, 256] - - [24, 63.96] - - - [10752, 512, 1, 256] - - [36, 48.931] - - - [19968, 4096, 1, 256] - - [62, 72.937] - - - [34304, 1024, 1, 256] - - [40, 69.634] - - - [41216, 8192, 1, 256] - - [42, 74.477] - - - [35840, 19712, 1, 256] - - [28, 75.823] - - - [43520, 27392, 1, 256] - - [39, 75.44] - - - [30720, 14848, 1, 256] - - [28, 75.633] - - - [38400, 22272, 1, 256] - - [39, 75.441] - - - [1536, 1024, 1, 256] - - [183, 44.029] - - - [40192, 1024, 1, 256] - - [62, 69.757] - - - [44800, 256, 1, 256] - - [59, 58.147] - - - [1536, 512, 1, 256] - - [124, 35.992] - - - [34560, 18432, 1, 256] - - [26, 74.758] - - - [1792, 1024, 1, 256] - - [54, 30.733] - - - [5376, 3584, 1, 256] - - [36, 67.587] - - - [30208, 1024, 1, 256] - - [64, 68.896] - - - [31232, 512, 1, 256] - - [35, 65.399] - - - [23040, 4096, 1, 256] - - [90, 73.309] - - - [35840, 4096, 1, 256] - - [28, 73.139] - - - [38144, 512, 1, 256] - - [37, 65.417] - - - [31744, 512, 1, 256] - - [77, 63.607] - - - [14592, 6912, 1, 256] - - [69, 73.121] - - - [19456, 11520, 1, 256] - - [28, 76.138] - - - [7168, 1024, 1, 256] - - [41, 53.439] - - - [18944, 11264, 1, 256] - - [43, 75.342] - - - [19712, 1024, 1, 256] - - [28, 63.902] - - - [26112, 9984, 1, 256] - - [38, 75.563] - - - [38656, 22784, 1, 256] - - [48, 74.655] - - - [24320, 8192, 1, 256] - - [48, 74.704] - - - [4864, 1024, 1, 256] - - [50, 52.923] - - - [20480, 4096, 1, 256] - - [38, 72.506] - - - [10240, 1024, 1, 256] - - [61, 62.923] - - - [31232, 15360, 1, 256] - - [25, 75.415] - - - [24320, 4096, 1, 256] - - [48, 72.813] - - - [33792, 1024, 1, 256] - - [59, 69.904] - - - [12032, 1024, 1, 256] - - [61, 63.985] - - - [39168, 512, 1, 256] - - [36, 64.9] - - - [16896, 4096, 1, 256] - - [28, 72.514] - - - [36096, 1024, 1, 256] - - [77, 66.757] - - - [28416, 12544, 1, 256] - - [52, 74.959] - - - [30720, 4096, 1, 256] - - [26, 73.127] - - - [19712, 4096, 1, 256] - - [42, 71.466] - - - [37120, 21248, 1, 256] - - [38, 74.681] - - - [16384, 4096, 1, 256] - - [28, 61.887] - - - [18688, 11008, 1, 256] - - [30, 75.091] - - - [38400, 8192, 1, 256] - - [38, 74.678] - - - [11264, 7424, 1, 256] - - [28, 74.924] - - - [23296, 512, 1, 256] - - [36, 61.999] - - - [25344, 512, 1, 256] - - [36, 61.71] - - - [44544, 256, 1, 256] - - [37, 57.866] - - - [43264, 4096, 1, 256] - - [55, 72.828] - - - [32512, 16640, 1, 256] - - [31, 75.357] - - - [39936, 8192, 1, 256] - - [30, 74.974] - - - [43264, 512, 1, 256] - - [36, 65.863] - - - [16640, 8704, 1, 256] - - [38, 74.811] - - - [26624, 8192, 1, 256] - - [38, 75.099] - - - [35328, 19456, 1, 256] - - [43, 75.247] - - - [42752, 26880, 1, 256] - - [26, 74.838] - - - [25344, 9216, 1, 256] - - [44, 73.445] - - - [34048, 8192, 1, 256] - - [44, 73.799] - - - [18688, 4096, 1, 256] - - [30, 72.394] - - - [37632, 8192, 1, 256] - - [55, 74.221] - - - [19968, 12032, 1, 256] - - [28, 75.67] - - - [8448, 4608, 1, 256] - - [50, 71.474] - - - [2048, 1536, 1, 256] - - [58, 49.785] - - - [31488, 15616, 1, 256] - - [38, 74.939] - - - [35328, 512, 1, 256] - - [84, 64.709] - - - [37376, 8192, 1, 256] - - [42, 74.732] - - - [33792, 8192, 1, 256] - - [38, 74.959] - - - [36608, 4096, 1, 256] - - [42, 72.898] - - - [28416, 8192, 1, 256] - - [48, 74.104] - - - [5632, 512, 1, 256] - - [70, 46.409] - - - [13568, 4096, 1, 256] - - [28, 71.849] - - - [17664, 9728, 1, 256] - - [58, 74.224] - - - [13568, 1024, 1, 256] - - [24, 64.791] - - - [8448, 512, 1, 256] - - [36, 46.517] - - - [22528, 4096, 1, 256] - - [30, 72.967] - - - [33536, 8192, 1, 256] - - [48, 74.444] - - - [23296, 1024, 1, 256] - - [24, 67.242] - - - [43520, 4096, 1, 256] - - [42, 73.78] - - - [39936, 23808, 1, 256] - - [38, 75.482] - - - [12544, 4096, 1, 256] - - [35, 71.348] - - - [22016, 4096, 1, 256] - - [75, 73.077] - - - [14592, 512, 1, 256] - - [24, 53.515] - - - [39936, 4096, 1, 256] - - [28, 73.242] - - - [18176, 1024, 1, 256] - - [41, 65.761] - - - [44800, 2048, 1, 256] - - [44, 70.237] - - - [14848, 4096, 1, 256] - - [28, 72.5] - - - [20224, 12288, 1, 256] - - [42, 74.421] - - - [16896, 8960, 1, 256] - - [38, 75.55] - - - [43264, 27392, 1, 256] - - [38, 74.802] - - - [24064, 16128, 1, 256] - - [48, 75.907] - - - [1024, 512, 1, 256] - - [110, 24.986] - - - [24576, 8448, 1, 256] - - [23, 70.765] - - - [25344, 9472, 1, 256] - - [29, 74.117] - - - [3328, 1536, 1, 256] - - [36, 54.557] - - - [31488, 4096, 1, 256] - - [28, 72.728] - - - [43008, 8192, 1, 256] - - [38, 74.904] - - - [28672, 12800, 1, 256] - - [26, 75.357] - - - [20736, 13056, 1, 256] - - [28, 75.359] - - - [17664, 9984, 1, 256] - - [58, 74.504] - - - [17920, 1024, 1, 256] - - [40, 65.342] - - - [11008, 1024, 1, 256] - - [60, 57.863] - - - [44800, 4096, 1, 256] - - [33, 72.511] - - - [29952, 14080, 1, 256] - - [25, 75.218] - - - [39168, 23296, 1, 256] - - [42, 74.594] - - - [9472, 512, 1, 256] - - [24, 51.769] - - - [27904, 8192, 1, 256] - - [29, 74.273] - - - [5120, 1024, 1, 256] - - [41, 54.163] - - - [15872, 7936, 1, 256] - - [26, 75.157] - - - [13568, 5632, 1, 256] - - [35, 73.703] - - - [17920, 9984, 1, 256] - - [38, 75.469] - - - [16640, 8960, 1, 256] - - [52, 74.67] - - - [41984, 4096, 1, 256] - - [28, 73.1] - - - [6912, 512, 1, 256] - - [54, 54.305] - - - [28416, 4096, 1, 256] - - [33, 72.453] - - - [27648, 11520, 1, 256] - - [30, 75.978] - - - [7680, 3840, 1, 256] - - [50, 70.372] - - - [34048, 4096, 1, 256] - - [43, 72.436] - - - [11264, 512, 1, 256] - - [35, 50.753] - - - [26368, 4096, 1, 256] - - [38, 72.565] - - - [21248, 13312, 1, 256] - - [28, 75.106] - - - [15104, 1024, 1, 256] - - [41, 64.971] - - - [35072, 18944, 1, 256] - - [28, 74.936] - - - [6144, 1024, 1, 256] - - [28, 54.464] - - - [44800, 8192, 1, 256] - - [38, 74.033] - - - [25088, 512, 1, 256] - - [61, 61.479] - - - [27904, 12032, 1, 256] - - [44, 74.855] - - - [27648, 1024, 1, 256] - - [41, 68.968] - - - [28928, 8192, 1, 256] - - [48, 74.128] - - - [29440, 13312, 1, 256] - - [31, 75.126] - - - [43264, 27136, 1, 256] - - [39, 74.753] - - - [23552, 512, 1, 256] - - [64, 62.539] - - - [26880, 10752, 1, 256] - - [30, 75.05] - - - [44032, 28160, 1, 256] - - [38, 75.553] - - - [36096, 512, 1, 256] - - [60, 63.214] - - - [4352, 2560, 1, 256] - - [56, 60.33] - - - [38912, 8192, 1, 256] - - [26, 74.954] - - - [12032, 4096, 1, 256] - - [26, 71.167] - - - [37632, 512, 1, 256] - - [41, 65.67] - - - [30208, 512, 1, 256] - - [36, 64.176] - - - [2304, 512, 1, 256] - - [122, 41.352] - - - [24320, 8448, 1, 256] - - [58, 74.743] - - - [39424, 512, 1, 256] - - [58, 65.209] - - - [37632, 21504, 1, 256] - - [39, 74.577] - - - [17152, 1024, 1, 256] - - [41, 66.735] - - - [22784, 15104, 1, 256] - - [52, 75.335] - - - [27904, 11776, 1, 256] - - [43, 74.855] - - - [43008, 26880, 1, 256] - - [39, 75.604] - - - [41728, 4096, 1, 256] - - [42, 72.671] - - - [25344, 8192, 1, 256] - - [42, 73.596] - - - [44800, 28928, 1, 256] - - [26, 74.343] - - - [38912, 22784, 1, 256] - - [39, 75.639] - - - [44032, 1024, 1, 256] - - [45, 70.885] - - - [30976, 4096, 1, 256] - - [29, 71.622] - - - [15872, 8192, 1, 256] - - [26, 75.001] - - - [40960, 4096, 1, 256] - - [33, 66.076] - - - [35584, 1024, 1, 256] - - [64, 69.271] - - - [18944, 4096, 1, 256] - - [77, 72.836] - - - [36096, 20224, 1, 256] - - [29, 74.266] - - - [11008, 7168, 1, 256] - - [86, 71.456] - - - [7936, 1024, 1, 256] - - [28, 56.833] - - - [44288, 1024, 1, 256] - - [40, 69.726] - - - [38656, 8192, 1, 256] - - [42, 74.287] - - - [38144, 1024, 1, 256] - - [30, 69.01] - - - [41984, 1024, 1, 256] - - [59, 70.499] - - - [20736, 512, 1, 256] - - [61, 63.106] - - - [32768, 16640, 1, 256] - - [89, 59.277] - - - [40960, 1024, 1, 256] - - [26, 64.607] - - - [25856, 9984, 1, 256] - - [26, 74.885] - - - [29696, 13824, 1, 256] - - [30, 75.882] - - - [37120, 4096, 1, 256] - - [42, 73.15] - - - [37120, 20992, 1, 256] - - [26, 74.854] - - - [35072, 512, 1, 256] - - [37, 66.521] - - - [38656, 1024, 1, 256] - - [37, 69.426] - - - [37376, 512, 1, 256] - - [24, 65.878] - - - [32000, 16128, 1, 256] - - [30, 74.91] - - - [41984, 25856, 1, 256] - - [39, 75.626] - - - [23040, 15104, 1, 256] - - [31, 76.021] - - - [31232, 15104, 1, 256] - - [52, 75.761] - - - [25088, 4096, 1, 256] - - [62, 73.174] - - - [15360, 7424, 1, 256] - - [26, 75.274] - - - [16384, 8448, 1, 256] - - [23, 63.356] - - - [26624, 4096, 1, 256] - - [28, 73.175] - - - [14080, 6400, 1, 256] - - [86, 73.029] - - - [16128, 4096, 1, 256] - - [28, 72.085] - - - [43776, 27904, 1, 256] - - [42, 74.364] - - - [15872, 512, 1, 256] - - [61, 57.075] - - - [43776, 8192, 1, 256] - - [42, 74.087] - - - [10496, 6656, 1, 256] - - [36, 73.241] - - - [13312, 512, 1, 256] - - [28, 58.378] - - - [29184, 512, 1, 256] - - [37, 62.86] - - - [15360, 7680, 1, 256] - - [26, 75.288] - - - [40192, 8192, 1, 256] - - [26, 74.473] - - - [34560, 8192, 1, 256] - - [30, 74.307] - - - [25856, 8192, 1, 256] - - [26, 74.455] - - - [32512, 16384, 1, 256] - - [43, 75.174] - - - [12288, 4352, 1, 256] - - [26, 73.324] - - - [29440, 13568, 1, 256] - - [52, 75.133] - - - [28160, 1024, 1, 256] - - [24, 68.878] - - - [32768, 4096, 1, 256] - - [67, 58.306] - - - [24832, 4096, 1, 256] - - [42, 72.684] - - - [39680, 23808, 1, 256] - - [38, 74.405] - - - [22784, 4096, 1, 256] - - [62, 72.548] - - - [7936, 4096, 1, 256] - - [41, 69.283] - - - [8704, 4864, 1, 256] - - [56, 72.098] - - - [29696, 512, 1, 256] - - [64, 63.904] - - - [39424, 23296, 1, 256] - - [23, 75.413] - - - [17408, 9472, 1, 256] - - [26, 75.849] - - - [33792, 4096, 1, 256] - - [55, 72.911] - - - [17920, 512, 1, 256] - - [54, 56.82] - - - [25856, 512, 1, 256] - - [36, 62.098] - - - [44288, 28160, 1, 256] - - [23, 74.632] - - - [40192, 4096, 1, 256] - - [75, 73.012] - - - [21248, 512, 1, 256] - - [61, 58.27] - - - [3072, 512, 1, 256] - - [184, 44.68] - - - [29184, 13312, 1, 256] - - [26, 75.113] - - - [44544, 1024, 1, 256] - - [61, 70.345] - - - [37888, 21760, 1, 256] - - [28, 75.792] - - - [33792, 17920, 1, 256] - - [28, 75.738] - - - [6912, 1024, 1, 256] - - [61, 60.997] - - - [41216, 512, 1, 256] - - [40, 66.051] - - - [42240, 26112, 1, 256] - - [30, 74.753] - - - [30720, 8192, 1, 256] - - [38, 75.059] - - - [11776, 1024, 1, 256] - - [36, 62.652] - - - [43008, 4096, 1, 256] - - [39, 73.053] - - - [34560, 18688, 1, 256] - - [30, 74.923] - - - [41984, 512, 1, 256] - - [45, 67.355] - - - [41728, 512, 1, 256] - - [30, 67.03] - - - [2560, 1792, 1, 256] - - [36, 49.744] - - - [36864, 8192, 1, 256] - - [28, 74.544] - - - [40704, 8192, 1, 256] - - [28, 74.341] - - - [30720, 14592, 1, 256] - - [26, 75.998] - - - [32256, 512, 1, 256] - - [36, 64.09] - - - [40192, 512, 1, 256] - - [28, 65.943] - - - [8960, 512, 1, 256] - - [50, 49.372] - - - [16640, 4096, 1, 256] - - [49, 71.565] - - - [30976, 15104, 1, 256] - - [42, 74.383] - - - [27136, 8192, 1, 256] - - [31, 74.986] - - - [30208, 8192, 1, 256] - - [48, 74.616] - - - [21504, 512, 1, 256] - - [28, 58.901] - - - [9728, 5888, 1, 256] - - [35, 73.206] - - - [38912, 23040, 1, 256] - - [39, 75.609] - - - [7424, 1024, 1, 256] - - [35, 54.579] - - - [38656, 22528, 1, 256] - - [29, 74.572] - - - [26880, 512, 1, 256] - - [36, 63.666] - - - [29184, 13056, 1, 256] - - [48, 75.458] - - - [44032, 27904, 1, 256] - - [26, 75.577] - - - [38144, 8192, 1, 256] - - [42, 74.149] - - - [29952, 512, 1, 256] - - [36, 63.804] - - - [18432, 4096, 1, 256] - - [28, 72.431] - - - [28160, 12288, 1, 256] - - [29, 74.924] - - - [29696, 1024, 1, 256] - - [54, 69.307] - - - [39936, 1024, 1, 256] - - [59, 70.178] - - - [25600, 512, 1, 256] - - [38, 61.911] - - - [40448, 24320, 1, 256] - - [25, 75.043] - - - [40448, 512, 1, 256] - - [45, 66.253] - - - [7424, 3584, 1, 256] - - [56, 69.863] - - - [5376, 512, 1, 256] - - [34, 43.993] - - - [27136, 4096, 1, 256] - - [44, 73.399] - - - [35840, 19968, 1, 256] - - [28, 75.549] - - - [18944, 11008, 1, 256] - - [60, 75.428] - - - [34816, 18688, 1, 256] - - [28, 75.739] - - - [38400, 1024, 1, 256] - - [59, 70.053] - - - [36352, 20480, 1, 256] - - [29, 75.176] - - - [36608, 20736, 1, 256] - - [23, 74.849] - - - [28672, 1024, 1, 256] - - [30, 67.129] - - - [42496, 26624, 1, 256] - - [42, 75.087] - - - [31488, 15360, 1, 256] - - [28, 74.816] - - - [20992, 4096, 1, 256] - - [29, 73.044] - - - [12544, 512, 1, 256] - - [24, 55.681] - - - [24064, 8192, 1, 256] - - [48, 74.93] - - - [26880, 8192, 1, 256] - - [28, 74.524] - - - [4352, 512, 1, 256] - - [73, 36.778] - - - [7680, 1024, 1, 256] - - [61, 55.641] - - - [16128, 8192, 1, 256] - - [26, 74.553] - - - [39168, 8192, 1, 256] - - [29, 74.205] - - - [29440, 4096, 1, 256] - - [48, 72.927] - - - [33536, 4096, 1, 256] - - [42, 73.191] - - - [33024, 17152, 1, 256] - - [48, 75.207] - - - [34816, 18944, 1, 256] - - [26, 75.801] - - - [22016, 512, 1, 256] - - [35, 59.438] - - - [14848, 6912, 1, 256] - - [26, 74.685] - - - [20736, 12800, 1, 256] - - [26, 75.324] - - - [32256, 16128, 1, 256] - - [31, 75.679] - - - [7680, 512, 1, 256] - - [41, 43.371] - - - [19968, 12288, 1, 256] - - [42, 75.413] - - - [29184, 4096, 1, 256] - - [75, 73.343] - - - [15616, 4096, 1, 256] - - [30, 72.242] - - - [44544, 28672, 1, 256] - - [23, 74.533] - - - [26112, 4096, 1, 256] - - [66, 73.758] - - - [26624, 10752, 1, 256] - - [30, 75.829] - - - [15104, 4096, 1, 256] - - [28, 71.987] - - - [23296, 4096, 1, 256] - - [30, 72.489] - - - [37888, 22016, 1, 256] - - [30, 75.62] - - - [11520, 7680, 1, 256] - - [60, 73.86] - - - [41728, 1024, 1, 256] - - [28, 69.281] - - - [2304, 1792, 1, 256] - - [36, 45.323] - - - [34048, 17920, 1, 256] - - [25, 74.538] - - - [1536, 768, 1, 256] - - [122, 40.855] - - - [33280, 8192, 1, 256] - - [25, 74.883] - - - [11264, 1024, 1, 256] - - [41, 61.449] - - - [21760, 1024, 1, 256] - - [45, 66.154] - - - [18432, 10496, 1, 256] - - [28, 75.833] - - - [41216, 4096, 1, 256] - - [44, 73.138] - - - [41472, 25344, 1, 256] - - [38, 75.153] - - - [17408, 1024, 1, 256] - - [54, 66.969] - - - [19456, 1024, 1, 256] - - [64, 65.93] - - - [36096, 19968, 1, 256] - - [42, 74.202] - - - [8704, 512, 1, 256] - - [61, 48.105] - - - [30464, 1024, 1, 256] - - [60, 66.822] - - - [8192, 1024, 1, 256] - - [28, 57.825] - - - [11520, 512, 1, 256] - - [26, 51.67] - - - [44544, 512, 1, 256] - - [32, 66.525] - - - [20736, 4096, 1, 256] - - [26, 72.521] - - - [42752, 8192, 1, 256] - - [30, 74.408] - - - [39936, 512, 1, 256] - - [41, 65.84] - - - [42496, 26368, 1, 256] - - [28, 75.25] - - - [28672, 4096, 1, 256] - - [38, 72.526] - - - [35840, 8192, 1, 256] - - [30, 74.949] - - - [17664, 1024, 1, 256] - - [37, 63.994] - - - [21248, 4096, 1, 256] - - [38, 72.544] - - - [1280, 768, 1, 256] - - [122, 35.43] - - - [28160, 512, 1, 256] - - [59, 65.774] - - - [34304, 18176, 1, 256] - - [43, 75.541] - - - [19200, 11520, 1, 256] - - [28, 75.066] - - - [25856, 9728, 1, 256] - - [28, 74.614] - - - [35328, 19200, 1, 256] - - [52, 75.548] - - - [29440, 8192, 1, 256] - - [25, 74.577] - - - [20992, 13056, 1, 256] - - [26, 75.803] - - - [21760, 512, 1, 256] - - [36, 59.218] - - - [12800, 512, 1, 256] - - [41, 56.426] - - - [28416, 12288, 1, 256] - - [31, 74.006] - - - [29696, 13568, 1, 256] - - [28, 75.794] - - - [21504, 4096, 1, 256] - - [29, 72.907] - - - [30464, 14592, 1, 256] - - [48, 74.281] - - - [13056, 5120, 1, 256] - - [50, 73.164] - - - [34560, 4096, 1, 256] - - [30, 72.865] - - - [32768, 16896, 1, 256] - - [89, 59.383] - - - [13824, 5888, 1, 256] - - [23, 73.544] - - - [33024, 8192, 1, 256] - - [44, 74.541] - - - [14080, 4096, 1, 256] - - [48, 70.417] - - - [43008, 1024, 1, 256] - - [37, 70.968] - - - [31744, 1024, 1, 256] - - [41, 69.565] - - - [11008, 512, 1, 256] - - [35, 49.571] - - - [24832, 8192, 1, 256] - - [32, 74.582] - - - [43776, 512, 1, 256] - - [34, 63.657] - - - [24064, 1024, 1, 256] - - [73, 68.455] - - - [12800, 4096, 1, 256] - - [75, 71.929] - - - [19456, 11776, 1, 256] - - [26, 75.91] - - - [22528, 14848, 1, 256] - - [28, 75.994] - - - [30208, 14080, 1, 256] - - [25, 75.797] - - - [40704, 1024, 1, 256] - - [73, 69.656] - - - [35584, 4096, 1, 256] - - [44, 73.157] - - - [26112, 8192, 1, 256] - - [30, 74.852] - - - [9472, 5632, 1, 256] - - [35, 72.733] - - - [15616, 512, 1, 256] - - [35, 56.189] - - - [34816, 4096, 1, 256] - - [30, 73.248] - - - [31232, 4096, 1, 256] - - [42, 73.472] - - - [9728, 1024, 1, 256] - - [50, 60.648] - - - [13312, 1024, 1, 256] - - [64, 64.256] - - - [20224, 1024, 1, 256] - - [61, 66.946] - - - [4864, 512, 1, 256] - - [56, 40.82] - - - [34304, 4096, 1, 256] - - [43, 73.894] - - - [43776, 1024, 1, 256] - - [77, 67.204] - - - [37120, 8192, 1, 256] - - [26, 74.285] - - - [33792, 512, 1, 256] - - [35, 65.761] - - - [42496, 512, 1, 256] - - [26, 66.06] - - - [9216, 512, 1, 256] - - [26, 50.463] - - - [14336, 4096, 1, 256] - - [28, 72.202] - - - [43008, 27136, 1, 256] - - [23, 75.512] - - - [35840, 512, 1, 256] - - [50, 64.842] - - - [40960, 25088, 1, 256] - - [39, 67.979] - - - [17408, 512, 1, 256] - - [50, 60.954] - - - [12288, 4096, 1, 256] - - [28, 71.721] - - - [6656, 512, 1, 256] - - [26, 52.846] - - - [40960, 24832, 1, 256] - - [39, 68.288] - - - [39168, 23040, 1, 256] - - [29, 74.625] - - - [512, 1, 1, 128] - - [206, 0.025] - - - [384, 1, 1, 384] - - [404, 0.028] - - - [256, 1, 1, 256] - - [206, 0.017] - - - [128, 1, 1, 128] - - [206, 0.006] - - - [640, 1, 1, 128] - - [206, 0.032] - - - [1, 128, 1, 256] - - [206, 0.009] - - - [512, 128, 1, 256] - - [114, 4.27] - - - [2049, 128, 1, 256] - - [181, 14.336] - - - [49, 128, 1, 256] - - [159, 0.348] - - - [1537, 128, 1, 256] - - [185, 11.345] - - - [257, 128, 1, 256] - - [159, 1.888] - - - [9728, 128, 1, 256] - - [122, 42.487] - - - [3840, 128, 1, 256] - - [115, 23.522] - - - [1280, 128, 1, 256] - - [114, 10.616] - - - [7168, 128, 1, 256] - - [109, 32.965] - - - [6656, 128, 1, 256] - - [122, 37.792] - - - [2561, 128, 1, 256] - - [119, 17.834] - - - [6912, 128, 1, 256] - - [109, 38.946] - - - [2048, 128, 1, 256] - - [117, 15.996] - - - [2304, 128, 1, 256] - - [117, 17.807] - - - [1536, 128, 1, 256] - - [167, 12.739] - - - [4864, 128, 1, 256] - - [109, 29.428] - - - [8448, 128, 1, 256] - - [109, 38.138] - - - [3072, 128, 1, 256] - - [117, 23.02] - - - [3329, 128, 1, 256] - - [119, 22.543] - - - [3328, 128, 1, 256] - - [111, 24.565] - - - [8960, 128, 1, 256] - - [110, 39.721] - - - [9216, 128, 1, 256] - - [186, 38.761] - - - [2817, 128, 1, 256] - - [151, 19.524] - - - [6400, 128, 1, 256] - - [172, 37.641] - - - [561, 128, 1, 256] - - [119, 3.87] - - - [2816, 128, 1, 256] - - [115, 21.21] - - - [3073, 128, 1, 256] - - [161, 20.073] - - - [2097, 128, 1, 256] - - [119, 14.071] - - - [768, 128, 1, 256] - - [167, 6.405] - - - [9984, 128, 1, 256] - - [115, 42.97] - - - [3584, 128, 1, 256] - - [110, 22.61] - - - [817, 128, 1, 256] - - [129, 5.636] - - - [5632, 128, 1, 256] - - [110, 33.525] - - - [9472, 128, 1, 256] - - [115, 41.248] - - - [2305, 128, 1, 256] - - [129, 15.9] - - - [1329, 128, 1, 256] - - [119, 9.168] - - - [5888, 128, 1, 256] - - [111, 34.907] - - - [7680, 128, 1, 256] - - [178, 33.944] - - - [4608, 128, 1, 256] - - [115, 28.462] - - - [2353, 128, 1, 256] - - [166, 15.789] - - - [5120, 128, 1, 256] - - [109, 30.601] - - - [769, 128, 1, 256] - - [185, 5.733] - - - [1792, 128, 1, 256] - - [119, 14.072] - - - [1073, 128, 1, 256] - - [119, 7.437] - - - [513, 128, 1, 256] - - [133, 3.806] - - - [4096, 128, 1, 256] - - [111, 25.3] - - - [7424, 128, 1, 256] - - [124, 34.356] - - - [4352, 128, 1, 256] - - [111, 26.77] - - - [1793, 128, 1, 256] - - [119, 12.725] - - - [8192, 128, 1, 256] - - [187, 35.992] - - - [1281, 128, 1, 256] - - [131, 9.551] - - - [305, 128, 1, 256] - - [170, 2.104] - - - [2560, 128, 1, 256] - - [117, 19.481] - - - [2609, 128, 1, 256] - - [119, 17.831] - - - [1585, 128, 1, 256] - - [151, 10.783] - - - [8704, 128, 1, 256] - - [110, 37.244] - - - [10240, 128, 1, 256] - - [188, 41.644] - - - [256, 128, 1, 256] - - [116, 2.135] - - - [1025, 128, 1, 256] - - [185, 7.604] - - - [2865, 128, 1, 256] - - [144, 19.051] - - - [5376, 128, 1, 256] - - [111, 32.001] - - - [1841, 128, 1, 256] - - [166, 12.41] - - - [7936, 128, 1, 256] - - [109, 36.158] - - - [6144, 128, 1, 256] - - [111, 35.292] - - - [1024, 128, 1, 256] - - [116, 8.54] - - - [36096, 1281, 1, 256] - - [91, 63.047] - - - [38656, 2816, 1, 256] - - [75, 72.503] - - - [35072, 2048, 1, 256] - - [75, 71.133] - - - [39424, 2865, 1, 256] - - [55, 70.508] - - - [39168, 3328, 1, 256] - - [29, 72.947] - - - [36096, 2865, 1, 256] - - [52, 68.618] - - - [39216, 5632, 1, 256] - - [59, 62.951] - - - [38144, 6144, 1, 256] - - [28, 74.027] - - - [35328, 3072, 1, 256] - - [48, 73.434] - - - [39936, 256, 1, 256] - - [42, 59.512] - - - [36864, 3328, 1, 256] - - [30, 72.842] - - - [39168, 6144, 1, 256] - - [42, 73.921] - - - [36352, 4352, 1, 256] - - [30, 74.346] - - - [37680, 10240, 1, 256] - - [42, 62.962] - - - [38144, 256, 1, 256] - - [59, 57.666] - - - [37632, 1281, 1, 256] - - [75, 63.847] - - - [35632, 1792, 1, 256] - - [59, 62.655] - - - [36096, 4096, 1, 256] - - [42, 72.242] - - - [36144, 2816, 1, 256] - - [37, 62.834] - - - [36352, 256, 1, 256] - - [26, 55.862] - - - [35888, 2865, 1, 256] - - [54, 61.552] - - - [38912, 1280, 1, 256] - - [30, 71.695] - - - [37120, 3072, 1, 256] - - [26, 73.092] - - - [38448, 10240, 1, 256] - - [62, 62.59] - - - [39936, 3328, 1, 256] - - [28, 73.571] - - - [39168, 10240, 1, 256] - - [29, 74.409] - - - [39680, 3329, 1, 256] - - [55, 69.552] - - - [37168, 2865, 1, 256] - - [54, 61.603] - - - [38144, 5888, 1, 256] - - [38, 73.803] - - - [37120, 1281, 1, 256] - - [75, 64.097] - - - [37376, 10240, 1, 256] - - [26, 74.986] - - - [38704, 5120, 1, 256] - - [45, 62.977] - - - [39168, 5376, 1, 256] - - [29, 73.875] - - - [38656, 2865, 1, 256] - - [55, 69.906] - - - [37376, 3584, 1, 256] - - [26, 74.017] - - - [35072, 6144, 1, 256] - - [26, 74.189] - - - [39936, 6144, 1, 256] - - [38, 74.671] - - - [37632, 5376, 1, 256] - - [26, 73.89] - - - [36352, 2304, 1, 256] - - [26, 73.025] - - - [35840, 2048, 1, 256] - - [87, 71.487] - - - [36608, 1280, 1, 256] - - [26, 70.589] - - - [39936, 1792, 1, 256] - - [28, 73.01] - - - [36608, 3329, 1, 256] - - [55, 69.48] - - - [35072, 3329, 1, 256] - - [55, 69.553] - - - [37168, 3584, 1, 256] - - [45, 62.89] - - - [36096, 1792, 1, 256] - - [32, 70.106] - - - [39424, 3329, 1, 256] - - [42, 70.143] - - - [39424, 2048, 1, 256] - - [42, 71.915] - - - [39984, 2865, 1, 256] - - [37, 61.483] - - - [38448, 256, 1, 256] - - [103, 51.782] - - - [35584, 256, 1, 256] - - [45, 55.093] - - - [36608, 10240, 1, 256] - - [30, 74.583] - - - [38960, 5376, 1, 256] - - [29, 62.631] - - - [36352, 2048, 1, 256] - - [42, 71.757] - - - [39680, 1281, 1, 256] - - [28, 64.293] - - - [36608, 2304, 1, 256] - - [30, 72.296] - - - [39936, 1280, 1, 256] - - [28, 71.595] - - - [39680, 5376, 1, 256] - - [38, 74.025] - - - [35584, 10240, 1, 256] - - [25, 74.399] - - - [36864, 512, 1, 256] - - [50, 65.314] - - - [39424, 2816, 1, 256] - - [52, 73.142] - - - [35840, 2816, 1, 256] - - [26, 73.659] - - - [38192, 2816, 1, 256] - - [37, 62.921] - - - [35584, 2048, 1, 256] - - [42, 70.786] - - - [37936, 2865, 1, 256] - - [30, 61.41] - - - [39936, 2865, 1, 256] - - [30, 70.99] - - - [38656, 10240, 1, 256] - - [42, 74.41] - - - [36608, 2048, 1, 256] - - [62, 70.999] - - - [35120, 2816, 1, 256] - - [54, 63.338] - - - [39424, 5888, 1, 256] - - [26, 74.33] - - - [37680, 2816, 1, 256] - - [37, 63.798] - - - [36096, 6144, 1, 256] - - [42, 73.136] - - - [38144, 1281, 1, 256] - - [92, 64.27] - - - [37632, 2048, 1, 256] - - [62, 70.581] - - - [39680, 256, 1, 256] - - [59, 59.321] - - - [37680, 3840, 1, 256] - - [54, 63.791] - - - [39168, 2816, 1, 256] - - [42, 72.679] - - - [38192, 2865, 1, 256] - - [54, 60.856] - - - [38912, 4608, 1, 256] - - [28, 74.158] - - - [37120, 2048, 1, 256] - - [42, 71.342] - - - [35376, 1536, 1, 256] - - [37, 61.119] - - - [38448, 4864, 1, 256] - - [37, 63.358] - - - [38192, 10240, 1, 256] - - [75, 62.37] - - - [37632, 2816, 1, 256] - - [30, 72.628] - - - [39424, 1024, 1, 256] - - [77, 70.073] - - - [39168, 256, 1, 256] - - [59, 58.604] - - - [39984, 6144, 1, 256] - - [42, 63.37] - - - [38144, 4608, 1, 256] - - [38, 73.45] - - - [35840, 2865, 1, 256] - - [28, 71.032] - - - [36352, 6144, 1, 256] - - [42, 74.483] - - - [36864, 768, 1, 256] - - [37, 67.906] - - - [37888, 3328, 1, 256] - - [26, 73.568] - - - [36912, 3328, 1, 256] - - [42, 61.472] - - - [37120, 3584, 1, 256] - - [26, 73.471] - - - [38912, 1281, 1, 256] - - [28, 64.419] - - - [39472, 256, 1, 256] - - [30, 52.657] - - - [39936, 1281, 1, 256] - - [29, 64.808] - - - [37376, 5120, 1, 256] - - [38, 74.278] - - - [37888, 2048, 1, 256] - - [29, 71.546] - - - [37632, 1280, 1, 256] - - [30, 70.731] - - - [35376, 2816, 1, 256] - - [59, 63.641] - - - [38656, 3329, 1, 256] - - [31, 69.446] - - - [36912, 256, 1, 256] - - [30, 51.073] - - - [39168, 768, 1, 256] - - [70, 68.628] - - - [37424, 256, 1, 256] - - [38, 51.135] - - - [38448, 2816, 1, 256] - - [54, 63.43] - - - [35840, 3840, 1, 256] - - [26, 74.545] - - - [38912, 2865, 1, 256] - - [23, 71.049] - - - [36096, 1280, 1, 256] - - [58, 68.901] - - - [35328, 1024, 1, 256] - - [77, 69.48] - - - [39680, 3328, 1, 256] - - [30, 72.712] - - - [36352, 2816, 1, 256] - - [26, 73.066] - - - [38912, 256, 1, 256] - - [30, 58.888] - - - [39424, 3328, 1, 256] - - [29, 73.378] - - - [35888, 2816, 1, 256] - - [59, 63.218] - - - [36096, 2816, 1, 256] - - [65, 71.255] - - - [38960, 10240, 1, 256] - - [42, 62.954] - - - [35840, 3584, 1, 256] - - [28, 74.213] - - - [39424, 5120, 1, 256] - - [38, 74.236] - - - [37376, 1024, 1, 256] - - [73, 69.945] - - - [37632, 4096, 1, 256] - - [28, 72.873] - - - [36400, 2865, 1, 256] - - [54, 61.927] - - - [36144, 2560, 1, 256] - - [54, 63.37] - - - [36864, 1281, 1, 256] - - [28, 63.521] - - - [39424, 5376, 1, 256] - - [38, 74.471] - - - [36400, 2816, 1, 256] - - [59, 63.396] - - - [38656, 6144, 1, 256] - - [31, 73.826] - - - [37888, 5632, 1, 256] - - [28, 75.071] - - - [36912, 2865, 1, 256] - - [28, 59.852] - - - [38656, 4352, 1, 256] - - [31, 73.699] - - - [37632, 1536, 1, 256] - - [38, 70.658] - - - [35072, 2865, 1, 256] - - [30, 70.225] - - - [35888, 2304, 1, 256] - - [37, 63.831] - - - [38912, 3329, 1, 256] - - [55, 70.513] - - - [37680, 4096, 1, 256] - - [62, 62.597] - - - [38400, 6144, 1, 256] - - [29, 74.375] - - - [37888, 3840, 1, 256] - - [28, 74.613] - - - [36608, 3328, 1, 256] - - [29, 72.821] - - - [35328, 256, 1, 256] - - [85, 55.159] - - - [36096, 3329, 1, 256] - - [29, 68.384] - - - [37888, 5888, 1, 256] - - [38, 74.629] - - - [36864, 3329, 1, 256] - - [55, 69.763] - - - [35632, 256, 1, 256] - - [45, 50.1] - - - [38656, 4864, 1, 256] - - [48, 73.807] - - - [37888, 2816, 1, 256] - - [30, 73.767] - - - [37120, 3328, 1, 256] - - [29, 72.974] - - - [35328, 1536, 1, 256] - - [32, 70.794] - - - [35328, 1280, 1, 256] - - [58, 70.894] - - - [35888, 10240, 1, 256] - - [29, 63.793] - - - [36400, 10240, 1, 256] - - [42, 62.71] - - - [35072, 10240, 1, 256] - - [26, 74.629] - - - [39680, 2816, 1, 256] - - [38, 72.727] - - - [35584, 3329, 1, 256] - - [74, 69.36] - - - [36656, 256, 1, 256] - - [59, 50.805] - - - [38144, 4096, 1, 256] - - [75, 73.033] - - - [39936, 2816, 1, 256] - - [26, 73.724] - - - [36864, 3072, 1, 256] - - [26, 73.23] - - - [37936, 2816, 1, 256] - - [54, 63.534] - - - [37632, 3584, 1, 256] - - [26, 73.496] - - - [39984, 10240, 1, 256] - - [42, 63.649] - - - [38656, 512, 1, 256] - - [54, 66.314] - - - [35328, 10240, 1, 256] - - [52, 74.983] - - - [36096, 2048, 1, 256] - - [29, 69.762] - - - [37120, 4864, 1, 256] - - [26, 73.988] - - - [35840, 10240, 1, 256] - - [30, 75.284] - - - [39680, 5632, 1, 256] - - [38, 74.307] - - - [38144, 4352, 1, 256] - - [30, 73.886] - - - [36400, 2560, 1, 256] - - [37, 63.434] - - - [35840, 3329, 1, 256] - - [55, 70.438] - - - [37424, 10240, 1, 256] - - [42, 62.738] - - - [38912, 10240, 1, 256] - - [26, 75.232] - - - [35072, 768, 1, 256] - - [45, 68.0] - - - [36096, 3840, 1, 256] - - [55, 72.331] - - - [36656, 3072, 1, 256] - - [28, 62.236] - - - [39680, 1536, 1, 256] - - [33, 70.578] - - - [36656, 2865, 1, 256] - - [37, 61.222] - - - [38912, 512, 1, 256] - - [28, 65.257] - - - [38400, 256, 1, 256] - - [26, 57.601] - - - [38704, 10240, 1, 256] - - [29, 62.954] - - - [38912, 5376, 1, 256] - - [26, 74.867] - - - [35120, 256, 1, 256] - - [59, 53.951] - - - [38656, 3328, 1, 256] - - [29, 72.917] - - - [37888, 1536, 1, 256] - - [38, 71.305] - - - [39216, 5376, 1, 256] - - [37, 63.737] - - - [37376, 3329, 1, 256] - - [33, 70.049] - - - [37680, 256, 1, 256] - - [54, 51.618] - - - [39680, 6144, 1, 256] - - [30, 74.112] - - - [38400, 2865, 1, 256] - - [55, 70.515] - - - [36608, 2865, 1, 256] - - [55, 69.97] - - - [38912, 768, 1, 256] - - [41, 69.012] - - - [35584, 1792, 1, 256] - - [35, 71.502] - - - [39424, 256, 1, 256] - - [26, 58.69] - - - [36352, 1281, 1, 256] - - [75, 64.906] - - - [38400, 2048, 1, 256] - - [29, 71.748] - - - [38144, 3329, 1, 256] - - [55, 69.524] - - - [39680, 2048, 1, 256] - - [54, 69.64] - - - [38656, 256, 1, 256] - - [26, 57.937] - - - [39728, 2816, 1, 256] - - [59, 63.085] - - - [36352, 3329, 1, 256] - - [30, 70.011] - - - [38400, 10240, 1, 256] - - [26, 74.925] - - - [39984, 6400, 1, 256] - - [42, 63.356] - - - [37888, 4352, 1, 256] - - [38, 74.699] - - - [37888, 4096, 1, 256] - - [38, 73.073] - - - [35584, 1536, 1, 256] - - [36, 70.284] - - - [36096, 256, 1, 256] - - [37, 55.392] - - - [36864, 2048, 1, 256] - - [28, 68.862] - - - [36144, 2865, 1, 256] - - [54, 61.056] - - - [35584, 3584, 1, 256] - - [38, 73.116] - - - [35072, 1024, 1, 256] - - [41, 69.616] - - - [36352, 3328, 1, 256] - - [42, 73.516] - - - [39424, 1281, 1, 256] - - [92, 65.011] - - - [39728, 10240, 1, 256] - - [42, 62.94] - - - [37632, 2865, 1, 256] - - [55, 70.226] - - - [37168, 3328, 1, 256] - - [37, 63.032] - - - [37376, 5376, 1, 256] - - [38, 74.538] - - - [35328, 2865, 1, 256] - - [71, 70.435] - - - [35584, 6144, 1, 256] - - [26, 73.895] - - - [38704, 2816, 1, 256] - - [54, 63.386] - - - [36608, 3072, 1, 256] - - [38, 73.008] - - - [39680, 1280, 1, 256] - - [26, 70.822] - - - [35328, 1281, 1, 256] - - [75, 64.856] - - - [36608, 512, 1, 256] - - [41, 65.036] - - - [39936, 1536, 1, 256] - - [33, 71.413] - - - [39728, 5888, 1, 256] - - [59, 63.022] - - - [39168, 1281, 1, 256] - - [79, 64.281] - - - [37120, 256, 1, 256] - - [59, 56.339] - - - [38960, 2865, 1, 256] - - [28, 61.03] - - - [39168, 5120, 1, 256] - - [25, 73.651] - - - [36864, 256, 1, 256] - - [59, 56.308] - - - [36912, 2816, 1, 256] - - [29, 61.151] - - - [36096, 2304, 1, 256] - - [58, 70.752] - - - [35840, 3328, 1, 256] - - [28, 73.553] - - - [38704, 2865, 1, 256] - - [59, 61.61] - - - [38144, 1792, 1, 256] - - [38, 71.784] - - - [36608, 2560, 1, 256] - - [30, 72.736] - - - [35376, 10240, 1, 256] - - [42, 63.016] - - - [35840, 2304, 1, 256] - - [38, 73.298] - - - [35840, 1280, 1, 256] - - [38, 71.379] - - - [37376, 1280, 1, 256] - - [38, 71.316] - - - [35584, 3328, 1, 256] - - [43, 72.682] - - - [35584, 2865, 1, 256] - - [33, 69.888] - - - [39936, 10240, 1, 256] - - [28, 75.215] - - - [38912, 5120, 1, 256] - - [38, 74.698] - - - [37632, 3329, 1, 256] - - [33, 69.618] - - - [37888, 1792, 1, 256] - - [28, 72.793] - - - [36608, 1281, 1, 256] - - [29, 64.424] - - - [38192, 4352, 1, 256] - - [59, 62.726] - - - [39936, 2048, 1, 256] - - [42, 72.048] - - - [35072, 1281, 1, 256] - - [79, 64.274] - - - [39472, 2816, 1, 256] - - [54, 63.357] - - - [39728, 2865, 1, 256] - - [37, 61.192] - - - [38400, 2816, 1, 256] - - [26, 73.171] - - - [38400, 4608, 1, 256] - - [43, 73.988] - - - [39216, 10240, 1, 256] - - [62, 62.289] - - - [35072, 3072, 1, 256] - - [26, 72.885] - - - [38400, 4352, 1, 256] - - [30, 74.322] - - - [39216, 2816, 1, 256] - - [59, 63.182] - - - [35840, 1792, 1, 256] - - [30, 72.46] - - - [35632, 2048, 1, 256] - - [37, 62.969] - - - [38704, 256, 1, 256] - - [37, 52.174] - - - [37888, 3329, 1, 256] - - [30, 70.547] - - - [37888, 6144, 1, 256] - - [28, 74.7] - - - [37376, 6144, 1, 256] - - [42, 74.51] - - - [37376, 256, 1, 256] - - [32, 56.495] - - - [36400, 256, 1, 256] - - [37, 50.53] - - - [37936, 4096, 1, 256] - - [42, 62.768] - - - [38144, 10240, 1, 256] - - [38, 74.524] - - - [35376, 1792, 1, 256] - - [37, 62.315] - - - [37168, 10240, 1, 256] - - [62, 62.52] - - - [39984, 2816, 1, 256] - - [45, 63.532] - - - [37168, 2816, 1, 256] - - [54, 62.931] - - - [39424, 5632, 1, 256] - - [28, 74.685] - - - [36352, 1280, 1, 256] - - [39, 70.957] - - - [39680, 10240, 1, 256] - - [30, 74.481] - - - [38144, 3328, 1, 256] - - [66, 72.575] - - - [39168, 2048, 1, 256] - - [42, 71.015] - - - [35328, 6144, 1, 256] - - [44, 74.535] - - - [35632, 2865, 1, 256] - - [54, 61.238] - - - [36656, 10240, 1, 256] - - [42, 63.069] - - - [36608, 4352, 1, 256] - - [38, 73.899] - - - [35120, 2865, 1, 256] - - [45, 61.404] - - - [36608, 6144, 1, 256] - - [28, 74.021] - - - [37888, 2865, 1, 256] - - [38, 70.976] - - - [39168, 1024, 1, 256] - - [70, 69.668] - - - [38704, 4864, 1, 256] - - [45, 63.525] - - - [39168, 2865, 1, 256] - - [30, 69.967] - - - [38960, 5120, 1, 256] - - [29, 62.521] - - - [36864, 2816, 1, 256] - - [26, 73.037] - - - [38656, 1280, 1, 256] - - [32, 70.603] - - - [35584, 1281, 1, 256] - - [40, 64.143] - - - [39216, 2865, 1, 256] - - [37, 61.369] - - - [35120, 1280, 1, 256] - - [59, 62.736] - - - [36096, 3328, 1, 256] - - [43, 71.742] - - - [38912, 6144, 1, 256] - - [28, 74.651] - - - [37376, 3840, 1, 256] - - [30, 74.191] - - - [37424, 2816, 1, 256] - - [54, 63.705] - - - [36864, 10240, 1, 256] - - [28, 74.71] - - - [35328, 3328, 1, 256] - - [44, 73.449] - - - [37632, 5632, 1, 256] - - [38, 74.204] - - - [35072, 1536, 1, 256] - - [38, 70.539] - - - [36864, 2865, 1, 256] - - [30, 70.301] - - - [36864, 4608, 1, 256] - - [30, 73.415] - - - [37888, 1280, 1, 256] - - [35, 71.371] - - - [36864, 4864, 1, 256] - - [30, 74.336] - - - [37632, 256, 1, 256] - - [35, 57.126] - - - [38912, 2816, 1, 256] - - [28, 73.697] - - - [38656, 5120, 1, 256] - - [25, 73.766] - - - [35072, 1280, 1, 256] - - [50, 70.409] - - - [38400, 3329, 1, 256] - - [55, 70.087] - - - [35840, 1281, 1, 256] - - [85, 64.22] - - - [39680, 2865, 1, 256] - - [55, 70.163] - - - [38192, 256, 1, 256] - - [54, 51.414] - - - [37632, 10240, 1, 256] - - [38, 74.479] - - - [39984, 256, 1, 256] - - [53, 53.316] - - - [37424, 2865, 1, 256] - - [37, 61.628] - - - [37888, 256, 1, 256] - - [45, 57.593] - - - [36864, 6144, 1, 256] - - [38, 74.004] - - - [38656, 1281, 1, 256] - - [90, 64.467] - - - [37936, 256, 1, 256] - - [59, 52.269] - - - [39168, 4864, 1, 256] - - [31, 73.766] - - - [35840, 256, 1, 256] - - [79, 55.143] - - - [37888, 10240, 1, 256] - - [30, 75.274] - - - [39728, 6144, 1, 256] - - [29, 62.67] - - - [39680, 5888, 1, 256] - - [38, 73.797] - - - [38144, 2816, 1, 256] - - [28, 72.629] - - - [39728, 256, 1, 256] - - [28, 53.166] - - - [37376, 2816, 1, 256] - - [26, 73.31] - - - [36352, 2865, 1, 256] - - [28, 70.628] - - - [39216, 256, 1, 256] - - [59, 52.562] - - - [37888, 1281, 1, 256] - - [75, 64.407] - - - [39472, 10240, 1, 256] - - [29, 63.129] - - - [37376, 2048, 1, 256] - - [44, 71.636] - - - [36096, 10240, 1, 256] - - [55, 73.83] - - - [35584, 1280, 1, 256] - - [50, 70.281] - - - [39168, 5632, 1, 256] - - [48, 73.916] - - - [39936, 5632, 1, 256] - - [28, 74.951] - - - [35072, 256, 1, 256] - - [37, 59.305] - - - [35376, 2865, 1, 256] - - [54, 61.863] - - - [38400, 4864, 1, 256] - - [38, 74.544] - - - [35888, 256, 1, 256] - - [45, 50.379] - - - [35072, 3328, 1, 256] - - [42, 72.749] - - - [37936, 10240, 1, 256] - - [42, 63.616] - - - [36352, 10240, 1, 256] - - [42, 75.006] - - - [38656, 2048, 1, 256] - - [42, 71.141] - - - [35632, 2816, 1, 256] - - [37, 63.573] - - - [36912, 10240, 1, 256] - - [29, 62.221] - - - [39936, 5888, 1, 256] - - [30, 74.727] - - - [38448, 2865, 1, 256] - - [45, 61.645] - - - [38144, 3840, 1, 256] - - [28, 73.821] - - - [37632, 6144, 1, 256] - - [38, 74.0] - - - [37376, 3328, 1, 256] - - [42, 73.555] - - - [36608, 2816, 1, 256] - - [28, 72.815] - - - [36912, 3072, 1, 256] - - [29, 60.786] - - - [37120, 2816, 1, 256] - - [29, 72.663] - - - [38144, 2865, 1, 256] - - [55, 70.012] - - - [38912, 2048, 1, 256] - - [42, 70.594] - - - [38192, 4608, 1, 256] - - [59, 63.111] - - - [37120, 5120, 1, 256] - - [28, 74.028] - - - [38400, 3328, 1, 256] - - [29, 73.36] - - - [35632, 10240, 1, 256] - - [29, 62.964] - - - [38912, 4864, 1, 256] - - [28, 74.806] - - - [37120, 10240, 1, 256] - - [28, 74.59] - - - [37120, 3329, 1, 256] - - [33, 69.606] - - - [35840, 6144, 1, 256] - - [28, 74.653] - - - [38400, 1281, 1, 256] - - [43, 64.99] - - - [36144, 10240, 1, 256] - - [42, 62.723] - - - [38144, 1280, 1, 256] - - [36, 70.558] - - - [39424, 10240, 1, 256] - - [28, 74.889] - - - [39424, 6144, 1, 256] - - [42, 74.434] - - - [39424, 1280, 1, 256] - - [35, 70.931] - - - [35328, 3329, 1, 256] - - [48, 70.11] - - - [39472, 5888, 1, 256] - - [54, 63.366] - - - [36352, 4096, 1, 256] - - [29, 73.815] - - - [38656, 4608, 1, 256] - - [29, 73.336] - - - [37168, 256, 1, 256] - - [59, 51.069] - - - [38144, 2048, 1, 256] - - [42, 71.056] - - - [35840, 1536, 1, 256] - - [38, 71.264] - - - [37120, 1280, 1, 256] - - [36, 70.842] - - - [37424, 3840, 1, 256] - - [37, 63.367] - - - [37424, 3584, 1, 256] - - [54, 63.462] - - - [36864, 2560, 1, 256] - - [23, 72.877] - - - [39936, 6400, 1, 256] - - [30, 75.294] - - - [36096, 2560, 1, 256] - - [58, 71.482] - - - [37120, 768, 1, 256] - - [45, 68.579] - - - [35328, 2048, 1, 256] - - [42, 71.615] - - - [36608, 4608, 1, 256] - - [29, 73.42] - - - [38400, 4096, 1, 256] - - [42, 73.807] - - - [35328, 2816, 1, 256] - - [52, 73.251] - - - [36144, 256, 1, 256] - - [28, 50.175] - - - [36608, 256, 1, 256] - - [56, 55.964] - - - [39168, 3329, 1, 256] - - [25, 69.417] - - - [38448, 4608, 1, 256] - - [59, 62.818] - - - [37632, 3328, 1, 256] - - [38, 72.662] - - - [37680, 2865, 1, 256] - - [54, 61.765] - - - [35120, 10240, 1, 256] - - [45, 62.831] - - - [37120, 6144, 1, 256] - - [42, 73.953] - - - [36656, 2816, 1, 256] - - [59, 63.274] - - - [39936, 3329, 1, 256] - - [55, 70.508] - - - [35328, 1792, 1, 256] - - [32, 72.387] - - - [35120, 1536, 1, 256] - - [37, 61.566] - - - [39472, 2865, 1, 256] - - [37, 61.639] - - - [37936, 4352, 1, 256] - - [26, 63.579] - - - [35888, 2048, 1, 256] - - [54, 62.81] - - - [37888, 3584, 1, 256] - - [28, 74.221] - - - [37376, 2865, 1, 256] - - [38, 70.609] - - - [36864, 1280, 1, 256] - - [56, 70.756] - - - [39472, 5632, 1, 256] - - [37, 63.752] - - - [37120, 1024, 1, 256] - - [61, 69.346] - - - [37120, 2865, 1, 256] - - [55, 70.15] - - - [38400, 1280, 1, 256] - - [35, 71.196] - - - [35584, 2816, 1, 256] - - [36, 72.373] - - - [37376, 1281, 1, 256] - - [77, 64.736] - - - [36352, 2560, 1, 256] - - [28, 73.332] - - - [36144, 2304, 1, 256] - - [45, 62.853] - - - [37632, 3840, 1, 256] - - [28, 73.728] - - - [38960, 2816, 1, 256] - - [30, 62.02] - - - [37376, 3072, 1, 256] - - [39, 73.241] - - - [35072, 2816, 1, 256] - - [28, 72.753] - - - [38912, 3328, 1, 256] - - [38, 73.576] - - - [38960, 256, 1, 256] - - [59, 52.833] - - - [35376, 256, 1, 256] - - [59, 49.669] - - - [39168, 1280, 1, 256] - - [28, 70.66] - - - [44032, 5888, 1, 256] - - [38, 74.602] - - - [40192, 2865, 1, 256] - - [26, 70.132] - - - [43312, 256, 1, 256] - - [37, 50.632] - - - [43520, 1280, 1, 256] - - [26, 71.545] - - - [41216, 2816, 1, 256] - - [26, 72.943] - - - [41520, 7936, 1, 256] - - [37, 63.037] - - - [43008, 2048, 1, 256] - - [29, 70.834] - - - [42496, 2048, 1, 256] - - [42, 72.118] - - - [40704, 3328, 1, 256] - - [42, 72.847] - - - [41776, 7936, 1, 256] - - [45, 63.282] - - - [40192, 1792, 1, 256] - - [28, 72.033] - - - [43520, 6144, 1, 256] - - [38, 74.517] - - - [42032, 2865, 1, 256] - - [37, 61.445] - - - [41472, 3329, 1, 256] - - [55, 70.049] - - - [41008, 7424, 1, 256] - - [29, 61.948] - - - [40448, 2865, 1, 256] - - [28, 70.49] - - - [41264, 2865, 1, 256] - - [45, 62.119] - - - [43312, 9728, 1, 256] - - [59, 62.838] - - - [40704, 2816, 1, 256] - - [28, 72.846] - - - [42544, 8704, 1, 256] - - [54, 62.483] - - - [40960, 7168, 1, 256] - - [39, 66.483] - - - [41216, 3329, 1, 256] - - [74, 69.564] - - - [41984, 6144, 1, 256] - - [28, 74.663] - - - [42240, 10240, 1, 256] - - [28, 74.534] - - - [42752, 2865, 1, 256] - - [55, 70.315] - - - [41216, 1280, 1, 256] - - [32, 71.322] - - - [40704, 7168, 1, 256] - - [26, 73.285] - - - [41216, 10240, 1, 256] - - [29, 74.599] - - - [40960, 256, 1, 256] - - [37, 60.229] - - - [40704, 2560, 1, 256] - - [26, 73.113] - - - [42752, 3329, 1, 256] - - [33, 69.627] - - - [43264, 3329, 1, 256] - - [55, 69.794] - - - [40192, 6144, 1, 256] - - [26, 74.232] - - - [43008, 10240, 1, 256] - - [26, 75.153] - - - [43520, 1281, 1, 256] - - [42, 65.119] - - - [42496, 8960, 1, 256] - - [30, 75.285] - - - [43312, 10240, 1, 256] - - [45, 62.292] - - - [44032, 6144, 1, 256] - - [26, 74.572] - - - [40192, 256, 1, 256] - - [54, 59.814] - - - [41984, 1536, 1, 256] - - [28, 71.539] - - - [41216, 768, 1, 256] - - [37, 68.62] - - - [40752, 256, 1, 256] - - [26, 54.047] - - - [44288, 1280, 1, 256] - - [58, 70.78] - - - [43520, 9216, 1, 256] - - [42, 74.767] - - - [42032, 8192, 1, 256] - - [42, 63.425] - - - [41728, 3584, 1, 256] - - [51, 71.854] - - - [40448, 1280, 1, 256] - - [28, 71.25] - - - [41216, 7168, 1, 256] - - [28, 73.415] - - - [42496, 1280, 1, 256] - - [26, 71.711] - - - [40448, 6656, 1, 256] - - [25, 74.479] - - - [40240, 256, 1, 256] - - [26, 53.619] - - - [41264, 2816, 1, 256] - - [37, 63.512] - - - [43264, 3328, 1, 256] - - [28, 72.954] - - - [43008, 9216, 1, 256] - - [26, 74.511] - - - [42240, 1281, 1, 256] - - [42, 64.663] - - - [42288, 2865, 1, 256] - - [54, 61.45] - - - [43008, 3328, 1, 256] - - [26, 73.593] - - - [40496, 256, 1, 256] - - [26, 53.635] - - - [43264, 8960, 1, 256] - - [30, 74.747] - - - [43056, 9472, 1, 256] - - [42, 62.988] - - - [40448, 3328, 1, 256] - - [44, 73.139] - - - [41776, 8192, 1, 256] - - [42, 62.693] - - - [40704, 6400, 1, 256] - - [26, 74.452] - - - [41984, 7680, 1, 256] - - [38, 75.275] - - - [43312, 9472, 1, 256] - - [45, 62.742] - - - [40192, 1280, 1, 256] - - [38, 71.066] - - - [43776, 5632, 1, 256] - - [42, 73.586] - - - [41984, 2865, 1, 256] - - [33, 71.09] - - - [40448, 2816, 1, 256] - - [28, 73.248] - - - [42240, 3328, 1, 256] - - [42, 73.03] - - - [42752, 2048, 1, 256] - - [75, 70.609] - - - [42240, 256, 1, 256] - - [59, 61.761] - - - [43008, 3329, 1, 256] - - [33, 70.532] - - - [44032, 5632, 1, 256] - - [26, 75.031] - - - [40192, 2048, 1, 256] - - [75, 71.401] - - - [41216, 256, 1, 256] - - [27, 60.115] - - - [44288, 9984, 1, 256] - - [28, 74.699] - - - [43008, 1280, 1, 256] - - [30, 71.961] - - - [41984, 2816, 1, 256] - - [26, 73.74] - - - [42752, 6144, 1, 256] - - [38, 74.06] - - - [43776, 3329, 1, 256] - - [29, 68.689] - - - [43008, 2865, 1, 256] - - [23, 71.08] - - - [43776, 9728, 1, 256] - - [29, 74.27] - - - [42240, 7936, 1, 256] - - [28, 74.473] - - - [41472, 7424, 1, 256] - - [26, 74.806] - - - [43776, 5376, 1, 256] - - [29, 73.395] - - - [43008, 6144, 1, 256] - - [28, 74.574] - - - [41216, 3072, 1, 256] - - [38, 73.182] - - - [42496, 8192, 1, 256] - - [38, 74.717] - - - [40704, 6144, 1, 256] - - [26, 74.159] - - - [44032, 3329, 1, 256] - - [55, 70.575] - - - [43520, 2048, 1, 256] - - [62, 71.975] - - - [43264, 2048, 1, 256] - - [75, 71.216] - - - [40448, 1281, 1, 256] - - [62, 64.998] - - - [40496, 2865, 1, 256] - - [45, 61.745] - - - [40448, 6144, 1, 256] - - [26, 74.257] - - - [41008, 10240, 1, 256] - - [29, 62.391] - - - [43056, 2865, 1, 256] - - [28, 60.62] - - - [43264, 1280, 1, 256] - - [30, 71.169] - - - [40192, 10240, 1, 256] - - [28, 74.615] - - - [41216, 7680, 1, 256] - - [28, 74.657] - - - [41008, 7168, 1, 256] - - [42, 62.179] - - - [44288, 2048, 1, 256] - - [29, 71.786] - - - [41472, 6144, 1, 256] - - [44, 74.195] - - - [43264, 2865, 1, 256] - - [55, 70.362] - - - [40448, 6912, 1, 256] - - [31, 74.755] - - - [41216, 6912, 1, 256] - - [38, 74.568] - - - [41984, 1792, 1, 256] - - [28, 72.966] - - - [40192, 1281, 1, 256] - - [75, 64.585] - - - [40960, 3329, 1, 256] - - [55, 62.664] - - - [41520, 10240, 1, 256] - - [75, 62.859] - - - [44032, 10240, 1, 256] - - [26, 75.084] - - - [43264, 2816, 1, 256] - - [38, 72.884] - - - [43008, 4608, 1, 256] - - [38, 74.106] - - - [43776, 1281, 1, 256] - - [62, 63.973] - - - [40240, 6656, 1, 256] - - [45, 63.007] - - - [43264, 9216, 1, 256] - - [30, 74.059] - - - [40704, 3329, 1, 256] - - [33, 69.77] - - - [42752, 3328, 1, 256] - - [38, 72.653] - - - [41984, 2048, 1, 256] - - [42, 71.786] - - - [44288, 3329, 1, 256] - - [29, 69.201] - - - [40192, 3328, 1, 256] - - [42, 72.906] - - - [40960, 10240, 1, 256] - - [39, 67.284] - - - [42496, 256, 1, 256] - - [64, 56.259] - - - [40496, 10240, 1, 256] - - [75, 62.619] - - - [40496, 6912, 1, 256] - - [54, 63.78] - - - [43776, 6144, 1, 256] - - [29, 73.627] - - - [40960, 1280, 1, 256] - - [26, 64.486] - - - [42288, 8704, 1, 256] - - [29, 62.698] - - - [42496, 3328, 1, 256] - - [29, 73.671] - - - [41216, 2865, 1, 256] - - [55, 70.173] - - - [42496, 3329, 1, 256] - - [33, 70.164] - - - [41984, 7936, 1, 256] - - [28, 75.243] - - - [41472, 1281, 1, 256] - - [62, 65.032] - - - [41776, 256, 1, 256] - - [54, 55.02] - - - [42752, 8960, 1, 256] - - [23, 74.694] - - - [41472, 7168, 1, 256] - - [26, 73.532] - - - [40240, 10240, 1, 256] - - [42, 62.723] - - - [41728, 1280, 1, 256] - - [60, 69.927] - - - [40752, 2865, 1, 256] - - [37, 61.872] - - - [40960, 2048, 1, 256] - - [28, 62.991] - - - [41472, 7680, 1, 256] - - [38, 74.81] - - - [41472, 10240, 1, 256] - - [26, 74.705] - - - [41264, 7680, 1, 256] - - [54, 63.108] - - - [42800, 8960, 1, 256] - - [54, 63.579] - - - [41728, 10240, 1, 256] - - [42, 74.303] - - - [44032, 3328, 1, 256] - - [30, 73.632] - - - [40704, 6912, 1, 256] - - [28, 74.49] - - - [41472, 2048, 1, 256] - - [49, 71.859] - - - [40960, 6144, 1, 256] - - [39, 67.029] - - - [43776, 3328, 1, 256] - - [29, 72.192] - - - [42496, 2865, 1, 256] - - [28, 70.925] - - - [40960, 3328, 1, 256] - - [39, 66.416] - - - [41728, 7936, 1, 256] - - [42, 73.453] - - - [41984, 3329, 1, 256] - - [26, 70.588] - - - [43008, 256, 1, 256] - - [28, 57.193] - - - [42240, 1280, 1, 256] - - [23, 70.817] - - - [43776, 10240, 1, 256] - - [29, 74.271] - - - [42752, 8448, 1, 256] - - [28, 74.38] - - - [42496, 1281, 1, 256] - - [42, 65.226] - - - [44032, 1536, 1, 256] - - [26, 71.765] - - - [40960, 2816, 1, 256] - - [23, 66.747] - - - [44288, 1792, 1, 256] - - [26, 71.649] - - - [43264, 1281, 1, 256] - - [75, 64.133] - - - [43008, 8704, 1, 256] - - [26, 75.246] - - - [41728, 1536, 1, 256] - - [58, 69.755] - - - [41728, 2048, 1, 256] - - [90, 69.934] - - - [43520, 9728, 1, 256] - - [28, 74.909] - - - [42032, 256, 1, 256] - - [37, 55.382] - - - [43776, 256, 1, 256] - - [59, 56.609] - - - [43008, 9472, 1, 256] - - [26, 75.471] - - - [44032, 1792, 1, 256] - - [38, 72.939] - - - [40704, 2865, 1, 256] - - [38, 70.238] - - - [42240, 1792, 1, 256] - - [28, 71.959] - - - [40704, 2304, 1, 256] - - [28, 72.693] - - - [42800, 9216, 1, 256] - - [29, 62.893] - - - [42240, 8704, 1, 256] - - [38, 74.597] - - - [42496, 6144, 1, 256] - - [26, 74.501] - - - [43568, 9728, 1, 256] - - [42, 63.01] - - - [40704, 2048, 1, 256] - - [29, 71.158] - - - [41472, 7936, 1, 256] - - [26, 74.77] - - - [42752, 2816, 1, 256] - - [28, 72.973] - - - [41008, 2865, 1, 256] - - [29, 58.776] - - - [40960, 6912, 1, 256] - - [39, 68.041] - - - [44032, 256, 1, 256] - - [37, 57.631] - - - [42496, 4352, 1, 256] - - [26, 74.491] - - - [42032, 8448, 1, 256] - - [29, 63.624] - - - [42752, 4608, 1, 256] - - [30, 73.474] - - - [44032, 1280, 1, 256] - - [28, 72.047] - - - [44288, 6144, 1, 256] - - [42, 74.003] - - - [42800, 2865, 1, 256] - - [37, 61.011] - - - [41008, 2816, 1, 256] - - [42, 61.755] - - - [41984, 8192, 1, 256] - - [26, 74.9] - - - [43264, 256, 1, 256] - - [61, 56.868] - - - [41728, 2865, 1, 256] - - [52, 68.836] - - - [43520, 5120, 1, 256] - - [26, 74.521] - - - [41984, 3584, 1, 256] - - [38, 74.444] - - - [41216, 3328, 1, 256] - - [29, 73.012] - - - [43520, 9472, 1, 256] - - [28, 75.262] - - - [43264, 9728, 1, 256] - - [26, 74.48] - - - [41728, 1281, 1, 256] - - [92, 64.039] - - - [40704, 1281, 1, 256] - - [42, 64.284] - - - [42288, 256, 1, 256] - - [26, 49.73] - - - [40960, 512, 1, 256] - - [28, 64.162] - - - [42752, 4352, 1, 256] - - [26, 74.11] - - - [40752, 10240, 1, 256] - - [29, 62.91] - - - [41728, 3328, 1, 256] - - [42, 72.265] - - - [43568, 2816, 1, 256] - - [37, 62.54] - - - [43008, 512, 1, 256] - - [45, 66.535] - - - [41216, 2048, 1, 256] - - [42, 71.246] - - - [42800, 256, 1, 256] - - [54, 49.994] - - - [43312, 2816, 1, 256] - - [37, 63.107] - - - [40192, 6400, 1, 256] - - [28, 74.557] - - - [41264, 7424, 1, 256] - - [37, 63.342] - - - [42544, 8960, 1, 256] - - [45, 63.039] - - - [41472, 256, 1, 256] - - [38, 61.092] - - - [42288, 10240, 1, 256] - - [42, 62.646] - - - [43520, 1024, 1, 256] - - [54, 70.335] - - - [42288, 8448, 1, 256] - - [45, 62.928] - - - [43776, 9472, 1, 256] - - [75, 74.186] - - - [43008, 1281, 1, 256] - - [26, 64.667] - - - [43008, 8960, 1, 256] - - [39, 75.419] - - - [41728, 256, 1, 256] - - [54, 61.757] - - - [41520, 7680, 1, 256] - - [29, 62.77] - - - [42240, 3329, 1, 256] - - [33, 69.539] - - - [41472, 2816, 1, 256] - - [48, 72.969] - - - [41216, 6144, 1, 256] - - [28, 74.096] - - - [40752, 2816, 1, 256] - - [45, 62.707] - - - [42496, 8704, 1, 256] - - [30, 75.149] - - - [40448, 6400, 1, 256] - - [38, 74.798] - - - [44032, 1281, 1, 256] - - [26, 64.951] - - - [41472, 1024, 1, 256] - - [64, 70.004] - - - [41216, 7424, 1, 256] - - [26, 74.529] - - - [43312, 2865, 1, 256] - - [37, 61.386] - - - [40960, 768, 1, 256] - - [30, 61.525] - - - [40240, 2865, 1, 256] - - [45, 61.619] - - - [43264, 768, 1, 256] - - [59, 68.7] - - - [40192, 3329, 1, 256] - - [38, 69.591] - - - [42800, 10240, 1, 256] - - [42, 62.897] - - - [42752, 512, 1, 256] - - [56, 65.629] - - - [40752, 6912, 1, 256] - - [37, 63.427] - - - [42240, 8192, 1, 256] - - [29, 74.303] - - - [42288, 2816, 1, 256] - - [37, 62.835] - - - [40960, 2865, 1, 256] - - [55, 63.581] - - - [42800, 2816, 1, 256] - - [59, 62.968] - - - [42496, 2816, 1, 256] - - [30, 73.572] - - - [41728, 7680, 1, 256] - - [29, 73.673] - - - [42240, 8448, 1, 256] - - [42, 74.494] - - - [41984, 1281, 1, 256] - - [75, 64.815] - - - [41984, 3328, 1, 256] - - [28, 73.564] - - - [40240, 6400, 1, 256] - - [45, 63.269] - - - [44288, 256, 1, 256] - - [24, 57.649] - - - [42496, 4096, 1, 256] - - [42, 73.815] - - - [43520, 3329, 1, 256] - - [55, 70.231] - - - [44288, 5888, 1, 256] - - [75, 73.788] - - - [42752, 1281, 1, 256] - - [38, 64.042] - - - [43776, 9984, 1, 256] - - [42, 73.921] - - - [41008, 256, 1, 256] - - [26, 53.831] - - - [40960, 1281, 1, 256] - - [39, 57.818] - - - [40704, 6656, 1, 256] - - [30, 74.314] - - - [40192, 2816, 1, 256] - - [28, 72.942] - - - [43264, 10240, 1, 256] - - [28, 74.641] - - - [44032, 9984, 1, 256] - - [28, 75.513] - - - [43520, 2865, 1, 256] - - [28, 70.933] - - - [42240, 3840, 1, 256] - - [26, 73.925] - - - [43056, 9216, 1, 256] - - [29, 62.923] - - - [43520, 10240, 1, 256] - - [26, 74.931] - - - [42544, 10240, 1, 256] - - [42, 62.521] - - - [40448, 2304, 1, 256] - - [30, 72.902] - - - [40704, 1280, 1, 256] - - [58, 70.86] - - - [43520, 2816, 1, 256] - - [26, 73.456] - - - [43520, 5376, 1, 256] - - [28, 74.646] - - - [41984, 256, 1, 256] - - [28, 61.95] - - - [43776, 1280, 1, 256] - - [62, 68.773] - - - [43568, 2865, 1, 256] - - [54, 61.386] - - - [41520, 256, 1, 256] - - [45, 55.016] - - - [41472, 3328, 1, 256] - - [44, 73.35] - - - [40192, 6656, 1, 256] - - [26, 74.325] - - - [40448, 2048, 1, 256] - - [29, 71.758] - - - [41520, 2816, 1, 256] - - [59, 63.286] - - - [43520, 9984, 1, 256] - - [26, 75.254] - - - [42544, 2865, 1, 256] - - [59, 61.581] - - - [42240, 2048, 1, 256] - - [42, 71.349] - - - [41472, 1280, 1, 256] - - [68, 71.256] - - - [40192, 5888, 1, 256] - - [28, 73.973] - - - [42240, 2865, 1, 256] - - [28, 70.024] - - - [41984, 10240, 1, 256] - - [26, 75.156] - - - [41264, 10240, 1, 256] - - [28, 62.503] - - - [42752, 10240, 1, 256] - - [28, 74.606] - - - [41216, 1024, 1, 256] - - [73, 69.891] - - - [41776, 10240, 1, 256] - - [42, 62.884] - - - [40960, 7424, 1, 256] - - [39, 67.903] - - - [40960, 2560, 1, 256] - - [23, 66.523] - - - [41216, 1281, 1, 256] - - [90, 64.548] - - - [41984, 1280, 1, 256] - - [28, 71.781] - - - [40448, 3329, 1, 256] - - [33, 69.924] - - - [41776, 2816, 1, 256] - - [54, 63.494] - - - [40704, 256, 1, 256] - - [37, 60.412] - - - [43264, 4864, 1, 256] - - [30, 74.253] - - - [42240, 6144, 1, 256] - - [28, 74.187] - - - [43520, 3328, 1, 256] - - [42, 73.577] - - - [42752, 256, 1, 256] - - [45, 56.12] - - - [40752, 7168, 1, 256] - - [29, 62.197] - - - [43776, 1536, 1, 256] - - [92, 67.546] - - - [42032, 10240, 1, 256] - - [29, 63.661] - - - [43008, 4864, 1, 256] - - [26, 74.853] - - - [40704, 10240, 1, 256] - - [28, 74.63] - - - [44288, 1281, 1, 256] - - [75, 64.844] - - - [41520, 2865, 1, 256] - - [45, 61.695] - - - [41264, 256, 1, 256] - - [104, 53.76] - - - [40496, 6656, 1, 256] - - [45, 63.334] - - - [42240, 4096, 1, 256] - - [42, 73.206] - - - [43568, 9984, 1, 256] - - [54, 63.212] - - - [43264, 9472, 1, 256] - - [28, 74.68] - - - [43008, 768, 1, 256] - - [37, 69.577] - - - [43776, 2816, 1, 256] - - [42, 72.397] - - - [43008, 2816, 1, 256] - - [23, 73.792] - - - [41984, 8448, 1, 256] - - [30, 75.147] - - - [43520, 256, 1, 256] - - [54, 56.961] - - - [43776, 2865, 1, 256] - - [29, 68.682] - - - [41984, 3840, 1, 256] - - [28, 74.662] - - - [42544, 256, 1, 256] - - [38, 49.843] - - - [43056, 2816, 1, 256] - - [42, 61.699] - - - [41472, 3072, 1, 256] - - [38, 73.12] - - - [41776, 2865, 1, 256] - - [54, 61.177] - - - [43056, 256, 1, 256] - - [38, 50.543] - - - [41728, 6144, 1, 256] - - [29, 73.428] - - - [42496, 8448, 1, 256] - - [29, 74.955] - - - [43568, 256, 1, 256] - - [59, 50.65] - - - [42752, 8704, 1, 256] - - [28, 74.639] - - - [42544, 2816, 1, 256] - - [54, 63.276] - - - [40448, 10240, 1, 256] - - [48, 74.672] - - - [41728, 2816, 1, 256] - - [92, 71.895] - - - [43568, 10240, 1, 256] - - [42, 62.816] - - - [44032, 2048, 1, 256] - - [42, 71.938] - - - [41472, 2865, 1, 256] - - [26, 70.643] - - - [40448, 256, 1, 256] - - [38, 60.316] - - - [41728, 3329, 1, 256] - - [29, 68.274] - - - [43264, 6144, 1, 256] - - [26, 74.082] - - - [40960, 6656, 1, 256] - - [23, 67.684] - - - [42752, 9216, 1, 256] - - [30, 74.194] - - - [40496, 2816, 1, 256] - - [45, 63.636] - - - [40704, 512, 1, 256] - - [54, 66.176] - - - [43056, 10240, 1, 256] - - [42, 63.048] - - - [44032, 9728, 1, 256] - - [28, 75.072] - - - [41728, 8192, 1, 256] - - [42, 74.078] - - - [43264, 1024, 1, 256] - - [54, 69.973] - - - [43776, 2048, 1, 256] - - [29, 70.652] - - - [40240, 2816, 1, 256] - - [59, 62.853] - - - [42752, 1280, 1, 256] - - [28, 71.138] - - - [44288, 10240, 1, 256] - - [42, 74.537] - - - [42240, 2816, 1, 256] - - [28, 72.834] - - - [41728, 7424, 1, 256] - - [29, 73.837] - - - [44288, 3328, 1, 256] - - [29, 72.991] - - - [43264, 5120, 1, 256] - - [28, 74.085] - - - [42032, 2816, 1, 256] - - [54, 63.145] - - - [11776, 6144, 1, 256] - - [38, 73.095] - - - [11264, 1792, 1, 256] - - [26, 67.19] - - - [4352, 2865, 1, 256] - - [26, 59.825] - - - [14640, 1536, 1, 256] - - [28, 59.341] - - - [4096, 2865, 1, 256] - - [60, 61.805] - - - [5168, 256, 1, 256] - - [122, 41.239] - - - [19968, 3328, 1, 256] - - [28, 72.493] - - - [12544, 3328, 1, 256] - - [36, 70.717] - - - [15408, 2816, 1, 256] - - [26, 62.707] - - - [16640, 3329, 1, 256] - - [26, 67.757] - - - [768, 768, 1, 256] - - [109, 28.462] - - - [3840, 512, 1, 256] - - [60, 32.591] - - - [7424, 5888, 1, 256] - - [28, 71.721] - - - [48, 49, 1, 256] - - [167, 0.141] - - - [16384, 768, 1, 256] - - [26, 54.969] - - - [15664, 2865, 1, 256] - - [26, 61.915] - - - [12544, 2048, 1, 256] - - [24, 67.949] - - - [7680, 4096, 1, 256] - - [28, 70.334] - - - [8240, 5376, 1, 256] - - [38, 65.489] - - - [11520, 256, 1, 256] - - [26, 47.463] - - - [12800, 256, 1, 256] - - [56, 51.019] - - - [10544, 2865, 1, 256] - - [28, 59.435] - - - [10032, 6912, 1, 256] - - [37, 63.949] - - - [3072, 3072, 1, 256] - - [56, 59.056] - - - [5888, 2865, 1, 256] - - [50, 64.896] - - - [8448, 3328, 1, 256] - - [35, 69.62] - - - [17920, 4096, 1, 256] - - [75, 72.746] - - - [19200, 5376, 1, 256] - - [26, 73.642] - - - [16432, 2865, 1, 256] - - [39, 64.164] - - - [12032, 3329, 1, 256] - - [36, 67.426] - - - [11776, 8704, 1, 256] - - [68, 74.604] - - - [11520, 1281, 1, 256] - - [61, 60.407] - - - [19760, 10240, 1, 256] - - [45, 62.731] - - - [15360, 1281, 1, 256] - - [26, 61.715] - - - [19712, 2865, 1, 256] - - [31, 67.478] - - - [9216, 6400, 1, 256] - - [26, 73.86] - - - [18944, 3329, 1, 256] - - [31, 69.341] - - - [5632, 2816, 1, 256] - - [56, 67.037] - - - [13872, 256, 1, 256] - - [36, 51.278] - - - [9984, 1280, 1, 256] - - [35, 61.779] - - - [19248, 10240, 1, 256] - - [37, 63.303] - - - [14128, 256, 1, 256] - - [36, 38.254] - - - [12080, 9216, 1, 256] - - [54, 63.728] - - - [18224, 5120, 1, 256] - - [59, 63.714] - - - [2352, 256, 1, 256] - - [124, 28.229] - - - [17712, 4608, 1, 256] - - [59, 63.487] - - - [8192, 5376, 1, 256] - - [36, 71.577] - - - [8752, 5888, 1, 256] - - [37, 63.141] - - - [11264, 3584, 1, 256] - - [50, 72.09] - - - [816, 256, 1, 256] - - [116, 12.953] - - - [5376, 3328, 1, 256] - - [26, 68.034] - - - [6144, 2560, 1, 256] - - [36, 67.197] - - - [9264, 256, 1, 256] - - [28, 38.214] - - - [8960, 5376, 1, 256] - - [58, 72.299] - - - [2608, 2353, 1, 256] - - [36, 50.213] - - - [2096, 256, 1, 256] - - [124, 25.259] - - - [9984, 7168, 1, 256] - - [35, 72.618] - - - [7424, 3329, 1, 256] - - [28, 65.354] - - - [2352, 2304, 1, 256] - - [37, 46.217] - - - [9984, 512, 1, 256] - - [35, 53.859] - - - [6656, 3840, 1, 256] - - [50, 69.201] - - - [17408, 3329, 1, 256] - - [38, 69.421] - - - [8496, 5376, 1, 256] - - [30, 64.353] - - - [11264, 3840, 1, 256] - - [35, 72.724] - - - [13312, 2865, 1, 256] - - [26, 69.024] - - - [3584, 768, 1, 256] - - [24, 44.664] - - - [11520, 6144, 1, 256] - - [60, 72.709] - - - [15360, 2048, 1, 256] - - [62, 69.024] - - - [7936, 3328, 1, 256] - - [64, 69.124] - - - [6144, 1281, 1, 256] - - [36, 55.503] - - - [19968, 6656, 1, 256] - - [38, 74.609] - - - [15152, 256, 1, 256] - - [50, 41.36] - - - [18432, 4608, 1, 256] - - [36, 73.713] - - - [1072, 256, 1, 256] - - [117, 16.399] - - - [6400, 4864, 1, 256] - - [56, 70.408] - - - [19712, 1281, 1, 256] - - [87, 59.967] - - - [1792, 1280, 1, 256] - - [60, 37.644] - - - [8192, 2865, 1, 256] - - [36, 65.783] - - - [3376, 256, 1, 256] - - [109, 37.471] - - - [10544, 2816, 1, 256] - - [59, 61.779] - - - [14336, 2816, 1, 256] - - [28, 71.657] - - - [16384, 1280, 1, 256] - - [23, 58.499] - - - [1280, 256, 1, 256] - - [117, 19.581] - - - [12544, 8960, 1, 256] - - [28, 74.392] - - - [13824, 1281, 1, 256] - - [54, 61.987] - - - [3072, 256, 1, 256] - - [110, 35.708] - - - [19760, 2816, 1, 256] - - [59, 63.795] - - - [8448, 5376, 1, 256] - - [50, 71.936] - - - [11824, 2865, 1, 256] - - [28, 60.194] - - - [6656, 3584, 1, 256] - - [36, 68.057] - - - [12288, 8704, 1, 256] - - [38, 75.086] - - - [11312, 256, 1, 256] - - [37, 44.077] - - - [15920, 2816, 1, 256] - - [54, 64.786] - - - [12032, 8448, 1, 256] - - [35, 73.765] - - - [14080, 2048, 1, 256] - - [93, 67.032] - - - [6400, 5120, 1, 256] - - [36, 70.545] - - - [7216, 2865, 1, 256] - - [28, 60.212] - - - [4400, 1280, 1, 256] - - [26, 47.188] - - - [5376, 3840, 1, 256] - - [35, 67.618] - - - [7168, 2816, 1, 256] - - [35, 67.248] - - - [19200, 5632, 1, 256] - - [28, 73.869] - - - [4144, 1024, 1, 256] - - [28, 44.266] - - - [12800, 3329, 1, 256] - - [68, 68.406] - - - [6400, 2865, 1, 256] - - [38, 64.555] - - - [12800, 5376, 1, 256] - - [58, 73.266] - - - [7168, 1536, 1, 256] - - [50, 59.413] - - - [19968, 1281, 1, 256] - - [59, 63.204] - - - [17664, 1281, 1, 256] - - [40, 62.694] - - - [11264, 3329, 1, 256] - - [30, 68.433] - - - [17712, 256, 1, 256] - - [24, 46.132] - - - [6656, 5376, 1, 256] - - [60, 71.81] - - - [13056, 5376, 1, 256] - - [35, 73.45] - - - [11568, 2865, 1, 256] - - [26, 58.964] - - - [3328, 1281, 1, 256] - - [50, 46.053] - - - [19968, 2048, 1, 256] - - [59, 70.791] - - - [2304, 2048, 1, 256] - - [36, 50.783] - - - [7728, 256, 1, 256] - - [54, 32.336] - - - [7424, 4352, 1, 256] - - [56, 71.451] - - - [5376, 2048, 1, 256] - - [36, 58.901] - - - [19456, 2816, 1, 256] - - [26, 72.622] - - - [7216, 2816, 1, 256] - - [26, 60.371] - - - [18688, 5376, 1, 256] - - [26, 73.843] - - - [4656, 1792, 1, 256] - - [28, 54.518] - - - [10240, 768, 1, 256] - - [45, 56.426] - - - [19456, 1280, 1, 256] - - [56, 69.213] - - - [18432, 3329, 1, 256] - - [30, 69.63] - - - [17920, 6144, 1, 256] - - [30, 74.033] - - - [1536, 1280, 1, 256] - - [48, 32.732] - - - [19456, 6400, 1, 256] - - [38, 75.291] - - - [15360, 6144, 1, 256] - - [26, 74.341] - - - [15664, 10240, 1, 256] - - [26, 63.248] - - - [3840, 256, 1, 256] - - [189, 34.149] - - - [4864, 3328, 1, 256] - - [56, 67.47] - - - [18224, 2865, 1, 256] - - [38, 61.253] - - - [13056, 9984, 1, 256] - - [26, 75.033] - - - [12288, 256, 1, 256] - - [50, 49.522] - - - [7168, 3840, 1, 256] - - [36, 70.01] - - - [17712, 4352, 1, 256] - - [59, 63.205] - - - [14592, 10240, 1, 256] - - [69, 73.199] - - - [8704, 5376, 1, 256] - - [50, 72.78] - - - [16128, 2816, 1, 256] - - [28, 71.212] - - - [4352, 3329, 1, 256] - - [50, 61.736] - - - [13568, 512, 1, 256] - - [36, 59.368] - - - [15872, 2865, 1, 256] - - [26, 69.974] - - - [12032, 1281, 1, 256] - - [35, 59.533] - - - [11520, 2048, 1, 256] - - [64, 66.369] - - - [12032, 2048, 1, 256] - - [59, 67.9] - - - [5632, 1281, 1, 256] - - [35, 51.805] - - - [13312, 9984, 1, 256] - - [30, 75.719] - - - [4912, 2865, 1, 256] - - [38, 54.733] - - - [15408, 2304, 1, 256] - - [28, 63.5] - - - [7472, 2816, 1, 256] - - [54, 61.407] - - - [18688, 10240, 1, 256] - - [30, 74.915] - - - [10752, 7936, 1, 256] - - [26, 74.542] - - - [2048, 1793, 1, 256] - - [36, 40.502] - - - [11776, 1280, 1, 256] - - [68, 64.065] - - - [10032, 256, 1, 256] - - [35, 39.665] - - - [17408, 1536, 1, 256] - - [50, 68.826] - - - [14080, 2865, 1, 256] - - [68, 68.381] - - - [16688, 3328, 1, 256] - - [45, 65.797] - - - [18944, 1024, 1, 256] - - [41, 67.129] - - - [2352, 2097, 1, 256] - - [38, 49.148] - - - [11008, 2048, 1, 256] - - [105, 64.317] - - - [10240, 6912, 1, 256] - - [50, 74.534] - - - [8448, 768, 1, 256] - - [73, 55.654] - - - [16640, 1024, 1, 256] - - [61, 63.0] - - - [11824, 8960, 1, 256] - - [54, 64.102] - - - [7936, 1280, 1, 256] - - [35, 62.118] - - - [6960, 3840, 1, 256] - - [30, 62.457] - - - [3328, 2048, 1, 256] - - [36, 58.869] - - - [16944, 2865, 1, 256] - - [38, 63.601] - - - [1024, 256, 1, 256] - - [117, 15.829] - - - [16944, 3840, 1, 256] - - [54, 64.63] - - - [3376, 2816, 1, 256] - - [30, 53.634] - - - [12288, 768, 1, 256] - - [37, 57.894] - - - [17152, 3329, 1, 256] - - [56, 69.147] - - - [6192, 2865, 1, 256] - - [26, 56.627] - - - [5888, 1281, 1, 256] - - [36, 53.849] - - - [11824, 256, 1, 256] - - [28, 45.599] - - - [18688, 1280, 1, 256] - - [35, 67.925] - - - [11520, 7936, 1, 256] - - [60, 73.686] - - - [15616, 1281, 1, 256] - - [37, 61.056] - - - [16944, 10240, 1, 256] - - [26, 64.198] - - - [12032, 4352, 1, 256] - - [36, 71.886] - - - [9984, 6656, 1, 256] - - [36, 73.374] - - - [17408, 1281, 1, 256] - - [62, 62.295] - - - [6912, 3329, 1, 256] - - [36, 65.222] - - - [16176, 2865, 1, 256] - - [30, 63.843] - - - [7936, 4864, 1, 256] - - [32, 71.543] - - - [7168, 256, 1, 256] - - [52, 30.817] - - - [9728, 6144, 1, 256] - - [28, 72.968] - - - [10752, 7680, 1, 256] - - [26, 74.394] - - - [13056, 5632, 1, 256] - - [38, 73.854] - - - [17152, 2865, 1, 256] - - [28, 68.888] - - - [4096, 512, 1, 256] - - [60, 34.82] - - - [3584, 2304, 1, 256] - - [35, 59.247] - - - [11264, 2048, 1, 256] - - [59, 67.528] - - - [18944, 5376, 1, 256] - - [68, 74.399] - - - [8960, 3329, 1, 256] - - [56, 66.204] - - - [7936, 1281, 1, 256] - - [41, 54.833] - - - [12848, 2816, 1, 256] - - [37, 61.496] - - - [9472, 3328, 1, 256] - - [41, 70.075] - - - [2816, 2816, 1, 256] - - [56, 56.977] - - - [15616, 10240, 1, 256] - - [30, 74.822] - - - [2816, 256, 1, 256] - - [110, 33.257] - - - [48, 256, 1, 256] - - [114, 0.779] - - - [17408, 1792, 1, 256] - - [56, 70.156] - - - [10032, 2865, 1, 256] - - [28, 59.522] - - - [3584, 2865, 1, 256] - - [36, 62.625] - - - [9472, 2816, 1, 256] - - [36, 69.497] - - - [2096, 2048, 1, 256] - - [36, 44.187] - - - [9216, 1536, 1, 256] - - [50, 65.881] - - - [5936, 256, 1, 256] - - [183, 41.044] - - - [11520, 1280, 1, 256] - - [50, 63.231] - - - [16896, 3328, 1, 256] - - [38, 72.404] - - - [7984, 4864, 1, 256] - - [26, 64.026] - - - [11008, 1280, 1, 256] - - [32, 63.427] - - - [18432, 6144, 1, 256] - - [26, 74.435] - - - [2096, 1841, 1, 256] - - [35, 40.144] - - - [8448, 1024, 1, 256] - - [24, 59.313] - - - [17968, 10240, 1, 256] - - [37, 63.221] - - - [1536, 1536, 1, 256] - - [35, 38.782] - - - [7728, 4864, 1, 256] - - [26, 61.971] - - - [18944, 3328, 1, 256] - - [32, 72.475] - - - [4608, 1792, 1, 256] - - [50, 58.808] - - - [8960, 6144, 1, 256] - - [36, 72.367] - - - [18736, 2816, 1, 256] - - [45, 62.4] - - - [8704, 5120, 1, 256] - - [36, 72.429] - - - [19456, 6144, 1, 256] - - [26, 74.371] - - - [19456, 1281, 1, 256] - - [62, 62.089] - - - [17200, 3840, 1, 256] - - [28, 64.105] - - - [2352, 2353, 1, 256] - - [37, 46.546] - - - [17408, 2816, 1, 256] - - [56, 72.23] - - - [13312, 2816, 1, 256] - - [35, 71.629] - - - [8960, 2816, 1, 256] - - [50, 70.027] - - - [2048, 1792, 1, 256] - - [50, 41.156] - - - [17152, 10240, 1, 256] - - [58, 74.66] - - - [16176, 10240, 1, 256] - - [28, 65.472] - - - [10288, 2865, 1, 256] - - [30, 60.769] - - - [8704, 2816, 1, 256] - - [35, 69.527] - - - [7424, 4096, 1, 256] - - [28, 70.152] - - - [6656, 1024, 1, 256] - - [28, 58.781] - - - [2304, 256, 1, 256] - - [109, 28.11] - - - [16384, 2865, 1, 256] - - [23, 59.821] - - - [7680, 2816, 1, 256] - - [26, 69.69] - - - [11520, 3329, 1, 256] - - [32, 66.836] - - - [10752, 1280, 1, 256] - - [35, 64.9] - - - [3120, 2816, 1, 256] - - [35, 56.127] - - - [15872, 1281, 1, 256] - - [24, 62.2] - - - [13824, 6144, 1, 256] - - [30, 73.534] - - - [6912, 3584, 1, 256] - - [35, 69.944] - - - [12032, 3328, 1, 256] - - [56, 70.445] - - - [11264, 1281, 1, 256] - - [54, 60.24] - - - [19456, 5632, 1, 256] - - [26, 74.96] - - - [17200, 2816, 1, 256] - - [59, 63.832] - - - [11520, 3840, 1, 256] - - [68, 71.601] - - - [11520, 2865, 1, 256] - - [56, 68.192] - - - [14848, 1280, 1, 256] - - [36, 66.953] - - - [16176, 256, 1, 256] - - [61, 43.839] - - - [16384, 256, 1, 256] - - [41, 45.395] - - - [4096, 768, 1, 256] - - [58, 49.726] - - - [4864, 2816, 1, 256] - - [35, 65.058] - - - [13568, 256, 1, 256] - - [41, 53.158] - - - [4608, 2048, 1, 256] - - [50, 58.79] - - - [9984, 6144, 1, 256] - - [36, 72.711] - - - [3632, 768, 1, 256] - - [45, 42.625] - - - [19200, 5888, 1, 256] - - [35, 73.564] - - - [5632, 2865, 1, 256] - - [36, 63.299] - - - [15360, 1280, 1, 256] - - [50, 67.381] - - - [12800, 1280, 1, 256] - - [35, 64.295] - - - [7168, 3328, 1, 256] - - [50, 68.647] - - - [11264, 8448, 1, 256] - - [38, 74.919] - - - [18176, 3328, 1, 256] - - [50, 72.027] - - - [4096, 2560, 1, 256] - - [56, 63.739] - - - [12544, 768, 1, 256] - - [54, 59.132] - - - [11568, 8448, 1, 256] - - [45, 63.978] - - - [8704, 1280, 1, 256] - - [35, 60.386] - - - [13056, 1536, 1, 256] - - [26, 65.618] - - - [2304, 1024, 1, 256] - - [58, 38.726] - - - [3072, 1281, 1, 256] - - [36, 43.281] - - - [6912, 1280, 1, 256] - - [36, 61.687] - - - [9216, 2816, 1, 256] - - [36, 69.713] - - - [17152, 6144, 1, 256] - - [28, 73.807] - - - [18992, 2865, 1, 256] - - [26, 60.345] - - - [10240, 2560, 1, 256] - - [26, 70.12] - - - [560, 256, 1, 256] - - [116, 8.889] - - - [2304, 1280, 1, 256] - - [36, 46.866] - - - [7680, 6144, 1, 256] - - [35, 72.08] - - - [15920, 2560, 1, 256] - - [38, 64.2] - - - [17456, 10240, 1, 256] - - [54, 63.79] - - - [14080, 3328, 1, 256] - - [58, 70.137] - - - [13360, 10240, 1, 256] - - [28, 64.032] - - - [8448, 5632, 1, 256] - - [50, 72.533] - - - [17408, 3584, 1, 256] - - [38, 73.531] - - - [6704, 2865, 1, 256] - - [26, 57.23] - - - [12592, 9472, 1, 256] - - [37, 64.017] - - - [18992, 10240, 1, 256] - - [45, 63.555] - - - [5376, 2865, 1, 256] - - [56, 63.726] - - - [18480, 5120, 1, 256] - - [38, 62.77] - - - [14336, 2048, 1, 256] - - [62, 68.004] - - - [7424, 3328, 1, 256] - - [41, 69.576] - - - [256, 49, 1, 256] - - [116, 0.778] - - - [12288, 1280, 1, 256] - - [28, 65.855] - - - [13568, 3329, 1, 256] - - [26, 67.986] - - - [15360, 1792, 1, 256] - - [50, 69.304] - - - [7168, 3584, 1, 256] - - [28, 69.608] - - - [10240, 3328, 1, 256] - - [26, 71.06] - - - [6400, 1281, 1, 256] - - [26, 57.395] - - - [11008, 6144, 1, 256] - - [76, 72.223] - - - [512, 513, 1, 256] - - [117, 15.777] - - - [19248, 256, 1, 256] - - [64, 48.424] - - - [2608, 256, 1, 256] - - [115, 30.8] - - - [16688, 3584, 1, 256] - - [30, 65.642] - - - [17920, 4864, 1, 256] - - [32, 74.051] - - - [18688, 1281, 1, 256] - - [41, 62.826] - - - [18224, 4864, 1, 256] - - [37, 63.328] - - - [10496, 2816, 1, 256] - - [64, 69.463] - - - [12288, 4864, 1, 256] - - [38, 73.366] - - - [9216, 2865, 1, 256] - - [28, 66.685] - - - [17664, 3329, 1, 256] - - [28, 68.974] - - - [3632, 512, 1, 256] - - [36, 29.845] - - - [11776, 3329, 1, 256] - - [58, 68.167] - - - [19456, 1792, 1, 256] - - [56, 70.657] - - - [12592, 256, 1, 256] - - [106, 47.498] - - - [10752, 3072, 1, 256] - - [28, 70.504] - - - [10800, 2816, 1, 256] - - [61, 63.108] - - - [6192, 3072, 1, 256] - - [26, 60.601] - - - [17152, 1536, 1, 256] - - [56, 68.551] - - - [2096, 2097, 1, 256] - - [26, 44.726] - - - [8192, 4608, 1, 256] - - [26, 70.244] - - - [13056, 6144, 1, 256] - - [28, 73.792] - - - [16640, 10240, 1, 256] - - [26, 74.759] - - - [12592, 9728, 1, 256] - - [59, 63.963] - - - [18176, 2816, 1, 256] - - [28, 71.857] - - - [18176, 4864, 1, 256] - - [26, 73.781] - - - [18944, 6144, 1, 256] - - [25, 74.1] - - - [12544, 256, 1, 256] - - [28, 50.485] - - - [13056, 1281, 1, 256] - - [64, 59.592] - - - [304, 49, 1, 256] - - [167, 0.914] - - - [17920, 2816, 1, 256] - - [65, 71.961] - - - [4656, 256, 1, 256] - - [115, 39.388] - - - [7728, 2865, 1, 256] - - [28, 59.378] - - - [15872, 1280, 1, 256] - - [58, 66.943] - - - [17456, 256, 1, 256] - - [35, 45.992] - - - [18176, 4608, 1, 256] - - [50, 73.101] - - - [7168, 5632, 1, 256] - - [50, 72.334] - - - [13616, 256, 1, 256] - - [37, 50.658] - - - [15104, 3329, 1, 256] - - [38, 68.888] - - - [19712, 3328, 1, 256] - - [68, 70.314] - - - [10032, 7168, 1, 256] - - [54, 63.028] - - - [11008, 3328, 1, 256] - - [76, 69.176] - - - [10496, 6144, 1, 256] - - [36, 72.906] - - - [6656, 2865, 1, 256] - - [60, 65.529] - - - [17664, 3840, 1, 256] - - [60, 72.723] - - - [6960, 4096, 1, 256] - - [45, 62.828] - - - [4608, 256, 1, 256] - - [176, 39.897] - - - [10496, 6912, 1, 256] - - [35, 73.557] - - - [16128, 2560, 1, 256] - - [60, 71.226] - - - [15872, 256, 1, 256] - - [35, 44.267] - - - [6656, 1281, 1, 256] - - [93, 53.011] - - - [3584, 512, 1, 256] - - [60, 31.134] - - - [11520, 1792, 1, 256] - - [35, 67.255] - - - [11264, 8192, 1, 256] - - [38, 74.546] - - - [10752, 2048, 1, 256] - - [41, 67.186] - - - [18688, 3329, 1, 256] - - [28, 69.295] - - - [4352, 768, 1, 256] - - [24, 51.77] - - - [18432, 512, 1, 256] - - [28, 58.069] - - - [18992, 256, 1, 256] - - [41, 47.829] - - - [13568, 2865, 1, 256] - - [26, 68.418] - - - [14640, 256, 1, 256] - - [50, 39.97] - - - [11264, 256, 1, 256] - - [35, 46.464] - - - [16896, 3329, 1, 256] - - [26, 69.544] - - - [18944, 5120, 1, 256] - - [48, 74.02] - - - [768, 513, 1, 256] - - [109, 22.833] - - - [14080, 10240, 1, 256] - - [25, 73.542] - - - [15872, 2560, 1, 256] - - [50, 71.856] - - - [6912, 5632, 1, 256] - - [50, 71.919] - - - [13360, 2865, 1, 256] - - [26, 61.513] - - - [6400, 3072, 1, 256] - - [36, 68.625] - - - [15616, 3329, 1, 256] - - [28, 69.058] - - - [9264, 2816, 1, 256] - - [38, 61.26] - - - [18176, 512, 1, 256] - - [41, 57.477] - - - [11264, 1280, 1, 256] - - [36, 66.843] - - - [1328, 1329, 1, 256] - - [50, 29.006] - - - [18736, 5376, 1, 256] - - [59, 62.844] - - - [5376, 1792, 1, 256] - - [36, 59.533] - - - [6144, 4608, 1, 256] - - [26, 70.913] - - - [6400, 3328, 1, 256] - - [35, 69.436] - - - [12032, 2865, 1, 256] - - [35, 68.265] - - - [12288, 4608, 1, 256] - - [28, 72.894] - - - [16128, 3072, 1, 256] - - [60, 71.509] - - - [2048, 256, 1, 256] - - [124, 25.3] - - - [4096, 256, 1, 256] - - [187, 36.316] - - - [5888, 2304, 1, 256] - - [36, 64.782] - - - [2816, 2561, 1, 256] - - [35, 52.938] - - - [3072, 1536, 1, 256] - - [36, 50.641] - - - [2304, 1281, 1, 256] - - [28, 46.337] - - - [15616, 2048, 1, 256] - - [64, 69.79] - - - [12800, 1024, 1, 256] - - [24, 62.114] - - - [8960, 3328, 1, 256] - - [36, 69.542] - - - [18432, 1280, 1, 256] - - [28, 67.656] - - - [8448, 2048, 1, 256] - - [41, 65.944] - - - [19712, 6400, 1, 256] - - [25, 73.172] - - - [14384, 1280, 1, 256] - - [30, 58.663] - - - [6448, 2816, 1, 256] - - [37, 58.008] - - - [18176, 2048, 1, 256] - - [24, 70.358] - - - [3072, 1792, 1, 256] - - [28, 49.282] - - - [12080, 8960, 1, 256] - - [37, 63.832] - - - [13312, 1281, 1, 256] - - [26, 60.158] - - - [16688, 2816, 1, 256] - - [28, 64.939] - - - [6400, 256, 1, 256] - - [184, 46.542] - - - [2048, 2048, 1, 256] - - [36, 46.271] - - - [14336, 256, 1, 256] - - [79, 41.029] - - - [11520, 2816, 1, 256] - - [68, 70.272] - - - [14384, 10240, 1, 256] - - [28, 63.156] - - - [7472, 256, 1, 256] - - [45, 31.265] - - - [1280, 1280, 1, 256] - - [122, 47.358] - - - [8704, 1024, 1, 256] - - [24, 61.245] - - - [9520, 2865, 1, 256] - - [45, 60.006] - - - [18480, 256, 1, 256] - - [61, 47.093] - - - [18176, 256, 1, 256] - - [41, 49.441] - - - [15872, 6144, 1, 256] - - [28, 74.514] - - - [304, 256, 1, 256] - - [117, 4.8] - - - [13568, 5888, 1, 256] - - [50, 73.35] - - - [3328, 3328, 1, 256] - - [35, 59.627] - - - [6656, 5120, 1, 256] - - [50, 71.447] - - - [9520, 2816, 1, 256] - - [59, 61.723] - - - [1536, 1537, 1, 256] - - [35, 38.752] - - - [3072, 2865, 1, 256] - - [81, 55.246] - - - [10032, 2816, 1, 256] - - [54, 62.235] - - - [12032, 9216, 1, 256] - - [55, 73.508] - - - [13872, 10240, 1, 256] - - [45, 63.139] - - - [13824, 2048, 1, 256] - - [45, 69.586] - - - [12544, 9728, 1, 256] - - [38, 74.151] - - - [17664, 4352, 1, 256] - - [58, 73.069] - - - [4096, 1281, 1, 256] - - [50, 47.228] - - - [17408, 1280, 1, 256] - - [50, 68.046] - - - [18432, 2816, 1, 256] - - [56, 72.467] - - - [5120, 256, 1, 256] - - [190, 41.991] - - - [18736, 2865, 1, 256] - - [59, 60.889] - - - [19200, 256, 1, 256] - - [35, 51.609] - - - [2048, 512, 1, 256] - - [175, 36.208] - - - [11008, 7680, 1, 256] - - [72, 73.369] - - - [5888, 3072, 1, 256] - - [36, 65.822] - - - [11776, 8192, 1, 256] - - [28, 74.076] - - - [5888, 512, 1, 256] - - [28, 48.316] - - - [7936, 2816, 1, 256] - - [56, 67.886] - - - [5120, 2865, 1, 256] - - [50, 62.345] - - - [12032, 2816, 1, 256] - - [50, 70.599] - - - [256, 257, 1, 256] - - [180, 4.192] - - - [13104, 2865, 1, 256] - - [45, 60.974] - - - [5680, 2865, 1, 256] - - [38, 57.715] - - - [15408, 10240, 1, 256] - - [30, 63.988] - - - [18432, 4864, 1, 256] - - [28, 74.451] - - - [17712, 2865, 1, 256] - - [28, 60.28] - - - [768, 256, 1, 256] - - [116, 12.668] - - - [9728, 3328, 1, 256] - - [30, 71.105] - - - [12848, 9728, 1, 256] - - [54, 64.542] - - - [2304, 2305, 1, 256] - - [58, 46.736] - - - [10240, 6144, 1, 256] - - [30, 73.496] - - - [13312, 1280, 1, 256] - - [50, 66.37] - - - [9008, 5888, 1, 256] - - [37, 63.331] - - - [7424, 3840, 1, 256] - - [50, 70.87] - - - [12032, 1280, 1, 256] - - [50, 65.217] - - - [18480, 2816, 1, 256] - - [30, 62.353] - - - [18432, 5120, 1, 256] - - [26, 74.354] - - - [7424, 4608, 1, 256] - - [24, 70.809] - - - [9776, 2865, 1, 256] - - [37, 58.455] - - - [5632, 2560, 1, 256] - - [35, 67.482] - - - [7680, 2048, 1, 256] - - [41, 65.594] - - - [6704, 2816, 1, 256] - - [59, 59.919] - - - [13872, 2816, 1, 256] - - [37, 63.15] - - - [17968, 2816, 1, 256] - - [45, 62.628] - - - [4144, 2865, 1, 256] - - [30, 58.002] - - - [14640, 1280, 1, 256] - - [28, 58.845] - - - [16432, 2816, 1, 256] - - [37, 67.138] - - - [16128, 1280, 1, 256] - - [35, 66.94] - - - [8240, 5120, 1, 256] - - [28, 65.285] - - - [4352, 2816, 1, 256] - - [36, 64.474] - - - [12544, 2865, 1, 256] - - [28, 68.603] - - - [6144, 2048, 1, 256] - - [64, 63.394] - - - [13616, 512, 1, 256] - - [41, 55.1] - - - [5632, 2048, 1, 256] - - [28, 60.737] - - - [13312, 2048, 1, 256] - - [24, 67.86] - - - [9728, 1281, 1, 256] - - [28, 59.385] - - - [7424, 1281, 1, 256] - - [36, 58.342] - - - [10800, 256, 1, 256] - - [37, 41.971] - - - [2048, 1281, 1, 256] - - [58, 42.578] - - - [5376, 1280, 1, 256] - - [35, 59.027] - - - [15664, 2816, 1, 256] - - [37, 63.151] - - - [256, 256, 1, 256] - - [167, 4.319] - - - [2048, 1280, 1, 256] - - [58, 42.717] - - - [9776, 256, 1, 256] - - [54, 39.723] - - - [4096, 3329, 1, 256] - - [60, 63.485] - - - [9728, 2304, 1, 256] - - [56, 68.567] - - - [19968, 2865, 1, 256] - - [26, 70.059] - - - [13568, 6144, 1, 256] - - [30, 73.621] - - - [15360, 2304, 1, 256] - - [56, 71.231] - - - [9264, 6400, 1, 256] - - [28, 63.431] - - - [19200, 2048, 1, 256] - - [24, 69.891] - - - [11520, 4096, 1, 256] - - [41, 70.668] - - - [18688, 5632, 1, 256] - - [26, 74.26] - - - [11776, 256, 1, 256] - - [37, 48.248] - - - [17152, 256, 1, 256] - - [41, 47.478] - - - [5120, 1280, 1, 256] - - [35, 57.075] - - - [14896, 1792, 1, 256] - - [45, 61.582] - - - [10288, 2816, 1, 256] - - [30, 60.366] - - - [7984, 2865, 1, 256] - - [38, 58.999] - - - [4864, 1281, 1, 256] - - [36, 54.304] - - - [7216, 256, 1, 256] - - [41, 30.145] - - - [5888, 3328, 1, 256] - - [36, 68.189] - - - [7424, 2816, 1, 256] - - [35, 68.231] - - - [15360, 3328, 1, 256] - - [36, 72.12] - - - [10544, 256, 1, 256] - - [35, 41.412] - - - [9776, 2816, 1, 256] - - [54, 61.887] - - - [8240, 2816, 1, 256] - - [28, 63.984] - - - [6656, 3072, 1, 256] - - [35, 66.992] - - - [18224, 10240, 1, 256] - - [54, 63.963] - - - [13824, 2865, 1, 256] - - [36, 69.183] - - - [5376, 1281, 1, 256] - - [36, 50.065] - - - [13568, 9984, 1, 256] - - [38, 74.874] - - - [18176, 4352, 1, 256] - - [38, 73.61] - - - [11776, 1281, 1, 256] - - [45, 58.768] - - - [15616, 6144, 1, 256] - - [38, 73.791] - - - [4400, 256, 1, 256] - - [124, 37.764] - - - [18992, 2816, 1, 256] - - [37, 63.728] - - - [14640, 10240, 1, 256] - - [59, 63.516] - - - [5120, 2048, 1, 256] - - [28, 61.987] - - - [19968, 10240, 1, 256] - - [28, 75.102] - - - [19200, 2865, 1, 256] - - [60, 69.04] - - - [15152, 2816, 1, 256] - - [54, 63.055] - - - [2560, 2560, 1, 256] - - [35, 57.421] - - - [8448, 2816, 1, 256] - - [56, 67.887] - - - [8704, 5632, 1, 256] - - [56, 72.624] - - - [1024, 769, 1, 256] - - [124, 35.338] - - - [17200, 4096, 1, 256] - - [45, 63.61] - - - [5376, 256, 1, 256] - - [184, 40.594] - - - [6656, 256, 1, 256] - - [183, 47.239] - - - [18688, 3328, 1, 256] - - [36, 71.918] - - - [13056, 256, 1, 256] - - [28, 51.76] - - - [13104, 2816, 1, 256] - - [37, 63.036] - - - [7424, 1792, 1, 256] - - [35, 63.557] - - - [14592, 2816, 1, 256] - - [60, 70.771] - - - [12336, 2865, 1, 256] - - [38, 59.727] - - - [17920, 256, 1, 256] - - [64, 49.106] - - - [12800, 2048, 1, 256] - - [37, 68.659] - - - [3632, 256, 1, 256] - - [122, 32.495] - - - [18688, 768, 1, 256] - - [54, 66.3] - - - [16384, 2816, 1, 256] - - [23, 62.061] - - - [14896, 10240, 1, 256] - - [26, 63.546] - - - [816, 817, 1, 256] - - [122, 30.034] - - - [9008, 2865, 1, 256] - - [28, 59.73] - - - [14848, 1024, 1, 256] - - [61, 64.645] - - - [16640, 256, 1, 256] - - [40, 45.255] - - - [7424, 256, 1, 256] - - [32, 31.555] - - - [10240, 3329, 1, 256] - - [30, 68.28] - - - [18176, 5120, 1, 256] - - [28, 73.841] - - - [6912, 2865, 1, 256] - - [26, 64.594] - - - [1024, 1025, 1, 256] - - [186, 35.087] - - - [5632, 4096, 1, 256] - - [41, 68.981] - - - [12544, 1024, 1, 256] - - [64, 61.314] - - - [2864, 2609, 1, 256] - - [28, 49.44] - - - [16896, 2048, 1, 256] - - [61, 69.794] - - - [3840, 1280, 1, 256] - - [50, 52.859] - - - [11008, 1281, 1, 256] - - [32, 56.944] - - - [15104, 1281, 1, 256] - - [24, 62.75] - - - [7168, 3329, 1, 256] - - [30, 66.789] - - - [12800, 5120, 1, 256] - - [60, 73.081] - - - [512, 257, 1, 256] - - [167, 8.338] - - - [12288, 8960, 1, 256] - - [28, 75.243] - - - [9728, 6912, 1, 256] - - [50, 74.008] - - - [9728, 6656, 1, 256] - - [35, 73.905] - - - [2560, 2304, 1, 256] - - [56, 52.115] - - - [10544, 7424, 1, 256] - - [28, 63.368] - - - [5888, 3329, 1, 256] - - [26, 64.312] - - - [3888, 2816, 1, 256] - - [28, 53.57] - - - [18944, 10240, 1, 256] - - [48, 75.238] - - - [17200, 10240, 1, 256] - - [30, 63.374] - - - [4144, 1280, 1, 256] - - [36, 53.132] - - - [9728, 1280, 1, 256] - - [35, 65.329] - - - [14896, 1536, 1, 256] - - [26, 60.265] - - - [5888, 4352, 1, 256] - - [35, 69.203] - - - [1024, 1024, 1, 256] - - [173, 36.317] - - - [4912, 2816, 1, 256] - - [38, 59.825] - - - [19456, 3329, 1, 256] - - [28, 69.935] - - - [7680, 4608, 1, 256] - - [56, 71.244] - - - [8496, 2865, 1, 256] - - [26, 59.372] - - - [3584, 2048, 1, 256] - - [35, 53.777] - - - [9984, 3329, 1, 256] - - [56, 66.875] - - - [10800, 7680, 1, 256] - - [45, 64.193] - - - [13616, 2816, 1, 256] - - [59, 62.517] - - - [15104, 10240, 1, 256] - - [28, 74.662] - - - [10240, 6656, 1, 256] - - [30, 74.233] - - - [16128, 1281, 1, 256] - - [70, 60.91] - - - [16896, 1280, 1, 256] - - [56, 68.869] - - - [12544, 9472, 1, 256] - - [26, 74.527] - - - [11008, 7424, 1, 256] - - [86, 73.158] - - - [9472, 3329, 1, 256] - - [38, 66.568] - - - [6912, 2816, 1, 256] - - [36, 67.987] - - - [2048, 1841, 1, 256] - - [41, 41.594] - - - [17152, 4096, 1, 256] - - [55, 72.113] - - - [12544, 5120, 1, 256] - - [50, 72.708] - - - [13824, 3328, 1, 256] - - [50, 71.534] - - - [6912, 2048, 1, 256] - - [41, 65.866] - - - [9472, 256, 1, 256] - - [58, 39.795] - - - [9216, 1281, 1, 256] - - [37, 56.597] - - - [7168, 1281, 1, 256] - - [50, 55.992] - - - [10752, 7424, 1, 256] - - [30, 74.342] - - - [16176, 3072, 1, 256] - - [38, 65.657] - - - [12288, 9216, 1, 256] - - [30, 74.52] - - - [14336, 512, 1, 256] - - [28, 52.927] - - - [14336, 3328, 1, 256] - - [26, 72.042] - - - [4864, 1280, 1, 256] - - [35, 54.748] - - - [19760, 2865, 1, 256] - - [37, 60.832] - - - [8240, 256, 1, 256] - - [35, 34.619] - - - [18688, 1024, 1, 256] - - [41, 66.573] - - - [16128, 10240, 1, 256] - - [28, 74.965] - - - [5632, 256, 1, 256] - - [184, 42.857] - - - [5680, 2560, 1, 256] - - [38, 57.404] - - - [7680, 1281, 1, 256] - - [28, 59.333] - - - [17408, 2048, 1, 256] - - [59, 69.586] - - - [10752, 2865, 1, 256] - - [26, 68.123] - - - [14848, 1281, 1, 256] - - [54, 62.3] - - - [560, 512, 1, 256] - - [110, 17.046] - - - [19968, 1280, 1, 256] - - [36, 68.465] - - - [16384, 10240, 1, 256] - - [23, 63.0] - - - [512, 305, 1, 256] - - [171, 9.788] - - - [19200, 6144, 1, 256] - - [30, 73.718] - - - [8448, 5120, 1, 256] - - [26, 72.211] - - - [13824, 3329, 1, 256] - - [30, 68.838] - - - [7984, 2816, 1, 256] - - [28, 61.507] - - - [17920, 3329, 1, 256] - - [52, 69.263] - - - [16688, 2865, 1, 256] - - [50, 63.796] - - - [12032, 256, 1, 256] - - [26, 49.22] - - - [7424, 2865, 1, 256] - - [28, 65.069] - - - [14336, 10240, 1, 256] - - [26, 75.425] - - - [17152, 2048, 1, 256] - - [64, 69.707] - - - [14896, 2816, 1, 256] - - [45, 62.421] - - - [16384, 2048, 1, 256] - - [39, 59.349] - - - [8192, 2816, 1, 256] - - [36, 68.313] - - - [6192, 256, 1, 256] - - [183, 42.027] - - - [2304, 768, 1, 256] - - [125, 48.821] - - - [18688, 256, 1, 256] - - [36, 50.462] - - - [8960, 1281, 1, 256] - - [50, 59.542] - - - [19968, 6400, 1, 256] - - [35, 74.803] - - - [8752, 2816, 1, 256] - - [30, 62.219] - - - [19456, 3328, 1, 256] - - [50, 72.717] - - - [2560, 2561, 1, 256] - - [35, 57.087] - - - [15920, 2865, 1, 256] - - [28, 61.767] - - - [12544, 6144, 1, 256] - - [26, 73.28] - - - [19200, 3328, 1, 256] - - [54, 71.858] - - - [3328, 2865, 1, 256] - - [35, 58.513] - - - [7936, 3329, 1, 256] - - [58, 65.631] - - - [11264, 2865, 1, 256] - - [30, 68.379] - - - [6144, 3329, 1, 256] - - [50, 66.13] - - - [16128, 3329, 1, 256] - - [26, 68.765] - - - [12800, 9728, 1, 256] - - [75, 74.84] - - - [512, 256, 1, 256] - - [167, 8.445] - - - [11264, 2816, 1, 256] - - [56, 70.812] - - - [12544, 3329, 1, 256] - - [28, 67.735] - - - [14848, 3329, 1, 256] - - [28, 69.124] - - - [1328, 256, 1, 256] - - [109, 19.803] - - - [3120, 256, 1, 256] - - [124, 35.025] - - - [1024, 768, 1, 256] - - [122, 35.569] - - - [7728, 2816, 1, 256] - - [24, 59.059] - - - [1024, 817, 1, 256] - - [122, 36.274] - - - [10288, 7424, 1, 256] - - [26, 63.45] - - - [19968, 6144, 1, 256] - - [42, 74.294] - - - [13616, 10240, 1, 256] - - [37, 63.511] - - - [1536, 1329, 1, 256] - - [70, 33.699] - - - [9984, 3328, 1, 256] - - [56, 70.3] - - - [9472, 5888, 1, 256] - - [30, 72.702] - - - [11264, 7936, 1, 256] - - [56, 74.907] - - - [8496, 256, 1, 256] - - [59, 35.102] - - - [17664, 1792, 1, 256] - - [35, 69.801] - - - [11824, 2816, 1, 256] - - [45, 62.525] - - - [16944, 2816, 1, 256] - - [30, 64.078] - - - [19968, 6912, 1, 256] - - [28, 74.983] - - - [3376, 2865, 1, 256] - - [28, 54.171] - - - [3840, 2560, 1, 256] - - [50, 60.629] - - - [11776, 8448, 1, 256] - - [28, 74.414] - - - [19248, 6144, 1, 256] - - [59, 63.548] - - - [14080, 512, 1, 256] - - [28, 61.021] - - - [16128, 3328, 1, 256] - - [50, 71.602] - - - [6656, 2048, 1, 256] - - [24, 63.698] - - - [15664, 256, 1, 256] - - [56, 42.258] - - - [17664, 1280, 1, 256] - - [36, 67.773] - - - [16384, 6144, 1, 256] - - [23, 63.174] - - - [9984, 256, 1, 256] - - [52, 41.767] - - - [14592, 1281, 1, 256] - - [35, 60.083] - - - [4608, 3329, 1, 256] - - [50, 64.498] - - - [8960, 2048, 1, 256] - - [35, 64.636] - - - [2864, 2865, 1, 256] - - [28, 53.819] - - - [2816, 2609, 1, 256] - - [35, 53.455] - - - [14080, 1281, 1, 256] - - [26, 60.65] - - - [1792, 1536, 1, 256] - - [32, 44.237] - - - [10240, 7424, 1, 256] - - [38, 74.721] - - - [5936, 2816, 1, 256] - - [38, 58.611] - - - [19712, 256, 1, 256] - - [41, 52.927] - - - [18944, 5888, 1, 256] - - [32, 74.241] - - - [9728, 3329, 1, 256] - - [28, 68.007] - - - [19248, 2816, 1, 256] - - [54, 62.64] - - - [13568, 1792, 1, 256] - - [50, 68.552] - - - [1584, 1585, 1, 256] - - [30, 38.322] - - - [8704, 2048, 1, 256] - - [41, 67.585] - - - [13056, 9728, 1, 256] - - [26, 74.687] - - - [12800, 2865, 1, 256] - - [32, 68.74] - - - [14336, 6144, 1, 256] - - [28, 74.178] - - - [5120, 1536, 1, 256] - - [36, 56.426] - - - [18432, 1281, 1, 256] - - [26, 61.726] - - - [10240, 256, 1, 256] - - [35, 42.778] - - - [12544, 9216, 1, 256] - - [38, 73.968] - - - [12800, 1281, 1, 256] - - [34, 61.721] - - - [8704, 5888, 1, 256] - - [26, 72.741] - - - [15360, 3329, 1, 256] - - [28, 69.087] - - - [11312, 8448, 1, 256] - - [26, 63.968] - - - [17152, 3328, 1, 256] - - [26, 71.81] - - - [16384, 3328, 1, 256] - - [39, 61.752] - - - [13824, 2816, 1, 256] - - [50, 71.392] - - - [560, 305, 1, 256] - - [114, 10.423] - - - [16432, 256, 1, 256] - - [24, 44.904] - - - [3632, 2865, 1, 256] - - [38, 51.217] - - - [3584, 3328, 1, 256] - - [36, 64.076] - - - [3840, 768, 1, 256] - - [59, 46.673] - - - [19504, 256, 1, 256] - - [61, 48.745] - - - [1280, 1073, 1, 256] - - [191, 40.001] - - - [17712, 10240, 1, 256] - - [54, 63.117] - - - [2816, 1536, 1, 256] - - [36, 47.445] - - - [12800, 6144, 1, 256] - - [60, 73.468] - - - [4656, 2816, 1, 256] - - [37, 57.374] - - - [17920, 10240, 1, 256] - - [38, 75.114] - - - [9984, 2816, 1, 256] - - [36, 69.679] - - - [4352, 256, 1, 256] - - [124, 38.47] - - - [11312, 2865, 1, 256] - - [26, 60.631] - - - [18432, 3328, 1, 256] - - [30, 72.629] - - - [4096, 1280, 1, 256] - - [107, 54.347] - - - [4864, 3329, 1, 256] - - [36, 63.146] - - - [14640, 2865, 1, 256] - - [38, 59.354] - - - [17152, 2816, 1, 256] - - [35, 71.888] - - - [7680, 1280, 1, 256] - - [36, 60.456] - - - [1584, 1536, 1, 256] - - [35, 37.94] - - - [14080, 1280, 1, 256] - - [36, 67.641] - - - [13824, 512, 1, 256] - - [28, 60.187] - - - [7936, 256, 1, 256] - - [27, 33.726] - - - [12592, 2865, 1, 256] - - [38, 60.455] - - - [2816, 2560, 1, 256] - - [50, 62.208] - - - [6912, 1536, 1, 256] - - [56, 64.344] - - - [12800, 9984, 1, 256] - - [26, 75.144] - - - [10496, 256, 1, 256] - - [41, 43.732] - - - [18176, 2865, 1, 256] - - [28, 69.649] - - - [4608, 1536, 1, 256] - - [36, 60.86] - - - [3328, 2816, 1, 256] - - [50, 58.508] - - - [3840, 1024, 1, 256] - - [36, 43.255] - - - [13824, 1280, 1, 256] - - [35, 67.666] - - - [3840, 1281, 1, 256] - - [50, 52.315] - - - [17152, 1281, 1, 256] - - [41, 62.238] - - - [13568, 1281, 1, 256] - - [41, 61.146] - - - [14848, 1792, 1, 256] - - [50, 70.07] - - - [13056, 9472, 1, 256] - - [26, 74.981] - - - [18176, 1281, 1, 256] - - [45, 62.024] - - - [5680, 256, 1, 256] - - [183, 39.74] - - - [13056, 2816, 1, 256] - - [36, 71.111] - - - [11824, 8704, 1, 256] - - [59, 64.185] - - - [7936, 4352, 1, 256] - - [58, 70.426] - - - [8704, 256, 1, 256] - - [60, 37.042] - - - [5424, 2304, 1, 256] - - [54, 56.662] - - - [14128, 768, 1, 256] - - [59, 52.887] - - - [10752, 1024, 1, 256] - - [41, 58.928] - - - [9264, 6144, 1, 256] - - [28, 63.829] - - - [4352, 3328, 1, 256] - - [36, 63.523] - - - [18944, 5632, 1, 256] - - [68, 74.526] - - - [12032, 8704, 1, 256] - - [50, 74.067] - - - [2048, 2049, 1, 256] - - [36, 45.631] - - - [6400, 3329, 1, 256] - - [36, 64.596] - - - [15616, 2560, 1, 256] - - [36, 71.001] - - - [7472, 2865, 1, 256] - - [30, 57.811] - - - [14848, 1536, 1, 256] - - [36, 68.167] - - - [18736, 10240, 1, 256] - - [59, 63.377] - - - [6400, 1024, 1, 256] - - [36, 57.281] - - - [7936, 5120, 1, 256] - - [38, 71.824] - - - [4656, 1536, 1, 256] - - [36, 47.999] - - - [3328, 256, 1, 256] - - [115, 37.36] - - - [3072, 1280, 1, 256] - - [36, 43.671] - - - [2864, 2816, 1, 256] - - [36, 53.201] - - - [9472, 6144, 1, 256] - - [56, 72.458] - - - [3840, 2304, 1, 256] - - [26, 62.014] - - - [17408, 2865, 1, 256] - - [26, 69.641] - - - [16384, 2560, 1, 256] - - [39, 60.625] - - - [16384, 3329, 1, 256] - - [30, 58.92] - - - [16688, 10240, 1, 256] - - [26, 65.624] - - - [18688, 2048, 1, 256] - - [64, 69.803] - - - [7936, 4608, 1, 256] - - [26, 70.813] - - - [9472, 6400, 1, 256] - - [35, 73.356] - - - [14336, 3329, 1, 256] - - [30, 69.103] - - - [4608, 1024, 1, 256] - - [56, 50.887] - - - [16896, 6144, 1, 256] - - [28, 74.464] - - - [10752, 3329, 1, 256] - - [28, 68.157] - - - [6704, 256, 1, 256] - - [182, 44.38] - - - [17408, 6144, 1, 256] - - [26, 74.582] - - - [9984, 2048, 1, 256] - - [64, 66.727] - - - [17968, 4864, 1, 256] - - [37, 63.918] - - - [5120, 3584, 1, 256] - - [35, 66.629] - - - [14336, 2865, 1, 256] - - [38, 69.431] - - - [18736, 256, 1, 256] - - [54, 47.713] - - - [13568, 2048, 1, 256] - - [41, 68.59] - - - [17456, 4352, 1, 256] - - [26, 64.048] - - - [5424, 2816, 1, 256] - - [54, 59.939] - - - [17664, 4608, 1, 256] - - [60, 72.479] - - - [7984, 256, 1, 256] - - [45, 33.496] - - - [6400, 3584, 1, 256] - - [50, 69.343] - - - [19712, 3329, 1, 256] - - [48, 67.597] - - - [10752, 2816, 1, 256] - - [60, 70.312] - - - [17152, 1280, 1, 256] - - [36, 67.098] - - - [560, 561, 1, 256] - - [115, 18.392] - - - [8192, 1281, 1, 256] - - [28, 55.243] - - - [4864, 1792, 1, 256] - - [35, 61.43] - - - [3632, 2816, 1, 256] - - [45, 57.042] - - - [11520, 8448, 1, 256] - - [30, 73.729] - - - [5168, 2865, 1, 256] - - [28, 57.066] - - - [13568, 10240, 1, 256] - - [30, 74.74] - - - [12544, 2816, 1, 256] - - [28, 70.656] - - - [19504, 6144, 1, 256] - - [30, 63.873] - - - [11776, 2048, 1, 256] - - [40, 67.352] - - - [18688, 2865, 1, 256] - - [35, 69.555] - - - [14336, 768, 1, 256] - - [59, 59.929] - - - [18688, 6144, 1, 256] - - [30, 74.125] - - - [11776, 2816, 1, 256] - - [32, 70.084] - - - [12288, 9472, 1, 256] - - [38, 75.314] - - - [5120, 1792, 1, 256] - - [28, 57.697] - - - [16128, 512, 1, 256] - - [50, 56.775] - - - [5376, 3329, 1, 256] - - [36, 63.193] - - - [9216, 3329, 1, 256] - - [35, 67.666] - - - [9008, 2816, 1, 256] - - [26, 59.908] - - - [6448, 3328, 1, 256] - - [28, 58.765] - - - [19968, 3329, 1, 256] - - [28, 69.474] - - - [11520, 8704, 1, 256] - - [68, 74.005] - - - [13824, 256, 1, 256] - - [36, 54.449] - - - [1584, 256, 1, 256] - - [117, 23.152] - - - [10496, 7168, 1, 256] - - [38, 72.807] - - - [5376, 2304, 1, 256] - - [35, 64.441] - - - [10752, 7168, 1, 256] - - [28, 73.268] - - - [18432, 2048, 1, 256] - - [79, 68.684] - - - [12080, 256, 1, 256] - - [37, 46.156] - - - [8704, 3328, 1, 256] - - [35, 69.305] - - - [4608, 1280, 1, 256] - - [35, 51.827] - - - [6192, 3328, 1, 256] - - [28, 60.553] - - - [8704, 3329, 1, 256] - - [56, 67.535] - - - [5424, 2560, 1, 256] - - [24, 61.133] - - - [11008, 2816, 1, 256] - - [86, 67.657] - - - [11776, 4352, 1, 256] - - [60, 72.393] - - - [11008, 1536, 1, 256] - - [58, 62.887] - - - [13312, 3328, 1, 256] - - [28, 71.901] - - - [7168, 4096, 1, 256] - - [28, 69.656] - - - [9216, 256, 1, 256] - - [41, 38.837] - - - [19504, 2865, 1, 256] - - [38, 61.604] - - - [5936, 2865, 1, 256] - - [26, 59.11] - - - [1840, 1792, 1, 256] - - [26, 47.791] - - - [19968, 2816, 1, 256] - - [50, 72.336] - - - [9008, 256, 1, 256] - - [59, 36.85] - - - [9728, 256, 1, 256] - - [26, 40.936] - - - [11056, 7936, 1, 256] - - [28, 64.522] - - - [7680, 3329, 1, 256] - - [38, 67.112] - - - [1792, 256, 1, 256] - - [115, 22.324] - - - [17664, 10240, 1, 256] - - [52, 74.207] - - - [11776, 2865, 1, 256] - - [32, 67.578] - - - [512, 512, 1, 256] - - [177, 15.996] - - - [16640, 768, 1, 256] - - [24, 59.572] - - - [4352, 2048, 1, 256] - - [26, 61.808] - - - [19504, 2816, 1, 256] - - [37, 63.411] - - - [12080, 2865, 1, 256] - - [28, 60.842] - - - [14080, 768, 1, 256] - - [61, 64.176] - - - [7936, 512, 1, 256] - - [41, 44.612] - - - [5376, 2560, 1, 256] - - [56, 64.677] - - - [5632, 3329, 1, 256] - - [28, 65.402] - - - [5120, 3840, 1, 256] - - [50, 68.426] - - - [6192, 2816, 1, 256] - - [28, 60.678] - - - [4608, 3072, 1, 256] - - [35, 67.182] - - - [19712, 6656, 1, 256] - - [60, 73.027] - - - [14896, 256, 1, 256] - - [35, 40.437] - - - [6400, 1280, 1, 256] - - [35, 58.678] - - - [12800, 9216, 1, 256] - - [75, 74.395] - - - [15616, 256, 1, 256] - - [24, 43.934] - - - [17920, 4608, 1, 256] - - [62, 73.278] - - - [7936, 2865, 1, 256] - - [32, 66.845] - - - [13312, 3329, 1, 256] - - [28, 68.915] - - - [5168, 2304, 1, 256] - - [26, 57.61] - - - [14128, 10240, 1, 256] - - [59, 63.75] - - - [3840, 2816, 1, 256] - - [35, 65.639] - - - [8960, 1536, 1, 256] - - [35, 64.968] - - - [3328, 3073, 1, 256] - - [28, 61.561] - - - [4096, 3328, 1, 256] - - [56, 64.747] - - - [14592, 2048, 1, 256] - - [40, 67.195] - - - [9728, 2048, 1, 256] - - [24, 65.56] - - - [13312, 5888, 1, 256] - - [56, 74.216] - - - [15616, 2304, 1, 256] - - [28, 71.067] - - - [19712, 2816, 1, 256] - - [52, 70.242] - - - [9216, 5888, 1, 256] - - [36, 73.129] - - - [7168, 4352, 1, 256] - - [28, 71.0] - - - [9520, 6400, 1, 256] - - [37, 62.601] - - - [13568, 3328, 1, 256] - - [35, 70.67] - - - [17408, 4352, 1, 256] - - [35, 74.236] - - - [11056, 2865, 1, 256] - - [30, 59.852] - - - [18480, 2865, 1, 256] - - [30, 60.402] - - - [13824, 768, 1, 256] - - [41, 63.874] - - - [17664, 6144, 1, 256] - - [68, 73.429] - - - [7216, 4352, 1, 256] - - [28, 62.523] - - - [14128, 2865, 1, 256] - - [38, 60.047] - - - [11520, 3328, 1, 256] - - [37, 69.856] - - - [18992, 5888, 1, 256] - - [45, 63.717] - - - [17408, 10240, 1, 256] - - [38, 75.574] - - - [15104, 6144, 1, 256] - - [28, 73.579] - - - [16640, 1280, 1, 256] - - [61, 65.558] - - - [13056, 2865, 1, 256] - - [28, 68.468] - - - [11776, 8960, 1, 256] - - [50, 74.797] - - - [11312, 2816, 1, 256] - - [28, 62.941] - - - [11264, 3328, 1, 256] - - [56, 71.228] - - - [8192, 512, 1, 256] - - [79, 44.765] - - - [14848, 6144, 1, 256] - - [38, 73.826] - - - [10496, 7680, 1, 256] - - [38, 73.763] - - - [2816, 2817, 1, 256] - - [36, 56.435] - - - [15104, 2865, 1, 256] - - [28, 69.417] - - - [18176, 3329, 1, 256] - - [26, 69.35] - - - [3328, 1792, 1, 256] - - [28, 52.483] - - - [6144, 3328, 1, 256] - - [36, 67.918] - - - [12288, 6144, 1, 256] - - [38, 73.86] - - - [8960, 5888, 1, 256] - - [58, 71.764] - - - [3584, 1280, 1, 256] - - [35, 50.073] - - - [7728, 4608, 1, 256] - - [54, 63.336] - - - [18176, 6144, 1, 256] - - [30, 74.114] - - - [16944, 256, 1, 256] - - [35, 45.256] - - - [3888, 768, 1, 256] - - [41, 45.092] - - - [8448, 3329, 1, 256] - - [30, 66.474] - - - [3072, 3073, 1, 256] - - [56, 57.596] - - - [4912, 256, 1, 256] - - [110, 40.173] - - - [5936, 3072, 1, 256] - - [28, 58.598] - - - [7168, 2865, 1, 256] - - [28, 66.573] - - - [19456, 10240, 1, 256] - - [26, 75.579] - - - [1840, 1585, 1, 256] - - [61, 43.249] - - - [18992, 5632, 1, 256] - - [45, 64.215] - - - [4912, 1792, 1, 256] - - [50, 56.232] - - - [8704, 6144, 1, 256] - - [35, 72.779] - - - [816, 768, 1, 256] - - [124, 29.143] - - - [18432, 2865, 1, 256] - - [26, 69.917] - - - [3120, 2865, 1, 256] - - [28, 51.19] - - - [6448, 2865, 1, 256] - - [36, 57.933] - - - [12080, 2816, 1, 256] - - [54, 62.716] - - - [10496, 3328, 1, 256] - - [61, 70.271] - - - [15920, 10240, 1, 256] - - [54, 64.072] - - - [15872, 2048, 1, 256] - - [59, 68.977] - - - [11568, 2816, 1, 256] - - [37, 61.211] - - - [19200, 10240, 1, 256] - - [38, 74.775] - - - [13312, 5632, 1, 256] - - [38, 74.331] - - - [15360, 2816, 1, 256] - - [30, 71.927] - - - [12288, 2865, 1, 256] - - [33, 68.287] - - - [19760, 6400, 1, 256] - - [45, 64.027] - - - [19968, 256, 1, 256] - - [45, 53.43] - - - [7680, 4352, 1, 256] - - [56, 71.265] - - - [11008, 3584, 1, 256] - - [76, 70.462] - - - [3072, 2817, 1, 256] - - [84, 54.528] - - - [11264, 6144, 1, 256] - - [30, 73.603] - - - [5424, 256, 1, 256] - - [111, 38.221] - - - [13568, 1280, 1, 256] - - [26, 66.879] - - - [3840, 2048, 1, 256] - - [28, 56.612] - - - [6144, 3072, 1, 256] - - [26, 67.753] - - - [19200, 1536, 1, 256] - - [58, 68.062] - - - [10240, 1280, 1, 256] - - [50, 62.93] - - - [3376, 512, 1, 256] - - [184, 44.698] - - - [12544, 1281, 1, 256] - - [41, 61.48] - - - [9776, 6656, 1, 256] - - [45, 63.826] - - - [7680, 2865, 1, 256] - - [26, 66.362] - - - [10544, 7680, 1, 256] - - [54, 63.447] - - - [15616, 3328, 1, 256] - - [26, 71.606] - - - [3328, 1280, 1, 256] - - [35, 46.673] - - - [2560, 1024, 1, 256] - - [35, 42.786] - - - [17456, 4096, 1, 256] - - [59, 64.105] - - - [6912, 3328, 1, 256] - - [56, 68.767] - - - [3584, 2816, 1, 256] - - [35, 62.134] - - - [17408, 3328, 1, 256] - - [26, 72.343] - - - [19200, 2816, 1, 256] - - [56, 71.241] - - - [15104, 1792, 1, 256] - - [36, 68.538] - - - [6144, 256, 1, 256] - - [183, 44.901] - - - [8192, 6144, 1, 256] - - [28, 72.019] - - - [12032, 4608, 1, 256] - - [36, 72.088] - - - [1840, 256, 1, 256] - - [109, 22.542] - - - [13312, 256, 1, 256] - - [35, 52.493] - - - [9216, 1792, 1, 256] - - [28, 65.175] - - - [14592, 3329, 1, 256] - - [58, 68.572] - - - [8448, 1280, 1, 256] - - [56, 64.286] - - - [11520, 8192, 1, 256] - - [58, 73.806] - - - [2608, 2560, 1, 256] - - [28, 53.501] - - - [5120, 2304, 1, 256] - - [50, 62.892] - - - [13056, 3328, 1, 256] - - [36, 70.928] - - - [11008, 8192, 1, 256] - - [76, 72.75] - - - [14896, 2865, 1, 256] - - [28, 60.701] - - - [6704, 3840, 1, 256] - - [28, 61.112] - - - [15872, 3329, 1, 256] - - [26, 69.435] - - - [7168, 1792, 1, 256] - - [56, 62.425] - - - [4656, 2865, 1, 256] - - [38, 57.651] - - - [18736, 5632, 1, 256] - - [59, 63.345] - - - [768, 512, 1, 256] - - [109, 23.02] - - - [16432, 3072, 1, 256] - - [39, 65.493] - - - [14848, 2865, 1, 256] - - [28, 68.873] - - - [4864, 1536, 1, 256] - - [36, 53.95] - - - [9472, 2865, 1, 256] - - [35, 66.928] - - - [10496, 2048, 1, 256] - - [61, 68.578] - - - [14336, 1024, 1, 256] - - [41, 63.079] - - - [18432, 256, 1, 256] - - [41, 50.137] - - - [16896, 3840, 1, 256] - - [68, 73.437] - - - [10240, 512, 1, 256] - - [41, 54.602] - - - [15664, 2304, 1, 256] - - [59, 62.137] - - - [10496, 3329, 1, 256] - - [38, 66.755] - - - [19456, 1536, 1, 256] - - [36, 69.248] - - - [17920, 1281, 1, 256] - - [34, 63.538] - - - [8960, 256, 1, 256] - - [35, 37.859] - - - [10496, 768, 1, 256] - - [54, 56.926] - - - [5120, 2816, 1, 256] - - [50, 66.912] - - - [12288, 2048, 1, 256] - - [38, 66.331] - - - [11568, 8704, 1, 256] - - [37, 63.361] - - - [10496, 1024, 1, 256] - - [41, 64.161] - - - [10288, 256, 1, 256] - - [35, 41.229] - - - [5168, 2048, 1, 256] - - [26, 59.001] - - - [11776, 3328, 1, 256] - - [58, 71.261] - - - [15152, 10240, 1, 256] - - [54, 63.573] - - - [14384, 2865, 1, 256] - - [28, 60.698] - - - [12288, 512, 1, 256] - - [85, 54.504] - - - [16688, 256, 1, 256] - - [35, 44.404] - - - [6912, 4096, 1, 256] - - [26, 69.673] - - - [4864, 2048, 1, 256] - - [36, 60.744] - - - [4096, 1024, 1, 256] - - [50, 45.746] - - - [12848, 9984, 1, 256] - - [54, 64.505] - - - [16896, 1281, 1, 256] - - [54, 61.455] - - - [768, 561, 1, 256] - - [115, 24.478] - - - [16896, 3584, 1, 256] - - [28, 73.171] - - - [14592, 6144, 1, 256] - - [68, 72.431] - - - [17664, 4096, 1, 256] - - [48, 71.584] - - - [8448, 2865, 1, 256] - - [36, 67.332] - - - [18432, 768, 1, 256] - - [54, 65.406] - - - [12032, 512, 1, 256] - - [24, 53.772] - - - [11008, 256, 1, 256] - - [80, 44.969] - - - [15360, 1536, 1, 256] - - [36, 67.238] - - - [5888, 2048, 1, 256] - - [41, 63.123] - - - [13104, 256, 1, 256] - - [64, 49.375] - - - [11264, 7680, 1, 256] - - [30, 74.867] - - - [19248, 2865, 1, 256] - - [59, 60.917] - - - [17200, 2865, 1, 256] - - [28, 61.583] - - - [8192, 2048, 1, 256] - - [38, 64.148] - - - [7472, 4608, 1, 256] - - [54, 61.625] - - - [7168, 2048, 1, 256] - - [24, 63.091] - - - [13360, 2816, 1, 256] - - [59, 62.643] - - - [17920, 4352, 1, 256] - - [58, 73.831] - - - [15408, 256, 1, 256] - - [41, 41.789] - - - [19200, 1281, 1, 256] - - [61, 63.126] - - - [15360, 256, 1, 256] - - [56, 43.165] - - - [9984, 6400, 1, 256] - - [36, 73.374] - - - [18944, 2865, 1, 256] - - [38, 69.777] - - - [3840, 2865, 1, 256] - - [36, 59.129] - - - [8192, 3328, 1, 256] - - [28, 68.008] - - - [5888, 256, 1, 256] - - [192, 43.899] - - - [15616, 2816, 1, 256] - - [28, 71.289] - - - [17664, 2865, 1, 256] - - [60, 69.269] - - - [14592, 768, 1, 256] - - [41, 59.758] - - - [18944, 1281, 1, 256] - - [73, 63.675] - - - [11264, 1536, 1, 256] - - [50, 66.394] - - - [8496, 5632, 1, 256] - - [35, 64.342] - - - [17664, 3328, 1, 256] - - [32, 71.465] - - - [14848, 2048, 1, 256] - - [41, 69.541] - - - [15408, 2865, 1, 256] - - [26, 61.565] - - - [4096, 2048, 1, 256] - - [60, 58.762] - - - [14128, 1024, 1, 256] - - [45, 55.462] - - - [1072, 817, 1, 256] - - [124, 30.424] - - - [17152, 3840, 1, 256] - - [26, 72.897] - - - [17664, 2048, 1, 256] - - [54, 68.715] - - - [16896, 256, 1, 256] - - [35, 46.769] - - - [2304, 2097, 1, 256] - - [50, 51.622] - - - [5888, 2560, 1, 256] - - [50, 64.841] - - - [9472, 1792, 1, 256] - - [26, 65.661] - - - [1328, 1280, 1, 256] - - [35, 36.117] - - - [19200, 1280, 1, 256] - - [58, 68.368] - - - [12544, 1280, 1, 256] - - [26, 66.492] - - - [16432, 3328, 1, 256] - - [29, 66.565] - - - [17920, 1280, 1, 256] - - [58, 68.79] - - - [8752, 5632, 1, 256] - - [30, 63.332] - - - [7936, 2048, 1, 256] - - [45, 63.392] - - - [9472, 1280, 1, 256] - - [28, 63.42] - - - [16896, 1024, 1, 256] - - [41, 66.25] - - - [6656, 3329, 1, 256] - - [60, 65.83] - - - [17456, 2865, 1, 256] - - [30, 62.344] - - - [5632, 2304, 1, 256] - - [50, 62.835] - - - [14080, 1024, 1, 256] - - [34, 65.233] - - - [15872, 3328, 1, 256] - - [35, 72.279] - - - [5168, 2816, 1, 256] - - [26, 56.571] - - - [13312, 9728, 1, 256] - - [26, 75.464] - - - [1584, 1329, 1, 256] - - [28, 33.18] - - - [15664, 2560, 1, 256] - - [38, 62.896] - - - [2048, 768, 1, 256] - - [118, 44.901] - - - [17712, 2816, 1, 256] - - [54, 63.326] - - - [16128, 2865, 1, 256] - - [28, 68.888] - - - [15872, 2816, 1, 256] - - [36, 72.252] - - - [18224, 2816, 1, 256] - - [45, 62.778] - - - [5632, 4352, 1, 256] - - [36, 69.961] - - - [1792, 1281, 1, 256] - - [50, 37.52] - - - [6656, 2816, 1, 256] - - [56, 66.455] - - - [16640, 1281, 1, 256] - - [66, 60.184] - - - [13056, 10240, 1, 256] - - [30, 74.986] - - - [17968, 256, 1, 256] - - [26, 46.195] - - - [5376, 4096, 1, 256] - - [28, 67.138] - - - [15152, 2048, 1, 256] - - [59, 60.877] - - - [13568, 2816, 1, 256] - - [35, 70.507] - - - [12800, 2816, 1, 256] - - [32, 70.881] - - - [6960, 2816, 1, 256] - - [30, 61.413] - - - [17968, 4608, 1, 256] - - [45, 63.285] - - - [15104, 3328, 1, 256] - - [24, 71.526] - - - [7472, 4352, 1, 256] - - [26, 61.464] - - - [15872, 2304, 1, 256] - - [36, 71.22] - - - [4400, 2816, 1, 256] - - [28, 59.033] - - - [16128, 6144, 1, 256] - - [28, 73.909] - - - [18944, 2816, 1, 256] - - [68, 72.175] - - - [5424, 2865, 1, 256] - - [45, 60.264] - - - [8192, 768, 1, 256] - - [35, 52.953] - - - [12848, 256, 1, 256] - - [41, 48.34] - - - [12288, 1281, 1, 256] - - [28, 60.225] - - - [13872, 512, 1, 256] - - [26, 55.706] - - - [5888, 1280, 1, 256] - - [50, 54.731] - - - [2816, 1281, 1, 256] - - [36, 41.044] - - - [19200, 3329, 1, 256] - - [68, 68.61] - - - [12800, 3328, 1, 256] - - [58, 71.351] - - - [15360, 2865, 1, 256] - - [28, 69.192] - - - [17152, 3584, 1, 256] - - [56, 72.518] - - - [17456, 2816, 1, 256] - - [30, 63.271] - - - [18176, 10240, 1, 256] - - [26, 74.987] - - - [6144, 2816, 1, 256] - - [50, 67.471] - - - [18176, 1280, 1, 256] - - [35, 68.702] - - - [16384, 1281, 1, 256] - - [28, 53.321] - - - [9216, 3328, 1, 256] - - [26, 69.932] - - - [14080, 2816, 1, 256] - - [68, 70.205] - - - [18688, 2816, 1, 256] - - [36, 71.742] - - - [15872, 10240, 1, 256] - - [26, 75.281] - - - [10800, 7936, 1, 256] - - [45, 64.944] - - - [9984, 1281, 1, 256] - - [41, 60.139] - - - [4144, 256, 1, 256] - - [111, 36.091] - - - [16640, 6144, 1, 256] - - [28, 73.468] - - - [11776, 4096, 1, 256] - - [75, 71.643] - - - [11056, 8192, 1, 256] - - [30, 64.782] - - - [5376, 2816, 1, 256] - - [56, 64.272] - - - [19712, 1280, 1, 256] - - [26, 67.903] - - - [4608, 2816, 1, 256] - - [56, 63.153] - - - [19456, 2865, 1, 256] - - [28, 70.001] - - - [14080, 256, 1, 256] - - [54, 54.728] - - - [7216, 4096, 1, 256] - - [30, 61.81] - - - [2816, 1280, 1, 256] - - [50, 54.728] - - - [10496, 3072, 1, 256] - - [35, 70.522] - - - [12544, 4864, 1, 256] - - [56, 72.899] - - - [9984, 6912, 1, 256] - - [35, 73.374] - - - [4912, 2048, 1, 256] - - [45, 56.406] - - - [9984, 2304, 1, 256] - - [36, 68.987] - - - [19248, 5888, 1, 256] - - [37, 63.654] - - - [19712, 2048, 1, 256] - - [85, 67.452] - - - [9728, 2816, 1, 256] - - [35, 69.216] - - - [19504, 6400, 1, 256] - - [45, 64.123] - - - [16896, 3072, 1, 256] - - [35, 72.467] - - - [15104, 1536, 1, 256] - - [36, 67.917] - - - [2608, 2609, 1, 256] - - [26, 46.653] - - - [14384, 256, 1, 256] - - [41, 39.048] - - - [17664, 2816, 1, 256] - - [32, 71.206] - - - [9776, 6912, 1, 256] - - [45, 64.424] - - - [1792, 512, 1, 256] - - [109, 33.381] - - - [13312, 1536, 1, 256] - - [50, 66.916] - - - [1072, 1073, 1, 256] - - [109, 36.546] - - - [9472, 2048, 1, 256] - - [41, 67.028] - - - [4608, 3328, 1, 256] - - [56, 65.809] - - - [9984, 2560, 1, 256] - - [36, 68.769] - - - [6912, 5376, 1, 256] - - [50, 71.656] - - - [16640, 2865, 1, 256] - - [35, 68.013] - - - [4352, 3072, 1, 256] - - [35, 63.961] - - - [5632, 3328, 1, 256] - - [36, 67.157] - - - [9216, 5632, 1, 256] - - [30, 73.616] - - - [3328, 3329, 1, 256] - - [35, 58.687] - - - [13824, 10240, 1, 256] - - [52, 74.807] - - - [12288, 3329, 1, 256] - - [28, 68.787] - - - [2864, 256, 1, 256] - - [111, 33.16] - - - [19712, 1792, 1, 256] - - [58, 68.68] - - - [6656, 1280, 1, 256] - - [50, 59.572] - - - [13056, 2048, 1, 256] - - [61, 68.872] - - - [6912, 1281, 1, 256] - - [50, 55.428] - - - [16176, 2816, 1, 256] - - [26, 65.223] - - - [14592, 3328, 1, 256] - - [58, 70.537] - - - [10496, 2865, 1, 256] - - [26, 66.963] - - - [9728, 6400, 1, 256] - - [35, 73.801] - - - [5888, 4608, 1, 256] - - [56, 68.523] - - - [16432, 10240, 1, 256] - - [55, 68.845] - - - [19456, 5888, 1, 256] - - [28, 74.612] - - - [3888, 256, 1, 256] - - [111, 34.576] - - - [12336, 2816, 1, 256] - - [28, 61.954] - - - [19456, 256, 1, 256] - - [41, 52.392] - - - [14384, 2816, 1, 256] - - [30, 62.004] - - - [14384, 1024, 1, 256] - - [26, 56.205] - - - [16640, 2816, 1, 256] - - [35, 70.493] - - - [3840, 3329, 1, 256] - - [36, 61.31] - - - [2304, 2304, 1, 256] - - [36, 56.085] - - - [10240, 2816, 1, 256] - - [28, 71.023] - - - [13104, 10240, 1, 256] - - [37, 64.67] - - - [1536, 256, 1, 256] - - [117, 22.905] - - - [11008, 2865, 1, 256] - - [86, 67.12] - - - [13104, 9984, 1, 256] - - [45, 64.266] - - - [10240, 7168, 1, 256] - - [26, 73.574] - - - [3888, 2865, 1, 256] - - [26, 53.93] - - - [8192, 4864, 1, 256] - - [26, 70.814] - - - [15920, 256, 1, 256] - - [24, 43.417] - - - [6448, 3584, 1, 256] - - [54, 61.758] - - - [16128, 2304, 1, 256] - - [36, 70.651] - - - [9728, 2865, 1, 256] - - [28, 68.808] - - - [6144, 4864, 1, 256] - - [38, 71.369] - - - [14848, 256, 1, 256] - - [61, 42.175] - - - [4352, 1024, 1, 256] - - [35, 48.505] - - - [15360, 10240, 1, 256] - - [30, 75.504] - - - [19504, 10240, 1, 256] - - [29, 63.61] - - - [3328, 3072, 1, 256] - - [35, 62.752] - - - [1536, 1281, 1, 256] - - [32, 32.663] - - - [19760, 6656, 1, 256] - - [59, 63.827] - - - [3584, 3329, 1, 256] - - [35, 62.683] - - - [14848, 2816, 1, 256] - - [35, 71.299] - - - [4400, 2865, 1, 256] - - [28, 54.738] - - - [3888, 1024, 1, 256] - - [45, 41.614] - - - [16640, 2048, 1, 256] - - [41, 68.188] - - - [4096, 2816, 1, 256] - - [60, 61.935] - - - [14640, 2816, 1, 256] - - [45, 62.79] - - - [9472, 1281, 1, 256] - - [35, 57.761] - - - [8192, 1280, 1, 256] - - [36, 62.359] - - - [8960, 2865, 1, 256] - - [50, 67.185] - - - [4144, 2816, 1, 256] - - [26, 57.389] - - - [10288, 7168, 1, 256] - - [26, 62.857] - - - [14592, 256, 1, 256] - - [41, 41.321] - - - [10240, 2048, 1, 256] - - [37, 66.868] - - - [17920, 2865, 1, 256] - - [48, 69.514] - - - [12592, 2816, 1, 256] - - [59, 62.814] - - - [14592, 1536, 1, 256] - - [36, 66.733] - - - [11568, 256, 1, 256] - - [30, 44.265] - - - [6704, 3584, 1, 256] - - [30, 61.656] - - - [5120, 3328, 1, 256] - - [50, 66.584] - - - [4400, 1536, 1, 256] - - [36, 53.849] - - - [18944, 256, 1, 256] - - [36, 51.106] - - - [19712, 5888, 1, 256] - - [42, 72.495] - - - [7984, 5120, 1, 256] - - [35, 64.029] - - - [8240, 2865, 1, 256] - - [36, 61.677] - - - [6144, 1280, 1, 256] - - [50, 56.718] - - - [8496, 2816, 1, 256] - - [38, 61.931] - - - [14592, 1024, 1, 256] - - [28, 62.882] - - - [14592, 2865, 1, 256] - - [32, 68.747] - - - [13360, 256, 1, 256] - - [54, 50.34] - - - [8448, 256, 1, 256] - - [60, 36.005] - - - [16896, 2816, 1, 256] - - [56, 72.165] - - - [15152, 2865, 1, 256] - - [38, 61.182] - - - [11056, 2816, 1, 256] - - [45, 62.578] - - - [15616, 1280, 1, 256] - - [50, 66.097] - - - [8192, 5120, 1, 256] - - [28, 70.894] - - - [17408, 256, 1, 256] - - [41, 48.142] - - - [18432, 10240, 1, 256] - - [38, 75.394] - - - [14592, 1280, 1, 256] - - [36, 65.738] - - - [3328, 512, 1, 256] - - [182, 48.048] - - - [14336, 1280, 1, 256] - - [50, 66.254] - - - [13616, 2865, 1, 256] - - [37, 60.21] - - - [8192, 256, 1, 256] - - [36, 35.021] - - - [10240, 1281, 1, 256] - - [50, 60.343] - - - [1840, 1841, 1, 256] - - [35, 35.183] - - - [12800, 9472, 1, 256] - - [30, 75.03] - - - [17664, 256, 1, 256] - - [41, 48.45] - - - [768, 769, 1, 256] - - [110, 28.381] - - - [19456, 2048, 1, 256] - - [37, 69.833] - - - [13056, 3329, 1, 256] - - [38, 67.803] - - - [11056, 256, 1, 256] - - [28, 43.423] - - - [7424, 6144, 1, 256] - - [36, 71.699] - - - [14848, 3328, 1, 256] - - [30, 71.856] - - - [6656, 3328, 1, 256] - - [54, 67.747] - - - [10752, 1281, 1, 256] - - [41, 59.249] - - - [9984, 2865, 1, 256] - - [28, 67.115] - - - [14080, 3329, 1, 256] - - [58, 68.263] - - - [17920, 3328, 1, 256] - - [77, 72.159] - - - [13312, 10240, 1, 256] - - [28, 75.296] - - - [16640, 3584, 1, 256] - - [35, 71.857] - - - [17408, 3840, 1, 256] - - [38, 73.865] - - - [12032, 8960, 1, 256] - - [35, 74.177] - - - [10800, 2865, 1, 256] - - [54, 61.201] - - - [3072, 2816, 1, 256] - - [35, 60.572] - - - [14128, 2816, 1, 256] - - [59, 61.885] - - - [11312, 8192, 1, 256] - - [30, 64.421] - - - [2560, 2305, 1, 256] - - [36, 51.575] - - - [16640, 3072, 1, 256] - - [28, 70.804] - - - [16128, 2048, 1, 256] - - [41, 68.171] - - - [6144, 512, 1, 256] - - [61, 49.726] - - - [18688, 4864, 1, 256] - - [36, 73.833] - - - [17200, 256, 1, 256] - - [41, 45.57] - - - [8752, 2865, 1, 256] - - [30, 58.991] - - - [18944, 1280, 1, 256] - - [50, 68.497] - - - [16640, 3328, 1, 256] - - [49, 70.911] - - - [304, 305, 1, 256] - - [120, 5.569] - - - [15104, 256, 1, 256] - - [41, 42.985] - - - [7680, 3328, 1, 256] - - [36, 68.466] - - - [12336, 9216, 1, 256] - - [26, 63.141] - - - [14080, 6144, 1, 256] - - [68, 72.26] - - - [7168, 5888, 1, 256] - - [28, 72.398] - - - [7424, 1280, 1, 256] - - [35, 59.134] - - - [4864, 3584, 1, 256] - - [56, 67.73] - - - [1280, 1025, 1, 256] - - [184, 38.019] - - - [10240, 2865, 1, 256] - - [28, 68.642] - - - [18480, 10240, 1, 256] - - [26, 62.84] - - - [7680, 256, 1, 256] - - [48, 32.732] - - - [9472, 6656, 1, 256] - - [35, 73.381] - - - [12032, 6144, 1, 256] - - [30, 72.815] - - - [5120, 3329, 1, 256] - - [35, 65.105] - - - [10752, 256, 1, 256] - - [28, 44.79] - - - [6960, 256, 1, 256] - - [35, 39.312] - - - [9008, 6144, 1, 256] - - [59, 63.58] - - - [7424, 2048, 1, 256] - - [61, 64.503] - - - [5632, 1280, 1, 256] - - [36, 61.021] - - - [19712, 10240, 1, 256] - - [42, 73.761] - - - [6400, 768, 1, 256] - - [64, 52.653] - - - [10752, 3328, 1, 256] - - [28, 71.314] - - - [18432, 5376, 1, 256] - - [28, 74.701] - - - [9520, 256, 1, 256] - - [41, 38.314] - - - [5680, 2816, 1, 256] - - [26, 60.823] - - - [11008, 3329, 1, 256] - - [76, 65.949] - - - [4608, 2865, 1, 256] - - [36, 62.836] - - - [6448, 256, 1, 256] - - [193, 43.664] - - - [3584, 256, 1, 256] - - [115, 33.276] - - - [12336, 9472, 1, 256] - - [38, 63.293] - - - [1280, 1281, 1, 256] - - [64, 36.52] - - - [7936, 6144, 1, 256] - - [60, 71.94] - - - [15152, 1792, 1, 256] - - [38, 60.278] - - - [4352, 1281, 1, 256] - - [28, 49.182] - - - [12848, 2865, 1, 256] - - [38, 59.68] - - - [16944, 3584, 1, 256] - - [37, 64.55] - - - [8752, 256, 1, 256] - - [35, 36.102] - - - [6912, 256, 1, 256] - - [125, 48.589] - - - [14336, 1281, 1, 256] - - [37, 59.694] - - - [2304, 2049, 1, 256] - - [36, 50.153] - - - [9216, 6144, 1, 256] - - [30, 73.135] - - - [1072, 1024, 1, 256] - - [111, 36.172] - - - [10752, 6144, 1, 256] - - [28, 73.602] - - - [1792, 1537, 1, 256] - - [36, 43.657] - - - [17968, 2865, 1, 256] - - [26, 61.071] - - - [8448, 4864, 1, 256] - - [35, 72.259] - - - [15408, 2048, 1, 256] - - [54, 62.18] - - - [15104, 1280, 1, 256] - - [35, 67.135] - - - [9264, 2865, 1, 256] - - [26, 59.195] - - - [15616, 2865, 1, 256] - - [36, 69.029] - - - [16896, 10240, 1, 256] - - [26, 75.376] - - - [15104, 2816, 1, 256] - - [61, 71.358] - - - [13872, 2865, 1, 256] - - [26, 60.894] - - - [13056, 1280, 1, 256] - - [35, 65.37] - - - [12288, 3328, 1, 256] - - [28, 71.562] - - - [3840, 3328, 1, 256] - - [35, 62.248] - - - [9216, 1280, 1, 256] - - [35, 62.921] - - - [8448, 1281, 1, 256] - - [35, 56.762] - - - [2560, 256, 1, 256] - - [124, 31.104] - - - [4608, 1281, 1, 256] - - [36, 51.593] - - - [6144, 2865, 1, 256] - - [26, 63.396] - - - [5888, 2816, 1, 256] - - [36, 65.221] - - - [3584, 1281, 1, 256] - - [35, 49.145] - - - [18688, 5120, 1, 256] - - [26, 73.865] - - - [12288, 2816, 1, 256] - - [26, 70.398] - - - [4864, 2865, 1, 256] - - [35, 65.382] - - - [9216, 2048, 1, 256] - - [59, 66.26] - - - [13872, 768, 1, 256] - - [37, 58.587] - - - [10496, 7424, 1, 256] - - [26, 73.594] - - - [16384, 512, 1, 256] - - [35, 54.525] - - - [14848, 10240, 1, 256] - - [28, 74.949] - - - [17920, 2048, 1, 256] - - [40, 69.882] - - - [11008, 7936, 1, 256] - - [72, 73.234] - - - [1792, 1792, 1, 256] - - [28, 49.797] - - - [7680, 4864, 1, 256] - - [35, 72.155] - - - [19760, 256, 1, 256] - - [59, 49.25] - - - [15616, 1792, 1, 256] - - [36, 69.504] - - - [1792, 1793, 1, 256] - - [26, 49.099] - - - [8192, 3329, 1, 256] - - [28, 66.799] - - - [2560, 1280, 1, 256] - - [35, 51.366] - - - [1328, 1073, 1, 256] - - [118, 38.13] - - - [16896, 2865, 1, 256] - - [38, 69.675] - - - [8960, 1280, 1, 256] - - [50, 61.258] - - - [6960, 2865, 1, 256] - - [28, 58.909] - - - [1280, 1024, 1, 256] - - [126, 42.945] - - - [6400, 2048, 1, 256] - - [56, 62.695] - - - [18480, 5376, 1, 256] - - [30, 62.704] - - - [18944, 2048, 1, 256] - - [41, 70.368] - - - [9520, 6656, 1, 256] - - [37, 63.438] - - - [4352, 1536, 1, 256] - - [36, 57.987] - - - [19712, 6144, 1, 256] - - [52, 72.546] - - - [6400, 2816, 1, 256] - - [30, 68.737] - - - [1792, 1585, 1, 256] - - [26, 44.837] - - - [13312, 6144, 1, 256] - - [26, 74.102] - - - [17408, 4096, 1, 256] - - [28, 72.71] - - - [16128, 256, 1, 256] - - [70, 44.519] - - - [15104, 2048, 1, 256] - - [61, 68.882] - - - [8704, 2865, 1, 256] - - [30, 66.521] - - - [6144, 768, 1, 256] - - [61, 50.604] - - - [10496, 1280, 1, 256] - - [36, 63.682] - - - [816, 561, 1, 256] - - [110, 21.462] - - - [6912, 3840, 1, 256] - - [56, 70.36] - - - [8704, 1281, 1, 256] - - [24, 58.499] - - - [13312, 1792, 1, 256] - - [36, 68.51] - - - [5120, 1281, 1, 256] - - [56, 55.721] - - - [10496, 1281, 1, 256] - - [61, 58.196] - - - [8448, 6144, 1, 256] - - [35, 72.701] - - - [2560, 2353, 1, 256] - - [35, 52.819] - - - [4352, 1280, 1, 256] - - [35, 49.744] - - - [12336, 256, 1, 256] - - [37, 47.326] - - - [21504, 10240, 1, 256] - - [30, 75.583] - - - [31744, 6144, 1, 256] - - [28, 74.524] - - - [27648, 1280, 1, 256] - - [35, 70.518] - - - [22272, 512, 1, 256] - - [24, 60.583] - - - [29184, 256, 1, 256] - - [35, 52.208] - - - [23808, 4096, 1, 256] - - [75, 72.707] - - - [30720, 7168, 1, 256] - - [38, 73.869] - - - [29440, 2865, 1, 256] - - [28, 70.006] - - - [25600, 5632, 1, 256] - - [28, 75.075] - - - [24832, 10240, 1, 256] - - [28, 74.967] - - - [22784, 2865, 1, 256] - - [30, 69.758] - - - [24368, 768, 1, 256] - - [59, 59.639] - - - [21760, 8192, 1, 256] - - [38, 74.528] - - - [29184, 10240, 1, 256] - - [25, 74.869] - - - [26368, 6144, 1, 256] - - [28, 74.073] - - - [23088, 10240, 1, 256] - - [54, 63.352] - - - [29952, 4096, 1, 256] - - [30, 72.803] - - - [24320, 6144, 1, 256] - - [31, 74.167] - - - [32256, 2048, 1, 256] - - [66, 71.243] - - - [29488, 2865, 1, 256] - - [38, 60.926] - - - [21808, 2816, 1, 256] - - [59, 63.226] - - - [32000, 7936, 1, 256] - - [28, 74.425] - - - [23040, 10240, 1, 256] - - [26, 75.226] - - - [31792, 10240, 1, 256] - - [29, 63.633] - - - [24320, 1281, 1, 256] - - [70, 64.001] - - - [27136, 3072, 1, 256] - - [52, 73.24] - - - [31488, 6144, 1, 256] - - [28, 74.077] - - - [34096, 2865, 1, 256] - - [45, 62.013] - - - [33024, 8960, 1, 256] - - [31, 74.681] - - - [28928, 1280, 1, 256] - - [35, 69.965] - - - [31488, 5632, 1, 256] - - [28, 74.077] - - - [27696, 4096, 1, 256] - - [26, 62.899] - - - [31488, 10240, 1, 256] - - [38, 74.538] - - - [28928, 10240, 1, 256] - - [30, 74.419] - - - [26160, 10240, 1, 256] - - [37, 62.983] - - - [26112, 3328, 1, 256] - - [90, 73.184] - - - [28928, 4864, 1, 256] - - [30, 73.672] - - - [27904, 3328, 1, 256] - - [82, 72.375] - - - [29184, 5376, 1, 256] - - [31, 74.353] - - - [29952, 1281, 1, 256] - - [34, 64.138] - - - [24832, 6144, 1, 256] - - [25, 74.183] - - - [28160, 4096, 1, 256] - - [43, 73.284] - - - [24320, 1280, 1, 256] - - [58, 69.334] - - - [34816, 768, 1, 256] - - [54, 68.672] - - - [34816, 1281, 1, 256] - - [28, 63.991] - - - [27136, 2816, 1, 256] - - [58, 73.176] - - - [32256, 8192, 1, 256] - - [31, 74.855] - - - [26624, 2865, 1, 256] - - [28, 70.664] - - - [23808, 3840, 1, 256] - - [28, 73.539] - - - [29440, 5376, 1, 256] - - [58, 74.171] - - - [30464, 10240, 1, 256] - - [25, 73.683] - - - [29232, 10240, 1, 256] - - [37, 62.969] - - - [27136, 1280, 1, 256] - - [58, 70.23] - - - [27904, 6144, 1, 256] - - [29, 73.746] - - - [33024, 2816, 1, 256] - - [43, 72.269] - - - [34816, 3329, 1, 256] - - [26, 70.685] - - - [34048, 1792, 1, 256] - - [32, 71.4] - - - [21248, 7424, 1, 256] - - [38, 74.819] - - - [29952, 256, 1, 256] - - [24, 53.356] - - - [34560, 256, 1, 256] - - [59, 59.115] - - - [26368, 3072, 1, 256] - - [38, 72.767] - - - [23600, 2865, 1, 256] - - [38, 61.367] - - - [30720, 512, 1, 256] - - [36, 64.809] - - - [30768, 10240, 1, 256] - - [42, 62.81] - - - [28928, 1024, 1, 256] - - [64, 68.461] - - - [26624, 256, 1, 256] - - [26, 57.116] - - - [26928, 10240, 1, 256] - - [54, 63.381] - - - [21248, 7936, 1, 256] - - [28, 74.601] - - - [34304, 2816, 1, 256] - - [60, 73.087] - - - [29696, 3840, 1, 256] - - [28, 74.502] - - - [27696, 10240, 1, 256] - - [28, 63.529] - - - [24064, 2048, 1, 256] - - [73, 70.68] - - - [33536, 6144, 1, 256] - - [31, 74.074] - - - [32512, 8704, 1, 256] - - [48, 74.767] - - - [21552, 2816, 1, 256] - - [54, 63.568] - - - [27648, 10240, 1, 256] - - [38, 75.468] - - - [22272, 2048, 1, 256] - - [34, 70.036] - - - [28976, 5632, 1, 256] - - [54, 63.367] - - - [30720, 10240, 1, 256] - - [26, 75.334] - - - [26112, 2816, 1, 256] - - [28, 73.015] - - - [20528, 10240, 1, 256] - - [29, 62.326] - - - [29696, 1536, 1, 256] - - [56, 70.945] - - - [31536, 2865, 1, 256] - - [54, 61.567] - - - [32000, 3328, 1, 256] - - [50, 72.394] - - - [20784, 2865, 1, 256] - - [26, 60.364] - - - [33280, 9984, 1, 256] - - [52, 75.53] - - - [25600, 3329, 1, 256] - - [33, 70.116] - - - [27904, 4096, 1, 256] - - [66, 72.935] - - - [29488, 256, 1, 256] - - [59, 48.225] - - - [32048, 10240, 1, 256] - - [28, 63.133] - - - [31280, 2865, 1, 256] - - [54, 61.474] - - - [32816, 2816, 1, 256] - - [29, 65.244] - - - [34096, 2816, 1, 256] - - [45, 63.755] - - - [20992, 3328, 1, 256] - - [26, 72.666] - - - [32768, 1281, 1, 256] - - [96, 51.657] - - - [24576, 4864, 1, 256] - - [39, 70.216] - - - [30464, 3328, 1, 256] - - [60, 71.563] - - - [28208, 256, 1, 256] - - [45, 46.063] - - - [23552, 1280, 1, 256] - - [50, 70.066] - - - [20528, 7168, 1, 256] - - [26, 61.991] - - - [34560, 2865, 1, 256] - - [38, 70.251] - - - [20736, 2816, 1, 256] - - [35, 72.019] - - - [26880, 3328, 1, 256] - - [26, 72.345] - - - [31536, 8192, 1, 256] - - [59, 62.985] - - - [31744, 8448, 1, 256] - - [26, 75.166] - - - [20224, 2865, 1, 256] - - [28, 69.451] - - - [22528, 2048, 1, 256] - - [62, 69.864] - - - [24320, 2048, 1, 256] - - [70, 70.443] - - - [32512, 8960, 1, 256] - - [48, 74.928] - - - [33072, 10240, 1, 256] - - [30, 64.214] - - - [24880, 10240, 1, 256] - - [30, 63.243] - - - [21040, 7680, 1, 256] - - [37, 64.019] - - - [26368, 10240, 1, 256] - - [26, 74.681] - - - [32304, 8704, 1, 256] - - [59, 63.345] - - - [33536, 1281, 1, 256] - - [75, 63.723] - - - [27136, 1024, 1, 256] - - [24, 68.778] - - - [33792, 1281, 1, 256] - - [42, 64.358] - - - [33584, 256, 1, 256] - - [28, 54.154] - - - [20528, 7424, 1, 256] - - [28, 62.655] - - - [28928, 2865, 1, 256] - - [30, 69.869] - - - [22016, 2048, 1, 256] - - [49, 70.212] - - - [29440, 3328, 1, 256] - - [62, 72.706] - - - [30208, 2048, 1, 256] - - [45, 71.22] - - - [20480, 2816, 1, 256] - - [28, 72.352] - - - [25904, 256, 1, 256] - - [35, 51.108] - - - [20736, 10240, 1, 256] - - [28, 74.9] - - - [32816, 256, 1, 256] - - [59, 53.638] - - - [33792, 3328, 1, 256] - - [26, 73.479] - - - [22272, 1281, 1, 256] - - [70, 63.01] - - - [25600, 1280, 1, 256] - - [30, 69.893] - - - [33280, 3329, 1, 256] - - [25, 70.393] - - - [22784, 1281, 1, 256] - - [40, 63.786] - - - [25392, 2816, 1, 256] - - [54, 63.703] - - - [33280, 3328, 1, 256] - - [30, 73.324] - - - [21760, 1280, 1, 256] - - [36, 68.663] - - - [33024, 768, 1, 256] - - [24, 65.517] - - - [25088, 1792, 1, 256] - - [56, 71.376] - - - [26368, 3329, 1, 256] - - [55, 69.349] - - - [34560, 3328, 1, 256] - - [62, 72.635] - - - [23040, 6144, 1, 256] - - [48, 74.371] - - - [30464, 2048, 1, 256] - - [90, 68.048] - - - [28672, 3328, 1, 256] - - [30, 72.77] - - - [30464, 6912, 1, 256] - - [31, 73.595] - - - [32048, 2816, 1, 256] - - [37, 63.364] - - - [33792, 9728, 1, 256] - - [26, 75.271] - - - [27392, 1536, 1, 256] - - [32, 66.59] - - - [24112, 512, 1, 256] - - [30, 57.899] - - - [28160, 256, 1, 256] - - [35, 59.165] - - - [34816, 2048, 1, 256] - - [29, 70.698] - - - [25648, 10240, 1, 256] - - [42, 63.438] - - - [20992, 10240, 1, 256] - - [26, 75.383] - - - [22528, 1281, 1, 256] - - [26, 63.117] - - - [25904, 2304, 1, 256] - - [54, 62.878] - - - [27952, 2865, 1, 256] - - [59, 61.685] - - - [30976, 768, 1, 256] - - [40, 66.489] - - - [20480, 3329, 1, 256] - - [33, 69.603] - - - [33072, 256, 1, 256] - - [26, 52.096] - - - [26624, 2560, 1, 256] - - [26, 73.237] - - - [28208, 2865, 1, 256] - - [59, 61.173] - - - [26672, 3328, 1, 256] - - [26, 61.969] - - - [26880, 2865, 1, 256] - - [55, 69.839] - - - [26112, 2304, 1, 256] - - [38, 72.648] - - - [29184, 5120, 1, 256] - - [38, 74.128] - - - [29744, 6144, 1, 256] - - [42, 62.89] - - - [30464, 3329, 1, 256] - - [71, 68.208] - - - [22272, 2560, 1, 256] - - [32, 72.039] - - - [25344, 2048, 1, 256] - - [40, 69.446] - - - [31792, 256, 1, 256] - - [45, 51.858] - - - [21248, 2816, 1, 256] - - [36, 71.886] - - - [32816, 10240, 1, 256] - - [75, 65.69] - - - [27136, 3840, 1, 256] - - [60, 74.215] - - - [34096, 10240, 1, 256] - - [42, 62.84] - - - [24576, 4608, 1, 256] - - [23, 69.32] - - - [32256, 1281, 1, 256] - - [40, 64.32] - - - [26928, 2865, 1, 256] - - [54, 60.727] - - - [20784, 7424, 1, 256] - - [37, 63.384] - - - [24112, 2816, 1, 256] - - [59, 64.168] - - - [22272, 256, 1, 256] - - [37, 49.692] - - - [30208, 1281, 1, 256] - - [70, 64.377] - - - [28720, 2816, 1, 256] - - [26, 61.361] - - - [20992, 1280, 1, 256] - - [35, 69.476] - - - [31488, 1536, 1, 256] - - [30, 70.299] - - - [21296, 8192, 1, 256] - - [54, 63.224] - - - [30512, 7168, 1, 256] - - [59, 62.559] - - - [27136, 2865, 1, 256] - - [26, 70.677] - - - [25088, 3329, 1, 256] - - [52, 69.78] - - - [29696, 3329, 1, 256] - - [38, 70.496] - - - [23040, 1280, 1, 256] - - [61, 69.267] - - - [30000, 256, 1, 256] - - [54, 48.765] - - - [20224, 3329, 1, 256] - - [38, 69.242] - - - [29232, 2816, 1, 256] - - [45, 63.791] - - - [31232, 7424, 1, 256] - - [48, 75.246] - - - [29488, 2816, 1, 256] - - [59, 63.082] - - - [25904, 2865, 1, 256] - - [37, 61.333] - - - [30512, 2816, 1, 256] - - [54, 63.905] - - - [20736, 768, 1, 256] - - [24, 65.491] - - - [20480, 256, 1, 256] - - [24, 53.911] - - - [28672, 6144, 1, 256] - - [28, 74.135] - - - [26624, 2816, 1, 256] - - [38, 73.242] - - - [28928, 768, 1, 256] - - [41, 66.936] - - - [27648, 256, 1, 256] - - [79, 58.412] - - - [32256, 6144, 1, 256] - - [48, 74.469] - - - [30720, 6144, 1, 256] - - [38, 74.708] - - - [32560, 2865, 1, 256] - - [54, 63.06] - - - [23088, 9728, 1, 256] - - [45, 63.913] - - - [22784, 9728, 1, 256] - - [29, 74.598] - - - [33024, 6144, 1, 256] - - [44, 73.922] - - - [27392, 2865, 1, 256] - - [42, 67.581] - - - [21504, 1280, 1, 256] - - [56, 68.976] - - - [30720, 6656, 1, 256] - - [28, 75.049] - - - [24880, 2865, 1, 256] - - [37, 62.105] - - - [25392, 1792, 1, 256] - - [45, 61.999] - - - [20224, 2816, 1, 256] - - [56, 71.778] - - - [20224, 256, 1, 256] - - [41, 53.677] - - - [25856, 3329, 1, 256] - - [26, 69.116] - - - [30976, 256, 1, 256] - - [41, 54.826] - - - [26880, 6144, 1, 256] - - [38, 74.125] - - - [26672, 2816, 1, 256] - - [38, 62.709] - - - [25600, 256, 1, 256] - - [81, 55.352] - - - [28160, 1281, 1, 256] - - [75, 64.386] - - - [20480, 10240, 1, 256] - - [28, 75.272] - - - [21504, 7936, 1, 256] - - [26, 75.534] - - - [20272, 7168, 1, 256] - - [45, 62.663] - - - [24880, 2816, 1, 256] - - [45, 64.271] - - - [23296, 9728, 1, 256] - - [30, 74.626] - - - [34816, 2865, 1, 256] - - [28, 71.003] - - - [31792, 2865, 1, 256] - - [30, 62.045] - - - [29488, 6144, 1, 256] - - [45, 62.997] - - - [23856, 2865, 1, 256] - - [37, 60.587] - - - [25088, 256, 1, 256] - - [54, 54.654] - - - [22016, 8960, 1, 256] - - [32, 75.62] - - - [23040, 3072, 1, 256] - - [28, 72.891] - - - [23856, 512, 1, 256] - - [26, 57.075] - - - [33792, 3329, 1, 256] - - [28, 70.407] - - - [22784, 9216, 1, 256] - - [29, 74.474] - - - [30720, 4864, 1, 256] - - [38, 74.739] - - - [32000, 8192, 1, 256] - - [28, 74.349] - - - [28160, 3329, 1, 256] - - [33, 69.566] - - - [28672, 256, 1, 256] - - [45, 51.858] - - - [27648, 1281, 1, 256] - - [29, 63.629] - - - [23808, 6144, 1, 256] - - [38, 74.136] - - - [23344, 10240, 1, 256] - - [37, 62.842] - - - [20736, 7680, 1, 256] - - [26, 74.828] - - - [33024, 9216, 1, 256] - - [43, 74.582] - - - [26160, 2816, 1, 256] - - [59, 63.826] - - - [24064, 10240, 1, 256] - - [48, 75.204] - - - [24320, 768, 1, 256] - - [73, 65.439] - - - [28208, 10240, 1, 256] - - [75, 62.743] - - - [34560, 1024, 1, 256] - - [41, 69.938] - - - [33792, 1792, 1, 256] - - [38, 72.448] - - - [30720, 2816, 1, 256] - - [30, 73.516] - - - [24624, 2816, 1, 256] - - [42, 62.429] - - - [20736, 3329, 1, 256] - - [30, 68.948] - - - [21760, 1792, 1, 256] - - [28, 70.593] - - - [21760, 8704, 1, 256] - - [28, 75.019] - - - [34608, 10240, 1, 256] - - [45, 63.342] - - - [22784, 9472, 1, 256] - - [68, 74.935] - - - [31536, 2816, 1, 256] - - [59, 63.566] - - - [27904, 4352, 1, 256] - - [60, 73.693] - - - [23552, 2865, 1, 256] - - [30, 70.547] - - - [24064, 256, 1, 256] - - [41, 52.765] - - - [34304, 2048, 1, 256] - - [42, 71.565] - - - [30464, 1280, 1, 256] - - [68, 68.821] - - - [29440, 5632, 1, 256] - - [32, 74.412] - - - [21808, 8704, 1, 256] - - [37, 63.111] - - - [30464, 6656, 1, 256] - - [25, 73.443] - - - [20736, 1024, 1, 256] - - [61, 68.087] - - - [24832, 1024, 1, 256] - - [70, 67.376] - - - [24576, 1024, 1, 256] - - [28, 64.881] - - - [29184, 2048, 1, 256] - - [75, 71.095] - - - [30976, 4864, 1, 256] - - [68, 72.422] - - - [25344, 1536, 1, 256] - - [58, 69.677] - - - [22016, 1280, 1, 256] - - [35, 69.22] - - - [32560, 8960, 1, 256] - - [54, 64.641] - - - [31536, 7936, 1, 256] - - [54, 64.014] - - - [26880, 3072, 1, 256] - - [38, 72.76] - - - [28464, 2865, 1, 256] - - [54, 61.458] - - - [20224, 6400, 1, 256] - - [28, 74.537] - - - [26624, 3328, 1, 256] - - [56, 73.411] - - - [24320, 512, 1, 256] - - [58, 63.719] - - - [34352, 768, 1, 256] - - [37, 61.293] - - - [30720, 768, 1, 256] - - [41, 67.188] - - - [34560, 10240, 1, 256] - - [38, 74.662] - - - [22016, 3328, 1, 256] - - [35, 72.823] - - - [20480, 1281, 1, 256] - - [28, 62.783] - - - [31232, 2816, 1, 256] - - [38, 72.918] - - - [31232, 6144, 1, 256] - - [31, 74.376] - - - [27136, 256, 1, 256] - - [28, 57.836] - - - [23344, 256, 1, 256] - - [28, 47.781] - - - [30208, 4352, 1, 256] - - [52, 74.258] - - - [32000, 6144, 1, 256] - - [30, 74.002] - - - [29184, 6144, 1, 256] - - [28, 74.248] - - - [29232, 5632, 1, 256] - - [59, 63.1] - - - [22576, 2816, 1, 256] - - [38, 62.212] - - - [31488, 1280, 1, 256] - - [35, 70.391] - - - [23856, 2816, 1, 256] - - [54, 63.075] - - - [29184, 2865, 1, 256] - - [58, 70.156] - - - [21248, 6144, 1, 256] - - [28, 74.121] - - - [30720, 4608, 1, 256] - - [38, 74.138] - - - [27952, 256, 1, 256] - - [30, 53.969] - - - [32512, 10240, 1, 256] - - [31, 74.736] - - - [31744, 3328, 1, 256] - - [26, 73.375] - - - [22528, 3328, 1, 256] - - [30, 73.174] - - - [34048, 3329, 1, 256] - - [74, 69.195] - - - [31744, 2816, 1, 256] - - [35, 73.414] - - - [27904, 256, 1, 256] - - [36, 58.921] - - - [21552, 256, 1, 256] - - [37, 45.306] - - - [29952, 6144, 1, 256] - - [25, 74.05] - - - [22784, 3328, 1, 256] - - [43, 72.078] - - - [20784, 256, 1, 256] - - [59, 51.079] - - - [30208, 2816, 1, 256] - - [92, 72.979] - - - [31232, 5376, 1, 256] - - [49, 74.579] - - - [30256, 256, 1, 256] - - [26, 48.965] - - - [21248, 1280, 1, 256] - - [56, 68.285] - - - [28160, 1280, 1, 256] - - [32, 70.64] - - - [30720, 3329, 1, 256] - - [38, 70.407] - - - [34560, 3329, 1, 256] - - [33, 69.597] - - - [31024, 2816, 1, 256] - - [59, 63.748] - - - [32000, 256, 1, 256] - - [59, 56.124] - - - [20528, 256, 1, 256] - - [35, 50.618] - - - [24624, 10240, 1, 256] - - [42, 63.948] - - - [21504, 7680, 1, 256] - - [26, 75.495] - - - [33536, 9728, 1, 256] - - [25, 74.661] - - - [33280, 6144, 1, 256] - - [52, 74.485] - - - [20480, 2865, 1, 256] - - [30, 70.005] - - - [30720, 1281, 1, 256] - - [28, 63.715] - - - [21760, 6144, 1, 256] - - [38, 74.339] - - - [30976, 6912, 1, 256] - - [42, 73.297] - - - [27648, 2816, 1, 256] - - [26, 73.351] - - - [20992, 3329, 1, 256] - - [30, 69.637] - - - [26672, 3072, 1, 256] - - [26, 62.488] - - - [24832, 2816, 1, 256] - - [58, 72.466] - - - [23552, 9728, 1, 256] - - [26, 75.537] - - - [26880, 1280, 1, 256] - - [50, 69.337] - - - [25088, 1280, 1, 256] - - [26, 70.249] - - - [33280, 9472, 1, 256] - - [25, 75.514] - - - [27136, 3328, 1, 256] - - [32, 73.302] - - - [28416, 2816, 1, 256] - - [60, 72.194] - - - [20480, 3328, 1, 256] - - [36, 72.255] - - - [31232, 256, 1, 256] - - [56, 55.016] - - - [33328, 9728, 1, 256] - - [59, 63.695] - - - [26416, 256, 1, 256] - - [28, 51.781] - - - [31744, 2865, 1, 256] - - [26, 70.877] - - - [22784, 6144, 1, 256] - - [25, 73.854] - - - [32000, 5888, 1, 256] - - [28, 73.811] - - - [28160, 4864, 1, 256] - - [26, 74.305] - - - [34352, 2865, 1, 256] - - [45, 61.36] - - - [29696, 256, 1, 256] - - [28, 52.782] - - - [26112, 2048, 1, 256] - - [57, 70.991] - - - [25088, 5376, 1, 256] - - [68, 74.491] - - - [29952, 3329, 1, 256] - - [38, 69.674] - - - [21296, 10240, 1, 256] - - [59, 63.537] - - - [31744, 1280, 1, 256] - - [30, 71.007] - - - [21760, 256, 1, 256] - - [38, 48.404] - - - [31488, 2048, 1, 256] - - [62, 70.757] - - - [30976, 1281, 1, 256] - - [90, 62.501] - - - [23040, 256, 1, 256] - - [85, 50.859] - - - [34304, 6144, 1, 256] - - [44, 74.528] - - - [31744, 3329, 1, 256] - - [55, 70.267] - - - [31744, 5888, 1, 256] - - [38, 74.707] - - - [29184, 1281, 1, 256] - - [75, 64.472] - - - [23856, 10240, 1, 256] - - [54, 63.188] - - - [23808, 1792, 1, 256] - - [50, 71.018] - - - [32000, 1792, 1, 256] - - [28, 71.451] - - - [26880, 2816, 1, 256] - - [36, 72.335] - - - [28416, 3328, 1, 256] - - [60, 72.264] - - - [27136, 6144, 1, 256] - - [28, 74.561] - - - [28416, 4608, 1, 256] - - [52, 72.828] - - - [33536, 1280, 1, 256] - - [35, 70.538] - - - [27440, 2865, 1, 256] - - [54, 61.881] - - - [25088, 2865, 1, 256] - - [25, 70.235] - - - [30976, 2816, 1, 256] - - [68, 70.984] - - - [26672, 10240, 1, 256] - - [42, 62.731] - - - [34048, 10240, 1, 256] - - [48, 74.096] - - - [34352, 2816, 1, 256] - - [59, 63.786] - - - [22064, 2865, 1, 256] - - [37, 61.128] - - - [28208, 4864, 1, 256] - - [45, 63.705] - - - [22528, 1280, 1, 256] - - [28, 70.17] - - - [26624, 3072, 1, 256] - - [38, 73.536] - - - [33072, 2865, 1, 256] - - [30, 62.888] - - - [22576, 256, 1, 256] - - [30, 46.957] - - - [34560, 2048, 1, 256] - - [79, 71.126] - - - [29440, 5888, 1, 256] - - [32, 73.964] - - - [34560, 1280, 1, 256] - - [26, 70.838] - - - [32000, 10240, 1, 256] - - [30, 74.514] - - - [32304, 2816, 1, 256] - - [59, 64.167] - - - [30976, 2865, 1, 256] - - [58, 68.745] - - - [30208, 6400, 1, 256] - - [48, 74.958] - - - [29232, 2865, 1, 256] - - [54, 61.737] - - - [33072, 2816, 1, 256] - - [54, 64.978] - - - [30512, 2865, 1, 256] - - [54, 61.791] - - - [20016, 2816, 1, 256] - - [59, 63.492] - - - [28416, 4352, 1, 256] - - [68, 73.602] - - - [25648, 2816, 1, 256] - - [59, 63.004] - - - [25344, 1280, 1, 256] - - [58, 70.141] - - - [24576, 10240, 1, 256] - - [39, 70.414] - - - [33024, 1281, 1, 256] - - [92, 63.336] - - - [33584, 10240, 1, 256] - - [26, 63.044] - - - [28416, 4864, 1, 256] - - [32, 73.651] - - - [23296, 3329, 1, 256] - - [33, 69.072] - - - [30464, 4352, 1, 256] - - [31, 72.6] - - - [29696, 5632, 1, 256] - - [38, 75.118] - - - [25136, 256, 1, 256] - - [28, 49.859] - - - [20528, 2865, 1, 256] - - [28, 60.85] - - - [27440, 2816, 1, 256] - - [45, 64.396] - - - [28160, 2048, 1, 256] - - [75, 70.872] - - - [24320, 2816, 1, 256] - - [60, 72.422] - - - [20736, 6144, 1, 256] - - [38, 73.931] - - - [28416, 5120, 1, 256] - - [32, 73.518] - - - [21552, 8448, 1, 256] - - [37, 64.011] - - - [20736, 1281, 1, 256] - - [41, 62.761] - - - [28464, 4864, 1, 256] - - [59, 63.713] - - - [30512, 10240, 1, 256] - - [42, 62.92] - - - [34304, 512, 1, 256] - - [58, 66.06] - - - [22784, 10240, 1, 256] - - [60, 74.611] - - - [25648, 2048, 1, 256] - - [45, 62.668] - - - [25856, 10240, 1, 256] - - [38, 74.853] - - - [32256, 8960, 1, 256] - - [25, 75.392] - - - [20736, 2865, 1, 256] - - [38, 69.727] - - - [20992, 7680, 1, 256] - - [38, 75.084] - - - [31024, 10240, 1, 256] - - [62, 62.465] - - - [26112, 256, 1, 256] - - [28, 56.263] - - - [30000, 2865, 1, 256] - - [54, 61.012] - - - [25904, 2560, 1, 256] - - [59, 63.309] - - - [24832, 768, 1, 256] - - [40, 65.797] - - - [25088, 6144, 1, 256] - - [52, 74.353] - - - [24624, 1280, 1, 256] - - [23, 61.013] - - - [22016, 8192, 1, 256] - - [38, 74.983] - - - [29952, 3328, 1, 256] - - [58, 72.644] - - - [31232, 2048, 1, 256] - - [44, 71.486] - - - [30256, 6656, 1, 256] - - [37, 63.547] - - - [20992, 2816, 1, 256] - - [38, 72.698] - - - [33792, 1536, 1, 256] - - [30, 71.307] - - - [20224, 1280, 1, 256] - - [35, 68.324] - - - [25600, 5888, 1, 256] - - [28, 74.704] - - - [26624, 768, 1, 256] - - [24, 66.916] - - - [32256, 2816, 1, 256] - - [68, 73.279] - - - [21760, 1281, 1, 256] - - [70, 63.106] - - - [25392, 10240, 1, 256] - - [59, 63.224] - - - [32768, 256, 1, 256] - - [35, 57.515] - - - [22528, 3329, 1, 256] - - [28, 70.282] - - - [23552, 3329, 1, 256] - - [28, 70.102] - - - [33024, 2865, 1, 256] - - [55, 69.197] - - - [29696, 2816, 1, 256] - - [26, 73.527] - - - [27392, 10240, 1, 256] - - [42, 74.262] - - - [23040, 2048, 1, 256] - - [70, 70.576] - - - [27648, 6144, 1, 256] - - [28, 74.648] - - - [22016, 2304, 1, 256] - - [35, 71.872] - - - [34560, 1281, 1, 256] - - [79, 64.272] - - - [27136, 1281, 1, 256] - - [34, 63.812] - - - [32000, 1281, 1, 256] - - [38, 63.609] - - - [27184, 3840, 1, 256] - - [59, 63.75] - - - [24880, 1536, 1, 256] - - [30, 61.665] - - - [28672, 768, 1, 256] - - [24, 65.658] - - - [34816, 2816, 1, 256] - - [23, 73.569] - - - [26160, 256, 1, 256] - - [45, 51.484] - - - [30464, 7168, 1, 256] - - [25, 72.343] - - - [30208, 3328, 1, 256] - - [44, 73.144] - - - [32304, 10240, 1, 256] - - [38, 63.119] - - - [26624, 1280, 1, 256] - - [36, 70.172] - - - [29696, 10240, 1, 256] - - [28, 75.448] - - - [32000, 8704, 1, 256] - - [38, 74.667] - - - [27392, 1281, 1, 256] - - [77, 62.129] - - - [26416, 2865, 1, 256] - - [37, 60.973] - - - [26160, 2560, 1, 256] - - [37, 63.321] - - - [28672, 3329, 1, 256] - - [28, 69.719] - - - [23808, 256, 1, 256] - - [60, 52.087] - - - [27184, 10240, 1, 256] - - [54, 63.078] - - - [33280, 2048, 1, 256] - - [44, 71.202] - - - [33280, 2816, 1, 256] - - [52, 73.287] - - - [23040, 9984, 1, 256] - - [28, 75.6] - - - [26112, 1280, 1, 256] - - [36, 70.457] - - - [33328, 9984, 1, 256] - - [37, 64.001] - - - [32560, 9216, 1, 256] - - [38, 63.849] - - - [22832, 9728, 1, 256] - - [59, 63.859] - - - [27904, 1280, 1, 256] - - [32, 70.022] - - - [33280, 1281, 1, 256] - - [48, 64.269] - - - [33280, 1280, 1, 256] - - [28, 70.983] - - - [32048, 256, 1, 256] - - [28, 52.152] - - - [27184, 2865, 1, 256] - - [37, 61.678] - - - [26880, 3329, 1, 256] - - [33, 69.209] - - - [20784, 7680, 1, 256] - - [54, 63.953] - - - [24832, 3329, 1, 256] - - [31, 69.453] - - - [25856, 1280, 1, 256] - - [36, 69.588] - - - [34560, 2816, 1, 256] - - [28, 72.636] - - - [20016, 256, 1, 256] - - [41, 49.813] - - - [23600, 256, 1, 256] - - [30, 48.338] - - - [22576, 9216, 1, 256] - - [42, 62.894] - - - [25344, 5632, 1, 256] - - [66, 73.279] - - - [28928, 5632, 1, 256] - - [28, 74.019] - - - [31024, 256, 1, 256] - - [36, 50.513] - - - [21552, 2865, 1, 256] - - [28, 61.492] - - - [29184, 3072, 1, 256] - - [26, 72.996] - - - [24320, 2865, 1, 256] - - [58, 70.104] - - - [20480, 6656, 1, 256] - - [28, 74.673] - - - [33536, 10240, 1, 256] - - [52, 74.79] - - - [20736, 1280, 1, 256] - - [56, 69.078] - - - [24832, 1280, 1, 256] - - [26, 69.452] - - - [29488, 10240, 1, 256] - - [38, 63.086] - - - [27392, 6144, 1, 256] - - [29, 73.273] - - - [29440, 3329, 1, 256] - - [48, 69.625] - - - [25856, 1281, 1, 256] - - [37, 62.913] - - - [34560, 768, 1, 256] - - [61, 68.277] - - - [31488, 7680, 1, 256] - - [26, 74.434] - - - [29184, 5632, 1, 256] - - [28, 74.567] - - - [32512, 512, 1, 256] - - [58, 62.875] - - - [26112, 2865, 1, 256] - - [38, 70.579] - - - [32512, 1280, 1, 256] - - [56, 70.012] - - - [20992, 1024, 1, 256] - - [41, 68.585] - - - [27904, 10240, 1, 256] - - [30, 74.585] - - - [29952, 6656, 1, 256] - - [52, 74.395] - - - [21248, 2048, 1, 256] - - [41, 70.141] - - - [34352, 256, 1, 256] - - [37, 53.882] - - - [24064, 512, 1, 256] - - [56, 63.612] - - - [32816, 2865, 1, 256] - - [29, 62.415] - - - [33840, 256, 1, 256] - - [36, 53.89] - - - [33792, 1280, 1, 256] - - [36, 71.087] - - - [21296, 7936, 1, 256] - - [45, 63.676] - - - [34096, 256, 1, 256] - - [37, 53.881] - - - [32256, 8704, 1, 256] - - [48, 75.201] - - - [30464, 1281, 1, 256] - - [60, 62.873] - - - [28464, 2816, 1, 256] - - [59, 63.429] - - - [25136, 2865, 1, 256] - - [37, 61.245] - - - [31792, 8448, 1, 256] - - [64, 64.158] - - - [24320, 4608, 1, 256] - - [58, 73.571] - - - [25088, 5120, 1, 256] - - [32, 74.212] - - - [31744, 2048, 1, 256] - - [62, 71.217] - - - [30720, 1280, 1, 256] - - [38, 70.942] - - - [34048, 256, 1, 256] - - [35, 58.331] - - - [28416, 512, 1, 256] - - [35, 61.95] - - - [22272, 10240, 1, 256] - - [48, 74.77] - - - [32512, 3328, 1, 256] - - [66, 72.772] - - - [29744, 10240, 1, 256] - - [42, 63.34] - - - [22784, 2048, 1, 256] - - [40, 70.375] - - - [23552, 2048, 1, 256] - - [42, 70.837] - - - [25344, 2816, 1, 256] - - [58, 71.619] - - - [27440, 3840, 1, 256] - - [59, 64.087] - - - [21552, 10240, 1, 256] - - [28, 63.385] - - - [21808, 256, 1, 256] - - [59, 45.844] - - - [24576, 6144, 1, 256] - - [33, 69.698] - - - [29744, 256, 1, 256] - - [45, 48.643] - - - [31488, 3328, 1, 256] - - [42, 72.631] - - - [33536, 3329, 1, 256] - - [26, 69.443] - - - [21040, 256, 1, 256] - - [24, 51.366] - - - [22272, 9216, 1, 256] - - [42, 74.35] - - - [27648, 4096, 1, 256] - - [42, 73.226] - - - [29440, 1280, 1, 256] - - [58, 70.202] - - - [31744, 7936, 1, 256] - - [28, 75.389] - - - [26624, 1281, 1, 256] - - [28, 63.119] - - - [28672, 2048, 1, 256] - - [38, 68.271] - - - [24064, 3328, 1, 256] - - [58, 72.942] - - - [25344, 3329, 1, 256] - - [48, 68.833] - - - [33280, 9728, 1, 256] - - [43, 75.157] - - - [22320, 8960, 1, 256] - - [54, 64.407] - - - [30464, 6144, 1, 256] - - [52, 72.887] - - - [34304, 2304, 1, 256] - - [32, 72.816] - - - [28928, 256, 1, 256] - - [26, 51.625] - - - [27392, 1280, 1, 256] - - [24, 67.051] - - - [26672, 2865, 1, 256] - - [30, 61.236] - - - [28720, 10240, 1, 256] - - [42, 62.448] - - - [25088, 2816, 1, 256] - - [35, 72.671] - - - [31280, 256, 1, 256] - - [54, 50.995] - - - [29488, 5888, 1, 256] - - [59, 63.407] - - - [30720, 2048, 1, 256] - - [75, 70.334] - - - [21808, 10240, 1, 256] - - [54, 63.114] - - - [24576, 2865, 1, 256] - - [55, 65.98] - - - [23808, 1280, 1, 256] - - [24, 69.635] - - - [33280, 1024, 1, 256] - - [73, 69.675] - - - [25856, 256, 1, 256] - - [91, 55.701] - - - [25648, 2304, 1, 256] - - [54, 63.621] - - - [29952, 2865, 1, 256] - - [26, 70.007] - - - [23040, 1024, 1, 256] - - [61, 66.825] - - - [34304, 3328, 1, 256] - - [43, 73.509] - - - [31792, 8192, 1, 256] - - [26, 63.352] - - - [24576, 2816, 1, 256] - - [39, 68.802] - - - [27648, 1536, 1, 256] - - [30, 70.642] - - - [23296, 9472, 1, 256] - - [28, 75.084] - - - [24624, 256, 1, 256] - - [54, 49.754] - - - [20736, 2048, 1, 256] - - [64, 70.538] - - - [28720, 5376, 1, 256] - - [42, 62.018] - - - [20480, 512, 1, 256] - - [61, 62.552] - - - [33840, 2865, 1, 256] - - [26, 62.278] - - - [24064, 2865, 1, 256] - - [48, 70.099] - - - [24064, 2816, 1, 256] - - [32, 72.665] - - - [20992, 256, 1, 256] - - [35, 55.358] - - - [33328, 256, 1, 256] - - [37, 53.917] - - - [28928, 5120, 1, 256] - - [26, 73.722] - - - [34304, 256, 1, 256] - - [35, 58.916] - - - [34304, 1281, 1, 256] - - [62, 64.906] - - - [31744, 1281, 1, 256] - - [29, 64.217] - - - [33584, 2816, 1, 256] - - [54, 63.774] - - - [24064, 4352, 1, 256] - - [60, 74.349] - - - [20224, 6912, 1, 256] - - [26, 74.537] - - - [21504, 1281, 1, 256] - - [77, 62.721] - - - [33536, 3328, 1, 256] - - [42, 72.67] - - - [34816, 3328, 1, 256] - - [38, 73.475] - - - [31024, 7680, 1, 256] - - [59, 63.222] - - - [22016, 3329, 1, 256] - - [30, 69.596] - - - [25344, 1281, 1, 256] - - [40, 63.325] - - - [31744, 7680, 1, 256] - - [28, 75.385] - - - [27952, 10240, 1, 256] - - [45, 63.351] - - - [23808, 2048, 1, 256] - - [61, 70.427] - - - [32768, 2816, 1, 256] - - [39, 59.065] - - - [34816, 256, 1, 256] - - [26, 59.726] - - - [27904, 2865, 1, 256] - - [32, 69.642] - - - [31232, 1280, 1, 256] - - [56, 70.796] - - - [22016, 1281, 1, 256] - - [73, 63.18] - - - [22528, 8704, 1, 256] - - [26, 75.669] - - - [22528, 9216, 1, 256] - - [28, 74.957] - - - [34816, 1280, 1, 256] - - [26, 71.284] - - - [23808, 10240, 1, 256] - - [28, 75.037] - - - [32512, 2048, 1, 256] - - [43, 70.782] - - - [34816, 1024, 1, 256] - - [24, 70.048] - - - [34048, 2048, 1, 256] - - [92, 70.237] - - - [30768, 2816, 1, 256] - - [26, 61.946] - - - [22272, 3329, 1, 256] - - [52, 68.958] - - - [25600, 3328, 1, 256] - - [35, 73.271] - - - [34048, 2816, 1, 256] - - [58, 72.312] - - - [22064, 8704, 1, 256] - - [37, 63.169] - - - [25648, 256, 1, 256] - - [59, 50.986] - - - [22784, 768, 1, 256] - - [41, 66.495] - - - [27904, 2048, 1, 256] - - [24, 70.333] - - - [22528, 9472, 1, 256] - - [28, 75.892] - - - [21504, 2865, 1, 256] - - [28, 70.242] - - - [28672, 5376, 1, 256] - - [28, 74.327] - - - [22576, 9472, 1, 256] - - [26, 62.863] - - - [24576, 256, 1, 256] - - [26, 53.728] - - - [28672, 5120, 1, 256] - - [28, 74.274] - - - [24576, 3328, 1, 256] - - [23, 68.462] - - - [32816, 9472, 1, 256] - - [29, 66.519] - - - [27440, 256, 1, 256] - - [30, 53.249] - - - [22272, 8704, 1, 256] - - [52, 74.842] - - - [30000, 2816, 1, 256] - - [54, 63.72] - - - [26928, 2816, 1, 256] - - [59, 63.237] - - - [22064, 2816, 1, 256] - - [59, 63.866] - - - [23552, 3328, 1, 256] - - [28, 73.147] - - - [28416, 256, 1, 256] - - [61, 51.395] - - - [28928, 6144, 1, 256] - - [26, 73.774] - - - [32768, 512, 1, 256] - - [30, 57.901] - - - [22272, 2865, 1, 256] - - [38, 69.352] - - - [26928, 256, 1, 256] - - [26, 52.393] - - - [21760, 10240, 1, 256] - - [28, 74.893] - - - [26368, 512, 1, 256] - - [56, 62.805] - - - [26672, 256, 1, 256] - - [54, 52.318] - - - [33328, 2865, 1, 256] - - [45, 62.95] - - - [30720, 3328, 1, 256] - - [35, 73.371] - - - [25856, 2865, 1, 256] - - [30, 69.722] - - - [25088, 3328, 1, 256] - - [68, 72.895] - - - [28416, 2560, 1, 256] - - [58, 72.27] - - - [33536, 9472, 1, 256] - - [25, 74.945] - - - [20480, 1280, 1, 256] - - [56, 68.503] - - - [30208, 6144, 1, 256] - - [31, 74.374] - - - [34864, 1024, 1, 256] - - [26, 61.22] - - - [33280, 256, 1, 256] - - [36, 56.664] - - - [23296, 3328, 1, 256] - - [38, 72.158] - - - [32560, 256, 1, 256] - - [64, 52.359] - - - [32560, 2816, 1, 256] - - [45, 64.998] - - - [33536, 256, 1, 256] - - [41, 58.138] - - - [34608, 768, 1, 256] - - [54, 60.915] - - - [24832, 5120, 1, 256] - - [26, 73.988] - - - [25856, 2048, 1, 256] - - [79, 70.029] - - - [30768, 256, 1, 256] - - [37, 50.495] - - - [30000, 6656, 1, 256] - - [45, 63.377] - - - [24320, 1024, 1, 256] - - [70, 68.601] - - - [33280, 9216, 1, 256] - - [43, 74.887] - - - [31488, 5376, 1, 256] - - [28, 73.859] - - - [28416, 1281, 1, 256] - - [73, 62.992] - - - [27392, 3584, 1, 256] - - [42, 71.904] - - - [26368, 2048, 1, 256] - - [75, 70.201] - - - [22528, 256, 1, 256] - - [36, 50.074] - - - [32768, 2048, 1, 256] - - [23, 56.064] - - - [30256, 6912, 1, 256] - - [37, 63.907] - - - [28672, 512, 1, 256] - - [35, 62.915] - - - [21760, 8448, 1, 256] - - [26, 74.631] - - - [34560, 6144, 1, 256] - - [30, 74.161] - - - [27696, 2816, 1, 256] - - [45, 63.638] - - - [29952, 2048, 1, 256] - - [24, 70.524] - - - [22576, 10240, 1, 256] - - [42, 62.743] - - - [25600, 1792, 1, 256] - - [35, 71.951] - - - [28976, 10240, 1, 256] - - [54, 62.845] - - - [29952, 1280, 1, 256] - - [56, 70.288] - - - [26368, 2816, 1, 256] - - [35, 72.464] - - - [26416, 3072, 1, 256] - - [59, 62.558] - - - [27648, 3329, 1, 256] - - [38, 70.276] - - - [34560, 2560, 1, 256] - - [30, 72.91] - - - [32048, 8448, 1, 256] - - [54, 63.167] - - - [30464, 2865, 1, 256] - - [74, 68.347] - - - [34048, 3328, 1, 256] - - [44, 72.364] - - - [23808, 2865, 1, 256] - - [26, 69.84] - - - [25600, 2816, 1, 256] - - [26, 73.099] - - - [20736, 6912, 1, 256] - - [26, 74.702] - - - [24576, 512, 1, 256] - - [59, 62.802] - - - [33792, 256, 1, 256] - - [54, 58.332] - - - [22576, 2865, 1, 256] - - [38, 60.853] - - - [30464, 256, 1, 256] - - [35, 54.235] - - - [24368, 2816, 1, 256] - - [59, 63.888] - - - [20224, 512, 1, 256] - - [24, 62.059] - - - [30512, 6912, 1, 256] - - [59, 64.4] - - - [20272, 2816, 1, 256] - - [37, 63.653] - - - [23296, 256, 1, 256] - - [30, 51.395] - - - [27904, 2816, 1, 256] - - [58, 72.231] - - - [29184, 1280, 1, 256] - - [58, 70.572] - - - [24112, 10240, 1, 256] - - [37, 63.496] - - - [31280, 7680, 1, 256] - - [42, 63.157] - - - [24064, 6144, 1, 256] - - [31, 74.387] - - - [26624, 6144, 1, 256] - - [30, 74.59] - - - [30768, 2865, 1, 256] - - [30, 61.038] - - - [20528, 2816, 1, 256] - - [26, 61.315] - - - [25392, 2865, 1, 256] - - [30, 61.179] - - - [22272, 6144, 1, 256] - - [31, 73.834] - - - [25088, 10240, 1, 256] - - [31, 75.251] - - - [25344, 2865, 1, 256] - - [32, 69.458] - - - [23552, 1792, 1, 256] - - [56, 71.439] - - - [23296, 3584, 1, 256] - - [35, 73.272] - - - [28160, 2816, 1, 256] - - [68, 72.65] - - - [20272, 2865, 1, 256] - - [30, 60.656] - - - [22832, 9472, 1, 256] - - [45, 64.024] - - - [21760, 7936, 1, 256] - - [30, 74.725] - - - [26928, 3328, 1, 256] - - [54, 63.435] - - - [33072, 9472, 1, 256] - - [45, 64.768] - - - [33024, 1280, 1, 256] - - [35, 69.175] - - - [34352, 512, 1, 256] - - [37, 60.082] - - - [26368, 2865, 1, 256] - - [28, 70.011] - - - [27952, 4352, 1, 256] - - [54, 63.569] - - - [21504, 8192, 1, 256] - - [28, 75.141] - - - [22320, 9216, 1, 256] - - [45, 63.672] - - - [31232, 2865, 1, 256] - - [28, 70.381] - - - [21248, 7680, 1, 256] - - [38, 74.765] - - - [24368, 256, 1, 256] - - [28, 49.499] - - - [25648, 2865, 1, 256] - - [28, 60.967] - - - [21248, 2865, 1, 256] - - [28, 69.655] - - - [28416, 2865, 1, 256] - - [26, 69.557] - - - [24320, 3329, 1, 256] - - [52, 69.541] - - - [27648, 2048, 1, 256] - - [87, 71.036] - - - [27648, 2865, 1, 256] - - [30, 70.739] - - - [26880, 2048, 1, 256] - - [45, 70.716] - - - [28672, 2560, 1, 256] - - [38, 72.702] - - - [24064, 1280, 1, 256] - - [56, 69.177] - - - [30256, 2865, 1, 256] - - [37, 61.639] - - - [22064, 10240, 1, 256] - - [28, 63.217] - - - [30464, 4608, 1, 256] - - [48, 72.193] - - - [22016, 6144, 1, 256] - - [30, 74.548] - - - [29440, 2816, 1, 256] - - [32, 72.647] - - - [25392, 2048, 1, 256] - - [45, 62.586] - - - [20992, 2048, 1, 256] - - [59, 70.296] - - - [33024, 3329, 1, 256] - - [33, 69.227] - - - [20224, 3328, 1, 256] - - [35, 71.953] - - - [28208, 4608, 1, 256] - - [37, 63.809] - - - [25344, 6144, 1, 256] - - [29, 72.892] - - - [30464, 512, 1, 256] - - [73, 63.734] - - - [21248, 3329, 1, 256] - - [28, 69.165] - - - [29696, 6144, 1, 256] - - [30, 74.778] - - - [20992, 7936, 1, 256] - - [26, 75.164] - - - [33024, 9472, 1, 256] - - [66, 74.823] - - - [32000, 3329, 1, 256] - - [55, 69.733] - - - [21248, 1281, 1, 256] - - [59, 63.451] - - - [24624, 1024, 1, 256] - - [29, 59.008] - - - [22272, 2816, 1, 256] - - [60, 71.99] - - - [29440, 1281, 1, 256] - - [40, 64.108] - - - [30464, 6400, 1, 256] - - [65, 73.497] - - - [25136, 10240, 1, 256] - - [54, 63.177] - - - [23040, 9472, 1, 256] - - [52, 75.599] - - - [33840, 2816, 1, 256] - - [64, 63.323] - - - [30976, 1024, 1, 256] - - [73, 68.274] - - - [34048, 6144, 1, 256] - - [52, 73.573] - - - [32000, 2048, 1, 256] - - [59, 69.97] - - - [32048, 2865, 1, 256] - - [59, 61.728] - - - [33328, 10240, 1, 256] - - [30, 63.577] - - - [25088, 1536, 1, 256] - - [35, 69.981] - - - [30512, 256, 1, 256] - - [45, 49.716] - - - [20480, 6912, 1, 256] - - [26, 74.992] - - - [34608, 2816, 1, 256] - - [45, 63.396] - - - [22064, 256, 1, 256] - - [28, 46.283] - - - [25600, 2865, 1, 256] - - [30, 70.481] - - - [26880, 1024, 1, 256] - - [24, 68.001] - - - [27392, 2048, 1, 256] - - [79, 69.256] - - - [30208, 10240, 1, 256] - - [48, 75.228] - - - [20016, 10240, 1, 256] - - [59, 63.271] - - - [26880, 10240, 1, 256] - - [38, 74.707] - - - [28160, 3328, 1, 256] - - [42, 72.965] - - - [33536, 2048, 1, 256] - - [85, 70.838] - - - [31232, 7936, 1, 256] - - [48, 75.186] - - - [31536, 10240, 1, 256] - - [45, 62.995] - - - [24832, 1536, 1, 256] - - [56, 69.465] - - - [32768, 768, 1, 256] - - [23, 56.214] - - - [29440, 6144, 1, 256] - - [52, 74.249] - - - [26112, 2560, 1, 256] - - [38, 73.095] - - - [33792, 6144, 1, 256] - - [38, 74.818] - - - [22528, 10240, 1, 256] - - [26, 75.441] - - - [20480, 768, 1, 256] - - [61, 64.292] - - - [22320, 256, 1, 256] - - [45, 46.987] - - - [23808, 3328, 1, 256] - - [28, 72.471] - - - [28464, 256, 1, 256] - - [56, 46.705] - - - [27136, 2048, 1, 256] - - [49, 70.97] - - - [29744, 6400, 1, 256] - - [54, 63.787] - - - [20480, 7168, 1, 256] - - [30, 73.728] - - - [22832, 256, 1, 256] - - [54, 47.124] - - - [21552, 8192, 1, 256] - - [28, 63.331] - - - [25856, 2560, 1, 256] - - [38, 72.187] - - - [28160, 6144, 1, 256] - - [42, 74.237] - - - [31280, 2816, 1, 256] - - [37, 63.623] - - - [23600, 10240, 1, 256] - - [29, 63.677] - - - [26368, 1281, 1, 256] - - [61, 63.562] - - - [24576, 1280, 1, 256] - - [26, 65.76] - - - [33536, 1536, 1, 256] - - [36, 70.649] - - - [23088, 2816, 1, 256] - - [54, 63.162] - - - [26624, 2048, 1, 256] - - [62, 70.108] - - - [29952, 2816, 1, 256] - - [26, 72.497] - - - [21760, 2048, 1, 256] - - [64, 70.422] - - - [30976, 6144, 1, 256] - - [29, 72.357] - - - [29696, 1280, 1, 256] - - [35, 70.63] - - - [30208, 4096, 1, 256] - - [42, 73.684] - - - [24832, 2865, 1, 256] - - [26, 69.873] - - - [31488, 1281, 1, 256] - - [41, 64.289] - - - [34304, 2865, 1, 256] - - [52, 70.546] - - - [32512, 256, 1, 256] - - [59, 56.034] - - - [25136, 1536, 1, 256] - - [28, 61.602] - - - [26112, 3329, 1, 256] - - [55, 70.013] - - - [24880, 1280, 1, 256] - - [54, 63.141] - - - [28208, 2816, 1, 256] - - [37, 63.499] - - - [29184, 5888, 1, 256] - - [48, 74.176] - - - [28160, 4352, 1, 256] - - [31, 74.11] - - - [34352, 10240, 1, 256] - - [29, 62.689] - - - [23856, 256, 1, 256] - - [45, 48.368] - - - [25344, 10240, 1, 256] - - [29, 73.908] - - - [20992, 1281, 1, 256] - - [34, 63.172] - - - [26624, 512, 1, 256] - - [35, 63.853] - - - [21040, 10240, 1, 256] - - [38, 62.806] - - - [23040, 3328, 1, 256] - - [79, 72.72] - - - [30976, 7168, 1, 256] - - [42, 72.418] - - - [25856, 2304, 1, 256] - - [26, 71.728] - - - [24368, 1024, 1, 256] - - [45, 63.2] - - - [33280, 2865, 1, 256] - - [52, 70.866] - - - [23296, 1536, 1, 256] - - [35, 69.712] - - - [21504, 6144, 1, 256] - - [28, 74.607] - - - [23552, 2816, 1, 256] - - [28, 73.152] - - - [30464, 2816, 1, 256] - - [32, 71.467] - - - [22832, 2865, 1, 256] - - [37, 61.219] - - - [24576, 2048, 1, 256] - - [28, 64.874] - - - [22272, 8448, 1, 256] - - [52, 74.457] - - - [32256, 1280, 1, 256] - - [68, 71.001] - - - [25856, 5888, 1, 256] - - [28, 73.834] - - - [30976, 5120, 1, 256] - - [42, 72.377] - - - [29184, 3329, 1, 256] - - [74, 69.841] - - - [24112, 2865, 1, 256] - - [59, 61.649] - - - [29744, 2816, 1, 256] - - [54, 63.337] - - - [21760, 2816, 1, 256] - - [36, 72.102] - - - [25600, 2048, 1, 256] - - [42, 71.033] - - - [32000, 1280, 1, 256] - - [35, 70.416] - - - [25856, 3328, 1, 256] - - [42, 72.256] - - - [20016, 6656, 1, 256] - - [54, 64.154] - - - [32256, 2865, 1, 256] - - [33, 70.537] - - - [22272, 3328, 1, 256] - - [68, 72.025] - - - [21504, 3328, 1, 256] - - [26, 72.919] - - - [31232, 5120, 1, 256] - - [25, 74.185] - - - [24112, 256, 1, 256] - - [28, 48.82] - - - [30208, 1280, 1, 256] - - [56, 70.586] - - - [22064, 8960, 1, 256] - - [59, 64.15] - - - [28160, 10240, 1, 256] - - [30, 74.999] - - - [21504, 1536, 1, 256] - - [50, 69.413] - - - [31744, 5632, 1, 256] - - [23, 74.894] - - - [20272, 6912, 1, 256] - - [59, 63.868] - - - [29952, 1792, 1, 256] - - [26, 71.606] - - - [25904, 10240, 1, 256] - - [29, 62.643] - - - [25344, 1792, 1, 256] - - [68, 70.475] - - - [32512, 8448, 1, 256] - - [66, 74.783] - - - [25088, 2048, 1, 256] - - [59, 70.772] - - - [23808, 9984, 1, 256] - - [30, 75.123] - - - [32768, 3329, 1, 256] - - [104, 56.152] - - - [34816, 6144, 1, 256] - - [38, 74.652] - - - [32256, 256, 1, 256] - - [32, 56.472] - - - [26368, 3328, 1, 256] - - [28, 72.308] - - - [23296, 1280, 1, 256] - - [50, 69.319] - - - [34608, 1024, 1, 256] - - [45, 62.214] - - - [30976, 1280, 1, 256] - - [60, 69.531] - - - [22528, 6144, 1, 256] - - [28, 74.61] - - - [21248, 10240, 1, 256] - - [28, 74.773] - - - [22528, 2865, 1, 256] - - [38, 70.368] - - - [22528, 768, 1, 256] - - [41, 66.36] - - - [22016, 8704, 1, 256] - - [48, 75.464] - - - [30720, 6912, 1, 256] - - [28, 75.396] - - - [33024, 2048, 1, 256] - - [43, 70.653] - - - [31232, 3329, 1, 256] - - [38, 69.855] - - - [33024, 3328, 1, 256] - - [44, 72.644] - - - [30976, 7424, 1, 256] - - [42, 73.418] - - - [27136, 3584, 1, 256] - - [26, 73.896] - - - [34048, 1280, 1, 256] - - [60, 70.127] - - - [34864, 1280, 1, 256] - - [26, 62.169] - - - [25600, 2304, 1, 256] - - [30, 72.842] - - - [21760, 3329, 1, 256] - - [30, 69.145] - - - [26928, 3584, 1, 256] - - [54, 63.394] - - - [28976, 2816, 1, 256] - - [54, 63.623] - - - [24832, 4864, 1, 256] - - [58, 74.103] - - - [21248, 1536, 1, 256] - - [35, 68.786] - - - [23808, 2816, 1, 256] - - [35, 72.377] - - - [32768, 9472, 1, 256] - - [23, 59.284] - - - [27392, 3328, 1, 256] - - [44, 71.508] - - - [26880, 3584, 1, 256] - - [26, 73.397] - - - [23552, 1281, 1, 256] - - [75, 63.035] - - - [27648, 3840, 1, 256] - - [38, 74.357] - - - [22016, 10240, 1, 256] - - [48, 75.337] - - - [34816, 2560, 1, 256] - - [26, 73.621] - - - [31536, 256, 1, 256] - - [54, 50.925] - - - [34816, 10240, 1, 256] - - [28, 75.289] - - - [27904, 1792, 1, 256] - - [58, 71.214] - - - [33792, 10240, 1, 256] - - [28, 75.354] - - - [23296, 2816, 1, 256] - - [28, 72.116] - - - [31024, 7424, 1, 256] - - [37, 63.587] - - - [22784, 1280, 1, 256] - - [35, 68.723] - - - [30976, 2048, 1, 256] - - [94, 68.691] - - - [27392, 4096, 1, 256] - - [44, 72.243] - - - [33792, 2816, 1, 256] - - [38, 73.562] - - - [32560, 10240, 1, 256] - - [38, 64.059] - - - [20736, 7424, 1, 256] - - [30, 74.802] - - - [28672, 2865, 1, 256] - - [26, 70.166] - - - [31488, 256, 1, 256] - - [24, 55.599] - - - [20992, 7424, 1, 256] - - [28, 75.243] - - - [21504, 1792, 1, 256] - - [56, 71.135] - - - [27696, 2865, 1, 256] - - [38, 61.694] - - - [33024, 1024, 1, 256] - - [24, 67.629] - - - [22016, 256, 1, 256] - - [59, 49.001] - - - [23088, 256, 1, 256] - - [59, 47.102] - - - [28976, 256, 1, 256] - - [56, 47.864] - - - [27392, 256, 1, 256] - - [37, 57.922] - - - [34304, 3329, 1, 256] - - [48, 69.933] - - - [32512, 9216, 1, 256] - - [44, 74.559] - - - [31488, 3329, 1, 256] - - [55, 69.544] - - - [20016, 2865, 1, 256] - - [37, 61.493] - - - [22016, 8448, 1, 256] - - [52, 75.105] - - - [31024, 2865, 1, 256] - - [54, 61.593] - - - [29440, 256, 1, 256] - - [61, 52.381] - - - [34608, 2865, 1, 256] - - [37, 61.135] - - - [20480, 2048, 1, 256] - - [28, 67.869] - - - [28160, 2865, 1, 256] - - [55, 69.887] - - - [28416, 2304, 1, 256] - - [56, 71.87] - - - [23552, 6144, 1, 256] - - [26, 74.767] - - - [21296, 256, 1, 256] - - [45, 44.768] - - - [28672, 4864, 1, 256] - - [26, 74.407] - - - [27648, 1792, 1, 256] - - [38, 72.137] - - - [31488, 7424, 1, 256] - - [38, 74.459] - - - [23040, 2865, 1, 256] - - [31, 69.926] - - - [30976, 3328, 1, 256] - - [32, 71.027] - - - [25856, 1792, 1, 256] - - [35, 71.058] - - - [33536, 9984, 1, 256] - - [52, 74.907] - - - [24832, 1281, 1, 256] - - [70, 63.319] - - - [29184, 3328, 1, 256] - - [62, 72.913] - - - [32000, 2816, 1, 256] - - [28, 72.454] - - - [34304, 768, 1, 256] - - [40, 68.64] - - - [24576, 1281, 1, 256] - - [28, 58.687] - - - [25088, 1281, 1, 256] - - [40, 63.796] - - - [29744, 2865, 1, 256] - - [54, 61.455] - - - [25136, 2816, 1, 256] - - [59, 63.662] - - - [29696, 1281, 1, 256] - - [29, 64.089] - - - [27392, 3329, 1, 256] - - [85, 67.992] - - - [31488, 2816, 1, 256] - - [38, 72.59] - - - [30976, 10240, 1, 256] - - [29, 73.744] - - - [26624, 3329, 1, 256] - - [26, 70.492] - - - [34304, 1280, 1, 256] - - [60, 71.143] - - - [25392, 256, 1, 256] - - [54, 50.845] - - - [26624, 10240, 1, 256] - - [30, 75.441] - - - [26112, 6144, 1, 256] - - [26, 74.593] - - - [29696, 3328, 1, 256] - - [30, 73.404] - - - [32304, 2865, 1, 256] - - [59, 62.87] - - - [24368, 2865, 1, 256] - - [45, 62.435] - - - [31488, 8192, 1, 256] - - [38, 74.293] - - - [20224, 6656, 1, 256] - - [26, 74.229] - - - [31232, 1281, 1, 256] - - [34, 64.494] - - - [21296, 2865, 1, 256] - - [54, 61.548] - - - [24112, 768, 1, 256] - - [45, 58.71] - - - [32000, 8448, 1, 256] - - [28, 74.291] - - - [23552, 1536, 1, 256] - - [30, 69.852] - - - [30976, 7680, 1, 256] - - [76, 72.959] - - - [31280, 10240, 1, 256] - - [42, 63.142] - - - [23344, 9984, 1, 256] - - [45, 63.465] - - - [21248, 8192, 1, 256] - - [30, 74.558] - - - [29696, 6400, 1, 256] - - [28, 75.374] - - - [32304, 8960, 1, 256] - - [59, 64.13] - - - [27184, 256, 1, 256] - - [59, 52.727] - - - [28464, 10240, 1, 256] - - [42, 62.755] - - - [20736, 256, 1, 256] - - [28, 54.594] - - - [31232, 10240, 1, 256] - - [52, 75.057] - - - [25856, 6144, 1, 256] - - [38, 74.009] - - - [27440, 10240, 1, 256] - - [38, 62.863] - - - [23088, 2865, 1, 256] - - [37, 61.571] - - - [29696, 3584, 1, 256] - - [26, 74.288] - - - [23040, 9728, 1, 256] - - [66, 75.267] - - - [31744, 10240, 1, 256] - - [28, 75.225] - - - [31744, 1792, 1, 256] - - [36, 72.252] - - - [24320, 256, 1, 256] - - [45, 53.207] - - - [27696, 256, 1, 256] - - [59, 53.65] - - - [29696, 2865, 1, 256] - - [28, 70.8] - - - [22784, 3072, 1, 256] - - [28, 72.397] - - - [29952, 5888, 1, 256] - - [60, 73.889] - - - [28928, 2816, 1, 256] - - [28, 72.431] - - - [30768, 7424, 1, 256] - - [38, 62.446] - - - [27440, 4096, 1, 256] - - [54, 63.269] - - - [24064, 4096, 1, 256] - - [92, 73.037] - - - [32256, 3329, 1, 256] - - [31, 70.183] - - - [30976, 3329, 1, 256] - - [60, 68.061] - - - [25600, 10240, 1, 256] - - [26, 75.488] - - - [20224, 6144, 1, 256] - - [26, 73.983] - - - [21040, 7936, 1, 256] - - [41, 64.061] - - - [26368, 2560, 1, 256] - - [50, 72.493] - - - [32512, 1281, 1, 256] - - [92, 63.536] - - - [28928, 3072, 1, 256] - - [28, 72.55] - - - [34864, 2865, 1, 256] - - [28, 61.273] - - - [23552, 9984, 1, 256] - - [28, 75.93] - - - [21040, 2865, 1, 256] - - [54, 61.177] - - - [34048, 1281, 1, 256] - - [90, 63.624] - - - [23296, 10240, 1, 256] - - [30, 74.732] - - - [32768, 6144, 1, 256] - - [23, 58.987] - - - [25904, 2816, 1, 256] - - [45, 63.446] - - - [31232, 1024, 1, 256] - - [61, 69.645] - - - [27648, 3328, 1, 256] - - [26, 73.377] - - - [34864, 256, 1, 256] - - [45, 54.082] - - - [21248, 256, 1, 256] - - [28, 47.768] - - - [26416, 10240, 1, 256] - - [54, 62.855] - - - [27184, 3584, 1, 256] - - [59, 63.863] - - - [23296, 2048, 1, 256] - - [54, 69.765] - - - [34048, 512, 1, 256] - - [36, 65.439] - - - [21760, 2865, 1, 256] - - [26, 69.561] - - - [28672, 2816, 1, 256] - - [26, 72.765] - - - [28672, 4608, 1, 256] - - [26, 73.631] - - - [34560, 512, 1, 256] - - [36, 66.38] - - - [32768, 2865, 1, 256] - - [104, 56.916] - - - [30208, 6912, 1, 256] - - [48, 75.246] - - - [32512, 6144, 1, 256] - - [31, 74.065] - - - [24832, 3328, 1, 256] - - [58, 72.518] - - - [27392, 2816, 1, 256] - - [77, 70.787] - - - [32768, 8704, 1, 256] - - [39, 59.402] - - - [23552, 10240, 1, 256] - - [28, 75.547] - - - [32816, 9216, 1, 256] - - [29, 66.618] - - - [33024, 10240, 1, 256] - - [44, 74.697] - - - [34608, 256, 1, 256] - - [26, 54.121] - - - [20736, 3328, 1, 256] - - [30, 72.065] - - - [31232, 7680, 1, 256] - - [48, 75.166] - - - [22528, 512, 1, 256] - - [35, 60.959] - - - [30208, 2865, 1, 256] - - [38, 70.349] - - - [22272, 2304, 1, 256] - - [35, 71.492] - - - [32512, 2816, 1, 256] - - [25, 72.51] - - - [31488, 7936, 1, 256] - - [38, 74.414] - - - [28416, 2048, 1, 256] - - [49, 69.489] - - - [22784, 3329, 1, 256] - - [30, 69.258] - - - [23040, 2816, 1, 256] - - [36, 72.336] - - - [24320, 3328, 1, 256] - - [60, 72.697] - - - [24064, 1281, 1, 256] - - [40, 64.011] - - - [33072, 9728, 1, 256] - - [54, 64.079] - - - [29440, 10240, 1, 256] - - [31, 74.925] - - - [30208, 6656, 1, 256] - - [31, 74.77] - - - [32768, 3328, 1, 256] - - [23, 58.272] - - - [28416, 6144, 1, 256] - - [28, 73.683] - - - [27904, 4608, 1, 256] - - [29, 73.253] - - - [27184, 2816, 1, 256] - - [59, 63.994] - - - [29184, 1024, 1, 256] - - [40, 69.176] - - - [31744, 1536, 1, 256] - - [56, 71.1] - - - [28416, 10240, 1, 256] - - [48, 74.509] - - - [24368, 10240, 1, 256] - - [30, 63.847] - - - [27904, 3329, 1, 256] - - [31, 69.138] - - - [25344, 3328, 1, 256] - - [60, 71.605] - - - [29952, 6400, 1, 256] - - [25, 74.528] - - - [29440, 2048, 1, 256] - - [75, 70.786] - - - [28928, 1281, 1, 256] - - [34, 63.378] - - - [30208, 3329, 1, 256] - - [38, 69.876] - - - [23088, 9984, 1, 256] - - [59, 64.288] - - - [29184, 2816, 1, 256] - - [25, 72.862] - - - [22528, 2560, 1, 256] - - [28, 72.954] - - - [33328, 2816, 1, 256] - - [59, 64.013] - - - [26368, 256, 1, 256] - - [45, 56.773] - - - [22832, 10240, 1, 256] - - [59, 63.212] - - - [31792, 2816, 1, 256] - - [41, 63.597] - - - [24832, 2048, 1, 256] - - [73, 70.487] - - - [24880, 256, 1, 256] - - [30, 49.526] - - - [33840, 10240, 1, 256] - - [42, 63.711] - - - [33584, 9984, 1, 256] - - [45, 63.613] - - - [28672, 10240, 1, 256] - - [28, 74.999] - - - [24832, 256, 1, 256] - - [58, 53.701] - - - [31488, 2865, 1, 256] - - [38, 70.05] - - - [30720, 7424, 1, 256] - - [30, 75.416] - - - [33536, 2816, 1, 256] - - [58, 72.7] - - - [30000, 6400, 1, 256] - - [59, 63.97] - - - [20224, 1281, 1, 256] - - [61, 63.716] - - - [22832, 2816, 1, 256] - - [54, 63.64] - - - [25600, 6144, 1, 256] - - [28, 74.742] - - - [24320, 4352, 1, 256] - - [58, 73.984] - - - [32768, 10240, 1, 256] - - [67, 58.859] - - - [26880, 768, 1, 256] - - [41, 66.503] - - - [24576, 3329, 1, 256] - - [33, 65.155] - - - [27904, 3840, 1, 256] - - [58, 73.388] - - - [30256, 2816, 1, 256] - - [54, 63.427] - - - [23296, 1281, 1, 256] - - [41, 62.497] - - - [26880, 256, 1, 256] - - [37, 57.446] - - - [23344, 2816, 1, 256] - - [37, 63.552] - - - [33792, 2048, 1, 256] - - [87, 71.385] - - - [21504, 3329, 1, 256] - - [30, 70.177] - - - [20272, 256, 1, 256] - - [61, 50.146] - - - [32768, 1280, 1, 256] - - [39, 56.265] - - - [32256, 10240, 1, 256] - - [52, 75.088] - - - [27952, 2816, 1, 256] - - [59, 63.766] - - - [28928, 5376, 1, 256] - - [48, 73.7] - - - [20992, 6144, 1, 256] - - [30, 74.373] - - - [20224, 2048, 1, 256] - - [24, 69.842] - - - [33280, 10240, 1, 256] - - [28, 75.116] - - - [24064, 3329, 1, 256] - - [52, 69.727] - - - [32768, 9216, 1, 256] - - [67, 58.928] - - - [20016, 6912, 1, 256] - - [54, 64.096] - - - [22320, 10240, 1, 256] - - [45, 63.577] - - - [22784, 256, 1, 256] - - [35, 50.256] - - - [34816, 512, 1, 256] - - [30, 66.676] - - - [32048, 8704, 1, 256] - - [54, 63.329] - - - [29232, 5888, 1, 256] - - [37, 63.527] - - - [24064, 768, 1, 256] - - [64, 65.425] - - - [33792, 9984, 1, 256] - - [30, 75.596] - - - [32512, 3329, 1, 256] - - [74, 69.597] - - - [21504, 2048, 1, 256] - - [42, 70.357] - - - [28160, 2304, 1, 256] - - [28, 72.19] - - - [20784, 10240, 1, 256] - - [38, 63.029] - - - [20224, 7168, 1, 256] - - [30, 73.51] - - - [28976, 2865, 1, 256] - - [45, 61.422] - - - [21296, 2816, 1, 256] - - [45, 63.29] - - - [23552, 256, 1, 256] - - [35, 51.689] - - - [26160, 2865, 1, 256] - - [37, 61.461] - - - [23600, 2816, 1, 256] - - [37, 63.383] - - - [20480, 7424, 1, 256] - - [26, 75.186] - - - [28928, 3329, 1, 256] - - [55, 69.267] - - - [20784, 2816, 1, 256] - - [54, 63.356] - - - [25344, 256, 1, 256] - - [54, 55.05] - - - [20224, 10240, 1, 256] - - [38, 74.867] - - - [28672, 1280, 1, 256] - - [50, 69.999] - - - [29232, 256, 1, 256] - - [45, 47.674] - - - [28720, 2865, 1, 256] - - [26, 59.965] - - - [22016, 2816, 1, 256] - - [28, 72.506] - - - [25600, 1536, 1, 256] - - [35, 70.369] - - - [26112, 10240, 1, 256] - - [28, 75.339] - - - [27136, 10240, 1, 256] - - [52, 75.178] - - - [31744, 8192, 1, 256] - - [28, 74.947] - - - [24320, 10240, 1, 256] - - [52, 74.908] - - - [29952, 10240, 1, 256] - - [31, 74.674] - - - [23296, 9984, 1, 256] - - [26, 75.04] - - - [34560, 2304, 1, 256] - - [36, 72.547] - - - [32000, 2865, 1, 256] - - [55, 70.214] - - - [25088, 1024, 1, 256] - - [64, 68.105] - - - [20272, 10240, 1, 256] - - [45, 63.816] - - - [25344, 5376, 1, 256] - - [43, 73.025] - - - [21760, 3328, 1, 256] - - [26, 72.312] - - - [32768, 8960, 1, 256] - - [39, 59.008] - - - [29952, 3840, 1, 256] - - [32, 73.676] - - - [32512, 2865, 1, 256] - - [74, 69.874] - - - [23344, 2865, 1, 256] - - [37, 60.816] - - - [24576, 768, 1, 256] - - [35, 60.724] - - - [27648, 3584, 1, 256] - - [26, 74.08] - - - [27952, 4608, 1, 256] - - [54, 63.343] - - - [29440, 3584, 1, 256] - - [28, 73.461] - - - [34096, 512, 1, 256] - - [54, 59.465] - - - [32304, 256, 1, 256] - - [54, 52.548] - - - [21040, 2816, 1, 256] - - [37, 64.179] - - - [22784, 1024, 1, 256] - - [34, 67.558] - - - [22784, 2816, 1, 256] - - [36, 71.985] - - - [25856, 2816, 1, 256] - - [50, 72.336] - - - [23296, 6144, 1, 256] - - [28, 74.154] - - - [28160, 4608, 1, 256] - - [92, 73.703] - - - [25136, 1792, 1, 256] - - [30, 62.163] - - - [30208, 256, 1, 256] - - [24, 53.789] - - - [23808, 1281, 1, 256] - - [61, 63.94] - - - [26368, 2304, 1, 256] - - [36, 71.856] - - - [27648, 4352, 1, 256] - - [28, 74.703] - - - [31280, 7936, 1, 256] - - [45, 63.951] - - - [22320, 2865, 1, 256] - - [26, 61.292] - - - [22320, 2816, 1, 256] - - [59, 63.845] - - - [28720, 5120, 1, 256] - - [29, 61.738] - - - [22272, 1280, 1, 256] - - [35, 69.431] - - - [31232, 3328, 1, 256] - - [42, 73.148] - - - [29696, 2048, 1, 256] - - [42, 71.313] - - - [34048, 9984, 1, 256] - - [52, 74.355] - - - [28416, 1280, 1, 256] - - [56, 69.731] - - - [21504, 2816, 1, 256] - - [50, 72.798] - - - [33536, 2865, 1, 256] - - [38, 69.879] - - - [23552, 3840, 1, 256] - - [38, 74.249] - - - [31744, 256, 1, 256] - - [28, 55.586] - - - [25600, 1281, 1, 256] - - [42, 63.258] - - - [30768, 7168, 1, 256] - - [29, 62.311] - - - [23808, 3329, 1, 256] - - [28, 69.505] - - - [32256, 3328, 1, 256] - - [25, 73.232] - - - [23040, 9216, 1, 256] - - [42, 74.942] - - - [33024, 256, 1, 256] - - [57, 54.622] - - - [33584, 2865, 1, 256] - - [59, 62.267] - - - [21504, 8448, 1, 256] - - [28, 75.477] - - - [27904, 1281, 1, 256] - - [41, 63.938] - - - [34304, 10240, 1, 256] - - [44, 75.096] - - - [20992, 2865, 1, 256] - - [38, 70.053] - - - [22528, 8960, 1, 256] - - [30, 75.783] - - - [28928, 3328, 1, 256] - - [26, 72.199] - - - [21808, 2865, 1, 256] - - [54, 61.674] - - - [26416, 2816, 1, 256] - - [54, 63.431] - - - [27392, 3840, 1, 256] - - [42, 72.145] - - - [26112, 1281, 1, 256] - - [34, 64.084] - - - [34864, 10240, 1, 256] - - [29, 62.895] - - - [29440, 1536, 1, 256] - - [36, 70.113] - - - [30256, 10240, 1, 256] - - [45, 62.69] - - - [22528, 2816, 1, 256] - - [38, 73.081] - - - [28928, 2048, 1, 256] - - [40, 69.845] - - - [28976, 5376, 1, 256] - - [37, 63.383] - - - [20736, 7168, 1, 256] - - [26, 73.681] - - - [22016, 2865, 1, 256] - - [28, 70.061] - - - [26368, 1280, 1, 256] - - [35, 69.846] - - - [24624, 2865, 1, 256] - - [56, 59.672] - - - [23040, 3329, 1, 256] - - [74, 69.455] - - - [23296, 2865, 1, 256] - - [28, 69.699] - - - [28416, 3329, 1, 256] - - [55, 69.052] - - - [23040, 1281, 1, 256] - - [57, 63.922] - - - [21808, 8448, 1, 256] - - [37, 64.072] - - - [30720, 2865, 1, 256] - - [28, 70.964] - - - [22272, 8960, 1, 256] - - [32, 74.925] - - - [34864, 2816, 1, 256] - - [30, 62.188] - - - [31232, 7168, 1, 256] - - [52, 73.728] - - - [27696, 4352, 1, 256] - - [26, 63.284] - - - [21504, 256, 1, 256] - - [50, 48.344] - - - [28672, 1281, 1, 256] - - [30, 63.048] - - - [29696, 1792, 1, 256] - - [28, 72.277] - - - [28464, 5120, 1, 256] - - [45, 63.253] - - - [27136, 3329, 1, 256] - - [48, 70.101] - - - [21248, 3328, 1, 256] - - [56, 72.083] - - - [26880, 1281, 1, 256] - - [59, 63.574] - - - [32256, 8448, 1, 256] - - [48, 75.036] - - - [20480, 6144, 1, 256] - - [28, 74.369] - - - [34048, 2865, 1, 256] - - [74, 69.561] - - - [29696, 5888, 1, 256] - - [38, 74.749] - - - [28720, 256, 1, 256] - - [30, 46.969] - - - [33792, 2865, 1, 256] - - [38, 71.066] - - - [22784, 8960, 1, 256] - - [28, 74.796] - - - [30720, 256, 1, 256] - - [61, 54.635] - - - [23808, 512, 1, 256] - - [41, 63.037] - - - [33024, 9728, 1, 256] - - [44, 74.736] - - - [42624, 13824, 1, 384] - - [29, 88.278] - - - [33024, 3840, 1, 384] - - [37, 89.14] - - - [33408, 15360, 1, 384] - - [30, 90.765] - - - [44160, 8832, 1, 384] - - [54, 90.868] - - - [31488, 2688, 1, 384] - - [37, 89.239] - - - [39168, 3072, 1, 384] - - [26, 89.312] - - - [31872, 5760, 1, 384] - - [38, 90.116] - - - [36096, 13440, 1, 384] - - [29, 89.824] - - - [41856, 1152, 1, 384] - - [54, 87.469] - - - [32256, 1153, 1, 384] - - [54, 78.124] - - - [44160, 1153, 1, 384] - - [59, 78.2] - - - [31488, 7296, 1, 384] - - [30, 90.342] - - - [43008, 9216, 1, 384] - - [29, 88.337] - - - [31872, 6144, 1, 384] - - [38, 89.989] - - - [32640, 7297, 1, 384] - - [55, 85.715] - - - [33792, 1152, 1, 384] - - [28, 85.923] - - - [43776, 13441, 1, 384] - - [55, 87.285] - - - [36480, 1153, 1, 384] - - [37, 78.52] - - - [37632, 1152, 1, 384] - - [37, 85.178] - - - [37248, 8448, 1, 384] - - [28, 90.158] - - - [31872, 7297, 1, 384] - - [30, 88.03] - - - [41856, 7296, 1, 384] - - [59, 90.634] - - - [39936, 7297, 1, 384] - - [28, 87.892] - - - [35712, 1153, 1, 384] - - [37, 77.05] - - - [35712, 3072, 1, 384] - - [38, 89.547] - - - [31488, 1153, 1, 384] - - [26, 76.43] - - - [36480, 1152, 1, 384] - - [54, 85.411] - - - [36864, 9216, 1, 384] - - [55, 86.953] - - - [42624, 15360, 1, 384] - - [29, 87.319] - - - [37632, 8832, 1, 384] - - [59, 90.755] - - - [32640, 1153, 1, 384] - - [79, 73.418] - - - [36864, 3072, 1, 384] - - [30, 87.687] - - - [32640, 6912, 1, 384] - - [38, 88.737] - - - [31872, 13440, 1, 384] - - [28, 90.709] - - - [39168, 3840, 1, 384] - - [30, 89.866] - - - [39168, 10368, 1, 384] - - [54, 90.644] - - - [33792, 3072, 1, 384] - - [38, 88.025] - - - [39552, 1536, 1, 384] - - [28, 87.432] - - - [38784, 7296, 1, 384] - - [37, 90.329] - - - [40320, 1153, 1, 384] - - [37, 78.245] - - - [42240, 1152, 1, 384] - - [45, 87.93] - - - [43776, 14976, 1, 384] - - [42, 89.905] - - - [38784, 9216, 1, 384] - - [59, 90.782] - - - [33024, 4224, 1, 384] - - [54, 89.263] - - - [43776, 7297, 1, 384] - - [55, 86.204] - - - [34560, 9216, 1, 384] - - [57, 90.365] - - - [43392, 8064, 1, 384] - - [28, 90.729] - - - [34944, 7296, 1, 384] - - [28, 90.377] - - - [38400, 7296, 1, 384] - - [38, 90.343] - - - [41856, 6912, 1, 384] - - [45, 90.762] - - - [40704, 3072, 1, 384] - - [30, 89.188] - - - [41472, 12672, 1, 384] - - [26, 91.04] - - - [36864, 1920, 1, 384] - - [38, 87.498] - - - [43008, 1920, 1, 384] - - [28, 88.654] - - - [43008, 13824, 1, 384] - - [30, 90.147] - - - [31104, 13441, 1, 384] - - [30, 89.23] - - - [41472, 12288, 1, 384] - - [62, 89.607] - - - [31488, 7297, 1, 384] - - [26, 87.888] - - - [35712, 6912, 1, 384] - - [26, 90.541] - - - [40704, 5376, 1, 384] - - [59, 90.373] - - - [36480, 9216, 1, 384] - - [45, 90.768] - - - [38784, 13440, 1, 384] - - [38, 90.948] - - - [36096, 15360, 1, 384] - - [31, 89.555] - - - [41856, 15360, 1, 384] - - [37, 91.147] - - - [37632, 2688, 1, 384] - - [45, 88.713] - - - [33792, 4608, 1, 384] - - [28, 88.828] - - - [38400, 13440, 1, 384] - - [38, 90.905] - - - [31104, 3072, 1, 384] - - [26, 88.396] - - - [33792, 13440, 1, 384] - - [26, 90.724] - - - [34176, 5376, 1, 384] - - [37, 89.995] - - - [31872, 3072, 1, 384] - - [26, 88.853] - - - [33792, 1920, 1, 384] - - [30, 88.614] - - - [34560, 1153, 1, 384] - - [37, 77.466] - - - [43392, 15360, 1, 384] - - [28, 90.778] - - - [39168, 4224, 1, 384] - - [26, 90.179] - - - [43776, 1153, 1, 384] - - [52, 76.234] - - - [41472, 6528, 1, 384] - - [59, 90.214] - - - [42240, 1153, 1, 384] - - [26, 78.784] - - - [36480, 13441, 1, 384] - - [28, 88.921] - - - [31488, 5760, 1, 384] - - [30, 90.215] - - - [34560, 13440, 1, 384] - - [26, 90.784] - - - [32256, 3072, 1, 384] - - [26, 88.739] - - - [37632, 15360, 1, 384] - - [26, 90.795] - - - [43776, 8448, 1, 384] - - [42, 89.663] - - - [37248, 13440, 1, 384] - - [38, 90.849] - - - [34944, 13440, 1, 384] - - [30, 90.894] - - - [41088, 3072, 1, 384] - - [33, 86.816] - - - [43008, 14208, 1, 384] - - [26, 90.513] - - - [33792, 7296, 1, 384] - - [26, 90.356] - - - [43392, 8448, 1, 384] - - [54, 90.947] - - - [31104, 7297, 1, 384] - - [38, 88.048] - - - [31104, 2304, 1, 384] - - [54, 88.538] - - - [35712, 1152, 1, 384] - - [37, 86.709] - - - [39552, 13440, 1, 384] - - [26, 90.967] - - - [37632, 2304, 1, 384] - - [54, 88.3] - - - [31872, 1153, 1, 384] - - [45, 77.391] - - - [39552, 3072, 1, 384] - - [26, 89.481] - - - [36864, 15360, 1, 384] - - [33, 88.677] - - - [33408, 4608, 1, 384] - - [59, 89.822] - - - [43392, 7297, 1, 384] - - [28, 88.191] - - - [32256, 7296, 1, 384] - - [28, 90.393] - - - [41472, 7296, 1, 384] - - [26, 90.726] - - - [38016, 9216, 1, 384] - - [37, 90.741] - - - [38784, 1153, 1, 384] - - [26, 77.728] - - - [34944, 2688, 1, 384] - - [45, 88.484] - - - [36864, 1152, 1, 384] - - [26, 85.952] - - - [39168, 7297, 1, 384] - - [38, 88.429] - - - [33024, 768, 1, 384] - - [87, 79.528] - - - [34560, 13441, 1, 384] - - [28, 89.15] - - - [33792, 7680, 1, 384] - - [30, 90.317] - - - [36864, 1153, 1, 384] - - [26, 76.423] - - - [40320, 4992, 1, 384] - - [54, 90.42] - - - [31488, 13440, 1, 384] - - [26, 90.896] - - - [39552, 10752, 1, 384] - - [37, 91.09] - - - [36096, 1152, 1, 384] - - [55, 83.95] - - - [44160, 1152, 1, 384] - - [38, 86.378] - - - [37632, 9216, 1, 384] - - [45, 90.767] - - - [37248, 15360, 1, 384] - - [38, 90.911] - - - [34944, 5760, 1, 384] - - [28, 90.259] - - - [41088, 15360, 1, 384] - - [62, 89.764] - - - [41088, 11904, 1, 384] - - [75, 89.957] - - - [35328, 6528, 1, 384] - - [30, 90.045] - - - [32640, 15360, 1, 384] - - [55, 89.211] - - - [33024, 7297, 1, 384] - - [33, 86.942] - - - [31104, 1153, 1, 384] - - [26, 75.891] - - - [40704, 1153, 1, 384] - - [37, 78.6] - - - [42240, 13440, 1, 384] - - [59, 91.201] - - - [41472, 7297, 1, 384] - - [38, 88.199] - - - [33408, 3072, 1, 384] - - [26, 89.593] - - - [40704, 13440, 1, 384] - - [26, 91.016] - - - [39168, 7296, 1, 384] - - [26, 90.527] - - - [34176, 9216, 1, 384] - - [57, 90.031] - - - [35328, 15360, 1, 384] - - [28, 90.572] - - - [38400, 1152, 1, 384] - - [54, 86.421] - - - [37248, 3072, 1, 384] - - [28, 89.373] - - - [31488, 2304, 1, 384] - - [54, 87.961] - - - [40704, 1152, 1, 384] - - [30, 85.429] - - - [39168, 768, 1, 384] - - [37, 85.104] - - - [34944, 1153, 1, 384] - - [54, 78.011] - - - [39936, 13440, 1, 384] - - [26, 90.764] - - - [43008, 7297, 1, 384] - - [26, 87.793] - - - [33024, 15360, 1, 384] - - [28, 90.317] - - - [34176, 1920, 1, 384] - - [26, 87.757] - - - [40320, 15360, 1, 384] - - [30, 90.659] - - - [37632, 3072, 1, 384] - - [26, 88.886] - - - [40320, 11136, 1, 384] - - [45, 90.855] - - - [34944, 1152, 1, 384] - - [59, 85.364] - - - [44160, 14976, 1, 384] - - [45, 90.918] - - - [33792, 1536, 1, 384] - - [30, 86.629] - - - [38016, 13441, 1, 384] - - [30, 88.959] - - - [37632, 7296, 1, 384] - - [38, 90.485] - - - [41856, 6528, 1, 384] - - [37, 90.77] - - - [36096, 6912, 1, 384] - - [52, 89.39] - - - [39936, 15360, 1, 384] - - [33, 89.281] - - - [43776, 9216, 1, 384] - - [75, 89.701] - - - [38400, 9600, 1, 384] - - [30, 90.744] - - - [39552, 15360, 1, 384] - - [38, 90.804] - - - [37248, 2304, 1, 384] - - [30, 88.9] - - - [33792, 1153, 1, 384] - - [28, 78.249] - - - [42624, 1152, 1, 384] - - [30, 85.746] - - - [35328, 3072, 1, 384] - - [26, 88.672] - - - [37632, 13440, 1, 384] - - [28, 90.935] - - - [38400, 3072, 1, 384] - - [26, 88.81] - - - [32640, 1152, 1, 384] - - [38, 82.416] - - - [31872, 1152, 1, 384] - - [45, 84.902] - - - [40320, 3072, 1, 384] - - [26, 89.67] - - - [38016, 15360, 1, 384] - - [30, 91.029] - - - [35712, 9216, 1, 384] - - [59, 90.628] - - - [33024, 13441, 1, 384] - - [33, 88.236] - - - [36096, 3072, 1, 384] - - [31, 87.825] - - - [36864, 13440, 1, 384] - - [26, 90.3] - - - [33408, 13441, 1, 384] - - [28, 89.408] - - - [37248, 9216, 1, 384] - - [30, 90.012] - - - [31488, 1152, 1, 384] - - [30, 84.3] - - - [31488, 3072, 1, 384] - - [28, 88.897] - - - [35328, 1152, 1, 384] - - [38, 86.19] - - - [37248, 7297, 1, 384] - - [25, 87.954] - - - [34944, 6144, 1, 384] - - [28, 89.899] - - - [36480, 1536, 1, 384] - - [30, 85.898] - - - [39168, 15360, 1, 384] - - [38, 90.611] - - - [43392, 13441, 1, 384] - - [26, 89.041] - - - [42624, 1536, 1, 384] - - [33, 86.032] - - - [36480, 7296, 1, 384] - - [59, 90.429] - - - [33792, 9216, 1, 384] - - [62, 88.885] - - - [36096, 768, 1, 384] - - [63, 81.715] - - - [33408, 1536, 1, 384] - - [28, 86.603] - - - [31872, 13441, 1, 384] - - [26, 88.734] - - - [43008, 13440, 1, 384] - - [38, 90.679] - - - [33024, 1152, 1, 384] - - [54, 83.603] - - - [34560, 5376, 1, 384] - - [37, 89.95] - - - [32640, 3840, 1, 384] - - [30, 87.392] - - - [33408, 1153, 1, 384] - - [28, 77.608] - - - [32256, 1152, 1, 384] - - [28, 85.845] - - - [41856, 13440, 1, 384] - - [45, 91.068] - - - [43776, 2688, 1, 384] - - [75, 87.926] - - - [34560, 8832, 1, 384] - - [30, 90.698] - - - [32256, 6528, 1, 384] - - [26, 89.983] - - - [33408, 13440, 1, 384] - - [26, 90.855] - - - [36096, 7296, 1, 384] - - [31, 89.202] - - - [43776, 3072, 1, 384] - - [31, 86.725] - - - [38784, 7297, 1, 384] - - [26, 87.841] - - - [39936, 7296, 1, 384] - - [28, 90.369] - - - [37632, 8448, 1, 384] - - [54, 90.624] - - - [43392, 9216, 1, 384] - - [59, 90.379] - - - [41856, 13056, 1, 384] - - [54, 91.0] - - - [30720, 13441, 1, 384] - - [30, 88.74] - - - [36864, 7680, 1, 384] - - [26, 89.447] - - - [41472, 1152, 1, 384] - - [45, 86.836] - - - [39168, 13440, 1, 384] - - [38, 90.989] - - - [43776, 2304, 1, 384] - - [75, 87.775] - - - [34176, 15360, 1, 384] - - [26, 90.734] - - - [36096, 7297, 1, 384] - - [31, 86.378] - - - [33792, 4992, 1, 384] - - [28, 90.089] - - - [35712, 15360, 1, 384] - - [30, 90.899] - - - [39168, 9984, 1, 384] - - [26, 90.689] - - - [36096, 9216, 1, 384] - - [75, 89.158] - - - [43008, 1536, 1, 384] - - [30, 86.668] - - - [33408, 9216, 1, 384] - - [63, 90.217] - - - [40704, 7296, 1, 384] - - [59, 90.655] - - - [38016, 2688, 1, 384] - - [37, 89.371] - - - [39168, 13441, 1, 384] - - [30, 89.187] - - - [39168, 9216, 1, 384] - - [38, 89.958] - - - [38400, 15360, 1, 384] - - [38, 90.558] - - - [43392, 2304, 1, 384] - - [45, 89.144] - - - [38400, 13441, 1, 384] - - [26, 88.937] - - - [43008, 1152, 1, 384] - - [26, 86.658] - - - [39936, 4608, 1, 384] - - [30, 88.688] - - - [43392, 14592, 1, 384] - - [28, 90.978] - - - [34176, 13441, 1, 384] - - [31, 88.89] - - - [38784, 9984, 1, 384] - - [30, 90.687] - - - [44160, 13441, 1, 384] - - [38, 88.866] - - - [31488, 5376, 1, 384] - - [45, 90.263] - - - [39936, 13441, 1, 384] - - [38, 88.989] - - - [34176, 1152, 1, 384] - - [38, 86.48] - - - [32640, 3072, 1, 384] - - [26, 86.224] - - - [34560, 15360, 1, 384] - - [30, 90.733] - - - [34944, 15360, 1, 384] - - [38, 90.744] - - - [37632, 13441, 1, 384] - - [28, 89.069] - - - [40320, 5376, 1, 384] - - [37, 90.308] - - - [41856, 12672, 1, 384] - - [45, 91.102] - - - [34176, 4992, 1, 384] - - [30, 89.892] - - - [42624, 7297, 1, 384] - - [52, 84.477] - - - [41856, 1153, 1, 384] - - [26, 78.384] - - - [41472, 9216, 1, 384] - - [75, 89.586] - - - [40704, 2304, 1, 384] - - [54, 88.393] - - - [36864, 8064, 1, 384] - - [26, 90.015] - - - [40704, 5760, 1, 384] - - [38, 90.322] - - - [41088, 7297, 1, 384] - - [55, 86.365] - - - [38784, 1152, 1, 384] - - [26, 87.044] - - - [38784, 3072, 1, 384] - - [30, 89.002] - - - [34560, 2304, 1, 384] - - [38, 88.169] - - - [36096, 1153, 1, 384] - - [63, 76.971] - - - [35712, 13440, 1, 384] - - [30, 90.934] - - - [39936, 1152, 1, 384] - - [28, 86.321] - - - [43392, 14208, 1, 384] - - [26, 90.644] - - - [39552, 1153, 1, 384] - - [37, 77.084] - - - [35712, 6528, 1, 384] - - [45, 90.358] - - - [31104, 5376, 1, 384] - - [59, 90.199] - - - [31104, 9216, 1, 384] - - [45, 90.311] - - - [33024, 9216, 1, 384] - - [62, 89.636] - - - [39936, 11136, 1, 384] - - [26, 90.513] - - - [43008, 3072, 1, 384] - - [26, 87.535] - - - [41856, 768, 1, 384] - - [54, 85.598] - - - [43776, 1152, 1, 384] - - [75, 85.56] - - - [34176, 7297, 1, 384] - - [25, 87.794] - - - [38016, 7297, 1, 384] - - [54, 88.208] - - - [36480, 7680, 1, 384] - - [59, 90.618] - - - [38400, 7297, 1, 384] - - [38, 88.136] - - - [44160, 2688, 1, 384] - - [45, 89.787] - - - [33792, 15360, 1, 384] - - [55, 89.277] - - - [40704, 2688, 1, 384] - - [37, 89.517] - - - [38784, 3840, 1, 384] - - [54, 90.014] - - - [44160, 7296, 1, 384] - - [59, 90.668] - - - [41088, 2688, 1, 384] - - [37, 88.083] - - - [38016, 3072, 1, 384] - - [30, 89.602] - - - [42240, 7296, 1, 384] - - [37, 90.815] - - - [41856, 9216, 1, 384] - - [59, 90.762] - - - [32640, 13440, 1, 384] - - [28, 89.719] - - - [40320, 13441, 1, 384] - - [28, 88.933] - - - [36480, 13440, 1, 384] - - [28, 90.863] - - - [41856, 7297, 1, 384] - - [59, 88.115] - - - [41088, 7296, 1, 384] - - [39, 89.193] - - - [33408, 1152, 1, 384] - - [37, 85.116] - - - [43392, 1920, 1, 384] - - [28, 88.388] - - - [31104, 1920, 1, 384] - - [38, 86.98] - - - [31488, 15360, 1, 384] - - [350, 91.336] - - - [31872, 7296, 1, 384] - - [26, 90.056] - - - [43008, 7680, 1, 384] - - [30, 90.127] - - - [35328, 13440, 1, 384] - - [26, 90.941] - - - [43776, 15360, 1, 384] - - [62, 89.613] - - - [34944, 3072, 1, 384] - - [26, 89.225] - - - [37248, 1153, 1, 384] - - [37, 77.474] - - - [31104, 1152, 1, 384] - - [59, 86.317] - - - [34560, 7297, 1, 384] - - [38, 88.375] - - - [43776, 14592, 1, 384] - - [42, 89.832] - - - [33408, 7296, 1, 384] - - [26, 90.328] - - - [33024, 7296, 1, 384] - - [30, 89.596] - - - [33024, 13440, 1, 384] - - [28, 90.417] - - - [31104, 7296, 1, 384] - - [26, 90.339] - - - [42240, 9216, 1, 384] - - [54, 91.084] - - - [34944, 13441, 1, 384] - - [38, 89.08] - - - [33792, 7297, 1, 384] - - [38, 88.085] - - - [35328, 13441, 1, 384] - - [26, 89.221] - - - [34176, 7296, 1, 384] - - [38, 90.146] - - - [40320, 1920, 1, 384] - - [54, 89.136] - - - [31872, 15360, 1, 384] - - [26, 90.709] - - - [39168, 1153, 1, 384] - - [54, 78.445] - - - [31104, 4992, 1, 384] - - [38, 89.792] - - - [41088, 1152, 1, 384] - - [75, 84.162] - - - [39552, 10368, 1, 384] - - [45, 90.976] - - - [40704, 11520, 1, 384] - - [26, 90.92] - - - [36864, 7297, 1, 384] - - [28, 87.1] - - - [42240, 15360, 1, 384] - - [37, 91.13] - - - [34560, 1152, 1, 384] - - [28, 84.888] - - - [31104, 13440, 1, 384] - - [30, 90.806] - - - [31488, 9216, 1, 384] - - [30, 89.809] - - - [34176, 3072, 1, 384] - - [30, 88.904] - - - [41088, 1153, 1, 384] - - [59, 76.435] - - - [43392, 1153, 1, 384] - - [54, 78.69] - - - [42240, 6912, 1, 384] - - [37, 90.914] - - - [43008, 15360, 1, 384] - - [55, 89.299] - - - [42240, 7297, 1, 384] - - [28, 88.344] - - - [43776, 7296, 1, 384] - - [42, 89.302] - - - [35712, 7296, 1, 384] - - [38, 90.409] - - - [38400, 9216, 1, 384] - - [57, 89.62] - - - [39936, 9216, 1, 384] - - [62, 89.072] - - - [32256, 6144, 1, 384] - - [28, 89.891] - - - [42624, 7680, 1, 384] - - [26, 88.093] - - - [33408, 4224, 1, 384] - - [37, 89.662] - - - [38784, 768, 1, 384] - - [45, 84.311] - - - [38016, 7296, 1, 384] - - [30, 90.545] - - - [34560, 5760, 1, 384] - - [30, 90.265] - - - [34944, 7297, 1, 384] - - [28, 87.974] - - - [38016, 8832, 1, 384] - - [38, 90.763] - - - [39936, 1920, 1, 384] - - [28, 88.364] - - - [40320, 11520, 1, 384] - - [28, 91.004] - - - [32256, 7297, 1, 384] - - [28, 88.391] - - - [33792, 13441, 1, 384] - - [26, 88.944] - - - [41472, 3072, 1, 384] - - [38, 89.299] - - - [33024, 1153, 1, 384] - - [59, 76.254] - - - [36864, 7296, 1, 384] - - [38, 89.855] - - - [38016, 1153, 1, 384] - - [28, 78.612] - - - [40320, 7297, 1, 384] - - [30, 88.106] - - - [42624, 13441, 1, 384] - - [31, 84.193] - - - [43008, 13441, 1, 384] - - [26, 88.6] - - - [39552, 9216, 1, 384] - - [45, 90.808] - - - [35328, 9216, 1, 384] - - [75, 89.367] - - - [42624, 3072, 1, 384] - - [26, 88.928] - - - [40320, 13440, 1, 384] - - [28, 91.065] - - - [42240, 13441, 1, 384] - - [30, 89.158] - - - [39936, 10752, 1, 384] - - [38, 90.046] - - - [41472, 6144, 1, 384] - - [30, 89.786] - - - [36864, 1536, 1, 384] - - [26, 85.867] - - - [33408, 7297, 1, 384] - - [28, 88.0] - - - [31872, 2688, 1, 384] - - [54, 88.748] - - - [41472, 1153, 1, 384] - - [28, 78.0] - - - [38400, 1153, 1, 384] - - [37, 77.123] - - - [38400, 3456, 1, 384] - - [37, 89.537] - - - [41856, 13441, 1, 384] - - [30, 88.889] - - - [43392, 1152, 1, 384] - - [37, 87.397] - - - [39552, 4608, 1, 384] - - [45, 89.957] - - - [40704, 15360, 1, 384] - - [30, 90.792] - - - [42240, 3072, 1, 384] - - [30, 89.467] - - - [32640, 3456, 1, 384] - - [28, 86.766] - - - [35712, 768, 1, 384] - - [45, 82.475] - - - [31104, 15360, 1, 384] - - [396, 91.256] - - - [40704, 13441, 1, 384] - - [28, 89.252] - - - [32640, 7296, 1, 384] - - [30, 88.652] - - - [34176, 8448, 1, 384] - - [94, 90.045] - - - [32640, 13441, 1, 384] - - [55, 87.325] - - - [36864, 13441, 1, 384] - - [26, 87.818] - - - [34176, 13440, 1, 384] - - [38, 90.629] - - - [37248, 1152, 1, 384] - - [54, 87.015] - - - [44160, 7297, 1, 384] - - [45, 88.025] - - - [41088, 6144, 1, 384] - - [62, 88.988] - - - [39936, 1536, 1, 384] - - [26, 86.0] - - - [44160, 15360, 1, 384] - - [45, 90.988] - - - [35712, 7297, 1, 384] - - [26, 88.087] - - - [35328, 6144, 1, 384] - - [30, 89.727] - - - [42624, 7296, 1, 384] - - [51, 88.397] - - - [33408, 7680, 1, 384] - - [30, 90.515] - - - [41472, 13441, 1, 384] - - [38, 89.163] - - - [43776, 8832, 1, 384] - - [75, 89.563] - - - [32256, 15360, 1, 384] - - [26, 90.612] - - - [32256, 9216, 1, 384] - - [28, 89.77] - - - [31872, 9216, 1, 384] - - [27, 90.157] - - - [37248, 7296, 1, 384] - - [38, 90.367] - - - [40320, 1152, 1, 384] - - [45, 87.104] - - - [34560, 8448, 1, 384] - - [63, 90.284] - - - [38784, 3456, 1, 384] - - [54, 89.687] - - - [41472, 15360, 1, 384] - - [26, 90.467] - - - [41856, 3072, 1, 384] - - [30, 89.414] - - - [41088, 13441, 1, 384] - - [55, 87.66] - - - [39936, 1153, 1, 384] - - [28, 77.349] - - - [37248, 1920, 1, 384] - - [28, 88.403] - - - [39552, 7296, 1, 384] - - [37, 90.452] - - - [40320, 2304, 1, 384] - - [45, 89.112] - - - [34560, 2688, 1, 384] - - [54, 89.033] - - - [42240, 13056, 1, 384] - - [38, 90.791] - - - [40320, 9216, 1, 384] - - [37, 90.595] - - - [40704, 7297, 1, 384] - - [54, 88.281] - - - [43776, 13440, 1, 384] - - [29, 90.065] - - - [39936, 4992, 1, 384] - - [28, 90.119] - - - [42624, 13440, 1, 384] - - [23, 87.56] - - - [37632, 1153, 1, 384] - - [45, 78.129] - - - [33024, 3072, 1, 384] - - [30, 88.041] - - - [40704, 9216, 1, 384] - - [54, 90.812] - - - [42624, 1153, 1, 384] - - [26, 77.163] - - - [43392, 13440, 1, 384] - - [26, 90.96] - - - [36480, 3072, 1, 384] - - [38, 88.881] - - - [41088, 12288, 1, 384] - - [75, 89.769] - - - [39168, 1152, 1, 384] - - [26, 85.477] - - - [39936, 3072, 1, 384] - - [55, 87.229] - - - [35712, 13441, 1, 384] - - [26, 89.034] - - - [41088, 13440, 1, 384] - - [38, 90.307] - - - [43392, 3072, 1, 384] - - [28, 89.924] - - - [33792, 8064, 1, 384] - - [38, 90.4] - - - [32256, 13440, 1, 384] - - [28, 90.922] - - - [35328, 7297, 1, 384] - - [26, 88.144] - - - [40704, 11904, 1, 384] - - [26, 90.964] - - - [33024, 6912, 1, 384] - - [38, 89.944] - - - [38784, 15360, 1, 384] - - [26, 90.615] - - - [42240, 768, 1, 384] - - [54, 86.132] - - - [44160, 13440, 1, 384] - - [37, 91.059] - - - [39552, 7297, 1, 384] - - [26, 88.337] - - - [32640, 768, 1, 384] - - [77, 78.277] - - - [44160, 9216, 1, 384] - - [59, 90.934] - - - [32640, 6528, 1, 384] - - [30, 88.489] - - - [39552, 13441, 1, 384] - - [26, 88.971] - - - [31488, 13441, 1, 384] - - [30, 89.238] - - - [43008, 7296, 1, 384] - - [30, 90.267] - - - [41088, 5760, 1, 384] - - [55, 88.944] - - - [41472, 13440, 1, 384] - - [26, 90.964] - - - [43392, 7296, 1, 384] - - [26, 90.595] - - - [34944, 9216, 1, 384] - - [37, 90.258] - - - [43008, 1153, 1, 384] - - [26, 78.043] - - - [32640, 9216, 1, 384] - - [29, 88.653] - - - [36096, 13441, 1, 384] - - [33, 87.233] - - - [39552, 1152, 1, 384] - - [37, 86.13] - - - [37632, 7297, 1, 384] - - [28, 88.148] - - - [42624, 9216, 1, 384] - - [48, 86.819] - - - [43008, 8064, 1, 384] - - [28, 90.492] - - - [38784, 9600, 1, 384] - - [45, 90.879] - - - [37248, 8064, 1, 384] - - [38, 90.624] - - - [30720, 15360, 1, 384] - - [350, 91.265] - - - [38016, 13440, 1, 384] - - [30, 90.967] - - - [34944, 8832, 1, 384] - - [26, 90.609] - - - [37248, 13441, 1, 384] - - [30, 88.869] - - - [34560, 7296, 1, 384] - - [38, 90.444] - - - [44160, 3072, 1, 384] - - [30, 89.316] - - - [40320, 7296, 1, 384] - - [38, 90.462] - - - [34176, 2304, 1, 384] - - [30, 88.599] - - - [41088, 9216, 1, 384] - - [75, 89.557] - - - [34176, 1153, 1, 384] - - [26, 76.738] - - - [39552, 4224, 1, 384] - - [45, 90.185] - - - [38784, 13441, 1, 384] - - [26, 88.737] - - - [36480, 7297, 1, 384] - - [26, 87.916] - - - [32256, 3456, 1, 384] - - [28, 89.567] - - - [34176, 8064, 1, 384] - - [26, 90.295] - - - [36480, 15360, 1, 384] - - [59, 91.186] - - - [34560, 3072, 1, 384] - - [28, 89.269] - - - [35328, 7296, 1, 384] - - [30, 90.268] - - - [32256, 13441, 1, 384] - - [26, 89.398] - - - [38016, 1152, 1, 384] - - [59, 85.896] - - - [35328, 1153, 1, 384] - - [26, 76.639] - - - [23040, 7296, 1, 384] - - [26, 89.767] - - - [12672, 7296, 1, 384] - - [26, 88.855] - - - [4224, 4225, 1, 384] - - [28, 76.404] - - - [19968, 13440, 1, 384] - - [30, 90.89] - - - [16128, 3072, 1, 384] - - [26, 86.309] - - - [19968, 9216, 1, 384] - - [54, 89.746] - - - [24576, 13440, 1, 384] - - [39, 87.615] - - - [17280, 3072, 1, 384] - - [28, 87.102] - - - [16512, 9216, 1, 384] - - [62, 87.679] - - - [21120, 1536, 1, 384] - - [38, 85.882] - - - [18432, 13441, 1, 384] - - [30, 88.684] - - - [21120, 9216, 1, 384] - - [57, 90.106] - - - [27264, 3072, 1, 384] - - [26, 88.364] - - - [12288, 4608, 1, 384] - - [28, 86.933] - - - [22272, 5376, 1, 384] - - [45, 89.631] - - - [7296, 6912, 1, 384] - - [28, 88.378] - - - [26880, 9216, 1, 384] - - [45, 90.633] - - - [3072, 2688, 1, 384] - - [36, 72.028] - - - [16512, 2688, 1, 384] - - [85, 82.82] - - - [8064, 7680, 1, 384] - - [28, 87.628] - - - [22656, 1153, 1, 384] - - [45, 73.989] - - - [24960, 8064, 1, 384] - - [30, 90.562] - - - [23808, 9216, 1, 384] - - [54, 90.726] - - - [29568, 15360, 1, 384] - - [52, 90.385] - - - [1920, 1152, 1, 384] - - [337, 54.68] - - - [11136, 10752, 1, 384] - - [28, 89.659] - - - [25728, 1152, 1, 384] - - [28, 84.086] - - - [19584, 3072, 1, 384] - - [38, 87.287] - - - [3840, 1153, 1, 384] - - [56, 58.679] - - - [15360, 7296, 1, 384] - - [354, 91.646] - - - [13056, 12673, 1, 384] - - [30, 89.234] - - - [5376, 5377, 1, 384] - - [30, 82.103] - - - [28416, 13440, 1, 384] - - [37, 91.161] - - - [11904, 4224, 1, 384] - - [45, 88.137] - - - [24576, 10752, 1, 384] - - [55, 86.561] - - - [20352, 7297, 1, 384] - - [26, 88.385] - - - [16512, 7296, 1, 384] - - [42, 87.442] - - - [17280, 13441, 1, 384] - - [30, 89.218] - - - [24192, 10368, 1, 384] - - [45, 90.498] - - - [20352, 6528, 1, 384] - - [45, 90.017] - - - [1920, 1536, 1, 384] - - [36, 55.125] - - - [15744, 8064, 1, 384] - - [392, 92.441] - - - [13056, 3072, 1, 384] - - [28, 85.097] - - - [20352, 7296, 1, 384] - - [30, 89.816] - - - [10368, 1152, 1, 384] - - [61, 76.853] - - - [16128, 1152, 1, 384] - - [28, 79.04] - - - [13440, 7297, 1, 384] - - [28, 87.271] - - - [19200, 13441, 1, 384] - - [28, 89.246] - - - [13440, 13441, 1, 384] - - [59, 88.895] - - - [7680, 7297, 1, 384] - - [26, 86.162] - - - [27648, 14208, 1, 384] - - [26, 90.692] - - - [23424, 9216, 1, 384] - - [28, 89.747] - - - [24960, 1153, 1, 384] - - [37, 76.623] - - - [28032, 2304, 1, 384] - - [37, 88.371] - - - [30720, 3072, 1, 384] - - [30, 88.071] - - - [11904, 1152, 1, 384] - - [26, 77.734] - - - [24576, 3072, 1, 384] - - [30, 84.927] - - - [26112, 1153, 1, 384] - - [59, 75.83] - - - [10368, 10369, 1, 384] - - [26, 87.889] - - - [14976, 1536, 1, 384] - - [26, 82.728] - - - [11520, 7296, 1, 384] - - [37, 89.153] - - - [5376, 5376, 1, 384] - - [26, 82.183] - - - [28800, 7296, 1, 384] - - [45, 90.21] - - - [22656, 3072, 1, 384] - - [28, 87.879] - - - [11904, 7296, 1, 384] - - [28, 88.549] - - - [13824, 3072, 1, 384] - - [28, 85.697] - - - [21504, 13440, 1, 384] - - [38, 90.62] - - - [28800, 13440, 1, 384] - - [45, 90.974] - - - [13824, 7296, 1, 384] - - [26, 89.863] - - - [28416, 13441, 1, 384] - - [28, 89.206] - - - [20736, 7296, 1, 384] - - [38, 90.329] - - - [4992, 4608, 1, 384] - - [26, 84.18] - - - [21888, 1153, 1, 384] - - [58, 73.218] - - - [6912, 3072, 1, 384] - - [353, 84.355] - - - [7680, 7680, 1, 384] - - [26, 88.391] - - - [11904, 11905, 1, 384] - - [26, 88.706] - - - [9600, 1920, 1, 384] - - [28, 79.204] - - - [25728, 2688, 1, 384] - - [37, 87.846] - - - [29568, 3840, 1, 384] - - [25, 89.362] - - - [9984, 7297, 1, 384] - - [26, 86.231] - - - [13056, 2688, 1, 384] - - [54, 84.981] - - - [3456, 1920, 1, 384] - - [355, 75.001] - - - [19200, 1152, 1, 384] - - [38, 80.905] - - - [15744, 2304, 1, 384] - - [26, 84.376] - - - [17664, 7296, 1, 384] - - [54, 89.781] - - - [3072, 3072, 1, 384] - - [230, 71.128] - - - [21888, 7296, 1, 384] - - [65, 87.508] - - - [16128, 13440, 1, 384] - - [28, 90.31] - - - [23040, 1153, 1, 384] - - [59, 74.924] - - - [21504, 9216, 1, 384] - - [62, 88.698] - - - [21120, 4608, 1, 384] - - [57, 89.236] - - - [10368, 1153, 1, 384] - - [37, 68.413] - - - [29184, 13441, 1, 384] - - [48, 89.21] - - - [8832, 1536, 1, 384] - - [26, 77.661] - - - [30336, 3072, 1, 384] - - [28, 89.232] - - - [24192, 1153, 1, 384] - - [37, 74.563] - - - [16128, 2304, 1, 384] - - [26, 85.472] - - - [20736, 13440, 1, 384] - - [28, 90.841] - - - [24960, 7297, 1, 384] - - [26, 88.598] - - - [18048, 1536, 1, 384] - - [26, 82.772] - - - [19200, 5760, 1, 384] - - [26, 89.282] - - - [13440, 13056, 1, 384] - - [26, 90.019] - - - [6144, 1152, 1, 384] - - [35, 73.359] - - - [1920, 1920, 1, 384] - - [229, 59.067] - - - [18816, 5376, 1, 384] - - [37, 88.84] - - - [28800, 2688, 1, 384] - - [54, 88.798] - - - [20352, 3840, 1, 384] - - [28, 88.528] - - - [3840, 3841, 1, 384] - - [30, 76.356] - - - [17280, 768, 1, 384] - - [54, 76.137] - - - [21888, 2304, 1, 384] - - [55, 85.589] - - - [28416, 14592, 1, 384] - - [59, 90.909] - - - [18816, 3072, 1, 384] - - [26, 86.239] - - - [25344, 13440, 1, 384] - - [30, 91.119] - - - [20736, 6912, 1, 384] - - [54, 89.894] - - - [26880, 1152, 1, 384] - - [38, 83.551] - - - [29952, 3072, 1, 384] - - [38, 88.052] - - - [24960, 8448, 1, 384] - - [54, 90.766] - - - [15360, 8064, 1, 384] - - [30, 89.763] - - - [27648, 1920, 1, 384] - - [26, 87.035] - - - [3456, 2304, 1, 384] - - [56, 69.753] - - - [23040, 6528, 1, 384] - - [37, 89.69] - - - [14208, 1153, 1, 384] - - [37, 69.98] - - - [27648, 1153, 1, 384] - - [38, 76.656] - - - [1920, 1921, 1, 384] - - [41, 49.571] - - - [19584, 13441, 1, 384] - - [28, 89.274] - - - [8448, 3072, 1, 384] - - [30, 82.728] - - - [16512, 13441, 1, 384] - - [33, 86.662] - - - [4992, 768, 1, 384] - - [24, 52.095] - - - [28416, 14976, 1, 384] - - [59, 90.971] - - - [8448, 1152, 1, 384] - - [36, 72.41] - - - [20352, 9216, 1, 384] - - [59, 90.003] - - - [19584, 1153, 1, 384] - - [28, 76.048] - - - [20736, 768, 1, 384] - - [54, 81.601] - - - [28416, 2688, 1, 384] - - [54, 88.293] - - - [27264, 13440, 1, 384] - - [350, 91.304] - - - [16128, 7296, 1, 384] - - [30, 89.025] - - - [27648, 13440, 1, 384] - - [350, 91.535] - - - [26880, 13056, 1, 384] - - [350, 90.919] - - - [6528, 1920, 1, 384] - - [50, 80.083] - - - [20352, 13441, 1, 384] - - [26, 89.041] - - - [12288, 7297, 1, 384] - - [26, 87.098] - - - [21120, 7680, 1, 384] - - [26, 90.502] - - - [13824, 13441, 1, 384] - - [28, 88.762] - - - [26112, 13440, 1, 384] - - [38, 90.947] - - - [16512, 7297, 1, 384] - - [33, 84.44] - - - [6144, 5761, 1, 384] - - [26, 82.18] - - - [24960, 1152, 1, 384] - - [26, 85.839] - - - [9600, 9216, 1, 384] - - [37, 88.337] - - - [22272, 1153, 1, 384] - - [59, 76.45] - - - [24960, 2304, 1, 384] - - [38, 88.378] - - - [11136, 7296, 1, 384] - - [28, 88.208] - - - [28800, 3072, 1, 384] - - [28, 87.63] - - - [6912, 2688, 1, 384] - - [28, 79.957] - - - [25728, 3072, 1, 384] - - [30, 88.902] - - - [15744, 13441, 1, 384] - - [28, 89.503] - - - [18816, 7296, 1, 384] - - [37, 89.387] - - - [18816, 7297, 1, 384] - - [26, 87.989] - - - [13440, 13440, 1, 384] - - [26, 90.071] - - - [29184, 3456, 1, 384] - - [26, 89.49] - - - [8064, 768, 1, 384] - - [36, 63.254] - - - [4992, 4609, 1, 384] - - [26, 78.626] - - - [26496, 13056, 1, 384] - - [350, 91.119] - - - [21504, 4608, 1, 384] - - [30, 88.464] - - - [18048, 9216, 1, 384] - - [396, 91.449] - - - [14592, 13441, 1, 384] - - [38, 89.353] - - - [22656, 1152, 1, 384] - - [45, 83.362] - - - [14976, 3072, 1, 384] - - [30, 86.849] - - - [24960, 13441, 1, 384] - - [37, 89.752] - - - [768, 768, 1, 384] - - [109, 32.393] - - - [12672, 4992, 1, 384] - - [26, 87.364] - - - [11136, 3072, 1, 384] - - [28, 86.907] - - - [19584, 1152, 1, 384] - - [28, 82.418] - - - [16896, 3456, 1, 384] - - [30, 87.275] - - - [23040, 1152, 1, 384] - - [28, 84.389] - - - [6528, 6528, 1, 384] - - [28, 86.8] - - - [25344, 3072, 1, 384] - - [38, 87.695] - - - [2688, 1536, 1, 384] - - [338, 66.65] - - - [5760, 1536, 1, 384] - - [36, 76.188] - - - [6144, 5760, 1, 384] - - [28, 85.977] - - - [21504, 8064, 1, 384] - - [38, 89.897] - - - [12288, 12288, 1, 384] - - [30, 88.86] - - - [16128, 13441, 1, 384] - - [28, 89.135] - - - [25344, 8448, 1, 384] - - [30, 90.218] - - - [23808, 7297, 1, 384] - - [59, 88.195] - - - [15744, 7296, 1, 384] - - [26, 89.763] - - - [16896, 13441, 1, 384] - - [26, 89.152] - - - [15360, 1920, 1, 384] - - [26, 83.634] - - - [21504, 1152, 1, 384] - - [30, 84.059] - - - [6912, 1152, 1, 384] - - [35, 69.651] - - - [16512, 3072, 1, 384] - - [33, 82.448] - - - [28800, 1153, 1, 384] - - [54, 76.181] - - - [21888, 8064, 1, 384] - - [51, 87.855] - - - [20736, 7297, 1, 384] - - [28, 88.128] - - - [10752, 10753, 1, 384] - - [30, 88.712] - - - [8832, 7297, 1, 384] - - [37, 86.031] - - - [28032, 7297, 1, 384] - - [30, 88.513] - - - [23424, 9600, 1, 384] - - [26, 90.657] - - - [23040, 13440, 1, 384] - - [38, 90.83] - - - [26880, 13441, 1, 384] - - [350, 88.02] - - - [4224, 4224, 1, 384] - - [36, 82.929] - - - [9600, 9600, 1, 384] - - [26, 88.649] - - - [26112, 1152, 1, 384] - - [59, 85.034] - - - [29568, 3456, 1, 384] - - [54, 88.517] - - - [28032, 9216, 1, 384] - - [54, 90.625] - - - [27648, 9216, 1, 384] - - [62, 88.785] - - - [17664, 1153, 1, 384] - - [54, 73.888] - - - [12672, 12289, 1, 384] - - [54, 87.114] - - - [21888, 1152, 1, 384] - - [30, 82.958] - - - [21888, 9216, 1, 384] - - [29, 87.296] - - - [10752, 10369, 1, 384] - - [38, 88.454] - - - [22656, 7296, 1, 384] - - [30, 90.219] - - - [13440, 13057, 1, 384] - - [45, 88.973] - - - [10752, 1153, 1, 384] - - [35, 70.09] - - - [12672, 3072, 1, 384] - - [26, 86.363] - - - [23424, 13440, 1, 384] - - [38, 90.874] - - - [29952, 3840, 1, 384] - - [54, 89.78] - - - [18432, 1920, 1, 384] - - [30, 85.746] - - - [26112, 7297, 1, 384] - - [30, 88.499] - - - [18816, 1153, 1, 384] - - [54, 73.484] - - - [17664, 4224, 1, 384] - - [28, 88.407] - - - [11520, 11521, 1, 384] - - [30, 88.63] - - - [30720, 1920, 1, 384] - - [26, 87.964] - - - [15360, 13441, 1, 384] - - [26, 89.017] - - - [17664, 13441, 1, 384] - - [54, 89.74] - - - [26496, 3072, 1, 384] - - [28, 87.645] - - - [20736, 4224, 1, 384] - - [59, 89.19] - - - [18816, 13441, 1, 384] - - [45, 89.24] - - - [18048, 13441, 1, 384] - - [37, 89.388] - - - [20352, 3072, 1, 384] - - [28, 88.042] - - - [1152, 768, 1, 384] - - [408, 43.128] - - - [16896, 7296, 1, 384] - - [26, 89.528] - - - [28800, 9216, 1, 384] - - [59, 90.373] - - - [9600, 1152, 1, 384] - - [54, 72.126] - - - [29952, 1153, 1, 384] - - [38, 76.04] - - - [20736, 1153, 1, 384] - - [37, 75.779] - - - [19584, 5760, 1, 384] - - [38, 89.61] - - - [29568, 7296, 1, 384] - - [54, 89.833] - - - [7296, 3072, 1, 384] - - [28, 82.095] - - - [27264, 1152, 1, 384] - - [26, 83.817] - - - [12288, 4992, 1, 384] - - [38, 86.87] - - - [5760, 5376, 1, 384] - - [38, 83.708] - - - [30720, 1152, 1, 384] - - [30, 85.402] - - - [14208, 13441, 1, 384] - - [45, 89.455] - - - [21504, 7296, 1, 384] - - [28, 89.468] - - - [7296, 6913, 1, 384] - - [28, 85.463] - - - [23808, 6912, 1, 384] - - [54, 90.007] - - - [20352, 768, 1, 384] - - [54, 79.924] - - - [2688, 2688, 1, 384] - - [229, 65.669] - - - [13056, 12672, 1, 384] - - [30, 90.482] - - - [29568, 13440, 1, 384] - - [37, 90.69] - - - [11904, 1153, 1, 384] - - [54, 70.203] - - - [2688, 2689, 1, 384] - - [50, 63.266] - - - [9984, 9985, 1, 384] - - [28, 87.492] - - - [22272, 13440, 1, 384] - - [28, 90.599] - - - [30336, 15360, 1, 384] - - [350, 91.394] - - - [21504, 7680, 1, 384] - - [30, 90.021] - - - [24192, 13441, 1, 384] - - [38, 89.274] - - - [15360, 1536, 1, 384] - - [28, 80.28] - - - [24576, 7297, 1, 384] - - [38, 83.448] - - - [11136, 3456, 1, 384] - - [28, 85.406] - - - [9600, 1153, 1, 384] - - [41, 70.514] - - - [18048, 7297, 1, 384] - - [30, 87.754] - - - [6144, 1153, 1, 384] - - [56, 61.339] - - - [23040, 9600, 1, 384] - - [26, 90.44] - - - [26880, 1153, 1, 384] - - [59, 75.08] - - - [10752, 7297, 1, 384] - - [26, 86.643] - - - [6912, 6529, 1, 384] - - [26, 85.275] - - - [29184, 9216, 1, 384] - - [57, 89.646] - - - [20736, 9216, 1, 384] - - [59, 90.051] - - - [23808, 1152, 1, 384] - - [45, 82.731] - - - [11136, 1153, 1, 384] - - [41, 72.462] - - - [25344, 1152, 1, 384] - - [26, 83.112] - - - [25344, 13441, 1, 384] - - [25, 89.36] - - - [14976, 7296, 1, 384] - - [350, 91.81] - - - [14592, 13440, 1, 384] - - [26, 90.678] - - - [7680, 7681, 1, 384] - - [26, 85.928] - - - [29568, 768, 1, 384] - - [73, 81.546] - - - [5760, 1152, 1, 384] - - [26, 68.56] - - - [21888, 13441, 1, 384] - - [29, 84.654] - - - [17664, 768, 1, 384] - - [54, 77.465] - - - [25728, 11904, 1, 384] - - [45, 90.922] - - - [9984, 2688, 1, 384] - - [37, 85.359] - - - [28416, 1153, 1, 384] - - [26, 75.623] - - - [17664, 3072, 1, 384] - - [26, 86.289] - - - [23040, 7297, 1, 384] - - [28, 88.179] - - - [8448, 8448, 1, 384] - - [54, 88.552] - - - [4608, 4225, 1, 384] - - [26, 76.866] - - - [4224, 2688, 1, 384] - - [50, 74.08] - - - [3072, 1152, 1, 384] - - [338, 68.197] - - - [29184, 1152, 1, 384] - - [26, 85.452] - - - [13440, 3072, 1, 384] - - [38, 87.428] - - - [6912, 6913, 1, 384] - - [38, 86.541] - - - [18432, 13440, 1, 384] - - [38, 90.518] - - - [14208, 7296, 1, 384] - - [350, 91.659] - - - [5376, 768, 1, 384] - - [36, 55.528] - - - [29184, 7296, 1, 384] - - [26, 90.039] - - - [20352, 1152, 1, 384] - - [28, 80.551] - - - [2304, 1153, 1, 384] - - [443, 53.269] - - - [23808, 9984, 1, 384] - - [26, 90.845] - - - [8448, 8065, 1, 384] - - [28, 86.48] - - - [24576, 1152, 1, 384] - - [26, 82.609] - - - [1536, 1537, 1, 384] - - [60, 44.725] - - - [4224, 3072, 1, 384] - - [28, 74.892] - - - [19968, 7296, 1, 384] - - [26, 90.127] - - - [19200, 5376, 1, 384] - - [37, 88.93] - - - [4608, 1152, 1, 384] - - [36, 69.296] - - - [18432, 4992, 1, 384] - - [26, 88.376] - - - [26880, 7297, 1, 384] - - [59, 88.499] - - - [15744, 3072, 1, 384] - - [28, 87.77] - - - [22272, 7296, 1, 384] - - [26, 89.771] - - - [20352, 6912, 1, 384] - - [28, 89.537] - - - [26880, 13440, 1, 384] - - [350, 91.071] - - - [4224, 3840, 1, 384] - - [26, 82.671] - - - [23424, 13441, 1, 384] - - [48, 88.888] - - - [16512, 13440, 1, 384] - - [33, 89.118] - - - [21120, 1152, 1, 384] - - [28, 82.711] - - - [10368, 3072, 1, 384] - - [28, 85.551] - - - [28032, 13440, 1, 384] - - [59, 90.785] - - - [14208, 6528, 1, 384] - - [37, 89.104] - - - [768, 769, 1, 384] - - [333, 32.652] - - - [3456, 1152, 1, 384] - - [35, 54.395] - - - [12672, 1152, 1, 384] - - [36, 75.651] - - - [7680, 3072, 1, 384] - - [26, 80.713] - - - [19200, 2304, 1, 384] - - [54, 86.548] - - - [13056, 1153, 1, 384] - - [37, 69.844] - - - [27264, 1153, 1, 384] - - [28, 75.153] - - - [29568, 1153, 1, 384] - - [59, 77.447] - - - [11520, 11136, 1, 384] - - [54, 89.532] - - - [9216, 9216, 1, 384] - - [26, 87.436] - - - [18048, 1153, 1, 384] - - [59, 75.368] - - - [8064, 1152, 1, 384] - - [26, 68.874] - - - [22272, 7297, 1, 384] - - [45, 88.518] - - - [22272, 13441, 1, 384] - - [59, 89.718] - - - [22656, 2688, 1, 384] - - [54, 88.542] - - - [19584, 6144, 1, 384] - - [59, 89.72] - - - [8064, 7297, 1, 384] - - [28, 85.223] - - - [8064, 7681, 1, 384] - - [26, 86.659] - - - [23808, 7296, 1, 384] - - [30, 90.104] - - - [24960, 7296, 1, 384] - - [28, 89.816] - - - [14208, 6912, 1, 384] - - [392, 91.738] - - - [19968, 6528, 1, 384] - - [45, 89.457] - - - [28416, 7296, 1, 384] - - [45, 90.064] - - - [29952, 13440, 1, 384] - - [38, 90.917] - - - [17280, 7297, 1, 384] - - [30, 87.711] - - - [1536, 1152, 1, 384] - - [183, 59.095] - - - [8832, 1153, 1, 384] - - [40, 65.553] - - - [28032, 1153, 1, 384] - - [28, 77.624] - - - [2688, 2305, 1, 384] - - [35, 64.521] - - - [8064, 3072, 1, 384] - - [28, 84.009] - - - [28032, 3072, 1, 384] - - [38, 89.145] - - - [3840, 3456, 1, 384] - - [36, 77.031] - - - [21888, 1920, 1, 384] - - [39, 83.085] - - - [11904, 11520, 1, 384] - - [26, 89.564] - - - [9600, 9601, 1, 384] - - [59, 88.003] - - - [21120, 13440, 1, 384] - - [30, 90.747] - - - [19584, 2688, 1, 384] - - [37, 86.563] - - - [6912, 6528, 1, 384] - - [37, 85.626] - - - [29568, 1152, 1, 384] - - [55, 85.277] - - - [23808, 3072, 1, 384] - - [38, 88.181] - - - [18816, 4992, 1, 384] - - [54, 88.521] - - - [29952, 9216, 1, 384] - - [59, 90.581] - - - [22656, 13440, 1, 384] - - [28, 90.994] - - - [20352, 3456, 1, 384] - - [54, 87.447] - - - [3456, 1153, 1, 384] - - [36, 53.265] - - - [3840, 3457, 1, 384] - - [56, 76.26] - - - [15744, 8448, 1, 384] - - [54, 90.086] - - - [26112, 3072, 1, 384] - - [30, 88.29] - - - [28032, 14208, 1, 384] - - [432, 91.174] - - - [21504, 1536, 1, 384] - - [38, 83.608] - - - [11520, 768, 1, 384] - - [54, 75.835] - - - [6528, 6144, 1, 384] - - [28, 85.453] - - - [18432, 1153, 1, 384] - - [28, 72.221] - - - [3072, 1920, 1, 384] - - [56, 62.213] - - - [25344, 9216, 1, 384] - - [59, 90.665] - - - [30336, 7297, 1, 384] - - [38, 87.864] - - - [8832, 1152, 1, 384] - - [50, 75.114] - - - [26112, 9216, 1, 384] - - [59, 90.379] - - - [29952, 7296, 1, 384] - - [28, 90.178] - - - [11520, 11137, 1, 384] - - [30, 89.105] - - - [16896, 13440, 1, 384] - - [28, 90.773] - - - [29568, 13441, 1, 384] - - [52, 88.47] - - - [30336, 9216, 1, 384] - - [54, 90.28] - - - [2688, 1152, 1, 384] - - [338, 61.949] - - - [10368, 10368, 1, 384] - - [59, 89.575] - - - [25344, 11520, 1, 384] - - [28, 91.069] - - - [24576, 1920, 1, 384] - - [30, 83.311] - - - [11904, 4608, 1, 384] - - [54, 87.354] - - - [12672, 5376, 1, 384] - - [54, 88.794] - - - [11520, 3072, 1, 384] - - [26, 85.887] - - - [3072, 3073, 1, 384] - - [28, 69.368] - - - [24960, 11136, 1, 384] - - [54, 90.638] - - - [9984, 9600, 1, 384] - - [26, 88.848] - - - [19200, 2688, 1, 384] - - [54, 87.33] - - - [26496, 7296, 1, 384] - - [26, 90.121] - - - [23040, 3072, 1, 384] - - [38, 87.388] - - - [5760, 5761, 1, 384] - - [28, 84.571] - - - [5760, 5377, 1, 384] - - [30, 83.087] - - - [26880, 768, 1, 384] - - [54, 81.248] - - - [13824, 7297, 1, 384] - - [30, 88.168] - - - [13440, 7296, 1, 384] - - [26, 89.037] - - - [16128, 8448, 1, 384] - - [37, 89.625] - - - [24960, 3072, 1, 384] - - [28, 88.551] - - - [6144, 6144, 1, 384] - - [28, 86.096] - - - [27648, 13441, 1, 384] - - [350, 88.393] - - - [10368, 7297, 1, 384] - - [37, 87.132] - - - [22272, 2304, 1, 384] - - [26, 87.15] - - - [30720, 1153, 1, 384] - - [38, 77.37] - - - [24192, 13440, 1, 384] - - [28, 90.816] - - - [9984, 9984, 1, 384] - - [28, 89.105] - - - [29952, 1152, 1, 384] - - [26, 84.039] - - - [26112, 12672, 1, 384] - - [350, 91.084] - - - [8448, 7296, 1, 384] - - [28, 87.405] - - - [19584, 13440, 1, 384] - - [38, 90.462] - - - [21120, 1153, 1, 384] - - [38, 76.613] - - - [8832, 8449, 1, 384] - - [54, 86.262] - - - [28032, 13441, 1, 384] - - [28, 89.389] - - - [7680, 1153, 1, 384] - - [24, 65.5] - - - [19584, 9216, 1, 384] - - [54, 89.89] - - - [28800, 1152, 1, 384] - - [59, 84.104] - - - [29952, 768, 1, 384] - - [54, 83.522] - - - [12288, 1152, 1, 384] - - [28, 80.098] - - - [9600, 9217, 1, 384] - - [63, 86.678] - - - [14976, 13441, 1, 384] - - [30, 89.511] - - - [25344, 8832, 1, 384] - - [28, 90.401] - - - [18432, 4608, 1, 384] - - [26, 87.45] - - - [2304, 1920, 1, 384] - - [50, 59.968] - - - [11520, 4224, 1, 384] - - [28, 88.351] - - - [26496, 1153, 1, 384] - - [54, 77.269] - - - [28416, 2304, 1, 384] - - [54, 87.819] - - - [19200, 3072, 1, 384] - - [38, 87.757] - - - [26112, 7296, 1, 384] - - [30, 90.453] - - - [21504, 7297, 1, 384] - - [26, 87.803] - - - [4224, 1152, 1, 384] - - [61, 64.294] - - - [17664, 3840, 1, 384] - - [59, 88.524] - - - [6144, 1536, 1, 384] - - [50, 70.378] - - - [28032, 14592, 1, 384] - - [37, 91.167] - - - [8064, 8064, 1, 384] - - [28, 87.259] - - - [11136, 1152, 1, 384] - - [28, 74.068] - - - [13056, 7297, 1, 384] - - [26, 87.919] - - - [19968, 3456, 1, 384] - - [54, 87.779] - - - [25344, 7297, 1, 384] - - [30, 88.445] - - - [17280, 3840, 1, 384] - - [28, 88.77] - - - [28416, 1152, 1, 384] - - [28, 83.942] - - - [21120, 3072, 1, 384] - - [26, 88.775] - - - [28416, 7297, 1, 384] - - [38, 88.08] - - - [6528, 6529, 1, 384] - - [28, 83.475] - - - [26496, 9216, 1, 384] - - [37, 90.68] - - - [14592, 7296, 1, 384] - - [350, 91.404] - - - [14208, 1152, 1, 384] - - [28, 76.797] - - - [24576, 1536, 1, 384] - - [38, 83.575] - - - [18048, 7296, 1, 384] - - [38, 89.317] - - - [4608, 3072, 1, 384] - - [36, 80.893] - - - [28800, 14976, 1, 384] - - [45, 90.941] - - - [17664, 1152, 1, 384] - - [37, 80.529] - - - [24576, 7680, 1, 384] - - [38, 86.674] - - - [16896, 9216, 1, 384] - - [45, 89.757] - - - [20736, 3840, 1, 384] - - [45, 88.309] - - - [27264, 9216, 1, 384] - - [26, 89.624] - - - [21888, 3072, 1, 384] - - [31, 84.822] - - - [24576, 11136, 1, 384] - - [23, 87.405] - - - [14592, 1153, 1, 384] - - [37, 71.604] - - - [23424, 7296, 1, 384] - - [26, 90.291] - - - [22272, 3072, 1, 384] - - [28, 88.543] - - - [8832, 8832, 1, 384] - - [38, 88.346] - - - [8064, 7296, 1, 384] - - [28, 88.017] - - - [22656, 8832, 1, 384] - - [59, 90.025] - - - [22272, 2688, 1, 384] - - [45, 87.261] - - - [6528, 1152, 1, 384] - - [50, 65.966] - - - [8832, 8833, 1, 384] - - [37, 87.997] - - - [28800, 15360, 1, 384] - - [37, 90.849] - - - [23424, 1153, 1, 384] - - [28, 75.606] - - - [13440, 1152, 1, 384] - - [54, 79.284] - - - [10752, 10368, 1, 384] - - [37, 89.934] - - - [3456, 3456, 1, 384] - - [56, 77.322] - - - [4608, 4608, 1, 384] - - [26, 83.853] - - - [4224, 1153, 1, 384] - - [36, 63.2] - - - [12672, 2304, 1, 384] - - [26, 83.227] - - - [25728, 7297, 1, 384] - - [59, 88.005] - - - [5376, 1153, 1, 384] - - [35, 63.522] - - - [30720, 4992, 1, 384] - - [26, 89.461] - - - [27264, 7297, 1, 384] - - [30, 87.552] - - - [21504, 1920, 1, 384] - - [28, 87.192] - - - [11136, 11136, 1, 384] - - [59, 90.048] - - - [22656, 6144, 1, 384] - - [59, 89.491] - - - [26496, 13440, 1, 384] - - [350, 91.203] - - - [9216, 7296, 1, 384] - - [30, 87.749] - - - [17280, 7296, 1, 384] - - [38, 90.133] - - - [23040, 13441, 1, 384] - - [30, 89.079] - - - [23808, 13441, 1, 384] - - [37, 89.697] - - - [30336, 4224, 1, 384] - - [30, 89.287] - - - [6144, 1920, 1, 384] - - [56, 76.159] - - - [11904, 11904, 1, 384] - - [59, 90.052] - - - [30336, 13441, 1, 384] - - [26, 89.106] - - - [11904, 1536, 1, 384] - - [38, 78.25] - - - [24576, 9216, 1, 384] - - [33, 84.679] - - - [9984, 2304, 1, 384] - - [59, 83.947] - - - [18048, 4608, 1, 384] - - [54, 88.225] - - - [18432, 7297, 1, 384] - - [30, 87.867] - - - [11136, 3840, 1, 384] - - [28, 87.16] - - - [12288, 11904, 1, 384] - - [38, 89.561] - - - [19584, 7296, 1, 384] - - [28, 89.672] - - - [3072, 2689, 1, 384] - - [35, 71.091] - - - [2304, 2305, 1, 384] - - [56, 56.221] - - - [26496, 7297, 1, 384] - - [30, 88.044] - - - [15744, 1152, 1, 384] - - [26, 77.976] - - - [6912, 6912, 1, 384] - - [30, 87.066] - - - [4992, 3072, 1, 384] - - [28, 79.07] - - - [15744, 13440, 1, 384] - - [30, 90.395] - - - [2688, 2304, 1, 384] - - [35, 65.324] - - - [8448, 7297, 1, 384] - - [28, 86.977] - - - [25344, 11904, 1, 384] - - [28, 90.852] - - - [18432, 7296, 1, 384] - - [28, 89.726] - - - [8448, 8449, 1, 384] - - [30, 86.577] - - - [30720, 1536, 1, 384] - - [28, 85.263] - - - [9216, 1153, 1, 384] - - [45, 68.145] - - - [24192, 9216, 1, 384] - - [59, 90.538] - - - [25344, 2688, 1, 384] - - [59, 88.692] - - - [24576, 1153, 1, 384] - - [30, 73.377] - - - [14208, 7297, 1, 384] - - [54, 87.622] - - - [12672, 1920, 1, 384] - - [38, 82.978] - - - [4608, 4224, 1, 384] - - [36, 83.371] - - - [27264, 1536, 1, 384] - - [28, 84.468] - - - [24576, 13441, 1, 384] - - [26, 84.434] - - - [21504, 4992, 1, 384] - - [28, 89.361] - - - [21888, 4992, 1, 384] - - [51, 87.042] - - - [18432, 3072, 1, 384] - - [38, 86.579] - - - [19968, 6144, 1, 384] - - [28, 88.894] - - - [24192, 1536, 1, 384] - - [28, 85.442] - - - [9600, 7297, 1, 384] - - [28, 86.663] - - - [13824, 6528, 1, 384] - - [353, 90.995] - - - [2304, 2304, 1, 384] - - [50, 69.873] - - - [23424, 9984, 1, 384] - - [30, 90.684] - - - [18816, 1152, 1, 384] - - [28, 79.623] - - - [1152, 769, 1, 384] - - [172, 35.891] - - - [23424, 768, 1, 384] - - [54, 83.094] - - - [17280, 1153, 1, 384] - - [37, 72.434] - - - [9600, 2304, 1, 384] - - [28, 81.095] - - - [29184, 7297, 1, 384] - - [30, 88.119] - - - [26880, 3072, 1, 384] - - [30, 88.803] - - - [11520, 11520, 1, 384] - - [26, 90.049] - - - [23040, 6144, 1, 384] - - [59, 89.401] - - - [18048, 13440, 1, 384] - - [59, 90.675] - - - [30336, 1536, 1, 384] - - [28, 86.95] - - - [14976, 7680, 1, 384] - - [353, 92.265] - - - [14976, 1152, 1, 384] - - [26, 80.267] - - - [15360, 7680, 1, 384] - - [353, 92.224] - - - [28800, 13441, 1, 384] - - [26, 88.768] - - - [28032, 1920, 1, 384] - - [26, 88.185] - - - [16128, 2688, 1, 384] - - [59, 84.65] - - - [6144, 6145, 1, 384] - - [26, 81.33] - - - [10368, 7296, 1, 384] - - [28, 89.3] - - - [5760, 3072, 1, 384] - - [392, 83.163] - - - [24960, 9216, 1, 384] - - [37, 90.634] - - - [14592, 768, 1, 384] - - [61, 72.813] - - - [14208, 768, 1, 384] - - [41, 70.96] - - - [6912, 1153, 1, 384] - - [41, 68.662] - - - [21888, 13440, 1, 384] - - [29, 88.568] - - - [13056, 5760, 1, 384] - - [28, 89.082] - - - [12288, 1920, 1, 384] - - [28, 80.703] - - - [13056, 13056, 1, 384] - - [28, 90.324] - - - [6528, 1153, 1, 384] - - [35, 64.455] - - - [22272, 8448, 1, 384] - - [59, 90.311] - - - [7296, 1153, 1, 384] - - [61, 62.16] - - - [17280, 3456, 1, 384] - - [37, 87.076] - - - [27264, 13441, 1, 384] - - [350, 88.182] - - - [9216, 7297, 1, 384] - - [28, 87.23] - - - [4992, 4992, 1, 384] - - [38, 85.007] - - - [16128, 7297, 1, 384] - - [28, 87.628] - - - [20352, 13440, 1, 384] - - [26, 90.851] - - - [30336, 1153, 1, 384] - - [37, 76.888] - - - [13056, 7296, 1, 384] - - [30, 89.592] - - - [27648, 1152, 1, 384] - - [30, 85.05] - - - [13824, 6144, 1, 384] - - [26, 88.162] - - - [9216, 1920, 1, 384] - - [28, 81.957] - - - [17280, 13440, 1, 384] - - [28, 90.76] - - - [21888, 5376, 1, 384] - - [31, 85.822] - - - [3456, 3072, 1, 384] - - [35, 78.283] - - - [13440, 1153, 1, 384] - - [37, 71.698] - - - [24192, 7680, 1, 384] - - [26, 90.064] - - - [29952, 4224, 1, 384] - - [28, 89.448] - - - [8832, 3072, 1, 384] - - [30, 81.695] - - - [5760, 5760, 1, 384] - - [26, 85.038] - - - [23424, 6912, 1, 384] - - [28, 90.398] - - - [24192, 3072, 1, 384] - - [26, 87.871] - - - [18048, 3072, 1, 384] - - [26, 87.997] - - - [27264, 7296, 1, 384] - - [30, 90.021] - - - [11520, 3840, 1, 384] - - [30, 86.779] - - - [18432, 1536, 1, 384] - - [26, 84.098] - - - [11136, 10753, 1, 384] - - [28, 87.819] - - - [9600, 7296, 1, 384] - - [28, 89.09] - - - [26496, 13441, 1, 384] - - [28, 89.289] - - - [29568, 9216, 1, 384] - - [37, 89.74] - - - [25728, 7296, 1, 384] - - [28, 89.917] - - - [6528, 3072, 1, 384] - - [447, 82.002] - - - [18816, 9216, 1, 384] - - [396, 91.738] - - - [1920, 1153, 1, 384] - - [35, 42.154] - - - [1152, 1153, 1, 384] - - [182, 46.328] - - - [16896, 1153, 1, 384] - - [45, 75.524] - - - [4992, 1153, 1, 384] - - [61, 60.197] - - - [22656, 13441, 1, 384] - - [59, 89.502] - - - [9984, 1152, 1, 384] - - [35, 74.729] - - - [26496, 768, 1, 384] - - [37, 80.366] - - - [25344, 2304, 1, 384] - - [28, 87.575] - - - [14592, 6912, 1, 384] - - [392, 92.027] - - - [9216, 8833, 1, 384] - - [30, 87.479] - - - [19584, 7297, 1, 384] - - [26, 88.16] - - - [8448, 1153, 1, 384] - - [36, 71.078] - - - [21120, 7297, 1, 384] - - [48, 88.024] - - - [11520, 7297, 1, 384] - - [30, 87.053] - - - [12288, 7296, 1, 384] - - [30, 89.211] - - - [4224, 3841, 1, 384] - - [50, 76.561] - - - [9984, 9601, 1, 384] - - [28, 88.172] - - - [2304, 1152, 1, 384] - - [336, 63.882] - - - [21120, 7296, 1, 384] - - [28, 89.903] - - - [15360, 1153, 1, 384] - - [54, 74.403] - - - [27648, 3072, 1, 384] - - [30, 86.696] - - - [19200, 1153, 1, 384] - - [28, 74.708] - - - [28032, 1152, 1, 384] - - [59, 86.259] - - - [12672, 12288, 1, 384] - - [54, 90.041] - - - [22272, 5760, 1, 384] - - [26, 89.478] - - - [26496, 1152, 1, 384] - - [37, 86.187] - - - [26880, 7296, 1, 384] - - [28, 90.428] - - - [6528, 2304, 1, 384] - - [35, 77.781] - - - [9984, 7296, 1, 384] - - [28, 88.426] - - - [19968, 1152, 1, 384] - - [59, 83.787] - - - [10368, 9984, 1, 384] - - [30, 89.268] - - - [3840, 3840, 1, 384] - - [36, 76.876] - - - [5376, 1152, 1, 384] - - [36, 64.654] - - - [24192, 7296, 1, 384] - - [26, 90.513] - - - [14592, 3072, 1, 384] - - [30, 87.402] - - - [27648, 7297, 1, 384] - - [28, 87.86] - - - [23424, 1152, 1, 384] - - [28, 85.461] - - - [3456, 3457, 1, 384] - - [35, 76.449] - - - [13056, 2304, 1, 384] - - [30, 85.22] - - - [23808, 768, 1, 384] - - [37, 78.629] - - - [18048, 1152, 1, 384] - - [28, 81.926] - - - [28416, 9216, 1, 384] - - [54, 90.634] - - - [21888, 7297, 1, 384] - - [68, 83.447] - - - [25728, 12288, 1, 384] - - [59, 90.662] - - - [21120, 4224, 1, 384] - - [28, 88.889] - - - [20736, 3072, 1, 384] - - [30, 87.451] - - - [3840, 2688, 1, 384] - - [56, 76.533] - - - [29568, 7297, 1, 384] - - [48, 87.324] - - - [13824, 1153, 1, 384] - - [37, 73.584] - - - [15744, 1153, 1, 384] - - [54, 71.28] - - - [11136, 768, 1, 384] - - [41, 73.494] - - - [17664, 7297, 1, 384] - - [30, 88.339] - - - [24192, 7297, 1, 384] - - [30, 88.31] - - - [25344, 1153, 1, 384] - - [37, 77.647] - - - [30720, 4608, 1, 384] - - [26, 88.764] - - - [25728, 9216, 1, 384] - - [54, 90.516] - - - [29184, 1153, 1, 384] - - [45, 77.332] - - - [30336, 1152, 1, 384] - - [45, 84.968] - - - [24960, 13440, 1, 384] - - [45, 90.738] - - - [18432, 9216, 1, 384] - - [353, 91.624] - - - [15360, 13440, 1, 384] - - [28, 90.359] - - - [12288, 1536, 1, 384] - - [26, 80.308] - - - [8832, 8448, 1, 384] - - [28, 88.439] - - - [19968, 7297, 1, 384] - - [38, 88.003] - - - [19968, 3072, 1, 384] - - [38, 86.678] - - - [24960, 1920, 1, 384] - - [28, 87.186] - - - [15360, 1152, 1, 384] - - [28, 81.733] - - - [30720, 7296, 1, 384] - - [28, 90.195] - - - [14976, 1153, 1, 384] - - [54, 73.21] - - - [25344, 7296, 1, 384] - - [28, 90.358] - - - [16512, 8832, 1, 384] - - [42, 87.877] - - - [26112, 13441, 1, 384] - - [30, 89.566] - - - [22272, 1152, 1, 384] - - [30, 82.343] - - - [27648, 1536, 1, 384] - - [28, 85.661] - - - [15744, 1920, 1, 384] - - [26, 85.498] - - - [5760, 1153, 1, 384] - - [61, 57.969] - - - [29952, 13441, 1, 384] - - [28, 89.176] - - - [12672, 1153, 1, 384] - - [45, 74.098] - - - [13440, 2688, 1, 384] - - [59, 83.727] - - - [18816, 13440, 1, 384] - - [45, 90.572] - - - [22656, 9216, 1, 384] - - [37, 90.68] - - - [9216, 1152, 1, 384] - - [36, 77.636] - - - [20736, 1152, 1, 384] - - [26, 81.729] - - - [8832, 7296, 1, 384] - - [28, 88.738] - - - [15744, 7297, 1, 384] - - [26, 88.247] - - - [16512, 1153, 1, 384] - - [79, 70.083] - - - [29952, 7297, 1, 384] - - [38, 88.269] - - - [11136, 7297, 1, 384] - - [28, 87.578] - - - [9600, 3072, 1, 384] - - [26, 83.769] - - - [28800, 7297, 1, 384] - - [26, 87.828] - - - [27648, 13824, 1, 384] - - [350, 91.418] - - - [23808, 10368, 1, 384] - - [54, 90.899] - - - [13824, 13440, 1, 384] - - [28, 90.055] - - - [9216, 1536, 1, 384] - - [26, 80.124] - - - [23808, 1153, 1, 384] - - [45, 76.937] - - - [15360, 3072, 1, 384] - - [28, 85.325] - - - [12288, 3072, 1, 384] - - [28, 86.147] - - - [28416, 3072, 1, 384] - - [28, 88.406] - - - [30336, 13440, 1, 384] - - [26, 90.835] - - - [1152, 1152, 1, 384] - - [173, 53.392] - - - [21504, 3072, 1, 384] - - [38, 86.952] - - - [23040, 9216, 1, 384] - - [59, 90.364] - - - [22656, 7297, 1, 384] - - [38, 88.163] - - - [22656, 5760, 1, 384] - - [26, 89.68] - - - [12288, 11905, 1, 384] - - [28, 88.072] - - - [28032, 7296, 1, 384] - - [30, 90.255] - - - [29184, 3072, 1, 384] - - [30, 88.836] - - - [7680, 1152, 1, 384] - - [50, 75.841] - - - [16896, 7297, 1, 384] - - [38, 88.076] - - - [13056, 5376, 1, 384] - - [28, 89.113] - - - [5376, 4993, 1, 384] - - [28, 80.619] - - - [17280, 9216, 1, 384] - - [59, 89.968] - - - [8448, 8064, 1, 384] - - [26, 88.899] - - - [4608, 1153, 1, 384] - - [41, 55.973] - - - [19200, 9216, 1, 384] - - [396, 91.726] - - - [30720, 7297, 1, 384] - - [26, 88.048] - - - [13440, 5760, 1, 384] - - [26, 89.431] - - - [9984, 3072, 1, 384] - - [38, 82.616] - - - [29952, 15360, 1, 384] - - [26, 90.909] - - - [3840, 1152, 1, 384] - - [50, 60.113] - - - [10368, 9985, 1, 384] - - [28, 87.627] - - - [14592, 7297, 1, 384] - - [350, 88.489] - - - [3456, 3073, 1, 384] - - [28, 68.565] - - - [22272, 9216, 1, 384] - - [54, 90.666] - - - [8064, 8065, 1, 384] - - [38, 86.659] - - - [1536, 1536, 1, 384] - - [228, 53.023] - - - [30336, 4608, 1, 384] - - [54, 89.816] - - - [26112, 12288, 1, 384] - - [54, 90.415] - - - [11904, 11521, 1, 384] - - [54, 89.158] - - - [13440, 6144, 1, 384] - - [37, 89.227] - - - [19200, 13440, 1, 384] - - [38, 90.448] - - - [17280, 1152, 1, 384] - - [28, 78.728] - - - [23424, 3072, 1, 384] - - [28, 88.984] - - - [2304, 1921, 1, 384] - - [36, 59.305] - - - [12672, 7297, 1, 384] - - [26, 86.974] - - - [16896, 1152, 1, 384] - - [26, 82.505] - - - [18432, 1152, 1, 384] - - [28, 83.095] - - - [27264, 13824, 1, 384] - - [350, 91.304] - - - [10752, 1152, 1, 384] - - [35, 79.228] - - - [30336, 7296, 1, 384] - - [30, 90.3] - - - [11904, 3072, 1, 384] - - [38, 84.928] - - - [2304, 768, 1, 384] - - [343, 59.574] - - - [14592, 1152, 1, 384] - - [28, 78.634] - - - [20736, 13441, 1, 384] - - [54, 89.398] - - - [10752, 10752, 1, 384] - - [28, 89.21] - - - [23808, 13440, 1, 384] - - [26, 90.934] - - - [5376, 4992, 1, 384] - - [38, 85.527] - - - [10752, 3072, 1, 384] - - [26, 84.298] - - - [24576, 7296, 1, 384] - - [28, 87.228] - - - [7296, 7296, 1, 384] - - [38, 87.616] - - - [19200, 7296, 1, 384] - - [30, 89.94] - - - [25728, 8832, 1, 384] - - [30, 90.353] - - - [18048, 4224, 1, 384] - - [26, 88.169] - - - [4992, 1152, 1, 384] - - [56, 60.473] - - - [22272, 8832, 1, 384] - - [26, 90.167] - - - [21504, 1153, 1, 384] - - [38, 74.005] - - - [14208, 13440, 1, 384] - - [38, 90.637] - - - [10752, 7296, 1, 384] - - [38, 88.717] - - - [24192, 1152, 1, 384] - - [26, 83.698] - - - [7296, 1152, 1, 384] - - [35, 72.559] - - - [16128, 1153, 1, 384] - - [32, 72.093] - - - [19200, 7297, 1, 384] - - [28, 87.529] - - - [4992, 4993, 1, 384] - - [28, 79.64] - - - [12672, 12673, 1, 384] - - [28, 89.458] - - - [14208, 3072, 1, 384] - - [28, 85.44] - - - [23424, 6528, 1, 384] - - [26, 89.85] - - - [24576, 8064, 1, 384] - - [23, 87.204] - - - [6528, 6145, 1, 384] - - [38, 82.886] - - - [1920, 1537, 1, 384] - - [35, 55.069] - - - [21888, 8448, 1, 384] - - [29, 87.485] - - - [3072, 1536, 1, 384] - - [367, 66.064] - - - [7680, 7296, 1, 384] - - [28, 86.882] - - - [16896, 3072, 1, 384] - - [26, 87.3] - - - [24960, 11520, 1, 384] - - [28, 90.87] - - - [13824, 1152, 1, 384] - - [30, 81.264] - - - [25728, 1153, 1, 384] - - [28, 75.264] - - - [19968, 13441, 1, 384] - - [28, 89.301] - - - [13056, 13057, 1, 384] - - [26, 89.144] - - - [29184, 13440, 1, 384] - - [30, 90.843] - - - [23424, 7297, 1, 384] - - [31, 87.659] - - - [9216, 8832, 1, 384] - - [38, 88.217] - - - [11520, 1153, 1, 384] - - [35, 68.201] - - - [19968, 1153, 1, 384] - - [37, 73.427] - - - [14976, 13440, 1, 384] - - [26, 90.603] - - - [9216, 3072, 1, 384] - - [28, 84.6] - - - [24192, 10752, 1, 384] - - [28, 90.661] - - - [16128, 8832, 1, 384] - - [38, 89.401] - - - [9984, 1153, 1, 384] - - [59, 66.306] - - - [8064, 1153, 1, 384] - - [60, 67.1] - - - [12672, 12672, 1, 384] - - [28, 89.862] - - - [25728, 13441, 1, 384] - - [37, 89.082] - - - [11520, 1152, 1, 384] - - [28, 76.213] - - - [26496, 12672, 1, 384] - - [37, 90.847] - - - [1920, 768, 1, 384] - - [442, 52.183] - - - [20352, 1153, 1, 384] - - [54, 74.649] - - - [10368, 2688, 1, 384] - - [37, 83.77] - - - [6912, 2304, 1, 384] - - [45, 81.625] - - - [17664, 13440, 1, 384] - - [54, 90.735] - - - [17664, 9216, 1, 384] - - [54, 90.066] - - - [25728, 13440, 1, 384] - - [37, 90.911] - - - [10752, 3456, 1, 384] - - [26, 86.158] - - - [6144, 3072, 1, 384] - - [353, 81.86] - - - [9216, 9217, 1, 384] - - [29, 84.232] - - - [3840, 2304, 1, 384] - - [50, 76.562] - - - [12288, 12289, 1, 384] - - [55, 83.752] - - - [11136, 11137, 1, 384] - - [26, 88.387] - - - [11904, 7297, 1, 384] - - [28, 86.522] - - - [29568, 3072, 1, 384] - - [52, 87.919] - - - [12288, 1153, 1, 384] - - [28, 71.726] - - - [18816, 1920, 1, 384] - - [28, 83.941] - - - [13056, 1152, 1, 384] - - [28, 77.549] - - - [8448, 768, 1, 384] - - [41, 67.489] - - - [18816, 2304, 1, 384] - - [45, 85.066] - - - [5376, 3072, 1, 384] - - [28, 77.57] - - - [16512, 1152, 1, 384] - - [46, 75.119] - - - [27648, 7296, 1, 384] - - [38, 90.303] - - - [7296, 2688, 1, 384] - - [28, 83.159] - - - [29184, 15360, 1, 384] - - [28, 90.741] - - - [4608, 4609, 1, 384] - - [28, 77.982] - - - [7296, 7297, 1, 384] - - [26, 84.469] - - - [30720, 9216, 1, 384] - - [75, 88.322] - - - [16384, 3072, 1, 256] - - [23, 62.582] - - - [42496, 10240, 1, 256] - - [30, 74.992] - - - [20992, 7168, 1, 256] - - [28, 73.905] - - - [8960, 5632, 1, 256] - - [60, 72.47] - - - [4864, 256, 1, 256] - - [194, 41.505] - - - [23552, 3584, 1, 256] - - [26, 73.921] - - - [2560, 1281, 1, 256] - - [95, 50.321] - - - [7168, 1280, 1, 256] - - [30, 57.663] - - - [1536, 1153, 1, 384] - - [52, 33.947] - - - [18224, 256, 1, 256] - - [41, 46.948] - - - [13441, 128, 1, 384] - - [195, 47.394] - - - [10753, 128, 1, 384] - - [196, 40.289] - - - [12289, 128, 1, 384] - - [197, 42.921] - - - [385, 128, 1, 384] - - [161, 3.33] - - - [11136, 128, 1, 384] - - [183, 48.351] - - - [13440, 128, 1, 384] - - [182, 55.421] - - - [1153, 128, 1, 384] - - [133, 10.092] - - - [6145, 128, 1, 384] - - [149, 37.382] - - - [4225, 128, 1, 384] - - [152, 26.504] - - - [1537, 128, 1, 384] - - [133, 13.779] - - - [8064, 128, 1, 384] - - [109, 42.016] - - - [3072, 128, 1, 384] - - [122, 28.226] - - - [3457, 128, 1, 384] - - [119, 26.58] - - - [5760, 128, 1, 384] - - [109, 39.611] - - - [8449, 128, 1, 384] - - [141, 38.339] - - - [2305, 128, 1, 384] - - [151, 18.904] - - - [11520, 128, 1, 384] - - [184, 49.823] - - - [11521, 128, 1, 384] - - [198, 42.238] - - - [6528, 128, 1, 384] - - [109, 44.071] - - - [14208, 128, 1, 384] - - [24, 34.859] - - - [768, 128, 1, 384] - - [111, 7.268] - - - [12672, 128, 1, 384] - - [112, 53.145] - - - [9216, 128, 1, 384] - - [126, 45.249] - - - [8448, 128, 1, 384] - - [122, 44.434] - - - [6144, 128, 1, 384] - - [115, 41.479] - - - [2689, 128, 1, 384] - - [119, 21.972] - - - [4224, 128, 1, 384] - - [111, 29.979] - - - [9601, 128, 1, 384] - - [141, 41.933] - - - [13056, 128, 1, 384] - - [199, 54.343] - - - [8065, 128, 1, 384] - - [128, 36.521] - - - [2304, 128, 1, 384] - - [109, 21.347] - - - [8833, 128, 1, 384] - - [152, 39.514] - - - [13824, 128, 1, 384] - - [200, 56.065] - - - [7680, 128, 1, 384] - - [126, 38.65] - - - [3840, 128, 1, 384] - - [110, 27.254] - - - [1920, 128, 1, 384] - - [115, 18.247] - - - [5761, 128, 1, 384] - - [138, 35.338] - - - [7681, 128, 1, 384] - - [201, 34.854] - - - [4608, 128, 1, 384] - - [124, 32.496] - - - [10369, 128, 1, 384] - - [160, 44.325] - - - [3841, 128, 1, 384] - - [152, 24.094] - - - [7296, 128, 1, 384] - - [111, 38.194] - - - [7297, 128, 1, 384] - - [149, 33.523] - - - [10752, 128, 1, 384] - - [145, 47.522] - - - [1536, 128, 1, 384] - - [167, 14.724] - - - [11137, 128, 1, 384] - - [202, 40.762] - - - [2688, 128, 1, 384] - - [115, 24.495] - - - [4609, 128, 1, 384] - - [164, 28.912] - - - [6529, 128, 1, 384] - - [149, 39.286] - - - [11905, 128, 1, 384] - - [202, 43.004] - - - [6912, 128, 1, 384] - - [115, 46.101] - - - [769, 128, 1, 384] - - [129, 6.811] - - - [12288, 128, 1, 384] - - [192, 52.227] - - - [15360, 128, 1, 384] - - [58, 37.498] - - - [9600, 128, 1, 384] - - [115, 48.868] - - - [13057, 128, 1, 384] - - [203, 46.04] - - - [10368, 128, 1, 384] - - [115, 51.018] - - - [12673, 128, 1, 384] - - [136, 45.262] - - - [9217, 128, 1, 384] - - [201, 38.95] - - - [4993, 128, 1, 384] - - [142, 31.057] - - - [9984, 128, 1, 384] - - [138, 49.458] - - - [6913, 128, 1, 384] - - [164, 40.93] - - - [8832, 128, 1, 384] - - [173, 45.694] - - - [3073, 128, 1, 384] - - [204, 23.96] - - - [14976, 128, 1, 384] - - [35, 36.622] - - - [384, 128, 1, 384] - - [110, 3.649] - - - [5377, 128, 1, 384] - - [149, 33.352] - - - [1152, 128, 1, 384] - - [167, 10.996] - - - [9985, 128, 1, 384] - - [160, 42.849] - - - [14592, 128, 1, 384] - - [40, 35.35] - - - [4992, 128, 1, 384] - - [109, 34.871] - - - [3456, 128, 1, 384] - - [173, 30.612] - - - [1921, 128, 1, 384] - - [166, 15.697] - - - [5376, 128, 1, 384] - - [110, 37.318] - - - [11904, 128, 1, 384] - - [192, 50.594] - - - [44544, 2048, 1, 384] - - [54, 88.554] - - - [39552, 512, 1, 384] - - [45, 79.797] - - - [38016, 22145, 1, 384] - - [38, 89.555] - - - [39552, 23297, 1, 384] - - [38, 89.357] - - - [39552, 23681, 1, 384] - - [30, 89.45] - - - [36864, 2048, 1, 384] - - [38, 82.377] - - - [44544, 28673, 1, 384] - - [33, 87.03] - - - [43776, 512, 1, 384] - - [27, 79.402] - - - [43392, 1024, 1, 384] - - [37, 86.883] - - - [42240, 4096, 1, 384] - - [37, 90.427] - - - [42624, 26369, 1, 384] - - [52, 83.329] - - - [35328, 1024, 1, 384] - - [59, 83.966] - - - [36096, 384, 1, 384] - - [387, 79.425] - - - [38784, 4096, 1, 384] - - [37, 89.674] - - - [39552, 384, 1, 384] - - [392, 78.502] - - - [42240, 8192, 1, 384] - - [54, 91.007] - - - [42240, 25985, 1, 384] - - [26, 89.72] - - - [38016, 4096, 1, 384] - - [59, 89.941] - - - [39168, 4096, 1, 384] - - [54, 89.71] - - - [35328, 19457, 1, 384] - - [33, 87.338] - - - [43392, 2048, 1, 384] - - [37, 88.584] - - - [38400, 4096, 1, 384] - - [45, 88.905] - - - [35712, 1024, 1, 384] - - [54, 84.814] - - - [36480, 2048, 1, 384] - - [45, 87.979] - - - [40704, 512, 1, 384] - - [28, 81.464] - - - [36864, 20609, 1, 384] - - [38, 88.091] - - - [37632, 21761, 1, 384] - - [38, 89.191] - - - [38016, 2048, 1, 384] - - [45, 87.947] - - - [44160, 2048, 1, 384] - - [54, 88.431] - - - [35328, 384, 1, 384] - - [393, 78.175] - - - [43392, 384, 1, 384] - - [350, 79.041] - - - [39168, 512, 1, 384] - - [28, 79.257] - - - [38784, 1024, 1, 384] - - [45, 84.387] - - - [35328, 2048, 1, 384] - - [59, 87.335] - - - [44544, 8192, 1, 384] - - [26, 89.941] - - - [40704, 384, 1, 384] - - [346, 80.262] - - - [39936, 512, 1, 384] - - [28, 80.271] - - - [41472, 25217, 1, 384] - - [30, 89.363] - - - [42240, 2048, 1, 384] - - [59, 89.372] - - - [37632, 512, 1, 384] - - [37, 81.473] - - - [37248, 1024, 1, 384] - - [37, 84.781] - - - [42240, 26369, 1, 384] - - [28, 89.63] - - - [43776, 384, 1, 384] - - [346, 78.769] - - - [44160, 8192, 1, 384] - - [59, 90.909] - - - [39936, 1024, 1, 384] - - [59, 86.121] - - - [43392, 27137, 1, 384] - - [26, 88.823] - - - [39936, 384, 1, 384] - - [353, 79.147] - - - [41472, 25601, 1, 384] - - [33, 87.287] - - - [36864, 4096, 1, 384] - - [33, 84.148] - - - [43392, 8192, 1, 384] - - [59, 90.761] - - - [36096, 512, 1, 384] - - [28, 78.538] - - - [36480, 4096, 1, 384] - - [37, 90.149] - - - [40320, 512, 1, 384] - - [38, 80.915] - - - [41088, 4096, 1, 384] - - [75, 88.437] - - - [43776, 27521, 1, 384] - - [55, 88.034] - - - [35328, 19073, 1, 384] - - [30, 89.484] - - - [44160, 384, 1, 384] - - [392, 80.024] - - - [36864, 8192, 1, 384] - - [33, 87.044] - - - [41088, 2048, 1, 384] - - [62, 86.645] - - - [38016, 21761, 1, 384] - - [26, 89.316] - - - [41856, 1024, 1, 384] - - [37, 87.124] - - - [39552, 8192, 1, 384] - - [59, 90.847] - - - [37632, 4096, 1, 384] - - [45, 90.013] - - - [41856, 384, 1, 384] - - [346, 81.998] - - - [44160, 28289, 1, 384] - - [28, 89.054] - - - [43008, 26753, 1, 384] - - [55, 88.861] - - - [38400, 512, 1, 384] - - [37, 82.08] - - - [39168, 384, 1, 384] - - [397, 78.302] - - - [37632, 1024, 1, 384] - - [45, 85.403] - - - [44544, 4096, 1, 384] - - [75, 88.698] - - - [42240, 512, 1, 384] - - [37, 83.781] - - - [43008, 2048, 1, 384] - - [75, 86.187] - - - [36480, 20609, 1, 384] - - [54, 89.24] - - - [36864, 512, 1, 384] - - [54, 80.085] - - - [43008, 384, 1, 384] - - [350, 78.278] - - - [43392, 4096, 1, 384] - - [37, 90.249] - - - [38400, 22145, 1, 384] - - [38, 89.347] - - - [39936, 23681, 1, 384] - - [26, 89.105] - - - [36096, 19841, 1, 384] - - [55, 87.621] - - - [44544, 512, 1, 384] - - [28, 82.624] - - - [38400, 2048, 1, 384] - - [37, 88.327] - - - [41856, 25985, 1, 384] - - [28, 89.011] - - - [42624, 2048, 1, 384] - - [38, 83.262] - - - [38400, 1024, 1, 384] - - [54, 86.389] - - - [36480, 512, 1, 384] - - [28, 79.706] - - - [42624, 26753, 1, 384] - - [52, 83.126] - - - [43776, 27905, 1, 384] - - [33, 88.04] - - - [37248, 2048, 1, 384] - - [59, 87.981] - - - [35712, 19841, 1, 384] - - [30, 89.178] - - - [43392, 27521, 1, 384] - - [26, 89.004] - - - [43008, 1024, 1, 384] - - [37, 86.036] - - - [42624, 512, 1, 384] - - [28, 79.899] - - - [41472, 384, 1, 384] - - [399, 81.365] - - - [40704, 2048, 1, 384] - - [54, 88.243] - - - [36096, 2048, 1, 384] - - [63, 85.766] - - - [39936, 4096, 1, 384] - - [62, 87.649] - - - [40320, 2048, 1, 384] - - [59, 89.135] - - - [41088, 8192, 1, 384] - - [75, 89.419] - - - [35328, 8192, 1, 384] - - [30, 90.273] - - - [40320, 4096, 1, 384] - - [59, 90.213] - - - [41856, 512, 1, 384] - - [59, 83.581] - - - [39552, 4096, 1, 384] - - [59, 90.197] - - - [35712, 2048, 1, 384] - - [37, 88.452] - - - [39936, 24065, 1, 384] - - [38, 88.775] - - - [36480, 20225, 1, 384] - - [30, 89.099] - - - [38016, 1024, 1, 384] - - [37, 85.94] - - - [43008, 512, 1, 384] - - [26, 80.274] - - - [40704, 24833, 1, 384] - - [38, 88.984] - - - [37248, 4096, 1, 384] - - [54, 89.751] - - - [41856, 4096, 1, 384] - - [37, 89.878] - - - [41472, 512, 1, 384] - - [54, 82.889] - - - [39552, 2048, 1, 384] - - [59, 89.065] - - - [41088, 384, 1, 384] - - [368, 80.517] - - - [36480, 8192, 1, 384] - - [59, 90.778] - - - [37632, 2048, 1, 384] - - [54, 88.738] - - - [40704, 8192, 1, 384] - - [54, 90.642] - - - [36864, 20993, 1, 384] - - [38, 87.812] - - - [35328, 512, 1, 384] - - [45, 77.624] - - - [40320, 384, 1, 384] - - [354, 80.003] - - - [36096, 1024, 1, 384] - - [94, 83.824] - - - [42624, 8192, 1, 384] - - [26, 89.493] - - - [38784, 22529, 1, 384] - - [33, 87.014] - - - [44160, 4096, 1, 384] - - [37, 90.077] - - - [41472, 4096, 1, 384] - - [94, 88.798] - - - [36480, 1024, 1, 384] - - [37, 85.97] - - - [38784, 2048, 1, 384] - - [54, 87.715] - - - [44544, 1024, 1, 384] - - [37, 85.806] - - - [41088, 24833, 1, 384] - - [55, 88.119] - - - [36864, 384, 1, 384] - - [392, 80.317] - - - [43392, 512, 1, 384] - - [28, 81.03] - - - [39168, 8192, 1, 384] - - [28, 90.446] - - - [42624, 4096, 1, 384] - - [26, 87.761] - - - [40320, 24065, 1, 384] - - [30, 89.226] - - - [44160, 512, 1, 384] - - [59, 82.101] - - - [38016, 384, 1, 384] - - [392, 76.367] - - - [38016, 512, 1, 384] - - [26, 82.165] - - - [37248, 512, 1, 384] - - [26, 80.906] - - - [43776, 2048, 1, 384] - - [62, 86.412] - - - [35712, 8192, 1, 384] - - [26, 90.561] - - - [38400, 384, 1, 384] - - [350, 77.045] - - - [42240, 1024, 1, 384] - - [54, 87.472] - - - [35712, 19457, 1, 384] - - [33, 86.829] - - - [41856, 2048, 1, 384] - - [59, 88.873] - - - [41472, 1024, 1, 384] - - [45, 86.378] - - - [37632, 384, 1, 384] - - [392, 75.71] - - - [40704, 1024, 1, 384] - - [45, 84.888] - - - [43008, 27137, 1, 384] - - [33, 88.817] - - - [40704, 4096, 1, 384] - - [59, 90.082] - - - [36096, 20225, 1, 384] - - [33, 87.899] - - - [39936, 8192, 1, 384] - - [75, 88.219] - - - [38784, 384, 1, 384] - - [346, 77.511] - - - [38784, 8192, 1, 384] - - [37, 90.697] - - - [42624, 384, 1, 384] - - [346, 77.558] - - - [35712, 4096, 1, 384] - - [59, 89.523] - - - [37632, 8192, 1, 384] - - [54, 90.488] - - - [38784, 22913, 1, 384] - - [28, 89.099] - - - [36864, 1024, 1, 384] - - [45, 84.734] - - - [37248, 384, 1, 384] - - [395, 81.153] - - - [39168, 23297, 1, 384] - - [28, 89.091] - - - [40704, 24449, 1, 384] - - [28, 89.212] - - - [41472, 2048, 1, 384] - - [59, 88.036] - - - [44160, 27905, 1, 384] - - [26, 88.95] - - - [44160, 1024, 1, 384] - - [54, 85.41] - - - [36480, 384, 1, 384] - - [350, 79.768] - - - [42240, 384, 1, 384] - - [346, 82.58] - - - [44544, 28289, 1, 384] - - [33, 89.099] - - - [37248, 21377, 1, 384] - - [26, 88.605] - - - [36096, 4096, 1, 384] - - [75, 87.498] - - - [38784, 512, 1, 384] - - [26, 78.508] - - - [35712, 384, 1, 384] - - [350, 78.71] - - - [43776, 1024, 1, 384] - - [75, 84.107] - - - [41088, 25217, 1, 384] - - [33, 88.117] - - - [40320, 8192, 1, 384] - - [26, 90.614] - - - [39168, 22913, 1, 384] - - [30, 89.058] - - - [38400, 8192, 1, 384] - - [38, 90.376] - - - [41088, 512, 1, 384] - - [25, 79.633] - - - [42624, 1024, 1, 384] - - [26, 83.626] - - - [39168, 2048, 1, 384] - - [37, 88.606] - - - [43008, 4096, 1, 384] - - [62, 87.372] - - - [35712, 512, 1, 384] - - [30, 78.178] - - - [41856, 8192, 1, 384] - - [37, 90.813] - - - [43008, 8192, 1, 384] - - [55, 87.95] - - - [41472, 8192, 1, 384] - - [30, 90.544] - - - [41088, 1024, 1, 384] - - [40, 83.173] - - - [37248, 20993, 1, 384] - - [28, 88.42] - - - [44544, 384, 1, 384] - - [350, 80.065] - - - [36096, 8192, 1, 384] - - [52, 88.991] - - - [43776, 8192, 1, 384] - - [62, 89.499] - - - [41856, 25601, 1, 384] - - [33, 87.17] - - - [37632, 21377, 1, 384] - - [38, 89.33] - - - [40320, 24449, 1, 384] - - [26, 89.215] - - - [43776, 4096, 1, 384] - - [75, 88.614] - - - [35328, 4096, 1, 384] - - [57, 88.908] - - - [39552, 1024, 1, 384] - - [59, 85.957] - - - [38016, 8192, 1, 384] - - [45, 90.868] - - - [38400, 22529, 1, 384] - - [55, 87.23] - - - [39936, 2048, 1, 384] - - [75, 86.374] - - - [39168, 1024, 1, 384] - - [54, 85.143] - - - [37248, 8192, 1, 384] - - [30, 90.594] - - - [40320, 1024, 1, 384] - - [59, 87.081] - - - [26112, 1024, 1, 384] - - [59, 85.092] - - - [24192, 2048, 1, 384] - - [59, 86.935] - - - [13440, 5761, 1, 384] - - [26, 87.024] - - - [3456, 384, 1, 384] - - [335, 51.148] - - - [21888, 4096, 1, 384] - - [25, 84.551] - - - [384, 384, 1, 384] - - [331, 10.812] - - - [21120, 1024, 1, 384] - - [54, 83.912] - - - [30336, 4096, 1, 384] - - [59, 89.998] - - - [31488, 512, 1, 384] - - [37, 81.143] - - - [2304, 1793, 1, 384] - - [56, 55.398] - - - [16896, 9217, 1, 384] - - [54, 86.173] - - - [9216, 1024, 1, 384] - - [41, 70.126] - - - [29568, 1024, 1, 384] - - [57, 83.999] - - - [27264, 11393, 1, 384] - - [26, 88.608] - - - [33408, 17537, 1, 384] - - [26, 89.304] - - - [18816, 1024, 1, 384] - - [37, 81.935] - - - [5760, 1024, 1, 384] - - [54, 61.788] - - - [31104, 14849, 1, 384] - - [38, 89.596] - - - [18816, 4096, 1, 384] - - [45, 88.971] - - - [11136, 1024, 1, 384] - - [54, 73.552] - - - [17664, 9985, 1, 384] - - [37, 88.697] - - - [9216, 512, 1, 384] - - [36, 61.98] - - - [17664, 1024, 1, 384] - - [37, 78.042] - - - [17664, 512, 1, 384] - - [54, 67.758] - - - [31488, 384, 1, 384] - - [389, 78.603] - - - [15744, 8065, 1, 384] - - [38, 88.222] - - - [5760, 3841, 1, 384] - - [26, 80.878] - - - [24192, 1024, 1, 384] - - [45, 84.078] - - - [20352, 384, 1, 384] - - [370, 74.349] - - - [21888, 2048, 1, 384] - - [26, 80.794] - - - [7680, 2048, 1, 384] - - [41, 80.651] - - - [2688, 512, 1, 384] - - [184, 48.391] - - - [13056, 1024, 1, 384] - - [37, 76.455] - - - [22656, 14977, 1, 384] - - [54, 89.781] - - - [10752, 6785, 1, 384] - - [26, 86.162] - - - [6912, 2048, 1, 384] - - [45, 80.824] - - - [15360, 512, 1, 384] - - [35, 68.193] - - - [31104, 384, 1, 384] - - [371, 77.703] - - - [30720, 14465, 1, 384] - - [26, 89.054] - - - [17280, 2048, 1, 384] - - [59, 85.776] - - - [34176, 1024, 1, 384] - - [37, 84.779] - - - [16896, 2048, 1, 384] - - [54, 84.012] - - - [17664, 384, 1, 384] - - [365, 73.225] - - - [21504, 512, 1, 384] - - [54, 71.604] - - - [18048, 10369, 1, 384] - - [38, 88.443] - - - [15744, 1024, 1, 384] - - [54, 82.108] - - - [33408, 4096, 1, 384] - - [54, 89.946] - - - [11904, 4096, 1, 384] - - [37, 85.832] - - - [18816, 512, 1, 384] - - [24, 71.129] - - - [34944, 4096, 1, 384] - - [59, 89.626] - - - [13824, 2048, 1, 384] - - [45, 84.822] - - - [3840, 512, 1, 384] - - [54, 37.763] - - - [4992, 1024, 1, 384] - - [64, 67.094] - - - [11136, 7553, 1, 384] - - [28, 87.11] - - - [16512, 1024, 1, 384] - - [108, 72.635] - - - [17280, 9217, 1, 384] - - [94, 85.933] - - - [29184, 1024, 1, 384] - - [37, 84.597] - - - [18048, 512, 1, 384] - - [61, 68.959] - - - [6528, 384, 1, 384] - - [341, 61.098] - - - [28416, 1024, 1, 384] - - [54, 82.714] - - - [2688, 1153, 1, 384] - - [443, 60.739] - - - [34560, 18305, 1, 384] - - [38, 89.206] - - - [20736, 384, 1, 384] - - [369, 75.297] - - - [11520, 512, 1, 384] - - [35, 61.863] - - - [26112, 8192, 1, 384] - - [45, 90.445] - - - [31872, 384, 1, 384] - - [390, 78.998] - - - [24192, 512, 1, 384] - - [54, 78.806] - - - [19968, 2048, 1, 384] - - [37, 86.677] - - - [32256, 8192, 1, 384] - - [38, 90.325] - - - [11520, 384, 1, 384] - - [341, 70.231] - - - [1920, 1409, 1, 384] - - [36, 50.649] - - - [25728, 9857, 1, 384] - - [30, 88.779] - - - [9216, 5633, 1, 384] - - [38, 84.683] - - - [28032, 12161, 1, 384] - - [30, 89.361] - - - [28800, 8192, 1, 384] - - [59, 90.586] - - - [28416, 12161, 1, 384] - - [54, 89.491] - - - [23040, 15361, 1, 384] - - [55, 87.18] - - - [31488, 15617, 1, 384] - - [38, 89.332] - - - [22272, 14209, 1, 384] - - [59, 89.567] - - - [1536, 512, 1, 384] - - [110, 41.102] - - - [1152, 257, 1, 384] - - [172, 21.611] - - - [21120, 2048, 1, 384] - - [27, 86.878] - - - [32256, 16001, 1, 384] - - [28, 89.634] - - - [9600, 6017, 1, 384] - - [30, 86.235] - - - [32640, 384, 1, 384] - - [391, 80.302] - - - [34176, 512, 1, 384] - - [59, 80.817] - - - [10368, 512, 1, 384] - - [35, 68.48] - - - [21120, 384, 1, 384] - - [372, 76.109] - - - [29568, 4096, 1, 384] - - [27, 88.237] - - - [31872, 2048, 1, 384] - - [54, 87.371] - - - [8832, 384, 1, 384] - - [360, 66.769] - - - [4224, 384, 1, 384] - - [340, 55.582] - - - [33408, 8192, 1, 384] - - [28, 90.609] - - - [768, 257, 1, 384] - - [172, 14.592] - - - [10368, 6401, 1, 384] - - [28, 86.647] - - - [13824, 384, 1, 384] - - [361, 72.389] - - - [29568, 512, 1, 384] - - [45, 77.505] - - - [28032, 1024, 1, 384] - - [37, 85.796] - - - [19200, 384, 1, 384] - - [371, 71.198] - - - [23040, 2048, 1, 384] - - [37, 85.758] - - - [8448, 4481, 1, 384] - - [28, 83.925] - - - [22272, 14593, 1, 384] - - [59, 89.97] - - - [26496, 10241, 1, 384] - - [59, 85.965] - - - [19584, 384, 1, 384] - - [356, 72.622] - - - [4992, 3457, 1, 384] - - [28, 80.037] - - - [22656, 384, 1, 384] - - [379, 76.384] - - - [15360, 1024, 1, 384] - - [37, 80.098] - - - [7296, 2048, 1, 384] - - [26, 77.098] - - - [30720, 384, 1, 384] - - [388, 77.062] - - - [6144, 2177, 1, 384] - - [35, 75.794] - - - [30720, 14849, 1, 384] - - [38, 88.592] - - - [23424, 2048, 1, 384] - - [45, 85.898] - - - [5760, 384, 1, 384] - - [336, 55.312] - - - [6144, 2561, 1, 384] - - [30, 73.855] - - - [12672, 384, 1, 384] - - [339, 68.074] - - - [16128, 8065, 1, 384] - - [350, 89.585] - - - [10752, 7169, 1, 384] - - [94, 83.986] - - - [2304, 384, 1, 384] - - [343, 44.899] - - - [18816, 2048, 1, 384] - - [45, 85.312] - - - [22272, 4096, 1, 384] - - [59, 89.256] - - - [12672, 4993, 1, 384] - - [26, 86.778] - - - [12288, 512, 1, 384] - - [61, 65.525] - - - [13056, 4993, 1, 384] - - [38, 85.015] - - - [19584, 512, 1, 384] - - [54, 73.656] - - - [30336, 14465, 1, 384] - - [28, 89.258] - - - [5376, 3841, 1, 384] - - [28, 81.167] - - - [17664, 9601, 1, 384] - - [45, 88.671] - - - [29952, 2048, 1, 384] - - [59, 86.831] - - - [8832, 512, 1, 384] - - [35, 60.337] - - - [9984, 512, 1, 384] - - [35, 66.79] - - - [19200, 1024, 1, 384] - - [37, 82.967] - - - [24192, 8321, 1, 384] - - [59, 88.409] - - - [26112, 10241, 1, 384] - - [33, 86.108] - - - [17280, 9601, 1, 384] - - [59, 88.482] - - - [7296, 384, 1, 384] - - [358, 56.905] - - - [16512, 8449, 1, 384] - - [33, 85.255] - - - [11904, 4225, 1, 384] - - [54, 85.165] - - - [24576, 4096, 1, 384] - - [26, 82.017] - - - [6912, 2945, 1, 384] - - [28, 79.917] - - - [33024, 16769, 1, 384] - - [33, 88.675] - - - [24576, 8705, 1, 384] - - [30, 83.163] - - - [16128, 2048, 1, 384] - - [54, 83.245] - - - [13824, 6145, 1, 384] - - [30, 84.121] - - - [28800, 512, 1, 384] - - [28, 75.38] - - - [33792, 8192, 1, 384] - - [62, 88.394] - - - [27648, 11393, 1, 384] - - [28, 88.687] - - - [21888, 384, 1, 384] - - [378, 73.944] - - - [12672, 4096, 1, 384] - - [37, 87.969] - - - [23040, 14977, 1, 384] - - [26, 89.321] - - - [11904, 384, 1, 384] - - [365, 64.469] - - - [7680, 3713, 1, 384] - - [30, 80.768] - - - [24576, 8192, 1, 384] - - [33, 84.517] - - - [34176, 384, 1, 384] - - [346, 76.209] - - - [17664, 2048, 1, 384] - - [45, 84.228] - - - [29952, 4096, 1, 384] - - [59, 89.17] - - - [9984, 6017, 1, 384] - - [38, 84.842] - - - [33408, 2048, 1, 384] - - [59, 88.753] - - - [21120, 4096, 1, 384] - - [57, 88.935] - - - [34560, 4096, 1, 384] - - [37, 89.877] - - - [19200, 11521, 1, 384] - - [38, 88.647] - - - [21120, 13057, 1, 384] - - [28, 89.01] - - - [25728, 384, 1, 384] - - [370, 77.116] - - - [28800, 12929, 1, 384] - - [30, 88.677] - - - [20736, 1024, 1, 384] - - [37, 83.369] - - - [18816, 10753, 1, 384] - - [26, 89.03] - - - [34560, 8192, 1, 384] - - [26, 90.331] - - - [23040, 512, 1, 384] - - [54, 75.666] - - - [30336, 2048, 1, 384] - - [59, 87.643] - - - [17280, 512, 1, 384] - - [45, 75.461] - - - [19200, 2048, 1, 384] - - [54, 86.532] - - - [12288, 4225, 1, 384] - - [28, 84.633] - - - [15744, 7681, 1, 384] - - [350, 88.512] - - - [30720, 4096, 1, 384] - - [75, 86.995] - - - [10752, 384, 1, 384] - - [341, 67.404] - - - [15744, 512, 1, 384] - - [28, 69.785] - - - [24960, 384, 1, 384] - - [370, 75.875] - - - [768, 384, 1, 384] - - [332, 21.625] - - - [6912, 3329, 1, 384] - - [26, 78.713] - - - [8064, 512, 1, 384] - - [35, 55.111] - - - [26496, 384, 1, 384] - - [385, 76.353] - - - [24960, 4096, 1, 384] - - [54, 89.572] - - - [19584, 11905, 1, 384] - - [59, 89.153] - - - [16512, 8833, 1, 384] - - [55, 85.805] - - - [18816, 384, 1, 384] - - [355, 70.284] - - - [23808, 1024, 1, 384] - - [45, 83.004] - - - [16512, 384, 1, 384] - - [371, 72.045] - - - [8448, 4865, 1, 384] - - [38, 83.744] - - - [34944, 1024, 1, 384] - - [59, 86.21] - - - [29184, 4096, 1, 384] - - [45, 88.774] - - - [8832, 2048, 1, 384] - - [59, 77.616] - - - [9984, 1024, 1, 384] - - [64, 75.292] - - - [22272, 1024, 1, 384] - - [59, 83.249] - - - [14592, 6913, 1, 384] - - [350, 88.818] - - - [9216, 2048, 1, 384] - - [24, 78.405] - - - [7296, 1024, 1, 384] - - [41, 65.639] - - - [26880, 8192, 1, 384] - - [54, 90.358] - - - [26880, 10625, 1, 384] - - [45, 89.218] - - - [28800, 12545, 1, 384] - - [30, 88.809] - - - [18048, 1024, 1, 384] - - [37, 79.175] - - - [27264, 11009, 1, 384] - - [28, 88.429] - - - [12288, 2048, 1, 384] - - [28, 79.298] - - - [19200, 4096, 1, 384] - - [54, 88.577] - - - [32256, 384, 1, 384] - - [377, 79.86] - - - [9216, 5249, 1, 384] - - [38, 84.69] - - - [29952, 14081, 1, 384] - - [31, 89.172] - - - [7680, 384, 1, 384] - - [358, 59.829] - - - [19200, 11137, 1, 384] - - [48, 88.991] - - - [14976, 1024, 1, 384] - - [45, 78.752] - - - [25728, 1024, 1, 384] - - [59, 83.763] - - - [3456, 1921, 1, 384] - - [36, 68.429] - - - [21120, 13441, 1, 384] - - [48, 89.157] - - - [15360, 2048, 1, 384] - - [62, 82.111] - - - [34560, 512, 1, 384] - - [28, 81.541] - - - [31872, 8192, 1, 384] - - [30, 90.406] - - - [32640, 16769, 1, 384] - - [33, 87.83] - - - [26496, 1024, 1, 384] - - [54, 81.816] - - - [12672, 1024, 1, 384] - - [54, 74.602] - - - [3072, 384, 1, 384] - - [349, 46.607] - - - [31104, 4096, 1, 384] - - [54, 89.842] - - - [25344, 4096, 1, 384] - - [37, 89.495] - - - [4224, 2689, 1, 384] - - [36, 73.263] - - - [24576, 1024, 1, 384] - - [33, 78.371] - - - [8448, 512, 1, 384] - - [35, 58.027] - - - [1536, 1025, 1, 384] - - [205, 50.72] - - - [14208, 6145, 1, 384] - - [57, 85.307] - - - [27264, 384, 1, 384] - - [373, 77.623] - - - [34560, 1024, 1, 384] - - [54, 85.683] - - - [14976, 6913, 1, 384] - - [28, 87.539] - - - [21504, 2048, 1, 384] - - [62, 84.642] - - - [14208, 4096, 1, 384] - - [54, 87.288] - - - [14592, 4096, 1, 384] - - [59, 86.938] - - - [6528, 2561, 1, 384] - - [28, 77.922] - - - [34176, 18305, 1, 384] - - [52, 89.147] - - - [19968, 384, 1, 384] - - [375, 73.39] - - - [30720, 8192, 1, 384] - - [75, 87.502] - - - [14592, 512, 1, 384] - - [41, 65.731] - - - [25728, 2048, 1, 384] - - [54, 86.383] - - - [23424, 4096, 1, 384] - - [28, 87.192] - - - [27264, 2048, 1, 384] - - [54, 84.397] - - - [21504, 1024, 1, 384] - - [54, 80.566] - - - [30336, 384, 1, 384] - - [370, 78.013] - - - [2688, 1024, 1, 384] - - [60, 51.497] - - - [22656, 4096, 1, 384] - - [37, 89.102] - - - [20352, 2048, 1, 384] - - [54, 84.991] - - - [33408, 384, 1, 384] - - [350, 74.672] - - - [15360, 4096, 1, 384] - - [38, 86.111] - - - [22272, 512, 1, 384] - - [38, 73.8] - - - [14208, 384, 1, 384] - - [360, 66.349] - - - [32640, 512, 1, 384] - - [77, 72.894] - - - [23808, 512, 1, 384] - - [26, 78.073] - - - [24960, 1024, 1, 384] - - [59, 81.728] - - - [4608, 512, 1, 384] - - [41, 45.066] - - - [25344, 2048, 1, 384] - - [37, 87.652] - - - [11904, 1024, 1, 384] - - [54, 78.276] - - - [28416, 12545, 1, 384] - - [30, 89.264] - - - [14208, 6529, 1, 384] - - [26, 87.106] - - - [13824, 5761, 1, 384] - - [30, 85.86] - - - [26112, 9857, 1, 384] - - [30, 89.098] - - - [9600, 2048, 1, 384] - - [59, 83.301] - - - [33024, 1024, 1, 384] - - [62, 83.484] - - - [34944, 18689, 1, 384] - - [26, 89.205] - - - [13824, 512, 1, 384] - - [64, 73.084] - - - [26880, 384, 1, 384] - - [386, 77.191] - - - [15744, 384, 1, 384] - - [360, 72.053] - - - [29568, 8192, 1, 384] - - [54, 89.975] - - - [24960, 9089, 1, 384] - - [59, 89.011] - - - [28032, 2048, 1, 384] - - [54, 88.256] - - - [19968, 11905, 1, 384] - - [38, 89.202] - - - [6528, 2945, 1, 384] - - [28, 75.792] - - - [20352, 12289, 1, 384] - - [94, 87.032] - - - [5376, 512, 1, 384] - - [60, 51.87] - - - [5376, 3457, 1, 384] - - [36, 79.024] - - - [21504, 384, 1, 384] - - [377, 72.944] - - - [11520, 1024, 1, 384] - - [59, 76.232] - - - [3840, 1921, 1, 384] - - [355, 70.526] - - - [18432, 4096, 1, 384] - - [62, 85.855] - - - [28416, 2048, 1, 384] - - [59, 87.239] - - - [3456, 512, 1, 384] - - [184, 59.209] - - - [2688, 384, 1, 384] - - [345, 42.63] - - - [28032, 4096, 1, 384] - - [59, 89.737] - - - [16128, 384, 1, 384] - - [370, 73.105] - - - [33792, 17537, 1, 384] - - [26, 89.18] - - - [2688, 1793, 1, 384] - - [50, 63.259] - - - [27648, 1024, 1, 384] - - [54, 84.344] - - - [13440, 1024, 1, 384] - - [37, 78.549] - - - [28032, 8192, 1, 384] - - [59, 90.546] - - - [34560, 18689, 1, 384] - - [38, 89.298] - - - [16896, 512, 1, 384] - - [56, 74.187] - - - [13056, 2048, 1, 384] - - [54, 84.795] - - - [3072, 1537, 1, 384] - - [365, 65.055] - - - [3072, 512, 1, 384] - - [182, 53.669] - - - [25344, 9089, 1, 384] - - [52, 88.616] - - - [9600, 384, 1, 384] - - [356, 65.429] - - - [26880, 512, 1, 384] - - [59, 78.239] - - - [33024, 512, 1, 384] - - [56, 76.942] - - - [21888, 1024, 1, 384] - - [26, 78.92] - - - [18048, 384, 1, 384] - - [373, 74.077] - - - [16896, 4096, 1, 384] - - [37, 87.888] - - - [23808, 384, 1, 384] - - [382, 75.445] - - - [26496, 4096, 1, 384] - - [54, 89.126] - - - [20736, 13057, 1, 384] - - [38, 89.631] - - - [24576, 512, 1, 384] - - [38, 78.16] - - - [14592, 6529, 1, 384] - - [38, 87.597] - - - [6528, 512, 1, 384] - - [54, 61.445] - - - [22656, 14593, 1, 384] - - [59, 89.381] - - - [26112, 2048, 1, 384] - - [54, 87.556] - - - [25728, 9473, 1, 384] - - [26, 88.602] - - - [15744, 2048, 1, 384] - - [59, 86.199] - - - [31488, 1024, 1, 384] - - [37, 86.085] - - - [11136, 2048, 1, 384] - - [59, 83.043] - - - [4608, 2689, 1, 384] - - [56, 71.478] - - - [30720, 1024, 1, 384] - - [54, 84.187] - - - [1920, 512, 1, 384] - - [124, 40.881] - - - [25728, 8192, 1, 384] - - [59, 90.683] - - - [31104, 2048, 1, 384] - - [37, 87.584] - - - [3456, 1024, 1, 384] - - [81, 64.265] - - - [25344, 384, 1, 384] - - [341, 76.552] - - - [27264, 8192, 1, 384] - - [26, 90.393] - - - [16128, 4096, 1, 384] - - [59, 87.503] - - - [20736, 12673, 1, 384] - - [38, 89.313] - - - [4224, 2305, 1, 384] - - [35, 71.826] - - - [27648, 11777, 1, 384] - - [28, 88.558] - - - [6144, 512, 1, 384] - - [36, 58.515] - - - [24576, 2048, 1, 384] - - [30, 79.488] - - - [15360, 384, 1, 384] - - [360, 70.199] - - - [34944, 19073, 1, 384] - - [26, 89.106] - - - [33792, 384, 1, 384] - - [350, 75.176] - - - [15360, 7681, 1, 384] - - [350, 88.864] - - - [34176, 17921, 1, 384] - - [25, 88.898] - - - [10368, 1024, 1, 384] - - [54, 77.886] - - - [34176, 8192, 1, 384] - - [28, 90.199] - - - [34176, 2048, 1, 384] - - [94, 88.035] - - - [7680, 4097, 1, 384] - - [45, 82.31] - - - [10752, 1024, 1, 384] - - [37, 71.496] - - - [9984, 2048, 1, 384] - - [45, 80.888] - - - [5760, 2048, 1, 384] - - [24, 75.92] - - - [30336, 1024, 1, 384] - - [54, 83.496] - - - [23424, 384, 1, 384] - - [381, 77.984] - - - [13440, 5377, 1, 384] - - [30, 85.376] - - - [14592, 2048, 1, 384] - - [45, 84.533] - - - [31872, 4096, 1, 384] - - [54, 89.55] - - - [6528, 2048, 1, 384] - - [38, 76.208] - - - [8064, 384, 1, 384] - - [358, 62.165] - - - [31872, 16001, 1, 384] - - [30, 89.19] - - - [16896, 1024, 1, 384] - - [45, 80.371] - - - [15360, 7297, 1, 384] - - [350, 88.861] - - - [33792, 4096, 1, 384] - - [75, 87.283] - - - [16896, 384, 1, 384] - - [372, 73.287] - - - [29952, 1024, 1, 384] - - [37, 82.722] - - - [768, 512, 1, 384] - - [110, 28.582] - - - [24576, 384, 1, 384] - - [382, 77.069] - - - [9984, 384, 1, 384] - - [356, 67.287] - - - [28416, 4096, 1, 384] - - [54, 89.71] - - - [11904, 7937, 1, 384] - - [28, 87.153] - - - [22656, 512, 1, 384] - - [26, 74.849] - - - [32640, 16385, 1, 384] - - [33, 85.468] - - - [14592, 1024, 1, 384] - - [37, 77.259] - - - [29952, 13697, 1, 384] - - [38, 89.409] - - - [32640, 1024, 1, 384] - - [75, 79.399] - - - [24960, 512, 1, 384] - - [54, 73.643] - - - [24192, 384, 1, 384] - - [383, 76.298] - - - [10752, 512, 1, 384] - - [35, 58.246] - - - [25344, 8192, 1, 384] - - [37, 90.336] - - - [32256, 16385, 1, 384] - - [55, 85.964] - - - [18432, 10753, 1, 384] - - [38, 88.005] - - - [27648, 512, 1, 384] - - [30, 79.871] - - - [28800, 4096, 1, 384] - - [45, 89.288] - - - [13440, 512, 1, 384] - - [54, 71.196] - - - [22272, 2048, 1, 384] - - [54, 86.157] - - - [29184, 2048, 1, 384] - - [54, 86.734] - - - [29952, 8192, 1, 384] - - [59, 90.427] - - - [384, 385, 1, 384] - - [331, 10.442] - - - [33408, 17153, 1, 384] - - [28, 89.285] - - - [27264, 512, 1, 384] - - [54, 78.612] - - - [33792, 1024, 1, 384] - - [45, 83.815] - - - [12288, 384, 1, 384] - - [367, 66.011] - - - [4224, 1024, 1, 384] - - [35, 58.094] - - - [13056, 5377, 1, 384] - - [30, 86.897] - - - [9600, 5633, 1, 384] - - [59, 85.785] - - - [30336, 512, 1, 384] - - [28, 79.244] - - - [7680, 1024, 1, 384] - - [41, 68.745] - - - [14976, 384, 1, 384] - - [341, 69.447] - - - [11904, 512, 1, 384] - - [28, 63.746] - - - [16128, 512, 1, 384] - - [30, 70.223] - - - [16128, 8449, 1, 384] - - [30, 88.103] - - - [18432, 2048, 1, 384] - - [62, 83.507] - - - [32256, 1024, 1, 384] - - [54, 84.016] - - - [16896, 8833, 1, 384] - - [38, 89.018] - - - [11136, 7169, 1, 384] - - [63, 84.227] - - - [8832, 4865, 1, 384] - - [28, 84.203] - - - [13440, 4096, 1, 384] - - [54, 87.546] - - - [10752, 2048, 1, 384] - - [59, 80.532] - - - [27264, 1024, 1, 384] - - [28, 82.207] - - - [1536, 384, 1, 384] - - [333, 34.137] - - - [20352, 1024, 1, 384] - - [37, 81.987] - - - [30720, 512, 1, 384] - - [45, 80.004] - - - [16512, 512, 1, 384] - - [34, 67.144] - - - [20736, 4096, 1, 384] - - [54, 88.305] - - - [23424, 15745, 1, 384] - - [38, 89.763] - - - [24960, 2048, 1, 384] - - [45, 86.756] - - - [32256, 2048, 1, 384] - - [37, 87.968] - - - [10368, 384, 1, 384] - - [355, 68.896] - - - [14976, 7297, 1, 384] - - [350, 88.83] - - - [23040, 4096, 1, 384] - - [59, 88.538] - - - [16512, 4096, 1, 384] - - [77, 83.999] - - - [20736, 512, 1, 384] - - [38, 77.072] - - - [34560, 384, 1, 384] - - [346, 76.833] - - - [23040, 1024, 1, 384] - - [37, 80.485] - - - [5376, 384, 1, 384] - - [355, 53.082] - - - [11136, 512, 1, 384] - - [35, 60.033] - - - [19200, 512, 1, 384] - - [54, 72.362] - - - [19584, 11521, 1, 384] - - [28, 89.051] - - - [21504, 4096, 1, 384] - - [75, 86.954] - - - [25728, 4096, 1, 384] - - [54, 89.093] - - - [4992, 512, 1, 384] - - [60, 48.394] - - - [26880, 4096, 1, 384] - - [54, 88.944] - - - [31488, 15233, 1, 384] - - [25, 89.348] - - - [2304, 1409, 1, 384] - - [102, 59.428] - - - [28800, 1024, 1, 384] - - [37, 83.054] - - - [25344, 9473, 1, 384] - - [26, 89.006] - - - [13824, 4096, 1, 384] - - [59, 87.375] - - - [18048, 2048, 1, 384] - - [54, 85.495] - - - [13056, 512, 1, 384] - - [26, 69.355] - - - [31104, 8192, 1, 384] - - [28, 90.454] - - - [1152, 641, 1, 384] - - [115, 39.307] - - - [8064, 1024, 1, 384] - - [40, 70.051] - - - [7296, 512, 1, 384] - - [85, 50.843] - - - [12672, 4609, 1, 384] - - [26, 84.709] - - - [27264, 4096, 1, 384] - - [28, 87.94] - - - [11520, 2048, 1, 384] - - [54, 80.803] - - - [15744, 4096, 1, 384] - - [45, 88.574] - - - [19968, 512, 1, 384] - - [54, 75.065] - - - [5760, 2177, 1, 384] - - [36, 71.734] - - - [3840, 384, 1, 384] - - [352, 51.55] - - - [30336, 8192, 1, 384] - - [59, 90.763] - - - [28416, 8192, 1, 384] - - [54, 90.413] - - - [25344, 512, 1, 384] - - [45, 74.705] - - - [7296, 3713, 1, 384] - - [28, 80.855] - - - [28416, 384, 1, 384] - - [370, 74.813] - - - [19584, 2048, 1, 384] - - [54, 85.165] - - - [10368, 2048, 1, 384] - - [45, 83.477] - - - [33024, 4096, 1, 384] - - [54, 88.724] - - - [4224, 512, 1, 384] - - [35, 41.356] - - - [26496, 8192, 1, 384] - - [37, 90.525] - - - [768, 385, 1, 384] - - [332, 21.231] - - - [23040, 384, 1, 384] - - [344, 77.403] - - - [11520, 7937, 1, 384] - - [26, 87.554] - - - [28800, 384, 1, 384] - - [341, 75.228] - - - [8064, 4481, 1, 384] - - [28, 82.664] - - - [28032, 384, 1, 384] - - [387, 79.031] - - - [31104, 512, 1, 384] - - [28, 80.719] - - - [23808, 16129, 1, 384] - - [37, 90.108] - - - [29184, 384, 1, 384] - - [370, 76.081] - - - [9600, 512, 1, 384] - - [35, 64.587] - - - [26112, 512, 1, 384] - - [26, 76.532] - - - [31488, 8192, 1, 384] - - [30, 90.292] - - - [8448, 384, 1, 384] - - [341, 64.378] - - - [34944, 8192, 1, 384] - - [38, 90.6] - - - [4608, 3073, 1, 384] - - [28, 72.603] - - - [30720, 2048, 1, 384] - - [75, 85.262] - - - [34944, 512, 1, 384] - - [30, 82.01] - - - [27648, 8192, 1, 384] - - [75, 88.499] - - - [33024, 2048, 1, 384] - - [54, 87.021] - - - [26112, 4096, 1, 384] - - [37, 88.996] - - - [17280, 384, 1, 384] - - [355, 74.66] - - - [33024, 17153, 1, 384] - - [33, 88.628] - - - [14208, 2048, 1, 384] - - [54, 82.778] - - - [13440, 2048, 1, 384] - - [54, 82.8] - - - [1536, 641, 1, 384] - - [115, 39.611] - - - [8064, 4097, 1, 384] - - [57, 81.905] - - - [26496, 10625, 1, 384] - - [26, 88.98] - - - [33024, 384, 1, 384] - - [354, 73.665] - - - [26112, 384, 1, 384] - - [385, 75.442] - - - [23424, 15361, 1, 384] - - [57, 85.936] - - - [34944, 2048, 1, 384] - - [59, 88.407] - - - [32256, 512, 1, 384] - - [28, 77.446] - - - [23808, 15745, 1, 384] - - [54, 90.117] - - - [5760, 512, 1, 384] - - [28, 54.858] - - - [16128, 1024, 1, 384] - - [26, 76.093] - - - [31488, 4096, 1, 384] - - [59, 89.699] - - - [29568, 13313, 1, 384] - - [55, 86.436] - - - [18816, 11137, 1, 384] - - [26, 89.01] - - - [26496, 2048, 1, 384] - - [54, 86.456] - - - [1920, 384, 1, 384] - - [340, 38.663] - - - [31872, 1024, 1, 384] - - [37, 83.409] - - - [12672, 512, 1, 384] - - [54, 67.275] - - - [13056, 4096, 1, 384] - - [45, 87.556] - - - [17280, 1024, 1, 384] - - [45, 82.053] - - - [12288, 1024, 1, 384] - - [59, 78.16] - - - [1152, 512, 1, 384] - - [109, 32.6] - - - [31104, 15233, 1, 384] - - [26, 89.393] - - - [4608, 384, 1, 384] - - [352, 59.458] - - - [21888, 512, 1, 384] - - [30, 71.776] - - - [33408, 1024, 1, 384] - - [37, 86.352] - - - [8448, 2048, 1, 384] - - [59, 80.242] - - - [7296, 3329, 1, 384] - - [30, 82.434] - - - [10368, 6785, 1, 384] - - [26, 86.949] - - - [8832, 1024, 1, 384] - - [28, 67.434] - - - [31104, 1024, 1, 384] - - [54, 85.253] - - - [11520, 7553, 1, 384] - - [30, 86.704] - - - [34176, 4096, 1, 384] - - [59, 89.32] - - - [20352, 512, 1, 384] - - [45, 76.379] - - - [18432, 512, 1, 384] - - [24, 69.917] - - - [31488, 2048, 1, 384] - - [59, 88.404] - - - [9984, 6401, 1, 384] - - [28, 85.589] - - - [6144, 2048, 1, 384] - - [45, 78.784] - - - [22656, 2048, 1, 384] - - [54, 87.419] - - - [2304, 512, 1, 384] - - [109, 47.349] - - - [21504, 13441, 1, 384] - - [26, 88.86] - - - [1920, 1025, 1, 384] - - [36, 37.494] - - - [24960, 8705, 1, 384] - - [54, 88.641] - - - [16512, 2048, 1, 384] - - [62, 80.148] - - - [26880, 11009, 1, 384] - - [26, 89.218] - - - [32256, 4096, 1, 384] - - [27, 88.284] - - - [14976, 2048, 1, 384] - - [54, 82.793] - - - [21120, 512, 1, 384] - - [26, 78.172] - - - [31872, 512, 1, 384] - - [45, 76.429] - - - [8064, 2048, 1, 384] - - [30, 76.349] - - - [3072, 1024, 1, 384] - - [28, 58.286] - - - [23808, 2048, 1, 384] - - [45, 85.843] - - - [12672, 2048, 1, 384] - - [59, 83.079] - - - [19968, 4096, 1, 384] - - [59, 88.078] - - - [14976, 512, 1, 384] - - [26, 66.951] - - - [25344, 1024, 1, 384] - - [37, 82.986] - - - [31872, 15617, 1, 384] - - [38, 89.141] - - - [20352, 12673, 1, 384] - - [26, 89.358] - - - [11136, 384, 1, 384] - - [341, 68.837] - - - [32640, 8192, 1, 384] - - [33, 88.408] - - - [28800, 2048, 1, 384] - - [45, 87.385] - - - [22656, 1024, 1, 384] - - [59, 84.226] - - - [17280, 4096, 1, 384] - - [54, 87.951] - - - [17664, 4096, 1, 384] - - [59, 87.84] - - - [32640, 2048, 1, 384] - - [29, 83.394] - - - [28032, 11777, 1, 384] - - [59, 89.285] - - - [20352, 4096, 1, 384] - - [59, 88.403] - - - [33792, 512, 1, 384] - - [54, 80.242] - - - [24192, 4096, 1, 384] - - [45, 89.825] - - - [9216, 384, 1, 384] - - [361, 67.895] - - - [6912, 512, 1, 384] - - [79, 64.604] - - - [14208, 1024, 1, 384] - - [37, 75.175] - - - [26496, 512, 1, 384] - - [54, 77.278] - - - [4992, 384, 1, 384] - - [345, 56.997] - - - [33408, 512, 1, 384] - - [38, 79.233] - - - [3456, 1537, 1, 384] - - [338, 61.569] - - - [21888, 14209, 1, 384] - - [42, 84.605] - - - [24576, 8321, 1, 384] - - [26, 83.293] - - - [33792, 17921, 1, 384] - - [26, 88.783] - - - [13440, 384, 1, 384] - - [369, 71.107] - - - [18432, 384, 1, 384] - - [374, 74.867] - - - [6912, 1024, 1, 384] - - [41, 73.539] - - - [22272, 384, 1, 384] - - [373, 75.395] - - - [3840, 2305, 1, 384] - - [35, 66.085] - - - [6144, 1024, 1, 384] - - [54, 65.55] - - - [7680, 512, 1, 384] - - [56, 53.268] - - - [19584, 4096, 1, 384] - - [45, 88.664] - - - [23808, 4096, 1, 384] - - [45, 88.714] - - - [29568, 384, 1, 384] - - [370, 76.901] - - - [29184, 512, 1, 384] - - [28, 76.749] - - - [13056, 384, 1, 384] - - [355, 69.63] - - - [28032, 512, 1, 384] - - [37, 80.796] - - - [26880, 2048, 1, 384] - - [37, 87.507] - - - [18048, 9985, 1, 384] - - [28, 88.583] - - - [29952, 512, 1, 384] - - [38, 78.472] - - - [27648, 2048, 1, 384] - - [75, 85.281] - - - [29568, 13697, 1, 384] - - [31, 88.429] - - - [19584, 1024, 1, 384] - - [54, 79.278] - - - [27648, 384, 1, 384] - - [381, 78.548] - - - [6912, 384, 1, 384] - - [352, 63.882] - - - [26880, 1024, 1, 384] - - [54, 82.929] - - - [24960, 8192, 1, 384] - - [54, 90.373] - - - [13824, 1024, 1, 384] - - [54, 80.611] - - - [11904, 2048, 1, 384] - - [59, 83.021] - - - [34560, 2048, 1, 384] - - [54, 87.843] - - - [12288, 4609, 1, 384] - - [28, 83.753] - - - [21504, 13825, 1, 384] - - [30, 88.505] - - - [29184, 8192, 1, 384] - - [30, 90.304] - - - [12288, 4096, 1, 384] - - [26, 85.445] - - - [23424, 1024, 1, 384] - - [54, 81.746] - - - [14208, 512, 1, 384] - - [77, 63.839] - - - [25728, 512, 1, 384] - - [26, 75.456] - - - [29568, 2048, 1, 384] - - [94, 86.694] - - - [9600, 1024, 1, 384] - - [54, 72.644] - - - [29952, 384, 1, 384] - - [370, 77.355] - - - [18048, 4096, 1, 384] - - [54, 87.568] - - - [30336, 14081, 1, 384] - - [30, 89.07] - - - [24192, 8192, 1, 384] - - [54, 90.674] - - - [33792, 2048, 1, 384] - - [62, 85.928] - - - [6144, 384, 1, 384] - - [356, 58.242] - - - [8448, 1024, 1, 384] - - [41, 74.445] - - - [6528, 1024, 1, 384] - - [36, 69.453] - - - [18432, 10369, 1, 384] - - [30, 88.481] - - - [19968, 1024, 1, 384] - - [54, 80.777] - - - [23424, 512, 1, 384] - - [54, 76.757] - - - [20736, 2048, 1, 384] - - [59, 86.572] - - - [29184, 12929, 1, 384] - - [52, 89.169] - - - [3072, 1153, 1, 384] - - [444, 62.354] - - - [28416, 512, 1, 384] - - [59, 75.226] - - - [14592, 384, 1, 384] - - [341, 67.808] - - - [18432, 1024, 1, 384] - - [37, 80.413] - - - [29184, 13313, 1, 384] - - [55, 86.922] - - - [32640, 4096, 1, 384] - - [92, 86.09] - - - [21888, 13825, 1, 384] - - [60, 85.031] - - - [5376, 1024, 1, 384] - - [36, 57.879] - - - [4608, 1024, 1, 384] - - [36, 63.474] - - - [8832, 5249, 1, 384] - - [28, 84.238] - - - [14976, 4096, 1, 384] - - [59, 86.86] - - - [3840, 1024, 1, 384] - - [64, 53.498] - - - [24192, 16129, 1, 384] - - [59, 89.733] - - - [19968, 12289, 1, 384] - - [45, 86.208] - - - [1152, 384, 1, 384] - - [332, 30.378] - - - [27648, 4096, 1, 384] - - [62, 87.529] - - - [4992, 3073, 1, 384] - - [38, 77.085] - - - [33024, 8192, 1, 384] - - [63, 89.435] - - - [34944, 384, 1, 384] - - [392, 77.299] - - - [32, 28672, 1, 32] - - [7, 23.634] - - - [32, 24576, 1, 32] - - [6, 23.263] - - - [32, 16384, 1, 32] - - [3, 22.424] - - - [32, 20480, 1, 32] - - [5, 22.915] - - - [32, 12288, 1, 32] - - [3, 20.484] - - - [32, 8192, 1, 32] - - [4, 16.315] - - - [32, 4096, 1, 32] - - [2, 8.373] - - - [32, 32768, 1, 32] - - [1, 23.586] - - - [4224, 3840, 1, 4096] - - [16, 98.172] - - - [5376, 4096, 1, 4096] - - [17, 94.215] - - - [7040, 4096, 1, 384] - - [18, 86.756] - - - [7040, 4096, 1, 768] - - [16, 93.117] - - - [7040, 4096, 1, 1536] - - [19, 96.418] - - - [3840, 4224, 1, 4096] - - [8, 79.968] - - - [3840, 4224, 1, 4224] - - [9, 79.975] - - - [3840, 4224, 1, 4320] - - [10, 79.983] - - - [7680, 8448, 1, 8192] - - [11, 81.886] - - - [7680, 8448, 1, 8448] - - [11, 81.886] - - - [7680, 8448, 1, 8640] - - [11, 81.886] - - - [4096, 7169, 1, 512] - - [350, 85.536] - - - [4096, 7681, 1, 512] - - [350, 88.592] - - - [4096, 8193, 1, 512] - - [350, 87.651] - - - [4608, 512, 1, 512] - - [340, 61.028] - - - [4608, 8193, 1, 512] - - [351, 85.78] - - - [4608, 8705, 1, 512] - - [350, 88.977] - - - [4608, 9217, 1, 512] - - [400, 87.937] - - - [5120, 512, 1, 512] - - [341, 64.621] - - - [5120, 9217, 1, 512] - - [400, 87.804] - - - [5120, 9729, 1, 512] - - [350, 90.757] - - - [5120, 10241, 1, 512] - - [400, 88.704] - - - [5632, 512, 1, 512] - - [361, 60.596] - - - [5632, 10241, 1, 512] - - [400, 88.774] - - - [5632, 10753, 1, 512] - - [350, 91.659] - - - [5632, 11265, 1, 512] - - [400, 89.158] - - - [6144, 512, 1, 512] - - [370, 64.806] - - - [6144, 11265, 1, 512] - - [400, 89.621] - - - [6144, 11777, 1, 512] - - [350, 90.914] - - - [6144, 12289, 1, 512] - - [400, 89.879] - - - [6656, 512, 1, 512] - - [415, 68.733] - - - [6656, 12289, 1, 512] - - [400, 90.159] - - - [6656, 12801, 1, 512] - - [350, 91.618] - - - [6656, 13313, 1, 512] - - [400, 90.46] - - - [7168, 512, 1, 512] - - [416, 63.335] - - - [7168, 13313, 1, 512] - - [400, 90.805] - - - [7168, 13825, 1, 512] - - [400, 91.872] - - - [7168, 14337, 1, 512] - - [400, 91.109] - - - [7680, 512, 1, 512] - - [417, 66.81] - - - [7680, 14337, 1, 512] - - [400, 91.315] - - - [7680, 14849, 1, 512] - - [350, 92.492] - - - [7680, 15361, 1, 512] - - [400, 91.716] - - - [8192, 512, 1, 512] - - [370, 69.324] - - - [8192, 15361, 1, 512] - - [400, 91.987] - - - [8192, 15873, 1, 512] - - [400, 92.658] - - - [8192, 16385, 1, 512] - - [400, 91.743] - - - [8704, 512, 1, 512] - - [341, 71.805] - - - [8704, 16385, 1, 512] - - [400, 91.765] - - - [8704, 16897, 1, 512] - - [400, 92.93] - - - [8704, 17409, 1, 512] - - [400, 91.896] - - - [9216, 512, 1, 512] - - [418, 67.247] - - - [9216, 17409, 1, 512] - - [400, 92.06] - - - [9216, 17921, 1, 512] - - [400, 93.035] - - - [9216, 18433, 1, 512] - - [400, 92.186] - - - [9728, 512, 1, 512] - - [358, 69.947] - - - [9728, 18433, 1, 512] - - [400, 92.306] - - - [9728, 18945, 1, 512] - - [400, 93.033] - - - [9728, 19457, 1, 512] - - [400, 92.201] - - - [10240, 512, 1, 512] - - [410, 73.227] - - - [10240, 19457, 1, 512] - - [400, 92.37] - - - [10240, 19969, 1, 512] - - [400, 93.205] - - - [10240, 20481, 1, 512] - - [400, 92.427] - - - [10752, 512, 1, 512] - - [358, 68.474] - - - [10752, 20481, 1, 512] - - [400, 92.448] - - - [10752, 20993, 1, 512] - - [400, 93.339] - - - [10752, 21505, 1, 512] - - [400, 92.35] - - - [11264, 512, 1, 512] - - [370, 71.121] - - - [11264, 21505, 1, 512] - - [400, 91.753] - - - [11264, 22017, 1, 512] - - [419, 93.268] - - - [11264, 22529, 1, 512] - - [400, 91.752] - - - [11776, 512, 1, 512] - - [358, 72.989] - - - [11776, 22529, 1, 512] - - [400, 92.456] - - - [11776, 23041, 1, 512] - - [400, 93.417] - - - [11776, 23553, 1, 512] - - [400, 92.571] - - - [12288, 512, 1, 512] - - [341, 74.712] - - - [12288, 23553, 1, 512] - - [400, 92.739] - - - [12288, 24065, 1, 512] - - [400, 93.524] - - - [12288, 24577, 1, 512] - - [400, 92.634] - - - [12800, 512, 1, 512] - - [420, 73.619] - - - [12800, 24577, 1, 512] - - [400, 92.516] - - - [12800, 25089, 1, 512] - - [419, 93.558] - - - [12800, 25601, 1, 512] - - [400, 92.628] - - - [13312, 512, 1, 512] - - [421, 75.137] - - - [13312, 25601, 1, 512] - - [400, 92.85] - - - [13312, 26113, 1, 512] - - [400, 93.698] - - - [13312, 26625, 1, 512] - - [400, 92.8] - - - [13824, 512, 1, 512] - - [353, 76.888] - - - [13824, 26625, 1, 512] - - [400, 92.725] - - - [13824, 27137, 1, 512] - - [400, 93.842] - - - [13824, 27649, 1, 512] - - [400, 92.875] - - - [14336, 512, 1, 512] - - [370, 72.686] - - - [14336, 27649, 1, 512] - - [400, 92.322] - - - [14336, 28161, 1, 512] - - [400, 93.692] - - - [14336, 28673, 1, 512] - - [400, 92.297] - - - [14848, 512, 1, 512] - - [370, 73.98] - - - [14848, 28673, 1, 512] - - [400, 92.863] - - - [14848, 29185, 1, 512] - - [400, 93.926] - - - [14848, 29697, 1, 512] - - [400, 92.832] - - - [15360, 512, 1, 512] - - [370, 76.564] - - - [15360, 29697, 1, 512] - - [400, 93.091] - - - [15360, 30209, 1, 512] - - [400, 94.077] - - - [15360, 30721, 1, 512] - - [400, 93.092] - - - [15872, 512, 1, 512] - - [347, 74.318] - - - [15872, 30721, 1, 512] - - [400, 93.106] - - - [15872, 31233, 1, 512] - - [400, 94.129] - - - [15872, 31745, 1, 512] - - [368, 92.849] - - - [16384, 512, 1, 512] - - [357, 76.807] - - - [16384, 31745, 1, 512] - - [400, 93.184] - - - [16384, 32257, 1, 512] - - [400, 94.208] - - - [16384, 32769, 1, 512] - - [368, 92.799] - - - [16896, 512, 1, 512] - - [392, 78.306] - - - [16896, 32769, 1, 512] - - [400, 92.7] - - - [16896, 33281, 1, 512] - - [419, 94.132] - - - [16896, 33793, 1, 512] - - [400, 92.887] - - - [17408, 512, 1, 512] - - [422, 79.555] - - - [17408, 33793, 1, 512] - - [400, 93.319] - - - [17408, 34305, 1, 512] - - [400, 94.243] - - - [17408, 34817, 1, 512] - - [400, 93.196] - - - [17920, 512, 1, 512] - - [423, 76.675] - - - [17920, 34817, 1, 512] - - [403, 92.761] - - - [17920, 35329, 1, 512] - - [400, 94.338] - - - [17920, 35841, 1, 512] - - [368, 93.146] - - - [18432, 512, 1, 512] - - [420, 76.888] - - - [18432, 35841, 1, 512] - - [400, 93.239] - - - [18432, 36353, 1, 512] - - [400, 94.316] - - - [18432, 36865, 1, 512] - - [400, 93.185] - - - [18944, 512, 1, 512] - - [370, 77.842] - - - [18944, 36865, 1, 512] - - [368, 93.139] - - - [18944, 37377, 1, 512] - - [400, 94.454] - - - [18944, 37889, 1, 512] - - [368, 93.15] - - - [19456, 512, 1, 512] - - [354, 77.259] - - - [19456, 37889, 1, 512] - - [400, 93.296] - - - [19456, 38401, 1, 512] - - [400, 94.441] - - - [19456, 38913, 1, 512] - - [368, 93.26] - - - [19968, 512, 1, 512] - - [347, 78.762] - - - [19968, 38913, 1, 512] - - [368, 92.615] - - - [19968, 39425, 1, 512] - - [419, 94.329] - - - [19968, 39937, 1, 512] - - [368, 92.875] - - - [20480, 512, 1, 512] - - [392, 80.62] - - - [20480, 39937, 1, 512] - - [368, 93.474] - - - [20480, 40449, 1, 512] - - [400, 94.532] - - - [20480, 40961, 1, 512] - - [368, 93.37] - - - [20992, 512, 1, 512] - - [424, 81.818] - - - [20992, 40961, 1, 512] - - [368, 93.332] - - - [20992, 41473, 1, 512] - - [400, 94.568] - - - [20992, 41985, 1, 512] - - [368, 93.318] - - - [21504, 512, 1, 512] - - [370, 76.725] - - - [21504, 41985, 1, 512] - - [368, 93.38] - - - [21504, 42497, 1, 512] - - [400, 94.572] - - - [21504, 43009, 1, 512] - - [368, 93.385] - - - [22016, 512, 1, 512] - - [418, 77.777] - - - [22016, 43009, 1, 512] - - [368, 93.137] - - - [22016, 43521, 1, 512] - - [400, 94.611] - - - [22016, 44033, 1, 512] - - [400, 93.604] - - - [22528, 512, 1, 512] - - [370, 79.113] - - - [22528, 44033, 1, 512] - - [368, 93.221] - - - [22528, 44545, 1, 512] - - [419, 94.489] - - - [22528, 45057, 1, 512] - - [403, 92.362] - - - [23040, 512, 1, 512] - - [422, 79.238] - - - [23040, 45057, 1, 512] - - [400, 93.403] - - - [23040, 45569, 1, 512] - - [400, 94.638] - - - [23040, 46081, 1, 512] - - [368, 93.413] - - - [23552, 512, 1, 512] - - [425, 80.553] - - - [23552, 46081, 1, 512] - - [368, 93.536] - - - [23552, 46593, 1, 512] - - [400, 94.678] - - - [23552, 47105, 1, 512] - - [394, 92.193] - - - [24064, 512, 1, 512] - - [354, 82.115] - - - [24064, 47105, 1, 512] - - [402, 92.929] - - - [24064, 47617, 1, 512] - - [400, 94.662] - - - [24064, 48129, 1, 512] - - [368, 92.516] - - - [24576, 512, 1, 512] - - [354, 82.668] - - - [24576, 48129, 1, 512] - - [402, 93.13] - - - [24576, 48641, 1, 512] - - [400, 94.711] - - - [24576, 49153, 1, 512] - - [368, 93.461] - - - [25088, 512, 1, 512] - - [423, 78.324] - - - [25088, 49153, 1, 512] - - [426, 90.302] - - - [25088, 49665, 1, 512] - - [400, 94.727] - - - [25088, 50177, 1, 512] - - [368, 92.839] - - - [25600, 512, 1, 512] - - [415, 78.778] - - - [25600, 50177, 1, 512] - - [368, 93.127] - - - [25600, 50689, 1, 512] - - [419, 94.652] - - - [25600, 51201, 1, 512] - - [368, 93.172] - - - [26112, 512, 1, 512] - - [370, 80.123] - - - [26112, 51201, 1, 512] - - [400, 93.64] - - - [26112, 51713, 1, 512] - - [400, 94.744] - - - [26112, 52225, 1, 512] - - [403, 93.146] - - - [26624, 512, 1, 512] - - [392, 81.126] - - - [26624, 52225, 1, 512] - - [402, 92.925] - - - [26624, 52737, 1, 512] - - [400, 94.753] - - - [26624, 53249, 1, 512] - - [402, 92.468] - - - [27136, 512, 1, 512] - - [350, 81.988] - - - [27136, 53249, 1, 512] - - [351, 91.081] - - - [27136, 53761, 1, 512] - - [400, 94.743] - - - [27136, 54273, 1, 512] - - [368, 92.403] - - - [27648, 512, 1, 512] - - [346, 83.513] - - - [27648, 54273, 1, 512] - - [402, 92.038] - - - [27648, 54785, 1, 512] - - [400, 94.792] - - - [27648, 55297, 1, 512] - - [368, 92.827] - - - [28160, 512, 1, 512] - - [425, 84.177] - - - [28160, 55297, 1, 512] - - [350, 91.328] - - - [28160, 55809, 1, 512] - - [427, 94.609] - - - [28160, 56321, 1, 512] - - [350, 91.689] - - - [28672, 512, 1, 512] - - [418, 78.955] - - - [28672, 56321, 1, 512] - - [350, 92.314] - - - [28672, 56833, 1, 512] - - [400, 94.769] - - - [28672, 57345, 1, 512] - - [428, 90.676] - - - [29184, 512, 1, 512] - - [418, 79.955] - - - [29184, 57345, 1, 512] - - [429, 90.551] - - - [29184, 57857, 1, 512] - - [400, 94.792] - - - [29184, 58369, 1, 512] - - [402, 92.013] - - - [29696, 512, 1, 512] - - [400, 81.132] - - - [29696, 58369, 1, 512] - - [368, 92.65] - - - [29696, 58881, 1, 512] - - [400, 94.781] - - - [29696, 59393, 1, 512] - - [350, 91.981] - - - [30208, 512, 1, 512] - - [430, 82.037] - - - [30208, 59393, 1, 512] - - [368, 91.735] - - - [30208, 59905, 1, 512] - - [400, 94.813] - - - [30208, 60417, 1, 512] - - [351, 92.037] - - - [30720, 512, 1, 512] - - [400, 83.409] - - - [30720, 60417, 1, 512] - - [368, 92.224] - - - [30720, 60929, 1, 512] - - [419, 94.757] - - - [30720, 61441, 1, 512] - - [351, 90.988] - - - [31232, 512, 1, 512] - - [347, 84.43] - - - [31232, 61441, 1, 512] - - [368, 91.213] - - - [31232, 61953, 1, 512] - - [419, 94.752] - - - [31232, 62465, 1, 512] - - [368, 92.432] - - - [31744, 512, 1, 512] - - [415, 78.916] - - - [31744, 62465, 1, 512] - - [350, 92.209] - - - [31744, 62977, 1, 512] - - [400, 94.803] - - - [31744, 63489, 1, 512] - - [351, 91.664] - - - [32256, 512, 1, 512] - - [418, 79.836] - - - [32256, 63489, 1, 512] - - [351, 91.629] - - - [32256, 64001, 1, 512] - - [419, 94.7] - - - [32256, 64513, 1, 512] - - [350, 91.751] - - - [32768, 512, 1, 512] - - [396, 81.599] - - - [32768, 64513, 1, 512] - - [350, 92.19] - - - [32768, 65025, 1, 512] - - [400, 94.825] - - - [32768, 65537, 1, 512] - - [431, 90.201] - - - [33280, 512, 1, 512] - - [353, 81.788] - - - [33280, 65537, 1, 512] - - [426, 90.25] - - - [33280, 66049, 1, 512] - - [400, 94.839] - - - [33280, 66561, 1, 512] - - [351, 92.013] - - - [33792, 512, 1, 512] - - [392, 83.219] - - - [33792, 66561, 1, 512] - - [351, 91.474] - - - [33792, 67073, 1, 512] - - [419, 94.765] - - - [33792, 67585, 1, 512] - - [350, 91.293] - - - [34304, 512, 1, 512] - - [432, 84.322] - - - [34304, 67585, 1, 512] - - [351, 91.61] - - - [34304, 68097, 1, 512] - - [419, 94.774] - - - [34304, 68609, 1, 512] - - [351, 92.073] - - - [34816, 512, 1, 512] - - [346, 85.563] - - - [34816, 68609, 1, 512] - - [351, 92.015] - - - [34816, 69121, 1, 512] - - [419, 94.787] - - - [34816, 69633, 1, 512] - - [350, 91.083] - - - [35328, 512, 1, 512] - - [400, 79.51] - - - [35328, 69633, 1, 512] - - [401, 90.898] - - - [35328, 70145, 1, 512] - - [419, 94.583] - - - [35328, 70657, 1, 512] - - [351, 91.977] - - - [35840, 512, 1, 512] - - [400, 81.188] - - - [35840, 70657, 1, 512] - - [350, 92.19] - - - [35840, 71169, 1, 512] - - [419, 94.724] - - - [35840, 71681, 1, 512] - - [350, 92.093] - - - [36352, 512, 1, 512] - - [396, 81.877] - - - [36352, 71681, 1, 512] - - [350, 91.525] - - - [36352, 72193, 1, 512] - - [419, 94.768] - - - [36352, 72705, 1, 512] - - [351, 91.204] - - - [36864, 512, 1, 512] - - [400, 83.284] - - - [36864, 72705, 1, 512] - - [350, 91.642] - - - [36864, 73217, 1, 512] - - [419, 94.751] - - - [36864, 73729, 1, 512] - - [401, 90.581] - - - [37376, 512, 1, 512] - - [433, 83.768] - - - [37376, 73729, 1, 512] - - [428, 90.822] - - - [37376, 74241, 1, 512] - - [419, 94.784] - - - [37376, 74753, 1, 512] - - [351, 92.247] - - - [37888, 512, 1, 512] - - [403, 85.126] - - - [37888, 74753, 1, 512] - - [350, 92.022] - - - [37888, 75265, 1, 512] - - [419, 94.731] - - - [37888, 75777, 1, 512] - - [351, 91.391] - - - [38400, 512, 1, 512] - - [347, 85.625] - - - [38400, 75777, 1, 512] - - [351, 91.621] - - - [38400, 76289, 1, 512] - - [419, 94.681] - - - [38400, 76801, 1, 512] - - [351, 91.801] - - - [38912, 512, 1, 512] - - [419, 81.231] - - - [38912, 76801, 1, 512] - - [351, 91.716] - - - [38912, 77313, 1, 512] - - [419, 94.613] - - - [38912, 77825, 1, 512] - - [428, 90.943] - - - [39424, 512, 1, 512] - - [430, 81.751] - - - [39424, 77825, 1, 512] - - [434, 90.725] - - - [39424, 78337, 1, 512] - - [433, 94.481] - - - [39424, 78849, 1, 512] - - [351, 91.106] - - - [39936, 512, 1, 512] - - [396, 83.266] - - - [39936, 78849, 1, 512] - - [351, 91.863] - - - [39936, 79361, 1, 512] - - [433, 94.554] - - - [39936, 79873, 1, 512] - - [351, 91.513] - - - [40448, 512, 1, 512] - - [433, 83.728] - - - [40448, 79873, 1, 512] - - [350, 91.538] - - - [40448, 80385, 1, 512] - - [419, 94.013] - - - [40448, 80897, 1, 512] - - [351, 91.42] - - - [40960, 512, 1, 512] - - [400, 84.197] - - - [40960, 80897, 1, 512] - - [351, 91.652] - - - [40960, 81409, 1, 512] - - [419, 94.542] - - - [40960, 81921, 1, 512] - - [426, 89.777] - - - [41472, 512, 1, 512] - - [346, 85.22] - - - [41472, 81921, 1, 512] - - [426, 90.029] - - - [41472, 82433, 1, 512] - - [433, 94.408] - - - [41472, 82945, 1, 512] - - [351, 91.515] - - - [41984, 512, 1, 512] - - [421, 86.377] - - - [41984, 82945, 1, 512] - - [428, 90.977] - - - [41984, 83457, 1, 512] - - [433, 94.462] - - - [41984, 83969, 1, 512] - - [428, 90.982] - - - [42496, 512, 1, 512] - - [400, 82.031] - - - [42496, 83969, 1, 512] - - [401, 90.965] - - - [42496, 84481, 1, 512] - - [433, 93.776] - - - [42496, 84993, 1, 512] - - [351, 91.505] - - - [43008, 512, 1, 512] - - [400, 83.182] - - - [43008, 84993, 1, 512] - - [351, 91.345] - - - [43008, 85505, 1, 512] - - [433, 93.68] - - - [43008, 86017, 1, 512] - - [401, 90.875] - - - [43520, 512, 1, 512] - - [368, 83.354] - - - [43520, 86017, 1, 512] - - [428, 90.711] - - - [43520, 86529, 1, 512] - - [433, 93.399] - - - [43520, 87041, 1, 512] - - [401, 90.98] - - - [44032, 512, 1, 512] - - [400, 84.913] - - - [44032, 87041, 1, 512] - - [351, 91.138] - - - [44032, 87553, 1, 512] - - [427, 92.587] - - - [44032, 88065, 1, 512] - - [351, 91.067] - - - [44544, 512, 1, 512] - - [432, 85.094] - - - [44544, 88065, 1, 512] - - [401, 90.952] - - - [44544, 88577, 1, 512] - - [351, 92.974] - - - [44544, 89089, 1, 512] - - [428, 90.98] - - - [45056, 512, 1, 512] - - [400, 85.864] - - - [45056, 89089, 1, 512] - - [428, 90.994] - - - [45056, 89601, 1, 512] - - [351, 92.561] - - - [45056, 90113, 1, 512] - - [435, 90.178] - - - [45568, 512, 1, 512] - - [350, 86.491] - - - [45568, 90113, 1, 512] - - [431, 90.432] - - - [45568, 90625, 1, 512] - - [351, 92.553] - - - [45568, 91137, 1, 512] - - [428, 90.981] - - - [46080, 512, 1, 512] - - [419, 83.186] - - - [46080, 91137, 1, 512] - - [401, 90.958] - - - [46080, 91649, 1, 512] - - [351, 92.372] - - - [46080, 92161, 1, 512] - - [428, 90.959] - - - [46592, 512, 1, 512] - - [419, 83.421] - - - [46592, 92161, 1, 512] - - [428, 90.961] - - - [46592, 92673, 1, 512] - - [422, 92.418] - - - [46592, 93185, 1, 512] - - [401, 90.931] - - - [47104, 512, 1, 512] - - [396, 84.84] - - - [47104, 93185, 1, 512] - - [428, 90.946] - - - [47104, 93697, 1, 512] - - [351, 92.316] - - - [47104, 94209, 1, 512] - - [431, 90.801] - - - [47616, 512, 1, 512] - - [403, 84.985] - - - [47616, 94209, 1, 512] - - [401, 90.871] - - - [47616, 94721, 1, 512] - - [427, 92.107] - - - [47616, 95233, 1, 512] - - [428, 90.92] - - - [48128, 512, 1, 512] - - [396, 85.938] - - - [48128, 95233, 1, 512] - - [428, 90.98] - - - [48128, 95745, 1, 512] - - [351, 92.194] - - - [48128, 96257, 1, 512] - - [428, 90.909] - - - [48640, 512, 1, 512] - - [396, 86.2] - - - [48640, 96257, 1, 512] - - [428, 90.919] - - - [48640, 96769, 1, 512] - - [436, 92.138] - - - [48640, 97281, 1, 512] - - [401, 90.928] - - - [49152, 512, 1, 512] - - [392, 85.833] - - - [49152, 97281, 1, 512] - - [401, 90.849] - - - [49152, 97793, 1, 512] - - [436, 91.949] - - - [49152, 98305, 1, 512] - - [426, 87.715] - - - [49664, 512, 1, 512] - - [427, 83.009] - - - [49664, 98305, 1, 512] - - [426, 88.652] - - - [49664, 98817, 1, 512] - - [436, 91.906] - - - [49664, 99329, 1, 512] - - [401, 90.805] - - - [50176, 512, 1, 512] - - [396, 84.679] - - - [50176, 99329, 1, 512] - - [434, 90.758] - - - [50176, 99841, 1, 512] - - [436, 91.898] - - - [50176, 100353, 1, 512] - - [434, 90.692] - - - [50688, 512, 1, 512] - - [400, 84.876] - - - [50688, 100353, 1, 512] - - [431, 90.704] - - - [50688, 100865, 1, 512] - - [436, 91.91] - - - [50688, 101377, 1, 512] - - [431, 90.709] - - - [51200, 512, 1, 512] - - [396, 85.978] - - - [51200, 101377, 1, 512] - - [431, 90.761] - - - [51200, 101889, 1, 512] - - [436, 91.709] - - - [51200, 102401, 1, 512] - - [431, 90.597] - - - [51712, 512, 1, 512] - - [400, 86.164] - - - [51712, 102401, 1, 512] - - [431, 90.686] - - - [51712, 102913, 1, 512] - - [436, 91.64] - - - [51712, 103425, 1, 512] - - [431, 90.761] - - - [52224, 512, 1, 512] - - [396, 87.177] - - - [52224, 103425, 1, 512] - - [431, 90.784] - - - [52224, 103937, 1, 512] - - [436, 91.567] - - - [52224, 104449, 1, 512] - - [431, 90.728] - - - [52736, 512, 1, 512] - - [392, 87.096] - - - [52736, 104449, 1, 512] - - [435, 90.681] - - - [52736, 104961, 1, 512] - - [436, 91.509] - - - [52736, 105473, 1, 512] - - [431, 90.712] - - - [53248, 512, 1, 512] - - [419, 84.232] - - - [53248, 105473, 1, 512] - - [431, 90.718] - - - [53248, 105985, 1, 512] - - [436, 91.411] - - - [53248, 106497, 1, 512] - - [431, 89.789] - - - [53760, 512, 1, 512] - - [419, 84.477] - - - [53760, 106497, 1, 512] - - [426, 88.672] - - - [53760, 107009, 1, 512] - - [437, 91.421] - - - [53760, 107521, 1, 512] - - [431, 90.688] - - - [54272, 512, 1, 512] - - [396, 85.577] - - - [54272, 107521, 1, 512] - - [435, 90.664] - - - [54272, 108033, 1, 512] - - [437, 91.426] - - - [54272, 108545, 1, 512] - - [435, 90.569] - - - [54784, 512, 1, 512] - - [396, 85.8] - - - [54784, 108545, 1, 512] - - [431, 90.245] - - - [54784, 109057, 1, 512] - - [437, 91.427] - - - [54784, 109569, 1, 512] - - [435, 90.553] - - - [55296, 512, 1, 512] - - [400, 86.948] - - - [55296, 109569, 1, 512] - - [435, 90.508] - - - [55296, 110081, 1, 512] - - [437, 91.43] - - - [55296, 110593, 1, 512] - - [427, 89.447] - - - [55808, 512, 1, 512] - - [419, 86.909] - - - [55808, 110593, 1, 512] - - [431, 90.408] - - - [55808, 111105, 1, 512] - - [437, 91.426] - - - [55808, 111617, 1, 512] - - [431, 90.145] - - - [56320, 512, 1, 512] - - [432, 87.521] - - - [56320, 111617, 1, 512] - - [438, 90.17] - - - [56320, 112129, 1, 512] - - [437, 91.446] - - - [56320, 112641, 1, 512] - - [439, 90.056] - - - [56832, 512, 1, 512] - - [403, 84.229] - - - [56832, 112641, 1, 512] - - [439, 90.033] - - - [56832, 113153, 1, 512] - - [437, 91.433] - - - [56832, 113665, 1, 512] - - [439, 90.236] - - - [57344, 512, 1, 512] - - [368, 84.701] - - - [57344, 113665, 1, 512] - - [438, 90.201] - - - [57344, 114177, 1, 512] - - [437, 91.436] - - - [57344, 114689, 1, 512] - - [419, 84.507] - - - [57856, 512, 1, 512] - - [400, 85.919] - - - [57856, 114689, 1, 512] - - [433, 84.504] - - - [57856, 115201, 1, 512] - - [437, 91.44] - - - [57856, 115713, 1, 512] - - [438, 89.892] - - - [58368, 512, 1, 512] - - [400, 86.81] - - - [58368, 115713, 1, 512] - - [438, 90.306] - - - [58368, 116225, 1, 512] - - [437, 91.434] - - - [58368, 116737, 1, 512] - - [438, 90.198] - - - [58880, 512, 1, 512] - - [400, 86.817] - - - [58880, 116737, 1, 512] - - [438, 90.052] - - - [58880, 117249, 1, 512] - - [437, 91.43] - - - [58880, 117761, 1, 512] - - [438, 90.221] - - - [59392, 512, 1, 512] - - [396, 87.66] - - - [59392, 117761, 1, 512] - - [438, 90.089] - - - [59392, 118273, 1, 512] - - [437, 91.442] - - - [59392, 118785, 1, 512] - - [427, 88.803] - - - [59904, 512, 1, 512] - - [400, 84.348] - - - [59904, 118785, 1, 512] - - [427, 88.885] - - - [59904, 119297, 1, 512] - - [437, 91.432] - - - [59904, 119809, 1, 512] - - [438, 89.959] - - - [60416, 512, 1, 512] - - [396, 85.374] - - - [60416, 119809, 1, 512] - - [438, 90.125] - - - [60416, 120321, 1, 512] - - [437, 91.436] - - - [60416, 120833, 1, 512] - - [438, 90.061] - - - [60928, 512, 1, 512] - - [400, 85.524] - - - [60928, 120833, 1, 512] - - [438, 89.972] - - - [60928, 121345, 1, 512] - - [437, 91.443] - - - [60928, 121857, 1, 512] - - [438, 90.213] - - - [61440, 512, 1, 512] - - [419, 86.645] - - - [61440, 121857, 1, 512] - - [439, 90.085] - - - [61440, 122369, 1, 512] - - [437, 91.453] - - - [61440, 122881, 1, 512] - - [436, 87.316] - - - [61952, 512, 1, 512] - - [364, 86.336] - - - [61952, 122881, 1, 512] - - [436, 87.335] - - - [61952, 123393, 1, 512] - - [437, 91.446] - - - [61952, 123905, 1, 512] - - [439, 89.936] - - - [62464, 512, 1, 512] - - [419, 87.618] - - - [62464, 123905, 1, 512] - - [439, 90.196] - - - [62464, 124417, 1, 512] - - [437, 91.458] - - - [62464, 124929, 1, 512] - - [438, 90.156] - - - [62976, 512, 1, 512] - - [427, 87.32] - - - [62976, 124929, 1, 512] - - [438, 90.203] - - - [62976, 125441, 1, 512] - - [437, 91.448] - - - [62976, 125953, 1, 512] - - [438, 90.056] - - - [63488, 512, 1, 512] - - [400, 85.346] - - - [63488, 125953, 1, 512] - - [438, 90.17] - - - [63488, 126465, 1, 512] - - [437, 91.441] - - - [63488, 126977, 1, 512] - - [440, 88.035] - - - [64000, 512, 1, 512] - - [396, 85.277] - - - [64000, 126977, 1, 512] - - [436, 88.054] - - - [64000, 127489, 1, 512] - - [13, 84.172] - - - [64000, 128001, 1, 512] - - [12, 84.304] - - - [64512, 512, 1, 512] - - [419, 86.545] - - - [64512, 128001, 1, 512] - - [13, 84.344] - - - [4096, 4096, 1, 4128] - - [14, 87.241] - - - [25600, 25600, 1, 512] - - [15, 85.988] - - - [512, 512, 1, 512] - - [489, 7.234] - - - [1024, 512, 1, 512] - - [405, 34.762] - - - [1536, 512, 1, 512] - - [408, 45.026] - - - [1536, 1024, 1, 512] - - [442, 57.88] - - - [2048, 512, 1, 512] - - [345, 46.26] - - - [2048, 1024, 1, 512] - - [355, 56.005] - - - [2560, 512, 1, 512] - - [413, 52.143] - - - [2560, 1024, 1, 512] - - [411, 65.745] - - - [2560, 1536, 1, 512] - - [456, 70.715] - - - [3072, 512, 1, 512] - - [340, 59.01] - - - [3072, 1024, 1, 512] - - [341, 63.502] - - - [3072, 1536, 1, 512] - - [370, 67.876] - - - [3584, 512, 1, 512] - - [414, 55.931] - - - [3584, 1536, 1, 512] - - [418, 68.18] - - - [3584, 2048, 1, 512] - - [447, 73.22] - - - [4096, 512, 1, 512] - - [355, 55.619] - - - [4096, 1536, 1, 512] - - [459, 73.838] - - - [4096, 2048, 1, 512] - - [347, 75.963] - - - [4608, 2048, 1, 512] - - [420, 76.997] - - - [4608, 2560, 1, 512] - - [456, 81.273] - - - [5120, 2048, 1, 512] - - [357, 79.822] - - - [5120, 2560, 1, 512] - - [456, 82.271] - - - [5632, 2560, 1, 512] - - [354, 84.973] - - - [5632, 3072, 1, 512] - - [447, 83.358] - - - [6144, 2560, 1, 512] - - [350, 84.183] - - - [6144, 3072, 1, 512] - - [448, 84.365] - - - [6656, 3072, 1, 512] - - [350, 82.798] - - - [6656, 3584, 1, 512] - - [463, 85.761] - - - [7168, 3072, 1, 512] - - [430, 83.182] - - - [7168, 3584, 1, 512] - - [419, 84.595] - - - [7680, 3584, 1, 512] - - [461, 85.627] - - - [7680, 4096, 1, 512] - - [460, 85.866] - - - [8192, 3584, 1, 512] - - [433, 86.287] - - - [8192, 4096, 1, 512] - - [353, 87.889] - - - [8704, 4096, 1, 512] - - [353, 87.735] - - - [8704, 4608, 1, 512] - - [430, 86.789] - - - [9216, 4096, 1, 512] - - [353, 89.25] - - - [9216, 4608, 1, 512] - - [419, 88.901] - - - [9728, 4608, 1, 512] - - [353, 89.287] - - - [9728, 5120, 1, 512] - - [347, 88.743] - - - [10240, 4608, 1, 512] - - [419, 88.386] - - - [10240, 5120, 1, 512] - - [403, 88.656] - - - [10752, 5120, 1, 512] - - [347, 88.806] - - - [10752, 5632, 1, 512] - - [400, 89.519] - - - [11264, 5120, 1, 512] - - [430, 90.695] - - - [11264, 5632, 1, 512] - - [419, 90.021] - - - [11776, 5632, 1, 512] - - [350, 89.927] - - - [11776, 6144, 1, 512] - - [433, 89.296] - - - [12288, 5632, 1, 512] - - [400, 90.641] - - - [12288, 6144, 1, 512] - - [396, 91.108] - - - [12800, 6144, 1, 512] - - [400, 89.896] - - - [12800, 6656, 1, 512] - - [403, 90.099] - - - [13312, 6144, 1, 512] - - [396, 90.824] - - - [13312, 6656, 1, 512] - - [403, 91.069] - - - [13824, 6656, 1, 512] - - [403, 90.058] - - - [13824, 7168, 1, 512] - - [430, 90.843] - - - [14336, 6656, 1, 512] - - [419, 91.813] - - - [14336, 7168, 1, 512] - - [396, 91.074] - - - [14848, 7168, 1, 512] - - [433, 90.72] - - - [14848, 7680, 1, 512] - - [403, 91.26] - - - [15360, 7168, 1, 512] - - [396, 91.246] - - - [15360, 7680, 1, 512] - - [419, 92.238] - - - [15872, 7680, 1, 512] - - [403, 91.457] - - - [15872, 8192, 1, 512] - - [403, 91.561] - - - [16384, 7680, 1, 512] - - [419, 92.935] - - - [16384, 8192, 1, 512] - - [403, 92.372] - - - [16896, 8192, 1, 512] - - [403, 92.027] - - - [16896, 8704, 1, 512] - - [400, 92.572] - - - [17408, 8192, 1, 512] - - [400, 92.204] - - - [17408, 8704, 1, 512] - - [400, 92.788] - - - [17920, 8704, 1, 512] - - [400, 92.745] - - - [17920, 9216, 1, 512] - - [419, 91.28] - - - [18432, 8704, 1, 512] - - [400, 93.141] - - - [18432, 9216, 1, 512] - - [419, 91.681] - - - [18944, 9216, 1, 512] - - [419, 91.341] - - - [18944, 9728, 1, 512] - - [403, 91.964] - - - [19456, 9216, 1, 512] - - [396, 91.834] - - - [19456, 9728, 1, 512] - - [433, 92.557] - - - [19968, 9728, 1, 512] - - [419, 92.288] - - - [19968, 10240, 1, 512] - - [400, 92.014] - - - [20480, 9728, 1, 512] - - [419, 92.945] - - - [20480, 10240, 1, 512] - - [400, 92.662] - - - [20992, 10240, 1, 512] - - [400, 92.508] - - - [20992, 10752, 1, 512] - - [400, 92.578] - - - [21504, 10240, 1, 512] - - [400, 92.743] - - - [21504, 10752, 1, 512] - - [400, 92.828] - - - [22016, 10752, 1, 512] - - [400, 92.555] - - - [22016, 11264, 1, 512] - - [400, 92.226] - - - [22528, 10752, 1, 512] - - [419, 92.891] - - - [22528, 11264, 1, 512] - - [400, 91.951] - - - [23040, 11264, 1, 512] - - [400, 92.215] - - - [23040, 11776, 1, 512] - - [433, 92.407] - - - [23552, 11264, 1, 512] - - [400, 92.527] - - - [23552, 11776, 1, 512] - - [419, 92.879] - - - [24064, 11776, 1, 512] - - [419, 92.483] - - - [24064, 12288, 1, 512] - - [403, 91.419] - - - [24576, 11776, 1, 512] - - [400, 92.731] - - - [24576, 12288, 1, 512] - - [396, 91.399] - - - [25088, 12288, 1, 512] - - [396, 91.447] - - - [25088, 12800, 1, 512] - - [433, 92.33] - - - [25600, 12288, 1, 512] - - [396, 91.528] - - - [25600, 12800, 1, 512] - - [419, 92.96] - - - [26112, 12800, 1, 512] - - [400, 92.689] - - - [26112, 13312, 1, 512] - - [403, 92.696] - - - [26624, 12800, 1, 512] - - [419, 93.074] - - - [26624, 13312, 1, 512] - - [400, 92.748] - - - [27136, 13312, 1, 512] - - [400, 92.702] - - - [27136, 13824, 1, 512] - - [400, 92.774] - - - [27648, 13312, 1, 512] - - [400, 92.898] - - - [27648, 13824, 1, 512] - - [419, 93.151] - - - [28160, 13824, 1, 512] - - [368, 92.411] - - - [28160, 14336, 1, 512] - - [419, 91.766] - - - [28672, 13824, 1, 512] - - [400, 93.122] - - - [28672, 14336, 1, 512] - - [427, 91.834] - - - [29184, 14336, 1, 512] - - [419, 92.103] - - - [29184, 14848, 1, 512] - - [419, 92.883] - - - [29696, 14336, 1, 512] - - [427, 92.148] - - - [29696, 14848, 1, 512] - - [419, 93.291] - - - [30208, 14848, 1, 512] - - [419, 92.905] - - - [30208, 15360, 1, 512] - - [403, 93.039] - - - [30720, 14848, 1, 512] - - [419, 93.41] - - - [30720, 15360, 1, 512] - - [400, 92.944] - - - [31232, 15360, 1, 512] - - [403, 92.645] - - - [31232, 15872, 1, 512] - - [419, 93.007] - - - [31744, 15360, 1, 512] - - [400, 93.237] - - - [31744, 15872, 1, 512] - - [400, 93.519] - - - [32256, 15872, 1, 512] - - [400, 93.008] - - - [32256, 16384, 1, 512] - - [400, 92.82] - - - [32768, 15872, 1, 512] - - [419, 93.627] - - - [32768, 16384, 1, 512] - - [396, 92.104] - - - [33280, 16384, 1, 512] - - [400, 92.88] - - - [33280, 16896, 1, 512] - - [400, 93.187] - - - [33792, 16384, 1, 512] - - [400, 92.876] - - - [33792, 16896, 1, 512] - - [419, 93.552] - - - [34304, 16896, 1, 512] - - [400, 93.18] - - - [34304, 17408, 1, 512] - - [396, 92.567] - - - [34816, 16896, 1, 512] - - [419, 93.617] - - - [34816, 17408, 1, 512] - - [427, 92.82] - - - [35328, 17408, 1, 512] - - [400, 92.452] - - - [35328, 17920, 1, 512] - - [400, 93.259] - - - [35840, 17408, 1, 512] - - [427, 92.734] - - - [35840, 17920, 1, 512] - - [419, 93.703] - - - [36352, 17920, 1, 512] - - [419, 93.206] - - - [36352, 18432, 1, 512] - - [400, 92.687] - - - [36864, 17920, 1, 512] - - [419, 93.588] - - - [36864, 18432, 1, 512] - - [400, 93.267] - - - [37376, 18432, 1, 512] - - [400, 93.397] - - - [37376, 18944, 1, 512] - - [400, 93.48] - - - [37888, 18432, 1, 512] - - [403, 93.251] - - - [37888, 18944, 1, 512] - - [419, 93.772] - - - [38400, 18944, 1, 512] - - [400, 93.336] - - - [38400, 19456, 1, 512] - - [419, 92.615] - - - [38912, 18944, 1, 512] - - [400, 93.686] - - - [38912, 19456, 1, 512] - - [400, 92.747] - - - [39424, 19456, 1, 512] - - [400, 92.99] - - - [39424, 19968, 1, 512] - - [400, 93.451] - - - [39936, 19456, 1, 512] - - [400, 92.831] - - - [39936, 19968, 1, 512] - - [419, 93.804] - - - [40448, 19968, 1, 512] - - [400, 93.329] - - - [40448, 20480, 1, 512] - - [400, 93.312] - - - [40960, 19968, 1, 512] - - [419, 93.409] - - - [40960, 20480, 1, 512] - - [400, 93.229] - - - [41472, 20480, 1, 512] - - [400, 93.477] - - - [41472, 20992, 1, 512] - - [400, 93.654] - - - [41984, 20480, 1, 512] - - [400, 93.242] - - - [41984, 20992, 1, 512] - - [419, 93.747] - - - [42496, 20992, 1, 512] - - [400, 93.312] - - - [42496, 21504, 1, 512] - - [400, 93.034] - - - [43008, 20992, 1, 512] - - [419, 93.959] - - - [43008, 21504, 1, 512] - - [403, 93.17] - - - [43520, 21504, 1, 512] - - [400, 93.399] - - - [43520, 22016, 1, 512] - - [400, 93.56] - - - [44032, 21504, 1, 512] - - [400, 93.59] - - - [44032, 22016, 1, 512] - - [419, 93.904] - - - [44544, 22016, 1, 512] - - [400, 93.517] - - - [44544, 22528, 1, 512] - - [400, 92.941] - - - [45056, 22016, 1, 512] - - [419, 93.872] - - - [45056, 22528, 1, 512] - - [400, 93.102] - - - [45568, 22528, 1, 512] - - [400, 92.817] - - - [45568, 23040, 1, 512] - - [400, 93.614] - - - [46080, 22528, 1, 512] - - [403, 92.833] - - - [46080, 23040, 1, 512] - - [400, 93.964] - - - [46592, 23040, 1, 512] - - [400, 93.575] - - - [46592, 23552, 1, 512] - - [403, 93.238] - - - [47104, 23040, 1, 512] - - [419, 93.985] - - - [47104, 23552, 1, 512] - - [400, 93.776] - - - [47616, 23552, 1, 512] - - [400, 93.091] - - - [47616, 24064, 1, 512] - - [432, 93.47] - - - [48128, 23552, 1, 512] - - [400, 93.639] - - - [48128, 24064, 1, 512] - - [419, 93.992] - - - [48640, 24064, 1, 512] - - [400, 93.787] - - - [48640, 24576, 1, 512] - - [400, 92.542] - - - [49152, 24064, 1, 512] - - [400, 93.163] - - - [49152, 24576, 1, 512] - - [396, 91.104] - - - [49664, 24576, 1, 512] - - [368, 92.154] - - - [49664, 25088, 1, 512] - - [400, 93.502] - - - [50176, 24576, 1, 512] - - [396, 92.383] - - - [50176, 25088, 1, 512] - - [419, 93.954] - - - [50688, 25088, 1, 512] - - [432, 93.581] - - - [50688, 25600, 1, 512] - - [400, 93.744] - - - [51200, 25088, 1, 512] - - [419, 94.029] - - - [51200, 25600, 1, 512] - - [400, 93.809] - - - [51712, 25600, 1, 512] - - [400, 93.511] - - - [51712, 26112, 1, 512] - - [400, 93.575] - - - [52224, 25600, 1, 512] - - [400, 93.831] - - - [52224, 26112, 1, 512] - - [419, 94.051] - - - [52736, 26112, 1, 512] - - [400, 93.713] - - - [52736, 26624, 1, 512] - - [400, 93.364] - - - [53248, 26112, 1, 512] - - [419, 93.695] - - - [53248, 26624, 1, 512] - - [400, 93.283] - - - [53760, 26624, 1, 512] - - [400, 93.254] - - - [53760, 27136, 1, 512] - - [419, 93.441] - - - [54272, 26624, 1, 512] - - [400, 93.58] - - - [54272, 27136, 1, 512] - - [419, 93.988] - - - [54784, 27136, 1, 512] - - [400, 93.467] - - - [54784, 27648, 1, 512] - - [400, 93.139] - - - [55296, 27136, 1, 512] - - [419, 94.025] - - - [55296, 27648, 1, 512] - - [427, 93.073] - - - [55808, 27648, 1, 512] - - [400, 92.986] - - - [55808, 28160, 1, 512] - - [400, 93.552] - - - [56320, 27648, 1, 512] - - [427, 93.184] - - - [56320, 28160, 1, 512] - - [400, 93.961] - - - [56832, 28160, 1, 512] - - [400, 93.56] - - - [56832, 28672, 1, 512] - - [400, 93.355] - - - [57344, 28160, 1, 512] - - [400, 93.192] - - - [57344, 28672, 1, 512] - - [400, 92.929] - - - [57856, 28672, 1, 512] - - [400, 93.41] - - - [57856, 29184, 1, 512] - - [400, 93.598] - - - [58368, 28672, 1, 512] - - [400, 93.422] - - - [58368, 29184, 1, 512] - - [419, 93.951] - - - [58880, 29184, 1, 512] - - [432, 93.57] - - - [58880, 29696, 1, 512] - - [400, 93.03] - - - [59392, 29184, 1, 512] - - [419, 94.034] - - - [59392, 29696, 1, 512] - - [427, 93.163] - - - [59904, 29696, 1, 512] - - [400, 93.044] - - - [59904, 30208, 1, 512] - - [400, 93.534] - - - [60416, 29696, 1, 512] - - [400, 93.272] - - - [60416, 30208, 1, 512] - - [419, 93.824] - - - [60928, 30208, 1, 512] - - [400, 93.508] - - - [60928, 30720, 1, 512] - - [400, 93.492] - - - [61440, 30208, 1, 512] - - [419, 93.935] - - - [61440, 30720, 1, 512] - - [400, 93.6] - - - [61952, 30720, 1, 512] - - [400, 93.47] - - - [61952, 31232, 1, 512] - - [400, 93.606] - - - [62464, 30720, 1, 512] - - [403, 93.379] - - - [62464, 31232, 1, 512] - - [419, 93.996] - - - [62976, 31232, 1, 512] - - [400, 93.564] - - - [62976, 31744, 1, 512] - - [400, 93.317] - - - [63488, 31232, 1, 512] - - [419, 93.948] - - - [63488, 31744, 1, 512] - - [400, 93.616] - - - [64000, 31744, 1, 512] - - [400, 93.41] - - - [64000, 32256, 1, 512] - - [400, 93.63] - - - [64512, 31744, 1, 512] - - [400, 93.343] - - - [64512, 32256, 1, 512] - - [419, 93.924] - - - [65024, 512, 1, 512] - - [460, 86.136] - - - [65024, 32256, 1, 512] - - [400, 93.523] - - - [65024, 32768, 1, 512] - - [368, 92.472] - - - [65536, 512, 1, 512] - - [464, 86.09] - - - [65536, 32256, 1, 512] - - [419, 92.92] - - - [65536, 32768, 1, 512] - - [403, 91.355] - - - [66048, 512, 1, 512] - - [400, 87.141] - - - [66048, 32768, 1, 512] - - [400, 92.817] - - - [66048, 33280, 1, 512] - - [403, 93.439] - - - [66560, 512, 1, 512] - - [419, 87.993] - - - [66560, 32768, 1, 512] - - [427, 92.351] - - - [66560, 33280, 1, 512] - - [400, 93.913] - - - [67072, 512, 1, 512] - - [427, 85.063] - - - [67072, 33280, 1, 512] - - [400, 93.459] - - - [67072, 33792, 1, 512] - - [400, 93.295] - - - [67584, 512, 1, 512] - - [396, 86.342] - - - [67584, 33280, 1, 512] - - [419, 93.85] - - - [67584, 33792, 1, 512] - - [427, 93.018] - - - [68096, 512, 1, 512] - - [400, 86.093] - - - [68096, 33792, 1, 512] - - [400, 93.243] - - - [68096, 34304, 1, 512] - - [400, 93.399] - - - [68608, 512, 1, 512] - - [396, 87.404] - - - [68608, 33792, 1, 512] - - [402, 92.898] - - - [68608, 34304, 1, 512] - - [419, 93.97] - - - [69120, 512, 1, 512] - - [419, 86.886] - - - [69120, 34304, 1, 512] - - [400, 93.581] - - - [69120, 34816, 1, 512] - - [400, 92.893] - - - [69632, 512, 1, 512] - - [396, 87.922] - - - [69632, 34304, 1, 512] - - [419, 93.88] - - - [69632, 34816, 1, 512] - - [427, 92.711] - - - [70144, 512, 1, 512] - - [472, 87.892] - - - [70144, 34816, 1, 512] - - [400, 92.998] - - - [70144, 35328, 1, 512] - - [400, 93.49] - - - [70656, 512, 1, 512] - - [419, 86.122] - - - [70656, 34816, 1, 512] - - [403, 92.804] - - - [70656, 35328, 1, 512] - - [419, 93.87] - - - [71168, 512, 1, 512] - - [433, 86.142] - - - [71168, 35328, 1, 512] - - [400, 93.28] - - - [71168, 35840, 1, 512] - - [403, 92.956] - - - [71680, 512, 1, 512] - - [419, 87.316] - - - [71680, 35328, 1, 512] - - [419, 93.885] - - - [71680, 35840, 1, 512] - - [427, 93.177] - - - [72192, 512, 1, 512] - - [424, 86.812] - - - [72192, 35840, 1, 512] - - [400, 93.372] - - - [72192, 36352, 1, 512] - - [400, 93.47] - - - [72704, 512, 1, 512] - - [419, 87.767] - - - [72704, 35840, 1, 512] - - [400, 93.54] - - - [72704, 36352, 1, 512] - - [419, 93.875] - - - [73216, 512, 1, 512] - - [419, 87.606] - - - [73216, 36352, 1, 512] - - [400, 93.463] - - - [73216, 36864, 1, 512] - - [402, 92.559] - - - [73728, 512, 1, 512] - - [467, 87.396] - - - [73728, 36352, 1, 512] - - [419, 92.903] - - - [73728, 36864, 1, 512] - - [368, 91.023] - - - [74240, 512, 1, 512] - - [436, 85.807] - - - [74240, 36864, 1, 512] - - [400, 93.004] - - - [74240, 37376, 1, 512] - - [400, 93.342] - - - [74752, 512, 1, 512] - - [396, 86.952] - - - [74752, 36864, 1, 512] - - [436, 92.363] - - - [74752, 37376, 1, 512] - - [400, 93.684] - - - [75264, 512, 1, 512] - - [400, 86.787] - - - [75264, 37376, 1, 512] - - [400, 93.328] - - - [75264, 37888, 1, 512] - - [403, 92.823] - - - [75776, 512, 1, 512] - - [419, 87.884] - - - [75776, 37376, 1, 512] - - [419, 93.845] - - - [75776, 37888, 1, 512] - - [427, 93.128] - - - [76288, 512, 1, 512] - - [430, 87.379] - - - [76288, 37888, 1, 512] - - [400, 92.806] - - - [76288, 38400, 1, 512] - - [400, 93.449] - - - [76800, 512, 1, 512] - - [396, 88.3] - - - [76800, 37888, 1, 512] - - [427, 92.473] - - - [76800, 38400, 1, 512] - - [419, 93.734] - - - [77312, 512, 1, 512] - - [467, 88.014] - - - [77312, 38400, 1, 512] - - [400, 93.393] - - - [77312, 38912, 1, 512] - - [368, 92.857] - - - [77824, 512, 1, 512] - - [396, 86.611] - - - [77824, 38400, 1, 512] - - [419, 93.758] - - - [77824, 38912, 1, 512] - - [427, 92.257] - - - [78336, 512, 1, 512] - - [424, 86.271] - - - [78336, 38912, 1, 512] - - [400, 93.263] - - - [78336, 39424, 1, 512] - - [400, 93.4] - - - [78848, 512, 1, 512] - - [396, 87.808] - - - [78848, 38912, 1, 512] - - [402, 92.813] - - - [78848, 39424, 1, 512] - - [433, 93.645] - - - [79360, 512, 1, 512] - - [465, 87.195] - - - [79360, 39424, 1, 512] - - [403, 93.18] - - - [79360, 39936, 1, 512] - - [400, 92.715] - - - [79872, 512, 1, 512] - - [396, 88.225] - - - [79872, 39424, 1, 512] - - [419, 93.761] - - - [79872, 39936, 1, 512] - - [427, 92.352] - - - [80384, 512, 1, 512] - - [473, 87.838] - - - [80384, 39936, 1, 512] - - [474, 92.304] - - - [80384, 40448, 1, 512] - - [433, 93.166] - - - [80896, 512, 1, 512] - - [419, 88.701] - - - [80896, 39936, 1, 512] - - [430, 92.198] - - - [80896, 40448, 1, 512] - - [419, 93.666] - - - [81408, 512, 1, 512] - - [427, 86.485] - - - [81408, 40448, 1, 512] - - [419, 93.092] - - - [81408, 40960, 1, 512] - - [419, 91.546] - - - [81920, 512, 1, 512] - - [429, 85.326] - - - [81920, 40448, 1, 512] - - [368, 91.005] - - - [81920, 40960, 1, 512] - - [400, 90.882] - - - [82432, 512, 1, 512] - - [396, 87.494] - - - [82432, 40960, 1, 512] - - [475, 91.126] - - - [82432, 41472, 1, 512] - - [400, 93.083] - - - [82944, 512, 1, 512] - - [400, 88.109] - - - [82944, 40960, 1, 512] - - [471, 91.475] - - - [82944, 41472, 1, 512] - - [433, 93.526] - - - [83456, 512, 1, 512] - - [476, 87.675] - - - [83456, 41472, 1, 512] - - [433, 93.084] - - - [83456, 41984, 1, 512] - - [400, 92.378] - - - [83968, 512, 1, 512] - - [400, 88.774] - - - [83968, 41472, 1, 512] - - [419, 93.627] - - - [83968, 41984, 1, 512] - - [427, 92.535] - - - [84480, 512, 1, 512] - - [465, 88.311] - - - [84480, 41984, 1, 512] - - [474, 92.137] - - - [84480, 42496, 1, 512] - - [400, 93.178] - - - [84992, 512, 1, 512] - - [424, 87.553] - - - [84992, 41984, 1, 512] - - [436, 92.48] - - - [84992, 42496, 1, 512] - - [475, 92.952] - - - [85504, 512, 1, 512] - - [396, 87.138] - - - [85504, 42496, 1, 512] - - [433, 92.912] - - - [85504, 43008, 1, 512] - - [436, 92.007] - - - [86016, 512, 1, 512] - - [400, 88.009] - - - [86016, 42496, 1, 512] - - [419, 93.332] - - - [86016, 43008, 1, 512] - - [430, 91.587] - - - [86528, 512, 1, 512] - - [473, 87.577] - - - [86528, 43008, 1, 512] - - [430, 91.805] - - - [86528, 43520, 1, 512] - - [400, 93.045] - - - [87040, 512, 1, 512] - - [396, 88.398] - - - [87040, 43008, 1, 512] - - [436, 92.175] - - - [87040, 43520, 1, 512] - - [433, 93.422] - - - [87552, 512, 1, 512] - - [477, 88.093] - - - [87552, 43520, 1, 512] - - [400, 93.043] - - - [87552, 44032, 1, 512] - - [427, 91.666] - - - [88064, 512, 1, 512] - - [432, 87.432] - - - [88064, 43520, 1, 512] - - [433, 93.474] - - - [88064, 44032, 1, 512] - - [436, 92.245] - - - [88576, 512, 1, 512] - - [403, 87.178] - - - [88576, 44032, 1, 512] - - [402, 92.339] - - - [88576, 44544, 1, 512] - - [433, 92.922] - - - [89088, 512, 1, 512] - - [400, 88.144] - - - [89088, 44032, 1, 512] - - [430, 91.858] - - - [89088, 44544, 1, 512] - - [433, 93.327] - - - [89600, 512, 1, 512] - - [432, 87.356] - - - [89600, 44544, 1, 512] - - [433, 92.725] - - - [89600, 45056, 1, 512] - - [430, 91.167] - - - [90112, 512, 1, 512] - - [467, 87.286] - - - [90112, 44544, 1, 512] - - [419, 92.324] - - - [90112, 45056, 1, 512] - - [427, 89.787] - - - [90624, 512, 1, 512] - - [400, 88.013] - - - [90624, 45056, 1, 512] - - [475, 91.636] - - - [90624, 45568, 1, 512] - - [433, 92.822] - - - [91136, 512, 1, 512] - - [424, 88.765] - - - [91136, 45056, 1, 512] - - [436, 91.594] - - - [91136, 45568, 1, 512] - - [475, 92.775] - - - [91648, 512, 1, 512] - - [430, 86.74] - - - [91648, 45568, 1, 512] - - [433, 92.886] - - - [91648, 46080, 1, 512] - - [436, 91.864] - - - [92160, 512, 1, 512] - - [427, 88.192] - - - [92160, 45568, 1, 512] - - [475, 92.11] - - - [92160, 46080, 1, 512] - - [471, 91.906] - - - [92672, 512, 1, 512] - - [396, 87.541] - - - [92672, 46080, 1, 512] - - [436, 91.569] - - - [92672, 46592, 1, 512] - - [433, 92.805] - - - [93184, 512, 1, 512] - - [396, 88.72] - - - [93184, 46080, 1, 512] - - [430, 91.706] - - - [93184, 46592, 1, 512] - - [433, 92.345] - - - [93696, 512, 1, 512] - - [465, 87.813] - - - [93696, 46592, 1, 512] - - [433, 92.716] - - - [93696, 47104, 1, 512] - - [436, 91.452] - - - [94208, 512, 1, 512] - - [396, 88.609] - - - [94208, 46592, 1, 512] - - [430, 91.358] - - - [94208, 47104, 1, 512] - - [430, 91.443] - - - [94720, 512, 1, 512] - - [468, 88.387] - - - [94720, 47104, 1, 512] - - [430, 91.423] - - - [94720, 47616, 1, 512] - - [433, 92.862] - - - [95232, 512, 1, 512] - - [432, 88.044] - - - [95232, 47104, 1, 512] - - [430, 91.756] - - - [95232, 47616, 1, 512] - - [430, 91.696] - - - [95744, 512, 1, 512] - - [478, 87.524] - - - [95744, 47616, 1, 512] - - [433, 92.801] - - - [95744, 48128, 1, 512] - - [430, 91.537] - - - [96256, 512, 1, 512] - - [396, 88.588] - - - [96256, 47616, 1, 512] - - [430, 91.813] - - - [96256, 48128, 1, 512] - - [398, 91.23] - - - [96768, 512, 1, 512] - - [433, 87.754] - - - [96768, 48128, 1, 512] - - [430, 91.454] - - - [96768, 48640, 1, 512] - - [419, 92.644] - - - [97280, 512, 1, 512] - - [419, 88.592] - - - [97280, 48128, 1, 512] - - [430, 91.291] - - - [97280, 48640, 1, 512] - - [430, 91.467] - - - [97792, 512, 1, 512] - - [480, 88.276] - - - [97792, 48640, 1, 512] - - [419, 92.611] - - - [97792, 49152, 1, 512] - - [475, 89.684] - - - [98304, 512, 1, 512] - - [396, 84.749] - - - [98304, 48640, 1, 512] - - [368, 89.348] - - - [98304, 49152, 1, 512] - - [403, 86.687] - - - [98816, 512, 1, 512] - - [433, 87.469] - - - [98816, 49152, 1, 512] - - [398, 89.055] - - - [98816, 49664, 1, 512] - - [475, 92.219] - - - [99328, 512, 1, 512] - - [419, 88.486] - - - [99328, 49152, 1, 512] - - [398, 88.746] - - - [99328, 49664, 1, 512] - - [430, 91.701] - - - [99840, 512, 1, 512] - - [472, 88.177] - - - [99840, 49664, 1, 512] - - [474, 92.029] - - - [99840, 50176, 1, 512] - - [436, 91.439] - - - [100352, 512, 1, 512] - - [400, 88.862] - - - [100352, 49664, 1, 512] - - [430, 91.506] - - - [100352, 50176, 1, 512] - - [430, 91.272] - - - [100864, 512, 1, 512] - - [480, 88.185] - - - [100864, 50176, 1, 512] - - [436, 91.327] - - - [100864, 50688, 1, 512] - - [433, 92.499] - - - [101376, 512, 1, 512] - - [427, 88.896] - - - [101376, 50176, 1, 512] - - [430, 90.999] - - - [101376, 50688, 1, 512] - - [479, 91.314] - - - [101888, 512, 1, 512] - - [481, 88.576] - - - [101888, 50688, 1, 512] - - [479, 91.334] - - - [101888, 51200, 1, 512] - - [481, 90.906] - - - [102400, 512, 1, 512] - - [424, 88.28] - - - [102400, 50688, 1, 512] - - [479, 91.07] - - - [102400, 51200, 1, 512] - - [430, 90.793] - - - [102912, 512, 1, 512] - - [400, 88.028] - - - [102912, 51200, 1, 512] - - [481, 90.866] - - - [102912, 51712, 1, 512] - - [479, 91.316] - - - [103424, 512, 1, 512] - - [396, 88.633] - - - [103424, 51200, 1, 512] - - [482, 90.382] - - - [103424, 51712, 1, 512] - - [479, 91.298] - - - [103936, 512, 1, 512] - - [481, 87.936] - - - [103936, 51712, 1, 512] - - [475, 91.925] - - - [103936, 52224, 1, 512] - - [430, 90.931] - - - [104448, 512, 1, 512] - - [400, 88.853] - - - [104448, 51712, 1, 512] - - [479, 91.265] - - - [104448, 52224, 1, 512] - - [430, 90.609] - - - [104960, 512, 1, 512] - - [484, 88.422] - - - [104960, 52224, 1, 512] - - [481, 90.901] - - - [104960, 52736, 1, 512] - - [433, 91.377] - - - [105472, 512, 1, 512] - - [396, 89.038] - - - [105472, 52224, 1, 512] - - [482, 90.553] - - - [105472, 52736, 1, 512] - - [479, 91.293] - - - [105984, 512, 1, 512] - - [427, 87.764] - - - [105984, 52736, 1, 512] - - [468, 91.284] - - - [105984, 53248, 1, 512] - - [481, 90.402] - - - [106496, 512, 1, 512] - - [424, 87.507] - - - [106496, 52736, 1, 512] - - [430, 89.015] - - - [106496, 53248, 1, 512] - - [398, 88.143] - - - [107008, 512, 1, 512] - - [400, 88.001] - - - [107008, 53248, 1, 512] - - [482, 89.987] - - - [107008, 53760, 1, 512] - - [479, 91.298] - - - [107520, 512, 1, 512] - - [419, 88.812] - - - [107520, 53248, 1, 512] - - [482, 89.961] - - - [107520, 53760, 1, 512] - - [479, 91.28] - - - [108032, 512, 1, 512] - - [480, 88.425] - - - [108032, 53760, 1, 512] - - [479, 91.321] - - - [108032, 54272, 1, 512] - - [481, 90.938] - - - [108544, 512, 1, 512] - - [400, 88.961] - - - [108544, 53760, 1, 512] - - [479, 91.293] - - - [108544, 54272, 1, 512] - - [482, 90.329] - - - [109056, 512, 1, 512] - - [482, 88.826] - - - [109056, 54272, 1, 512] - - [481, 90.955] - - - [109056, 54784, 1, 512] - - [479, 91.318] - - - [109568, 512, 1, 512] - - [427, 88.742] - - - [109568, 54272, 1, 512] - - [485, 90.284] - - - [109568, 54784, 1, 512] - - [479, 91.218] - - - [110080, 512, 1, 512] - - [403, 88.435] - - - [110080, 54784, 1, 512] - - [479, 91.289] - - - [110080, 55296, 1, 512] - - [482, 90.341] - - - [110592, 512, 1, 512] - - [419, 88.973] - - - [110592, 54784, 1, 512] - - [479, 91.129] - - - [110592, 55296, 1, 512] - - [481, 89.821] - - - [111104, 512, 1, 512] - - [478, 88.213] - - - [111104, 55296, 1, 512] - - [481, 90.687] - - - [111104, 55808, 1, 512] - - [479, 91.283] - - - [111616, 512, 1, 512] - - [396, 88.979] - - - [111616, 55296, 1, 512] - - [482, 90.383] - - - [111616, 55808, 1, 512] - - [468, 91.191] - - - [112128, 512, 1, 512] - - [481, 88.639] - - - [112128, 55808, 1, 512] - - [479, 91.293] - - - [112128, 56320, 1, 512] - - [481, 90.908] - - - [112640, 512, 1, 512] - - [419, 89.062] - - - [112640, 55808, 1, 512] - - [479, 91.241] - - - [112640, 56320, 1, 512] - - [482, 90.27] - - - [113152, 512, 1, 512] - - [472, 88.303] - - - [113152, 56320, 1, 512] - - [482, 90.638] - - - [113152, 56832, 1, 512] - - [468, 91.238] - - - [113664, 512, 1, 512] - - [419, 88.925] - - - [113664, 56320, 1, 512] - - [482, 90.353] - - - [113664, 56832, 1, 512] - - [479, 91.236] - - - [114176, 512, 1, 512] - - [419, 88.216] - - - [114176, 56832, 1, 512] - - [479, 91.289] - - - [114176, 57344, 1, 512] - - [485, 89.421] - - - [114688, 512, 1, 512] - - [419, 83.784] - - - [114688, 56832, 1, 512] - - [403, 84.181] - - - [114688, 57344, 1, 512] - - [402, 82.979] - - - [115200, 512, 1, 512] - - [480, 88.648] - - - [115200, 57344, 1, 512] - - [482, 88.756] - - - [115200, 57856, 1, 512] - - [479, 91.285] - - - [115712, 512, 1, 512] - - [400, 88.81] - - - [115712, 57344, 1, 512] - - [482, 88.202] - - - [115712, 57856, 1, 512] - - [479, 91.256] - - - [116224, 512, 1, 512] - - [400, 88.439] - - - [116224, 57856, 1, 512] - - [468, 91.301] - - - [116224, 58368, 1, 512] - - [481, 90.722] - - - [116736, 512, 1, 512] - - [436, 88.937] - - - [116736, 57856, 1, 512] - - [479, 91.205] - - - [116736, 58368, 1, 512] - - [482, 90.22] - - - [117248, 512, 1, 512] - - [478, 88.393] - - - [117248, 58368, 1, 512] - - [481, 90.744] - - - [117248, 58880, 1, 512] - - [479, 91.276] - - - [117760, 512, 1, 512] - - [400, 89.177] - - - [117760, 58368, 1, 512] - - [482, 90.375] - - - [117760, 58880, 1, 512] - - [479, 91.261] - - - [118272, 512, 1, 512] - - [484, 88.453] - - - [118272, 58880, 1, 512] - - [468, 91.246] - - - [118272, 59392, 1, 512] - - [481, 90.904] - - - [118784, 512, 1, 512] - - [419, 89.028] - - - [118784, 58880, 1, 512] - - [479, 91.064] - - - [118784, 59392, 1, 512] - - [486, 89.516] - - - [119296, 512, 1, 512] - - [479, 88.927] - - - [119296, 59392, 1, 512] - - [481, 90.655] - - - [119296, 59904, 1, 512] - - [479, 91.226] - - - [119808, 512, 1, 512] - - [400, 88.732] - - - [119808, 59392, 1, 512] - - [482, 90.247] - - - [119808, 59904, 1, 512] - - [479, 91.228] - - - [120320, 512, 1, 512] - - [419, 88.341] - - - [120320, 59904, 1, 512] - - [479, 91.223] - - - [120320, 60416, 1, 512] - - [481, 90.714] - - - [120832, 512, 1, 512] - - [400, 88.888] - - - [120832, 59904, 1, 512] - - [479, 91.225] - - - [120832, 60416, 1, 512] - - [486, 90.241] - - - [121344, 512, 1, 512] - - [480, 88.482] - - - [121344, 60416, 1, 512] - - [481, 90.756] - - - [121344, 60928, 1, 512] - - [479, 91.262] - - - [121856, 512, 1, 512] - - [400, 88.763] - - - [121856, 60416, 1, 512] - - [482, 90.266] - - - [121856, 60928, 1, 512] - - [479, 91.222] - - - [122368, 512, 1, 512] - - [468, 88.758] - - - [122368, 60928, 1, 512] - - [479, 91.237] - - - [122368, 61440, 1, 512] - - [485, 90.324] - - - [122880, 512, 1, 512] - - [396, 87.586] - - - [122880, 60928, 1, 512] - - [487, 87.702] - - - [122880, 61440, 1, 512] - - [485, 86.814] - - - [123392, 512, 1, 512] - - [427, 88.222] - - - [123392, 61440, 1, 512] - - [481, 90.487] - - - [123392, 61952, 1, 512] - - [479, 91.234] - - - [123904, 512, 1, 512] - - [400, 89.209] - - - [123904, 61440, 1, 512] - - [486, 89.908] - - - [123904, 61952, 1, 512] - - [479, 91.189] - - - [124416, 512, 1, 512] - - [427, 88.638] - - - [124416, 61952, 1, 512] - - [479, 91.245] - - - [124416, 62464, 1, 512] - - [481, 90.814] - - - [124928, 512, 1, 512] - - [400, 89.121] - - - [124928, 61952, 1, 512] - - [479, 91.211] - - - [124928, 62464, 1, 512] - - [482, 90.238] - - - [125440, 512, 1, 512] - - [468, 88.725] - - - [125440, 62464, 1, 512] - - [481, 90.751] - - - [125440, 62976, 1, 512] - - [479, 91.243] - - - [125952, 512, 1, 512] - - [480, 88.779] - - - [125952, 62464, 1, 512] - - [482, 90.251] - - - [125952, 62976, 1, 512] - - [479, 91.214] - - - [126464, 512, 1, 512] - - [480, 89.097] - - - [126464, 62976, 1, 512] - - [479, 91.242] - - - [126464, 63488, 1, 512] - - [481, 90.586] - - - [126976, 512, 1, 512] - - [396, 88.859] - - - [126976, 62976, 1, 512] - - [479, 91.026] - - - [126976, 63488, 1, 512] - - [482, 89.555] - - - [127488, 512, 1, 512] - - [396, 88.375] - - - [127488, 63488, 1, 512] - - [482, 90.491] - - - [127488, 64000, 1, 512] - - [468, 91.215] - - - [128000, 512, 1, 512] - - [419, 88.991] - - - [128000, 63488, 1, 512] - - [482, 90.261] - - - [128000, 64000, 1, 512] - - [479, 91.193] - - - [4096, 1537, 1, 512] - - [456, 71.839] - - - [4096, 2049, 1, 512] - - [346, 73.546] - - - [4608, 2049, 1, 512] - - [460, 71.194] - - - [5120, 2049, 1, 512] - - [447, 74.526] - - - [5120, 2561, 1, 512] - - [461, 79.954] - - - [5632, 2561, 1, 512] - - [462, 80.796] - - - [6144, 2561, 1, 512] - - [462, 81.039] - - - [6144, 3073, 1, 512] - - [428, 80.48] - - - [6656, 3073, 1, 512] - - [401, 80.508] - - - [7168, 3073, 1, 512] - - [350, 80.597] - - - [7168, 3585, 1, 512] - - [464, 83.124] - - - [7680, 3585, 1, 512] - - [463, 83.831] - - - [8192, 3585, 1, 512] - - [465, 84.408] - - - [8192, 4097, 1, 512] - - [400, 82.235] - - - [8704, 4097, 1, 512] - - [466, 82.752] - - - [9216, 4097, 1, 512] - - [400, 83.385] - - - [9216, 4609, 1, 512] - - [433, 85.135] - - - [9728, 4609, 1, 512] - - [465, 86.124] - - - [10240, 4609, 1, 512] - - [400, 86.709] - - - [10240, 5121, 1, 512] - - [400, 85.605] - - - [10752, 5121, 1, 512] - - [400, 85.27] - - - [11264, 5121, 1, 512] - - [400, 86.045] - - - [11264, 5633, 1, 512] - - [400, 88.3] - - - [11776, 5633, 1, 512] - - [465, 87.66] - - - [12288, 5633, 1, 512] - - [419, 88.184] - - - [12288, 6145, 1, 512] - - [368, 86.632] - - - [12800, 6145, 1, 512] - - [400, 86.611] - - - [13312, 6145, 1, 512] - - [403, 86.931] - - - [13312, 6657, 1, 512] - - [400, 88.893] - - - [13824, 6657, 1, 512] - - [400, 88.205] - - - [14336, 6657, 1, 512] - - [400, 89.331] - - - [14336, 7169, 1, 512] - - [467, 86.882] - - - [14848, 7169, 1, 512] - - [400, 87.71] - - - [15360, 7169, 1, 512] - - [400, 87.86] - - - [15360, 7681, 1, 512] - - [400, 90.205] - - - [15872, 7681, 1, 512] - - [468, 89.271] - - - [16384, 7681, 1, 512] - - [419, 90.97] - - - [16384, 8193, 1, 512] - - [400, 88.904] - - - [16896, 8193, 1, 512] - - [400, 88.772] - - - [17408, 8193, 1, 512] - - [400, 89.274] - - - [17408, 8705, 1, 512] - - [400, 90.992] - - - [17920, 8705, 1, 512] - - [403, 90.678] - - - [18432, 8705, 1, 512] - - [400, 91.273] - - - [18432, 9217, 1, 512] - - [403, 88.781] - - - [18944, 9217, 1, 512] - - [400, 89.134] - - - [19456, 9217, 1, 512] - - [403, 88.745] - - - [19456, 9729, 1, 512] - - [400, 90.978] - - - [19968, 9729, 1, 512] - - [400, 90.266] - - - [20480, 9729, 1, 512] - - [400, 91.138] - - - [20480, 10241, 1, 512] - - [400, 88.863] - - - [20992, 10241, 1, 512] - - [368, 88.854] - - - [21504, 10241, 1, 512] - - [428, 88.293] - - - [21504, 10753, 1, 512] - - [400, 91.091] - - - [22016, 10753, 1, 512] - - [400, 90.448] - - - [22528, 10753, 1, 512] - - [419, 91.216] - - - [22528, 11265, 1, 512] - - [400, 88.429] - - - [23040, 11265, 1, 512] - - [368, 88.578] - - - [23552, 11265, 1, 512] - - [434, 88.164] - - - [23552, 11777, 1, 512] - - [400, 91.138] - - - [24064, 11777, 1, 512] - - [400, 90.72] - - - [24576, 11777, 1, 512] - - [400, 91.116] - - - [24576, 12289, 1, 512] - - [400, 89.211] - - - [25088, 12289, 1, 512] - - [400, 88.664] - - - [25600, 12289, 1, 512] - - [428, 88.041] - - - [25600, 12801, 1, 512] - - [419, 91.174] - - - [26112, 12801, 1, 512] - - [400, 90.789] - - - [26624, 12801, 1, 512] - - [419, 91.364] - - - [26624, 13313, 1, 512] - - [467, 88.426] - - - [27136, 13313, 1, 512] - - [400, 89.531] - - - [27648, 13313, 1, 512] - - [403, 88.82] - - - [27648, 13825, 1, 512] - - [400, 91.58] - - - [28160, 13825, 1, 512] - - [400, 91.046] - - - [28672, 13825, 1, 512] - - [400, 91.488] - - - [28672, 14337, 1, 512] - - [401, 88.425] - - - [29184, 14337, 1, 512] - - [403, 89.499] - - - [29696, 14337, 1, 512] - - [400, 89.646] - - - [29696, 14849, 1, 512] - - [419, 91.824] - - - [30208, 14849, 1, 512] - - [400, 91.298] - - - [30720, 14849, 1, 512] - - [419, 91.892] - - - [30720, 15361, 1, 512] - - [428, 88.531] - - - [31232, 15361, 1, 512] - - [400, 89.424] - - - [31744, 15361, 1, 512] - - [401, 88.704] - - - [31744, 15873, 1, 512] - - [400, 92.089] - - - [32256, 15873, 1, 512] - - [400, 91.671] - - - [32768, 15873, 1, 512] - - [419, 92.132] - - - [32768, 16385, 1, 512] - - [469, 87.258] - - - [33280, 16385, 1, 512] - - [400, 89.643] - - - [33792, 16385, 1, 512] - - [428, 87.619] - - - [33792, 16897, 1, 512] - - [419, 92.091] - - - [34304, 16897, 1, 512] - - [433, 91.568] - - - [34816, 16897, 1, 512] - - [419, 92.142] - - - [34816, 17409, 1, 512] - - [467, 89.095] - - - [35328, 17409, 1, 512] - - [400, 89.971] - - - [35840, 17409, 1, 512] - - [401, 88.922] - - - [35840, 17921, 1, 512] - - [400, 92.279] - - - [36352, 17921, 1, 512] - - [419, 91.663] - - - [36864, 17921, 1, 512] - - [419, 92.039] - - - [36864, 18433, 1, 512] - - [403, 89.783] - - - [37376, 18433, 1, 512] - - [400, 90.515] - - - [37888, 18433, 1, 512] - - [467, 88.945] - - - [37888, 18945, 1, 512] - - [400, 92.412] - - - [38400, 18945, 1, 512] - - [400, 91.855] - - - [38912, 18945, 1, 512] - - [400, 92.34] - - - [38912, 19457, 1, 512] - - [467, 89.012] - - - [39424, 19457, 1, 512] - - [400, 90.343] - - - [39936, 19457, 1, 512] - - [434, 89.164] - - - [39936, 19969, 1, 512] - - [419, 92.477] - - - [40448, 19969, 1, 512] - - [400, 91.996] - - - [40960, 19969, 1, 512] - - [419, 91.92] - - - [40960, 20481, 1, 512] - - [401, 88.014] - - - [41472, 20481, 1, 512] - - [368, 89.728] - - - [41984, 20481, 1, 512] - - [470, 88.273] - - - [41984, 20993, 1, 512] - - [419, 92.319] - - - [42496, 20993, 1, 512] - - [400, 91.791] - - - [43008, 20993, 1, 512] - - [400, 92.638] - - - [43008, 21505, 1, 512] - - [428, 89.085] - - - [43520, 21505, 1, 512] - - [467, 89.693] - - - [44032, 21505, 1, 512] - - [434, 89.237] - - - [44032, 22017, 1, 512] - - [419, 92.605] - - - [44544, 22017, 1, 512] - - [400, 91.969] - - - [45056, 22017, 1, 512] - - [419, 92.6] - - - [45056, 22529, 1, 512] - - [401, 88.547] - - - [45568, 22529, 1, 512] - - [428, 89.709] - - - [46080, 22529, 1, 512] - - [434, 89.139] - - - [46080, 23041, 1, 512] - - [400, 92.662] - - - [46592, 23041, 1, 512] - - [400, 92.198] - - - [47104, 23041, 1, 512] - - [419, 92.773] - - - [47104, 23553, 1, 512] - - [434, 89.14] - - - [47616, 23553, 1, 512] - - [467, 89.731] - - - [48128, 23553, 1, 512] - - [428, 89.255] - - - [48128, 24065, 1, 512] - - [419, 92.802] - - - [48640, 24065, 1, 512] - - [400, 92.292] - - - [49152, 24065, 1, 512] - - [400, 91.57] - - - [49152, 24577, 1, 512] - - [428, 87.402] - - - [49664, 24577, 1, 512] - - [401, 89.119] - - - [50176, 24577, 1, 512] - - [403, 88.619] - - - [50176, 25089, 1, 512] - - [419, 92.722] - - - [50688, 25089, 1, 512] - - [400, 92.528] - - - [51200, 25089, 1, 512] - - [419, 92.84] - - - [51200, 25601, 1, 512] - - [428, 89.44] - - - [51712, 25601, 1, 512] - - [434, 89.843] - - - [52224, 25601, 1, 512] - - [467, 89.407] - - - [52224, 26113, 1, 512] - - [419, 92.82] - - - [52736, 26113, 1, 512] - - [400, 92.389] - - - [53248, 26113, 1, 512] - - [400, 92.643] - - - [53248, 26625, 1, 512] - - [435, 88.478] - - - [53760, 26625, 1, 512] - - [400, 90.518] - - - [54272, 26625, 1, 512] - - [401, 88.733] - - - [54272, 27137, 1, 512] - - [419, 92.79] - - - [54784, 27137, 1, 512] - - [400, 92.359] - - - [55296, 27137, 1, 512] - - [419, 92.891] - - - [55296, 27649, 1, 512] - - [467, 89.381] - - - [55808, 27649, 1, 512] - - [400, 90.523] - - - [56320, 27649, 1, 512] - - [401, 89.235] - - - [56320, 28161, 1, 512] - - [400, 92.785] - - - [56832, 28161, 1, 512] - - [400, 92.08] - - - [57344, 28161, 1, 512] - - [400, 91.675] - - - [57344, 28673, 1, 512] - - [428, 87.774] - - - [57856, 28673, 1, 512] - - [401, 89.327] - - - [58368, 28673, 1, 512] - - [471, 86.002] - - - [58368, 29185, 1, 512] - - [419, 92.825] - - - [58880, 29185, 1, 512] - - [400, 92.272] - - - [59392, 29185, 1, 512] - - [419, 92.883] - - - [59392, 29697, 1, 512] - - [401, 88.962] - - - [59904, 29697, 1, 512] - - [400, 90.464] - - - [60416, 29697, 1, 512] - - [467, 88.978] - - - [60416, 30209, 1, 512] - - [400, 92.798] - - - [60928, 30209, 1, 512] - - [400, 92.179] - - - [61440, 30209, 1, 512] - - [419, 92.782] - - - [61440, 30721, 1, 512] - - [469, 88.467] - - - [61952, 30721, 1, 512] - - [400, 90.638] - - - [62464, 30721, 1, 512] - - [401, 89.048] - - - [62464, 31233, 1, 512] - - [419, 92.892] - - - [62976, 31233, 1, 512] - - [400, 92.119] - - - [63488, 31233, 1, 512] - - [419, 92.915] - - - [63488, 31745, 1, 512] - - [428, 89.217] - - - [64000, 31745, 1, 512] - - [400, 90.576] - - - [64512, 31745, 1, 512] - - [428, 89.205] - - - [64512, 32257, 1, 512] - - [419, 92.811] - - - [65024, 32257, 1, 512] - - [400, 92.214] - - - [65536, 32257, 1, 512] - - [400, 91.415] - - - [65536, 32769, 1, 512] - - [402, 88.499] - - - [66048, 32769, 1, 512] - - [426, 85.277] - - - [66560, 32769, 1, 512] - - [427, 84.944] - - - [66560, 33281, 1, 512] - - [419, 92.737] - - - [67072, 33281, 1, 512] - - [419, 92.099] - - - [67584, 33281, 1, 512] - - [400, 92.826] - - - [67584, 33793, 1, 512] - - [428, 89.262] - - - [68096, 33793, 1, 512] - - [400, 90.745] - - - [68608, 33793, 1, 512] - - [401, 89.002] - - - [68608, 34305, 1, 512] - - [419, 92.844] - - - [69120, 34305, 1, 512] - - [400, 92.172] - - - [69632, 34305, 1, 512] - - [419, 92.759] - - - [69632, 34817, 1, 512] - - [428, 89.017] - - - [70144, 34817, 1, 512] - - [368, 90.209] - - - [70656, 34817, 1, 512] - - [401, 89.147] - - - [70656, 35329, 1, 512] - - [419, 92.812] - - - [71168, 35329, 1, 512] - - [400, 92.064] - - - [71680, 35329, 1, 512] - - [419, 92.809] - - - [71680, 35841, 1, 512] - - [428, 89.369] - - - [72192, 35841, 1, 512] - - [434, 89.954] - - - [72704, 35841, 1, 512] - - [401, 89.035] - - - [72704, 36353, 1, 512] - - [419, 92.779] - - - [73216, 36353, 1, 512] - - [400, 92.194] - - - [73728, 36353, 1, 512] - - [419, 91.445] - - - [73728, 36865, 1, 512] - - [403, 88.642] - - - [74240, 36865, 1, 512] - - [400, 89.188] - - - [74752, 36865, 1, 512] - - [428, 88.904] - - - [74752, 37377, 1, 512] - - [419, 92.664] - - - [75264, 37377, 1, 512] - - [403, 91.933] - - - [75776, 37377, 1, 512] - - [419, 92.804] - - - [75776, 37889, 1, 512] - - [434, 89.37] - - - [76288, 37889, 1, 512] - - [400, 89.886] - - - [76800, 37889, 1, 512] - - [428, 89.129] - - - [76800, 38401, 1, 512] - - [419, 92.661] - - - [77312, 38401, 1, 512] - - [400, 91.955] - - - [77824, 38401, 1, 512] - - [419, 92.651] - - - [77824, 38913, 1, 512] - - [401, 88.742] - - - [78336, 38913, 1, 512] - - [434, 89.733] - - - [78848, 38913, 1, 512] - - [428, 89.053] - - - [78848, 39425, 1, 512] - - [419, 92.581] - - - [79360, 39425, 1, 512] - - [433, 91.852] - - - [79872, 39425, 1, 512] - - [419, 92.691] - - - [79872, 39937, 1, 512] - - [428, 89.149] - - - [80384, 39937, 1, 512] - - [401, 89.875] - - - [80896, 39937, 1, 512] - - [428, 88.799] - - - [80896, 40449, 1, 512] - - [419, 92.597] - - - [81408, 40449, 1, 512] - - [403, 92.001] - - - [81920, 40449, 1, 512] - - [400, 89.596] - - - [81920, 40961, 1, 512] - - [368, 88.427] - - - [82432, 40961, 1, 512] - - [431, 87.975] - - - [82944, 40961, 1, 512] - - [471, 85.404] - - - [82944, 41473, 1, 512] - - [419, 92.475] - - - [83456, 41473, 1, 512] - - [403, 91.836] - - - [83968, 41473, 1, 512] - - [419, 92.607] - - - [83968, 41985, 1, 512] - - [428, 88.995] - - - [84480, 41985, 1, 512] - - [434, 89.701] - - - [84992, 41985, 1, 512] - - [428, 88.944] - - - [84992, 42497, 1, 512] - - [433, 92.418] - - - [85504, 42497, 1, 512] - - [433, 91.694] - - - [86016, 42497, 1, 512] - - [419, 92.28] - - - [86016, 43009, 1, 512] - - [428, 87.969] - - - [86528, 43009, 1, 512] - - [428, 89.279] - - - [87040, 43009, 1, 512] - - [401, 88.671] - - - [87040, 43521, 1, 512] - - [419, 92.37] - - - [87552, 43521, 1, 512] - - [403, 91.739] - - - [88064, 43521, 1, 512] - - [419, 92.428] - - - [88064, 44033, 1, 512] - - [428, 88.439] - - - [88576, 44033, 1, 512] - - [401, 89.435] - - - [89088, 44033, 1, 512] - - [428, 88.126] - - - [89088, 44545, 1, 512] - - [474, 91.871] - - - [89600, 44545, 1, 512] - - [403, 91.713] - - - [90112, 44545, 1, 512] - - [419, 90.886] - - - [90112, 45057, 1, 512] - - [471, 84.357] - - - [90624, 45057, 1, 512] - - [428, 88.305] - - - [91136, 45057, 1, 512] - - [401, 87.41] - - - [91136, 45569, 1, 512] - - [433, 92.177] - - - [91648, 45569, 1, 512] - - [433, 91.635] - - - [92160, 45569, 1, 512] - - [433, 92.32] - - - [92160, 46081, 1, 512] - - [401, 88.6] - - - [92672, 46081, 1, 512] - - [401, 89.023] - - - [93184, 46081, 1, 512] - - [469, 87.289] - - - [93184, 46593, 1, 512] - - [474, 91.668] - - - [93696, 46593, 1, 512] - - [400, 91.443] - - - [94208, 46593, 1, 512] - - [474, 91.742] - - - [94208, 47105, 1, 512] - - [471, 85.878] - - - [94720, 47105, 1, 512] - - [401, 88.803] - - - [95232, 47105, 1, 512] - - [436, 86.083] - - - [95232, 47617, 1, 512] - - [474, 91.769] - - - [95744, 47617, 1, 512] - - [433, 91.539] - - - [96256, 47617, 1, 512] - - [475, 91.876] - - - [96256, 48129, 1, 512] - - [428, 88.105] - - - [96768, 48129, 1, 512] - - [401, 88.471] - - - [97280, 48129, 1, 512] - - [428, 87.888] - - - [97280, 48641, 1, 512] - - [479, 90.792] - - - [97792, 48641, 1, 512] - - [419, 91.248] - - - [98304, 48641, 1, 512] - - [400, 88.251] - - - [98304, 49153, 1, 512] - - [402, 84.436] - - - [98816, 49153, 1, 512] - - [471, 83.519] - - - [99328, 49153, 1, 512] - - [471, 83.897] - - - [99328, 49665, 1, 512] - - [475, 90.831] - - - [99840, 49665, 1, 512] - - [433, 91.438] - - - [100352, 49665, 1, 512] - - [479, 90.719] - - - [100352, 50177, 1, 512] - - [428, 87.027] - - - [100864, 50177, 1, 512] - - [401, 87.702] - - - [101376, 50177, 1, 512] - - [401, 86.842] - - - [101376, 50689, 1, 512] - - [479, 90.8] - - - [101888, 50689, 1, 512] - - [433, 91.28] - - - [102400, 50689, 1, 512] - - [479, 90.356] - - - [102400, 51201, 1, 512] - - [436, 85.491] - - - [102912, 51201, 1, 512] - - [401, 87.554] - - - [103424, 51201, 1, 512] - - [471, 85.608] - - - [103424, 51713, 1, 512] - - [479, 90.769] - - - [103936, 51713, 1, 512] - - [433, 91.189] - - - [104448, 51713, 1, 512] - - [483, 90.489] - - - [104448, 52225, 1, 512] - - [471, 85.736] - - - [104960, 52225, 1, 512] - - [401, 87.557] - - - [105472, 52225, 1, 512] - - [436, 85.505] - - - [105472, 52737, 1, 512] - - [479, 90.706] - - - [105984, 52737, 1, 512] - - [479, 90.75] - - - [106496, 52737, 1, 512] - - [402, 88.805] - - - [106496, 53249, 1, 512] - - [426, 84.089] - - - [107008, 53249, 1, 512] - - [471, 84.808] - - - [107520, 53249, 1, 512] - - [471, 85.345] - - - [107520, 53761, 1, 512] - - [479, 90.734] - - - [108032, 53761, 1, 512] - - [479, 90.865] - - - [108544, 53761, 1, 512] - - [479, 90.651] - - - [108544, 54273, 1, 512] - - [436, 85.361] - - - [109056, 54273, 1, 512] - - [485, 87.03] - - - [109568, 54273, 1, 512] - - [436, 85.218] - - - [109568, 54785, 1, 512] - - [479, 90.692] - - - [110080, 54785, 1, 512] - - [479, 90.838] - - - [110592, 54785, 1, 512] - - [479, 90.355] - - - [110592, 55297, 1, 512] - - [471, 84.982] - - - [111104, 55297, 1, 512] - - [351, 85.341] - - - [111616, 55297, 1, 512] - - [471, 85.098] - - - [111616, 55809, 1, 512] - - [479, 90.695] - - - [112128, 55809, 1, 512] - - [479, 90.771] - - - [112640, 55809, 1, 512] - - [479, 90.514] - - - [112640, 56321, 1, 512] - - [471, 84.954] - - - [113152, 56321, 1, 512] - - [485, 86.354] - - - [113664, 56321, 1, 512] - - [482, 85.326] - - - [113664, 56833, 1, 512] - - [479, 90.723] - - - [114176, 56833, 1, 512] - - [479, 90.834] - - - [114688, 56833, 1, 512] - - [400, 83.086] - - - [114688, 57345, 1, 512] - - [402, 81.132] - - - [115200, 57345, 1, 512] - - [471, 83.764] - - - [115712, 57345, 1, 512] - - [471, 84.064] - - - [115712, 57857, 1, 512] - - [479, 90.664] - - - [116224, 57857, 1, 512] - - [479, 90.843] - - - [116736, 57857, 1, 512] - - [479, 90.659] - - - [116736, 58369, 1, 512] - - [471, 84.791] - - - [117248, 58369, 1, 512] - - [482, 84.993] - - - [117760, 58369, 1, 512] - - [482, 85.091] - - - [117760, 58881, 1, 512] - - [479, 90.715] - - - [118272, 58881, 1, 512] - - [479, 90.83] - - - [118784, 58881, 1, 512] - - [479, 90.284] - - - [118784, 59393, 1, 512] - - [471, 84.315] - - - [119296, 59393, 1, 512] - - [482, 84.813] - - - [119808, 59393, 1, 512] - - [486, 84.83] - - - [119808, 59905, 1, 512] - - [479, 90.677] - - - [120320, 59905, 1, 512] - - [479, 90.746] - - - [120832, 59905, 1, 512] - - [479, 90.678] - - - [120832, 60417, 1, 512] - - [486, 84.893] - - - [121344, 60417, 1, 512] - - [486, 85.333] - - - [121856, 60417, 1, 512] - - [482, 85.035] - - - [121856, 60929, 1, 512] - - [479, 90.653] - - - [122368, 60929, 1, 512] - - [479, 90.767] - - - [122880, 60929, 1, 512] - - [487, 85.297] - - - [122880, 61441, 1, 512] - - [471, 83.057] - - - [123392, 61441, 1, 512] - - [471, 83.823] - - - [123904, 61441, 1, 512] - - [471, 84.156] - - - [123904, 61953, 1, 512] - - [479, 90.644] - - - [124416, 61953, 1, 512] - - [479, 90.785] - - - [124928, 61953, 1, 512] - - [479, 90.688] - - - [124928, 62465, 1, 512] - - [482, 84.834] - - - [125440, 62465, 1, 512] - - [482, 84.899] - - - [125952, 62465, 1, 512] - - [486, 84.952] - - - [125952, 62977, 1, 512] - - [479, 90.662] - - - [126464, 62977, 1, 512] - - [479, 90.758] - - - [126976, 62977, 1, 512] - - [479, 90.325] - - - [126976, 63489, 1, 512] - - [471, 83.703] - - - [127488, 63489, 1, 512] - - [482, 84.879] - - - [128000, 63489, 1, 512] - - [486, 84.895] - - - [3584, 6657, 1, 512] - - [350, 85.27] - - - [3584, 6145, 1, 512] - - [350, 83.428] - - - [3072, 5633, 1, 512] - - [350, 85.079] - - - [3072, 5121, 1, 512] - - [350, 84.102] - - - [2560, 4609, 1, 512] - - [350, 81.413] - - - [2560, 4097, 1, 512] - - [350, 81.209] - - - [2048, 3585, 1, 512] - - [412, 73.4] - - - [2048, 3073, 1, 512] - - [411, 71.781] - - - [1536, 2561, 1, 512] - - [409, 71.299] - - - [1536, 2049, 1, 512] - - [342, 65.956] - - - [1024, 1537, 1, 512] - - [406, 58.384] - - - [1024, 1025, 1, 512] - - [336, 43.557] - - - [512, 513, 1, 512] - - [331, 21.043] - - - [512, 1, 1, 512] - - [441, 0.033] - - - [7040, 4096, 1, 512] - - [16, 90.075] - - - [8448, 3840, 1, 512] - - [20, 90.593] - - - [7680, 4224, 1, 512] - - [20, 90.538] - - - [1024, 513, 1, 512] - - [405, 34.73] - - - [1536, 513, 1, 512] - - [454, 45.566] - - - [2048, 513, 1, 512] - - [455, 46.087] - - - [2048, 1025, 1, 512] - - [355, 55.418] - - - [2560, 1025, 1, 512] - - [411, 64.823] - - - [3072, 1025, 1, 512] - - [339, 62.097] - - - [3072, 1537, 1, 512] - - [457, 66.837] - - - [3584, 1537, 1, 512] - - [458, 65.449] - - - [1024, 1, 1, 512] - - [488, 0.083] - - - [1152, 385, 1, 384] - - [335, 25.842] - - - [1536, 385, 1, 384] - - [406, 31.295] - - - [1536, 769, 1, 384] - - [345, 45.831] - - - [1920, 769, 1, 384] - - [442, 51.617] - - - [2304, 769, 1, 384] - - [334, 50.263] - - - [3456, 1536, 1, 384] - - [382, 72.964] - - - [3840, 1536, 1, 384] - - [338, 70.006] - - - [3840, 1537, 1, 384] - - [338, 66.494] - - - [3840, 1920, 1, 384] - - [382, 72.328] - - - [4224, 1920, 1, 384] - - [383, 77.41] - - - [4224, 1921, 1, 384] - - [445, 70.99] - - - [4224, 2304, 1, 384] - - [370, 76.309] - - - [4608, 1920, 1, 384] - - [385, 77.995] - - - [4608, 1921, 1, 384] - - [355, 72.599] - - - [4608, 2304, 1, 384] - - [446, 79.294] - - - [4608, 2305, 1, 384] - - [447, 70.803] - - - [4992, 2304, 1, 384] - - [448, 78.114] - - - [4992, 2305, 1, 384] - - [355, 75.055] - - - [5376, 2304, 1, 384] - - [449, 81.13] - - - [5376, 2305, 1, 384] - - [383, 74.072] - - - [5376, 2689, 1, 384] - - [354, 74.765] - - - [5760, 2689, 1, 384] - - [378, 79.323] - - - [6144, 2689, 1, 384] - - [350, 77.648] - - - [6144, 3073, 1, 384] - - [450, 77.718] - - - [6528, 3073, 1, 384] - - [350, 77.092] - - - [6528, 3456, 1, 384] - - [447, 84.688] - - - [6912, 3073, 1, 384] - - [351, 76.338] - - - [6912, 3456, 1, 384] - - [447, 84.657] - - - [6912, 3457, 1, 384] - - [350, 81.719] - - - [7296, 3456, 1, 384] - - [353, 86.904] - - - [7296, 3457, 1, 384] - - [350, 80.963] - - - [7296, 3840, 1, 384] - - [392, 85.978] - - - [7680, 3456, 1, 384] - - [350, 86.029] - - - [7680, 3457, 1, 384] - - [350, 81.059] - - - [7680, 3840, 1, 384] - - [392, 85.844] - - - [7680, 3841, 1, 384] - - [350, 84.237] - - - [8064, 3840, 1, 384] - - [447, 86.207] - - - [8064, 3841, 1, 384] - - [350, 83.724] - - - [8064, 4224, 1, 384] - - [350, 88.4] - - - [8448, 3840, 1, 384] - - [353, 88.253] - - - [8448, 3841, 1, 384] - - [350, 84.707] - - - [8448, 4224, 1, 384] - - [350, 88.495] - - - [8448, 4225, 1, 384] - - [350, 84.242] - - - [8832, 4224, 1, 384] - - [350, 88.111] - - - [8832, 4225, 1, 384] - - [350, 84.026] - - - [8832, 4608, 1, 384] - - [353, 88.499] - - - [9216, 4224, 1, 384] - - [354, 88.106] - - - [9216, 4225, 1, 384] - - [350, 83.965] - - - [9216, 4608, 1, 384] - - [392, 88.713] - - - [9216, 4609, 1, 384] - - [357, 84.074] - - - [9600, 4608, 1, 384] - - [392, 89.086] - - - [9600, 4609, 1, 384] - - [350, 83.994] - - - [9600, 4992, 1, 384] - - [353, 89.375] - - - [9984, 4608, 1, 384] - - [392, 89.104] - - - [9984, 4609, 1, 384] - - [357, 84.4] - - - [9984, 4992, 1, 384] - - [392, 89.739] - - - [9984, 4993, 1, 384] - - [350, 85.913] - - - [10368, 4992, 1, 384] - - [354, 89.826] - - - [10368, 4993, 1, 384] - - [350, 86.288] - - - [10368, 5376, 1, 384] - - [353, 90.339] - - - [10752, 4992, 1, 384] - - [353, 90.283] - - - [10752, 4993, 1, 384] - - [351, 86.662] - - - [10752, 5376, 1, 384] - - [350, 89.04] - - - [10752, 5377, 1, 384] - - [350, 87.141] - - - [11136, 5376, 1, 384] - - [350, 89.666] - - - [11136, 5377, 1, 384] - - [350, 86.733] - - - [11136, 5760, 1, 384] - - [392, 90.553] - - - [11520, 5376, 1, 384] - - [350, 89.94] - - - [11520, 5377, 1, 384] - - [350, 86.466] - - - [11520, 5760, 1, 384] - - [350, 90.895] - - - [11520, 5761, 1, 384] - - [350, 87.71] - - - [11904, 5760, 1, 384] - - [350, 90.02] - - - [11904, 5761, 1, 384] - - [350, 87.481] - - - [11904, 6144, 1, 384] - - [353, 90.696] - - - [12288, 5760, 1, 384] - - [354, 90.587] - - - [12288, 5761, 1, 384] - - [350, 87.223] - - - [12288, 6144, 1, 384] - - [353, 91.044] - - - [12288, 6145, 1, 384] - - [393, 83.886] - - - [12672, 6144, 1, 384] - - [396, 90.26] - - - [12672, 6145, 1, 384] - - [350, 83.927] - - - [12672, 6528, 1, 384] - - [392, 91.559] - - - [13056, 6144, 1, 384] - - [396, 90.527] - - - [13056, 6145, 1, 384] - - [393, 84.102] - - - [13056, 6528, 1, 384] - - [353, 91.088] - - - [13056, 6529, 1, 384] - - [350, 87.957] - - - [13440, 6528, 1, 384] - - [392, 91.612] - - - [13440, 6529, 1, 384] - - [353, 88.327] - - - [13440, 6912, 1, 384] - - [392, 91.726] - - - [13824, 6529, 1, 384] - - [350, 87.915] - - - [13824, 6912, 1, 384] - - [354, 91.148] - - - [13824, 6913, 1, 384] - - [353, 88.788] - - - [14208, 6913, 1, 384] - - [350, 88.641] - - - [15744, 7680, 1, 384] - - [392, 91.966] - - - [16128, 7680, 1, 384] - - [353, 92.502] - - - [16128, 7681, 1, 384] - - [350, 89.081] - - - [16128, 8064, 1, 384] - - [353, 92.163] - - - [16512, 8064, 1, 384] - - [353, 92.576] - - - [16512, 8065, 1, 384] - - [350, 89.059] - - - [16512, 8448, 1, 384] - - [392, 92.29] - - - [16896, 8064, 1, 384] - - [353, 92.728] - - - [16896, 8065, 1, 384] - - [353, 89.808] - - - [16896, 8448, 1, 384] - - [392, 92.243] - - - [16896, 8449, 1, 384] - - [350, 89.753] - - - [17280, 8448, 1, 384] - - [353, 92.603] - - - [17280, 8449, 1, 384] - - [350, 89.245] - - - [17280, 8832, 1, 384] - - [350, 92.37] - - - [17664, 8448, 1, 384] - - [353, 92.313] - - - [17664, 8449, 1, 384] - - [350, 89.38] - - - [17664, 8832, 1, 384] - - [350, 92.084] - - - [17664, 8833, 1, 384] - - [353, 89.083] - - - [18048, 8832, 1, 384] - - [354, 92.239] - - - [18048, 8833, 1, 384] - - [353, 88.905] - - - [18432, 8832, 1, 384] - - [354, 92.279] - - - [18432, 8833, 1, 384] - - [350, 89.533] - - - [18432, 9217, 1, 384] - - [451, 86.121] - - - [18816, 9217, 1, 384] - - [368, 85.981] - - - [18816, 9600, 1, 384] - - [350, 91.918] - - - [19200, 9217, 1, 384] - - [451, 86.267] - - - [19200, 9600, 1, 384] - - [350, 91.858] - - - [19200, 9601, 1, 384] - - [351, 88.014] - - - [19584, 9600, 1, 384] - - [350, 91.939] - - - [19584, 9601, 1, 384] - - [357, 88.486] - - - [19584, 9984, 1, 384] - - [357, 91.543] - - - [19968, 9600, 1, 384] - - [350, 91.799] - - - [19968, 9601, 1, 384] - - [350, 88.948] - - - [19968, 9984, 1, 384] - - [353, 91.583] - - - [19968, 9985, 1, 384] - - [350, 88.915] - - - [20352, 9984, 1, 384] - - [350, 91.682] - - - [20352, 9985, 1, 384] - - [350, 88.309] - - - [20352, 10368, 1, 384] - - [350, 91.279] - - - [20736, 9984, 1, 384] - - [350, 91.604] - - - [20736, 9985, 1, 384] - - [350, 88.39] - - - [20736, 10368, 1, 384] - - [357, 91.162] - - - [20736, 10369, 1, 384] - - [350, 88.539] - - - [21120, 10368, 1, 384] - - [424, 91.166] - - - [21120, 10369, 1, 384] - - [354, 88.325] - - - [21120, 10752, 1, 384] - - [350, 91.33] - - - [21504, 10368, 1, 384] - - [350, 91.455] - - - [21504, 10369, 1, 384] - - [350, 88.356] - - - [21504, 10752, 1, 384] - - [354, 91.327] - - - [21504, 10753, 1, 384] - - [350, 88.154] - - - [21888, 10752, 1, 384] - - [350, 91.611] - - - [21888, 10753, 1, 384] - - [350, 87.83] - - - [21888, 11136, 1, 384] - - [350, 91.289] - - - [22272, 10752, 1, 384] - - [350, 91.463] - - - [22272, 10753, 1, 384] - - [364, 88.01] - - - [22272, 11136, 1, 384] - - [357, 91.315] - - - [22272, 11137, 1, 384] - - [364, 87.926] - - - [22656, 11136, 1, 384] - - [346, 91.177] - - - [22656, 11137, 1, 384] - - [350, 88.649] - - - [22656, 11520, 1, 384] - - [357, 91.179] - - - [23040, 11136, 1, 384] - - [350, 91.356] - - - [23040, 11137, 1, 384] - - [350, 88.282] - - - [23040, 11520, 1, 384] - - [350, 91.245] - - - [23040, 11521, 1, 384] - - [350, 88.303] - - - [23424, 11520, 1, 384] - - [424, 91.243] - - - [23424, 11521, 1, 384] - - [368, 87.54] - - - [23424, 11904, 1, 384] - - [354, 90.966] - - - [23808, 11520, 1, 384] - - [350, 91.3] - - - [23808, 11521, 1, 384] - - [350, 88.103] - - - [23808, 11904, 1, 384] - - [350, 91.16] - - - [23808, 11905, 1, 384] - - [350, 88.247] - - - [24192, 11904, 1, 384] - - [350, 91.323] - - - [24192, 11905, 1, 384] - - [350, 88.012] - - - [24192, 12288, 1, 384] - - [396, 90.795] - - - [24576, 11904, 1, 384] - - [350, 91.311] - - - [24576, 11905, 1, 384] - - [350, 88.315] - - - [24576, 12288, 1, 384] - - [396, 90.484] - - - [24576, 12289, 1, 384] - - [368, 85.023] - - - [24960, 12288, 1, 384] - - [396, 90.759] - - - [24960, 12289, 1, 384] - - [368, 85.275] - - - [24960, 12672, 1, 384] - - [350, 91.247] - - - [25344, 12288, 1, 384] - - [396, 90.611] - - - [25344, 12289, 1, 384] - - [368, 85.196] - - - [25344, 12672, 1, 384] - - [353, 90.776] - - - [25344, 12673, 1, 384] - - [350, 87.985] - - - [25728, 12672, 1, 384] - - [350, 91.052] - - - [25728, 12673, 1, 384] - - [350, 87.996] - - - [25728, 13056, 1, 384] - - [357, 90.995] - - - [26112, 12673, 1, 384] - - [350, 88.228] - - - [26112, 13056, 1, 384] - - [350, 91.056] - - - [26112, 13057, 1, 384] - - [350, 88.283] - - - [26496, 13057, 1, 384] - - [350, 87.928] - - - [26880, 13057, 1, 384] - - [350, 88.129] - - - [27648, 13825, 1, 384] - - [350, 88.36] - - - [28032, 13824, 1, 384] - - [424, 91.403] - - - [28032, 13825, 1, 384] - - [350, 88.273] - - - [28416, 13824, 1, 384] - - [350, 90.799] - - - [28416, 13825, 1, 384] - - [350, 88.0] - - - [28416, 14208, 1, 384] - - [353, 90.783] - - - [28416, 14209, 1, 384] - - [350, 88.173] - - - [28800, 14208, 1, 384] - - [350, 91.131] - - - [28800, 14209, 1, 384] - - [364, 88.228] - - - [28800, 14592, 1, 384] - - [350, 91.452] - - - [29184, 14208, 1, 384] - - [346, 91.188] - - - [29184, 14209, 1, 384] - - [350, 88.588] - - - [29184, 14592, 1, 384] - - [350, 91.517] - - - [29184, 14593, 1, 384] - - [350, 88.702] - - - [29568, 14592, 1, 384] - - [424, 91.293] - - - [29568, 14593, 1, 384] - - [350, 88.347] - - - [29568, 14976, 1, 384] - - [350, 91.324] - - - [29952, 14592, 1, 384] - - [350, 91.475] - - - [29952, 14593, 1, 384] - - [350, 88.632] - - - [29952, 14976, 1, 384] - - [350, 91.522] - - - [29952, 14977, 1, 384] - - [350, 88.653] - - - [30336, 14976, 1, 384] - - [350, 91.671] - - - [30336, 14977, 1, 384] - - [350, 88.548] - - - [30720, 14976, 1, 384] - - [350, 91.58] - - - [30720, 14977, 1, 384] - - [350, 89.121] - - - [30720, 15361, 1, 384] - - [368, 86.135] - - - [31104, 15361, 1, 384] - - [368, 86.151] - - - [31104, 15744, 1, 384] - - [357, 91.61] - - - [31488, 15361, 1, 384] - - [368, 86.127] - - - [31488, 15744, 1, 384] - - [350, 91.688] - - - [31488, 15745, 1, 384] - - [357, 88.898] - - - [31872, 15744, 1, 384] - - [350, 91.668] - - - [31872, 15745, 1, 384] - - [350, 88.612] - - - [31872, 16128, 1, 384] - - [350, 91.488] - - - [32256, 15744, 1, 384] - - [350, 91.819] - - - [32256, 15745, 1, 384] - - [350, 89.132] - - - [32256, 16128, 1, 384] - - [357, 91.569] - - - [32256, 16129, 1, 384] - - [350, 89.205] - - - [32640, 16128, 1, 384] - - [357, 91.429] - - - [32640, 16129, 1, 384] - - [364, 88.446] - - - [32640, 16512, 1, 384] - - [350, 91.572] - - - [33024, 16128, 1, 384] - - [350, 91.716] - - - [33024, 16129, 1, 384] - - [350, 89.063] - - - [33024, 16512, 1, 384] - - [350, 92.026] - - - [33024, 16513, 1, 384] - - [350, 89.142] - - - [33408, 16512, 1, 384] - - [350, 92.013] - - - [33408, 16513, 1, 384] - - [350, 89.22] - - - [33408, 16896, 1, 384] - - [424, 91.925] - - - [33792, 16512, 1, 384] - - [357, 91.649] - - - [33792, 16513, 1, 384] - - [350, 89.273] - - - [33792, 16896, 1, 384] - - [357, 91.94] - - - [33792, 16897, 1, 384] - - [350, 88.953] - - - [34176, 16896, 1, 384] - - [346, 91.932] - - - [34176, 16897, 1, 384] - - [350, 88.825] - - - [34176, 17280, 1, 384] - - [350, 91.865] - - - [34560, 16896, 1, 384] - - [346, 91.904] - - - [34560, 16897, 1, 384] - - [396, 88.876] - - - [34560, 17280, 1, 384] - - [350, 92.057] - - - [34560, 17281, 1, 384] - - [350, 89.095] - - - [34944, 17280, 1, 384] - - [350, 92.054] - - - [34944, 17281, 1, 384] - - [357, 89.086] - - - [34944, 17664, 1, 384] - - [350, 92.034] - - - [35328, 17280, 1, 384] - - [350, 92.072] - - - [35328, 17281, 1, 384] - - [364, 89.303] - - - [35328, 17664, 1, 384] - - [350, 92.069] - - - [35328, 17665, 1, 384] - - [396, 89.223] - - - [35712, 17664, 1, 384] - - [350, 92.115] - - - [35712, 17665, 1, 384] - - [350, 89.037] - - - [35712, 18048, 1, 384] - - [350, 91.968] - - - [36096, 17664, 1, 384] - - [350, 92.098] - - - [36096, 17665, 1, 384] - - [364, 88.984] - - - [36096, 18048, 1, 384] - - [350, 92.005] - - - [36096, 18049, 1, 384] - - [364, 89.245] - - - [36480, 18048, 1, 384] - - [396, 91.869] - - - [36480, 18049, 1, 384] - - [350, 89.011] - - - [36480, 18432, 1, 384] - - [396, 91.913] - - - [36864, 18048, 1, 384] - - [350, 91.835] - - - [36864, 18049, 1, 384] - - [350, 89.344] - - - [36864, 18432, 1, 384] - - [396, 91.758] - - - [36864, 18433, 1, 384] - - [368, 86.562] - - - [37248, 18432, 1, 384] - - [396, 91.883] - - - [37248, 18433, 1, 384] - - [368, 86.88] - - - [37248, 18816, 1, 384] - - [350, 92.075] - - - [37632, 18432, 1, 384] - - [396, 91.919] - - - [37632, 18433, 1, 384] - - [368, 86.87] - - - [37632, 18816, 1, 384] - - [350, 92.194] - - - [37632, 18817, 1, 384] - - [364, 89.402] - - - [38016, 18816, 1, 384] - - [350, 92.074] - - - [38016, 18817, 1, 384] - - [364, 89.423] - - - [38016, 19200, 1, 384] - - [357, 92.175] - - - [38400, 18816, 1, 384] - - [350, 92.243] - - - [38400, 18817, 1, 384] - - [350, 89.459] - - - [38400, 19200, 1, 384] - - [350, 92.299] - - - [38400, 19201, 1, 384] - - [350, 89.413] - - - [38784, 19200, 1, 384] - - [350, 92.382] - - - [38784, 19201, 1, 384] - - [396, 89.403] - - - [38784, 19584, 1, 384] - - [350, 92.443] - - - [39168, 19200, 1, 384] - - [350, 92.106] - - - [39168, 19201, 1, 384] - - [350, 89.486] - - - [39168, 19584, 1, 384] - - [350, 92.125] - - - [39168, 19585, 1, 384] - - [350, 89.654] - - - [39552, 19584, 1, 384] - - [350, 92.135] - - - [39552, 19585, 1, 384] - - [396, 89.465] - - - [39552, 19968, 1, 384] - - [396, 92.279] - - - [39936, 19584, 1, 384] - - [350, 92.302] - - - [39936, 19585, 1, 384] - - [350, 89.695] - - - [39936, 19968, 1, 384] - - [396, 92.219] - - - [39936, 19969, 1, 384] - - [350, 89.486] - - - [40320, 19968, 1, 384] - - [432, 92.272] - - - [40320, 19969, 1, 384] - - [364, 89.327] - - - [40320, 20352, 1, 384] - - [350, 92.459] - - - [40704, 19968, 1, 384] - - [346, 92.201] - - - [40704, 19969, 1, 384] - - [396, 89.49] - - - [40704, 20352, 1, 384] - - [350, 92.38] - - - [40704, 20353, 1, 384] - - [396, 89.581] - - - [41088, 20352, 1, 384] - - [350, 92.459] - - - [41088, 20353, 1, 384] - - [364, 89.517] - - - [41088, 20736, 1, 384] - - [350, 92.414] - - - [41472, 20352, 1, 384] - - [350, 92.543] - - - [41472, 20353, 1, 384] - - [364, 89.817] - - - [41472, 20736, 1, 384] - - [350, 92.483] - - - [41472, 20737, 1, 384] - - [350, 89.685] - - - [41856, 20736, 1, 384] - - [424, 92.401] - - - [41856, 20737, 1, 384] - - [350, 89.732] - - - [41856, 21120, 1, 384] - - [350, 92.435] - - - [42240, 20736, 1, 384] - - [396, 92.052] - - - [42240, 20737, 1, 384] - - [364, 89.58] - - - [42240, 21120, 1, 384] - - [350, 92.159] - - - [42240, 21121, 1, 384] - - [364, 89.758] - - - [42624, 21120, 1, 384] - - [353, 92.149] - - - [42624, 21121, 1, 384] - - [364, 89.34] - - - [42624, 21504, 1, 384] - - [396, 92.171] - - - [43008, 21120, 1, 384] - - [350, 92.576] - - - [43008, 21121, 1, 384] - - [364, 89.874] - - - [43008, 21504, 1, 384] - - [396, 92.288] - - - [43008, 21505, 1, 384] - - [368, 86.799] - - - [43392, 21504, 1, 384] - - [396, 92.382] - - - [43392, 21505, 1, 384] - - [368, 87.137] - - - [43392, 21888, 1, 384] - - [350, 92.287] - - - [43776, 21504, 1, 384] - - [427, 91.976] - - - [43776, 21505, 1, 384] - - [368, 86.984] - - - [43776, 21888, 1, 384] - - [350, 92.347] - - - [43776, 21889, 1, 384] - - [364, 89.686] - - - [44160, 21888, 1, 384] - - [350, 92.401] - - - [44160, 21889, 1, 384] - - [364, 89.74] - - - [44160, 22272, 1, 384] - - [350, 92.627] - - - [44544, 21888, 1, 384] - - [350, 92.467] - - - [44544, 21889, 1, 384] - - [364, 89.572] - - - [44544, 22272, 1, 384] - - [350, 92.614] - - - [44544, 22273, 1, 384] - - [364, 89.762] - - - [44928, 384, 1, 384] - - [350, 80.973] - - - [44928, 22272, 1, 384] - - [424, 92.486] - - - [44928, 22273, 1, 384] - - [364, 89.716] - - - [44928, 22656, 1, 384] - - [350, 92.447] - - - [45312, 384, 1, 384] - - [354, 81.335] - - - [45312, 22272, 1, 384] - - [350, 92.34] - - - [45312, 22273, 1, 384] - - [364, 89.598] - - - [45312, 22656, 1, 384] - - [350, 92.399] - - - [45312, 22657, 1, 384] - - [396, 89.766] - - - [45696, 384, 1, 384] - - [346, 81.693] - - - [45696, 22656, 1, 384] - - [350, 92.588] - - - [45696, 22657, 1, 384] - - [364, 89.978] - - - [45696, 23040, 1, 384] - - [350, 92.587] - - - [46080, 384, 1, 384] - - [350, 82.269] - - - [46080, 22656, 1, 384] - - [350, 92.664] - - - [46080, 22657, 1, 384] - - [350, 90.03] - - - [46080, 23040, 1, 384] - - [350, 92.542] - - - [46080, 23041, 1, 384] - - [396, 89.954] - - - [46464, 384, 1, 384] - - [346, 82.511] - - - [46464, 23040, 1, 384] - - [396, 92.586] - - - [46464, 23041, 1, 384] - - [364, 89.607] - - - [46464, 23424, 1, 384] - - [350, 92.62] - - - [46848, 384, 1, 384] - - [350, 83.171] - - - [46848, 23040, 1, 384] - - [396, 92.521] - - - [46848, 23041, 1, 384] - - [364, 89.579] - - - [46848, 23424, 1, 384] - - [350, 92.651] - - - [46848, 23425, 1, 384] - - [364, 89.859] - - - [47232, 384, 1, 384] - - [350, 79.078] - - - [47232, 23424, 1, 384] - - [350, 92.643] - - - [47232, 23425, 1, 384] - - [364, 89.874] - - - [47232, 23808, 1, 384] - - [350, 92.477] - - - [47616, 384, 1, 384] - - [346, 79.421] - - - [47616, 23424, 1, 384] - - [350, 92.52] - - - [47616, 23425, 1, 384] - - [396, 90.104] - - - [47616, 23808, 1, 384] - - [357, 92.486] - - - [47616, 23809, 1, 384] - - [396, 90.029] - - - [48000, 384, 1, 384] - - [350, 79.861] - - - [48000, 23808, 1, 384] - - [396, 92.441] - - - [48000, 23809, 1, 384] - - [364, 89.709] - - - [48000, 24192, 1, 384] - - [350, 92.673] - - - [48384, 384, 1, 384] - - [350, 80.884] - - - [48384, 23808, 1, 384] - - [396, 92.434] - - - [48384, 23809, 1, 384] - - [364, 89.817] - - - [48384, 24192, 1, 384] - - [350, 92.568] - - - [48384, 24193, 1, 384] - - [364, 89.954] - - - [48768, 384, 1, 384] - - [346, 80.915] - - - [48768, 24192, 1, 384] - - [350, 92.688] - - - [48768, 24193, 1, 384] - - [396, 89.862] - - - [48768, 24576, 1, 384] - - [396, 92.291] - - - [49152, 384, 1, 384] - - [392, 80.946] - - - [49152, 24192, 1, 384] - - [350, 92.27] - - - [49152, 24193, 1, 384] - - [364, 89.276] - - - [49152, 24576, 1, 384] - - [396, 91.934] - - - [49152, 24577, 1, 384] - - [368, 85.798] - - - [49536, 384, 1, 384] - - [392, 81.72] - - - [49536, 24576, 1, 384] - - [396, 92.409] - - - [49536, 24577, 1, 384] - - [368, 86.414] - - - [49536, 24960, 1, 384] - - [350, 92.77] - - - [49920, 384, 1, 384] - - [354, 82.109] - - - [49920, 24576, 1, 384] - - [396, 92.254] - - - [49920, 24577, 1, 384] - - [368, 86.502] - - - [49920, 24960, 1, 384] - - [350, 92.737] - - - [49920, 24961, 1, 384] - - [396, 89.981] - - - [50304, 384, 1, 384] - - [346, 82.394] - - - [50304, 24960, 1, 384] - - [350, 92.61] - - - [50304, 24961, 1, 384] - - [364, 89.922] - - - [50304, 25344, 1, 384] - - [424, 92.579] - - - [50688, 384, 1, 384] - - [392, 82.88] - - - [50688, 24960, 1, 384] - - [350, 92.604] - - - [50688, 24961, 1, 384] - - [396, 90.367] - - - [50688, 25344, 1, 384] - - [350, 92.533] - - - [50688, 25345, 1, 384] - - [396, 90.051] - - - [51072, 384, 1, 384] - - [392, 83.16] - - - [51072, 25344, 1, 384] - - [396, 92.604] - - - [51072, 25345, 1, 384] - - [396, 90.112] - - - [51072, 25728, 1, 384] - - [350, 92.465] - - - [51456, 384, 1, 384] - - [346, 83.786] - - - [51456, 25344, 1, 384] - - [350, 92.591] - - - [51456, 25345, 1, 384] - - [396, 89.944] - - - [51456, 25728, 1, 384] - - [350, 92.57] - - - [51456, 25729, 1, 384] - - [396, 90.093] - - - [51840, 384, 1, 384] - - [397, 79.741] - - - [51840, 25728, 1, 384] - - [350, 92.601] - - - [51840, 25729, 1, 384] - - [364, 90.107] - - - [51840, 26112, 1, 384] - - [396, 92.708] - - - [52224, 384, 1, 384] - - [392, 80.276] - - - [52224, 25728, 1, 384] - - [350, 92.625] - - - [52224, 25729, 1, 384] - - [350, 89.896] - - - [52224, 26112, 1, 384] - - [350, 92.693] - - - [52224, 26113, 1, 384] - - [350, 89.779] - - - [52608, 384, 1, 384] - - [350, 80.866] - - - [52608, 26112, 1, 384] - - [424, 92.704] - - - [52608, 26113, 1, 384] - - [396, 89.9] - - - [52608, 26496, 1, 384] - - [350, 92.693] - - - [52992, 384, 1, 384] - - [392, 81.4] - - - [52992, 26112, 1, 384] - - [396, 92.633] - - - [52992, 26113, 1, 384] - - [364, 89.988] - - - [52992, 26496, 1, 384] - - [350, 92.604] - - - [52992, 26497, 1, 384] - - [364, 90.16] - - - [53376, 384, 1, 384] - - [346, 81.406] - - - [53376, 26496, 1, 384] - - [350, 92.663] - - - [53376, 26497, 1, 384] - - [396, 90.11] - - - [53376, 26880, 1, 384] - - [350, 92.711] - - - [53760, 384, 1, 384] - - [397, 81.954] - - - [53760, 26496, 1, 384] - - [350, 92.736] - - - [53760, 26497, 1, 384] - - [364, 90.183] - - - [53760, 26880, 1, 384] - - [350, 92.708] - - - [53760, 26881, 1, 384] - - [364, 90.044] - - - [54144, 384, 1, 384] - - [350, 82.464] - - - [54144, 26880, 1, 384] - - [350, 92.717] - - - [54144, 26881, 1, 384] - - [396, 90.005] - - - [54144, 27264, 1, 384] - - [350, 92.712] - - - [54528, 384, 1, 384] - - [353, 82.67] - - - [54528, 26880, 1, 384] - - [350, 92.698] - - - [54528, 26881, 1, 384] - - [396, 90.069] - - - [54528, 27264, 1, 384] - - [350, 92.768] - - - [54528, 27265, 1, 384] - - [396, 90.155] - - - [54912, 384, 1, 384] - - [350, 83.176] - - - [54912, 27264, 1, 384] - - [350, 92.795] - - - [54912, 27265, 1, 384] - - [364, 90.216] - - - [54912, 27648, 1, 384] - - [396, 92.605] - - - [55296, 384, 1, 384] - - [350, 83.434] - - - [55296, 27264, 1, 384] - - [350, 92.75] - - - [55296, 27265, 1, 384] - - [364, 90.245] - - - [55296, 27648, 1, 384] - - [396, 92.543] - - - [55296, 27649, 1, 384] - - [368, 87.284] - - - [55680, 384, 1, 384] - - [346, 83.615] - - - [55680, 27648, 1, 384] - - [396, 92.624] - - - [55680, 27649, 1, 384] - - [368, 87.185] - - - [55680, 28032, 1, 384] - - [350, 92.829] - - - [56064, 384, 1, 384] - - [392, 84.059] - - - [56064, 27648, 1, 384] - - [396, 92.551] - - - [56064, 27649, 1, 384] - - [368, 87.563] - - - [56064, 28032, 1, 384] - - [350, 92.732] - - - [56064, 28033, 1, 384] - - [364, 90.277] - - - [56448, 384, 1, 384] - - [350, 80.771] - - - [56448, 28032, 1, 384] - - [350, 92.709] - - - [56448, 28033, 1, 384] - - [396, 90.135] - - - [56448, 28416, 1, 384] - - [396, 92.662] - - - [56832, 384, 1, 384] - - [350, 81.129] - - - [56832, 28032, 1, 384] - - [350, 92.746] - - - [56832, 28033, 1, 384] - - [396, 90.399] - - - [56832, 28416, 1, 384] - - [396, 92.68] - - - [56832, 28417, 1, 384] - - [396, 90.316] - - - [57216, 384, 1, 384] - - [353, 81.589] - - - [57216, 28416, 1, 384] - - [396, 92.736] - - - [57216, 28417, 1, 384] - - [364, 89.941] - - - [57216, 28800, 1, 384] - - [350, 92.804] - - - [57600, 384, 1, 384] - - [392, 82.031] - - - [57600, 28416, 1, 384] - - [350, 92.69] - - - [57600, 28417, 1, 384] - - [364, 90.114] - - - [57600, 28800, 1, 384] - - [350, 92.808] - - - [57600, 28801, 1, 384] - - [396, 90.262] - - - [57984, 384, 1, 384] - - [350, 82.348] - - - [57984, 28800, 1, 384] - - [350, 92.782] - - - [57984, 28801, 1, 384] - - [364, 90.182] - - - [57984, 29184, 1, 384] - - [396, 92.757] - - - [58368, 384, 1, 384] - - [350, 82.682] - - - [58368, 28800, 1, 384] - - [350, 92.831] - - - [58368, 28801, 1, 384] - - [364, 90.36] - - - [58368, 29184, 1, 384] - - [396, 92.748] - - - [58368, 29185, 1, 384] - - [396, 90.318] - - - [58752, 384, 1, 384] - - [346, 83.031] - - - [58752, 29184, 1, 384] - - [396, 92.778] - - - [58752, 29185, 1, 384] - - [364, 89.99] - - - [58752, 29568, 1, 384] - - [357, 92.661] - - - [59136, 384, 1, 384] - - [392, 83.432] - - - [59136, 29184, 1, 384] - - [396, 92.74] - - - [59136, 29185, 1, 384] - - [364, 89.751] - - - [59136, 29568, 1, 384] - - [396, 92.643] - - - [59136, 29569, 1, 384] - - [396, 89.979] - - - [59520, 384, 1, 384] - - [392, 83.618] - - - [59520, 29568, 1, 384] - - [350, 92.661] - - - [59520, 29569, 1, 384] - - [364, 90.294] - - - [59520, 29952, 1, 384] - - [350, 92.813] - - - [59904, 384, 1, 384] - - [353, 83.927] - - - [59904, 29568, 1, 384] - - [350, 92.695] - - - [59904, 29569, 1, 384] - - [364, 90.431] - - - [59904, 29952, 1, 384] - - [350, 92.796] - - - [59904, 29953, 1, 384] - - [364, 90.332] - - - [60288, 384, 1, 384] - - [350, 84.358] - - - [60288, 29952, 1, 384] - - [350, 92.889] - - - [60288, 29953, 1, 384] - - [364, 90.249] - - - [60288, 30336, 1, 384] - - [350, 92.819] - - - [60672, 384, 1, 384] - - [392, 84.664] - - - [60672, 29952, 1, 384] - - [350, 92.781] - - - [60672, 29953, 1, 384] - - [364, 90.025] - - - [60672, 30336, 1, 384] - - [350, 92.838] - - - [60672, 30337, 1, 384] - - [364, 90.171] - - - [61056, 384, 1, 384] - - [350, 81.437] - - - [61056, 30336, 1, 384] - - [350, 92.901] - - - [61056, 30337, 1, 384] - - [396, 90.196] - - - [61056, 30720, 1, 384] - - [396, 92.619] - - - [61440, 384, 1, 384] - - [353, 81.802] - - - [61440, 30336, 1, 384] - - [350, 92.666] - - - [61440, 30337, 1, 384] - - [396, 90.373] - - - [61440, 30720, 1, 384] - - [396, 92.565] - - - [61440, 30721, 1, 384] - - [368, 87.292] - - - [61824, 384, 1, 384] - - [346, 82.198] - - - [61824, 30720, 1, 384] - - [396, 92.711] - - - [61824, 30721, 1, 384] - - [368, 87.654] - - - [61824, 31104, 1, 384] - - [350, 92.84] - - - [62208, 384, 1, 384] - - [353, 82.775] - - - [62208, 30720, 1, 384] - - [396, 92.632] - - - [62208, 30721, 1, 384] - - [368, 87.638] - - - [62208, 31104, 1, 384] - - [350, 92.868] - - - [62208, 31105, 1, 384] - - [364, 90.357] - - - [62592, 384, 1, 384] - - [350, 83.02] - - - [62592, 31104, 1, 384] - - [350, 92.808] - - - [62592, 31105, 1, 384] - - [364, 90.05] - - - [62592, 31488, 1, 384] - - [396, 92.817] - - - [62976, 384, 1, 384] - - [350, 83.247] - - - [62976, 31104, 1, 384] - - [350, 92.88] - - - [62976, 31105, 1, 384] - - [364, 90.455] - - - [62976, 31488, 1, 384] - - [350, 92.745] - - - [62976, 31489, 1, 384] - - [396, 90.367] - - - [63360, 384, 1, 384] - - [350, 83.406] - - - [63360, 31488, 1, 384] - - [396, 92.822] - - - [63360, 31489, 1, 384] - - [396, 90.223] - - - [63360, 31872, 1, 384] - - [350, 92.803] - - - [63744, 384, 1, 384] - - [350, 83.629] - - - [63744, 31488, 1, 384] - - [396, 92.775] - - - [63744, 31489, 1, 384] - - [364, 90.239] - - - [63744, 31872, 1, 384] - - [350, 92.888] - - - [63744, 31873, 1, 384] - - [364, 90.419] - - - [64128, 384, 1, 384] - - [350, 84.017] - - - [64128, 31872, 1, 384] - - [350, 92.772] - - - [64128, 31873, 1, 384] - - [364, 90.323] - - - [64128, 32256, 1, 384] - - [396, 92.869] - - - [64512, 384, 1, 384] - - [350, 84.321] - - - [64512, 31872, 1, 384] - - [350, 92.76] - - - [64512, 31873, 1, 384] - - [396, 90.27] - - - [64512, 32256, 1, 384] - - [396, 92.768] - - - [64512, 32257, 1, 384] - - [396, 90.267] - - - [64896, 384, 1, 384] - - [346, 84.889] - - - [64896, 32256, 1, 384] - - [396, 92.868] - - - [64896, 32257, 1, 384] - - [364, 90.235] - - - [64896, 32640, 1, 384] - - [350, 92.927] - - - [65280, 384, 1, 384] - - [354, 84.99] - - - [65280, 32256, 1, 384] - - [396, 92.847] - - - [65280, 32257, 1, 384] - - [396, 90.296] - - - [65280, 32640, 1, 384] - - [350, 92.841] - - - [65280, 32641, 1, 384] - - [364, 90.43] - - - [65664, 384, 1, 384] - - [364, 84.941] - - - [65664, 32640, 1, 384] - - [396, 92.751] - - - [65664, 32641, 1, 384] - - [364, 90.325] - - - [65664, 33024, 1, 384] - - [396, 92.846] - - - [66048, 384, 1, 384] - - [392, 82.514] - - - [66048, 32640, 1, 384] - - [350, 92.957] - - - [66048, 32641, 1, 384] - - [396, 90.384] - - - [66048, 33024, 1, 384] - - [350, 92.942] - - - [66048, 33025, 1, 384] - - [396, 90.301] - - - [66432, 384, 1, 384] - - [346, 82.885] - - - [66432, 33024, 1, 384] - - [350, 92.962] - - - [66432, 33025, 1, 384] - - [364, 90.271] - - - [66432, 33408, 1, 384] - - [350, 92.87] - - - [66816, 384, 1, 384] - - [350, 83.349] - - - [66816, 33024, 1, 384] - - [350, 92.825] - - - [66816, 33025, 1, 384] - - [364, 90.244] - - - [66816, 33408, 1, 384] - - [396, 92.817] - - - [66816, 33409, 1, 384] - - [396, 90.415] - - - [67200, 384, 1, 384] - - [350, 83.717] - - - [67200, 33408, 1, 384] - - [396, 92.81] - - - [67200, 33409, 1, 384] - - [364, 90.442] - - - [67200, 33792, 1, 384] - - [396, 92.674] - - - [67584, 384, 1, 384] - - [350, 83.96] - - - [67584, 33408, 1, 384] - - [350, 92.831] - - - [67584, 33409, 1, 384] - - [364, 90.379] - - - [67584, 33792, 1, 384] - - [396, 92.707] - - - [67584, 33793, 1, 384] - - [368, 87.603] - - - [67968, 384, 1, 384] - - [392, 83.966] - - - [67968, 33792, 1, 384] - - [396, 92.661] - - - [67968, 33793, 1, 384] - - [368, 87.83] - - - [67968, 34176, 1, 384] - - [350, 92.933] - - - [68352, 384, 1, 384] - - [353, 84.283] - - - [68352, 33792, 1, 384] - - [396, 92.675] - - - [68352, 33793, 1, 384] - - [368, 87.827] - - - [68352, 34176, 1, 384] - - [350, 92.858] - - - [68352, 34177, 1, 384] - - [364, 90.422] - - - [68736, 384, 1, 384] - - [392, 84.553] - - - [68736, 34176, 1, 384] - - [350, 92.941] - - - [68736, 34177, 1, 384] - - [364, 90.47] - - - [68736, 34560, 1, 384] - - [350, 92.962] - - - [69120, 384, 1, 384] - - [353, 84.696] - - - [69120, 34176, 1, 384] - - [350, 92.977] - - - [69120, 34177, 1, 384] - - [364, 90.503] - - - [69120, 34560, 1, 384] - - [350, 92.94] - - - [69120, 34561, 1, 384] - - [364, 90.307] - - - [69504, 384, 1, 384] - - [353, 84.979] - - - [69504, 34560, 1, 384] - - [350, 92.956] - - - [69504, 34561, 1, 384] - - [396, 90.277] - - - [69504, 34944, 1, 384] - - [350, 92.943] - - - [69888, 384, 1, 384] - - [350, 85.354] - - - [69888, 34560, 1, 384] - - [350, 92.787] - - - [69888, 34561, 1, 384] - - [364, 90.257] - - - [69888, 34944, 1, 384] - - [350, 92.838] - - - [69888, 34945, 1, 384] - - [364, 90.444] - - - [70272, 384, 1, 384] - - [419, 85.384] - - - [70272, 34944, 1, 384] - - [350, 92.931] - - - [70272, 34945, 1, 384] - - [364, 90.458] - - - [70272, 35328, 1, 384] - - [396, 92.864] - - - [70656, 384, 1, 384] - - [392, 82.851] - - - [70656, 34944, 1, 384] - - [350, 92.847] - - - [70656, 34945, 1, 384] - - [364, 90.384] - - - [70656, 35328, 1, 384] - - [396, 92.751] - - - [70656, 35329, 1, 384] - - [364, 90.151] - - - [71040, 384, 1, 384] - - [419, 83.008] - - - [71040, 35328, 1, 384] - - [396, 92.824] - - - [71040, 35329, 1, 384] - - [396, 89.889] - - - [71040, 35712, 1, 384] - - [350, 92.854] - - - [71424, 384, 1, 384] - - [353, 83.766] - - - [71424, 35328, 1, 384] - - [396, 92.746] - - - [71424, 35329, 1, 384] - - [396, 90.22] - - - [71424, 35712, 1, 384] - - [350, 92.86] - - - [71424, 35713, 1, 384] - - [396, 90.385] - - - [71808, 384, 1, 384] - - [346, 83.89] - - - [71808, 35712, 1, 384] - - [350, 92.901] - - - [71808, 35713, 1, 384] - - [364, 90.293] - - - [71808, 36096, 1, 384] - - [350, 92.851] - - - [72192, 384, 1, 384] - - [392, 84.162] - - - [72192, 35712, 1, 384] - - [350, 92.942] - - - [72192, 35713, 1, 384] - - [364, 90.416] - - - [72192, 36096, 1, 384] - - [350, 92.81] - - - [72192, 36097, 1, 384] - - [364, 90.18] - - - [72576, 384, 1, 384] - - [346, 84.313] - - - [72576, 36096, 1, 384] - - [396, 92.702] - - - [72576, 36097, 1, 384] - - [364, 89.861] - - - [72576, 36480, 1, 384] - - [350, 92.717] - - - [72960, 384, 1, 384] - - [346, 84.774] - - - [72960, 36096, 1, 384] - - [350, 92.784] - - - [72960, 36097, 1, 384] - - [364, 90.133] - - - [72960, 36480, 1, 384] - - [350, 92.855] - - - [72960, 36481, 1, 384] - - [396, 90.317] - - - [73344, 384, 1, 384] - - [392, 84.702] - - - [73344, 36480, 1, 384] - - [350, 92.892] - - - [73344, 36481, 1, 384] - - [364, 90.332] - - - [73344, 36864, 1, 384] - - [396, 92.519] - - - [73728, 384, 1, 384] - - [350, 84.137] - - - [73728, 36480, 1, 384] - - [350, 92.265] - - - [73728, 36481, 1, 384] - - [364, 89.417] - - - [73728, 36864, 1, 384] - - [396, 92.016] - - - [73728, 36865, 1, 384] - - [368, 85.793] - - - [74112, 384, 1, 384] - - [350, 85.278] - - - [74112, 36864, 1, 384] - - [396, 92.543] - - - [74112, 36865, 1, 384] - - [368, 87.079] - - - [74112, 37248, 1, 384] - - [350, 92.779] - - - [74496, 384, 1, 384] - - [350, 85.527] - - - [74496, 36864, 1, 384] - - [396, 92.415] - - - [74496, 36865, 1, 384] - - [368, 86.996] - - - [74496, 37248, 1, 384] - - [357, 92.667] - - - [74496, 37249, 1, 384] - - [396, 90.13] - - - [74880, 384, 1, 384] - - [392, 85.584] - - - [74880, 37248, 1, 384] - - [350, 92.734] - - - [74880, 37249, 1, 384] - - [364, 90.294] - - - [74880, 37632, 1, 384] - - [350, 92.865] - - - [75264, 384, 1, 384] - - [353, 83.37] - - - [75264, 37248, 1, 384] - - [350, 92.623] - - - [75264, 37249, 1, 384] - - [364, 90.303] - - - [75264, 37632, 1, 384] - - [350, 92.695] - - - [75264, 37633, 1, 384] - - [364, 90.139] - - - [75648, 384, 1, 384] - - [353, 83.614] - - - [75648, 37632, 1, 384] - - [350, 92.819] - - - [75648, 37633, 1, 384] - - [364, 90.13] - - - [75648, 38016, 1, 384] - - [350, 92.795] - - - [76032, 384, 1, 384] - - [346, 84.221] - - - [76032, 37632, 1, 384] - - [350, 92.73] - - - [76032, 37633, 1, 384] - - [364, 90.052] - - - [76032, 38016, 1, 384] - - [350, 92.697] - - - [76032, 38017, 1, 384] - - [364, 90.129] - - - [76416, 384, 1, 384] - - [350, 84.435] - - - [76416, 38016, 1, 384] - - [350, 92.79] - - - [76416, 38017, 1, 384] - - [364, 90.176] - - - [76416, 38400, 1, 384] - - [350, 92.688] - - - [76800, 384, 1, 384] - - [350, 84.732] - - - [76800, 38016, 1, 384] - - [350, 92.718] - - - [76800, 38017, 1, 384] - - [364, 90.105] - - - [76800, 38400, 1, 384] - - [396, 92.489] - - - [76800, 38401, 1, 384] - - [396, 89.875] - - - [77184, 384, 1, 384] - - [353, 84.832] - - - [77184, 38400, 1, 384] - - [350, 92.69] - - - [77184, 38401, 1, 384] - - [364, 89.9] - - - [77184, 38784, 1, 384] - - [350, 92.795] - - - [77568, 384, 1, 384] - - [346, 84.931] - - - [77568, 38400, 1, 384] - - [396, 92.576] - - - [77568, 38401, 1, 384] - - [364, 89.875] - - - [77568, 38784, 1, 384] - - [350, 92.609] - - - [77568, 38785, 1, 384] - - [364, 90.039] - - - [77952, 384, 1, 384] - - [350, 85.099] - - - [77952, 38784, 1, 384] - - [350, 92.668] - - - [77952, 38785, 1, 384] - - [364, 90.14] - - - [77952, 39168, 1, 384] - - [350, 92.526] - - - [78336, 384, 1, 384] - - [350, 85.182] - - - [78336, 38784, 1, 384] - - [350, 92.689] - - - [78336, 38785, 1, 384] - - [364, 90.179] - - - [78336, 39168, 1, 384] - - [350, 92.46] - - - [78336, 39169, 1, 384] - - [364, 89.957] - - - [78720, 384, 1, 384] - - [396, 85.307] - - - [78720, 39168, 1, 384] - - [350, 92.596] - - - [78720, 39169, 1, 384] - - [364, 90.024] - - - [78720, 39552, 1, 384] - - [350, 92.71] - - - [79104, 384, 1, 384] - - [353, 85.695] - - - [79104, 39168, 1, 384] - - [350, 92.444] - - - [79104, 39169, 1, 384] - - [364, 89.917] - - - [79104, 39552, 1, 384] - - [350, 92.623] - - - [79104, 39553, 1, 384] - - [364, 90.021] - - - [79488, 384, 1, 384] - - [353, 85.761] - - - [79488, 39552, 1, 384] - - [350, 92.688] - - - [79488, 39553, 1, 384] - - [364, 89.984] - - - [79488, 39936, 1, 384] - - [396, 92.242] - - - [79872, 384, 1, 384] - - [392, 83.795] - - - [79872, 39552, 1, 384] - - [350, 92.651] - - - [79872, 39553, 1, 384] - - [364, 90.127] - - - [79872, 39936, 1, 384] - - [396, 92.049] - - - [79872, 39937, 1, 384] - - [368, 87.255] - - - [80256, 384, 1, 384] - - [354, 84.198] - - - [80256, 39936, 1, 384] - - [396, 92.24] - - - [80256, 39937, 1, 384] - - [368, 87.442] - - - [80256, 40320, 1, 384] - - [350, 92.641] - - - [80640, 384, 1, 384] - - [392, 84.427] - - - [80640, 39936, 1, 384] - - [396, 92.276] - - - [80640, 39937, 1, 384] - - [368, 87.459] - - - [80640, 40320, 1, 384] - - [350, 92.489] - - - [80640, 40321, 1, 384] - - [364, 90.018] - - - [81024, 384, 1, 384] - - [350, 84.936] - - - [81024, 40320, 1, 384] - - [350, 92.623] - - - [81024, 40321, 1, 384] - - [364, 89.964] - - - [81024, 40704, 1, 384] - - [350, 92.556] - - - [81408, 384, 1, 384] - - [350, 85.057] - - - [81408, 40320, 1, 384] - - [350, 92.579] - - - [81408, 40321, 1, 384] - - [364, 90.021] - - - [81408, 40704, 1, 384] - - [350, 92.435] - - - [81408, 40705, 1, 384] - - [364, 89.723] - - - [81792, 384, 1, 384] - - [346, 84.858] - - - [81792, 40704, 1, 384] - - [350, 92.591] - - - [81792, 40705, 1, 384] - - [364, 89.875] - - - [81792, 41088, 1, 384] - - [350, 92.485] - - - [82176, 384, 1, 384] - - [354, 85.176] - - - [82176, 40704, 1, 384] - - [350, 92.41] - - - [82176, 40705, 1, 384] - - [364, 89.69] - - - [82176, 41088, 1, 384] - - [350, 92.392] - - - [82176, 41089, 1, 384] - - [364, 89.835] - - - [82560, 384, 1, 384] - - [392, 85.335] - - - [82560, 41088, 1, 384] - - [350, 92.473] - - - [82560, 41089, 1, 384] - - [364, 89.901] - - - [82560, 41472, 1, 384] - - [350, 92.421] - - - [82944, 384, 1, 384] - - [353, 85.506] - - - [82944, 41088, 1, 384] - - [350, 92.362] - - - [82944, 41089, 1, 384] - - [364, 89.797] - - - [82944, 41472, 1, 384] - - [396, 92.15] - - - [82944, 41473, 1, 384] - - [364, 89.49] - - - [83328, 384, 1, 384] - - [353, 85.424] - - - [83328, 41472, 1, 384] - - [396, 92.359] - - - [83328, 41473, 1, 384] - - [364, 89.789] - - - [83328, 41856, 1, 384] - - [350, 92.43] - - - [83712, 384, 1, 384] - - [350, 85.699] - - - [83712, 41472, 1, 384] - - [396, 92.291] - - - [83712, 41473, 1, 384] - - [364, 89.674] - - - [83712, 41856, 1, 384] - - [350, 92.349] - - - [83712, 41857, 1, 384] - - [364, 89.812] - - - [84096, 384, 1, 384] - - [353, 86.0] - - - [84096, 41856, 1, 384] - - [350, 92.448] - - - [84096, 41857, 1, 384] - - [364, 89.895] - - - [84096, 42240, 1, 384] - - [350, 92.435] - - - [84480, 384, 1, 384] - - [350, 86.061] - - - [84480, 41856, 1, 384] - - [350, 92.41] - - - [84480, 41857, 1, 384] - - [364, 89.859] - - - [84480, 42240, 1, 384] - - [350, 92.271] - - - [84480, 42241, 1, 384] - - [364, 89.59] - - - [84864, 384, 1, 384] - - [392, 84.636] - - - [84864, 42240, 1, 384] - - [350, 92.41] - - - [84864, 42241, 1, 384] - - [364, 89.706] - - - [84864, 42624, 1, 384] - - [350, 92.398] - - - [85248, 384, 1, 384] - - [396, 84.625] - - - [85248, 42240, 1, 384] - - [350, 92.316] - - - [85248, 42241, 1, 384] - - [364, 89.63] - - - [85248, 42624, 1, 384] - - [350, 92.266] - - - [85248, 42625, 1, 384] - - [364, 89.617] - - - [85632, 384, 1, 384] - - [350, 85.184] - - - [85632, 42624, 1, 384] - - [350, 92.322] - - - [85632, 42625, 1, 384] - - [364, 89.686] - - - [85632, 43008, 1, 384] - - [430, 91.97] - - - [86016, 384, 1, 384] - - [353, 84.931] - - - [86016, 42624, 1, 384] - - [350, 92.239] - - - [86016, 42625, 1, 384] - - [364, 89.556] - - - [86016, 43008, 1, 384] - - [396, 91.676] - - - [86016, 43009, 1, 384] - - [368, 86.876] - - - [86400, 384, 1, 384] - - [354, 85.221] - - - [86400, 43008, 1, 384] - - [396, 91.785] - - - [86400, 43009, 1, 384] - - [368, 87.132] - - - [86400, 43392, 1, 384] - - [350, 92.206] - - - [86784, 384, 1, 384] - - [392, 85.574] - - - [86784, 43008, 1, 384] - - [430, 91.875] - - - [86784, 43009, 1, 384] - - [368, 87.13] - - - [86784, 43392, 1, 384] - - [350, 92.131] - - - [86784, 43393, 1, 384] - - [364, 89.685] - - - [87168, 384, 1, 384] - - [353, 85.484] - - - [87168, 43392, 1, 384] - - [350, 92.243] - - - [87168, 43393, 1, 384] - - [364, 89.702] - - - [87168, 43776, 1, 384] - - [350, 92.09] - - - [87552, 384, 1, 384] - - [353, 85.683] - - - [87552, 43392, 1, 384] - - [350, 92.159] - - - [87552, 43393, 1, 384] - - [364, 89.707] - - - [87552, 43776, 1, 384] - - [350, 92.182] - - - [87552, 43777, 1, 384] - - [364, 89.525] - - - [87936, 384, 1, 384] - - [353, 85.656] - - - [87936, 43776, 1, 384] - - [350, 92.156] - - - [87936, 43777, 1, 384] - - [364, 89.487] - - - [87936, 44160, 1, 384] - - [350, 92.237] - - - [88320, 384, 1, 384] - - [353, 85.904] - - - [88320, 43776, 1, 384] - - [350, 92.084] - - - [88320, 43777, 1, 384] - - [364, 89.544] - - - [88320, 44160, 1, 384] - - [350, 92.1] - - - [88320, 44161, 1, 384] - - [364, 89.629] - - - [88704, 384, 1, 384] - - [350, 85.913] - - - [88704, 44160, 1, 384] - - [350, 92.232] - - - [88704, 44161, 1, 384] - - [364, 89.701] - - - [88704, 44544, 1, 384] - - [430, 92.084] - - - [89088, 384, 1, 384] - - [350, 85.872] - - - [89088, 44160, 1, 384] - - [350, 92.063] - - - [89088, 44161, 1, 384] - - [364, 89.534] - - - [89088, 44544, 1, 384] - - [430, 91.779] - - - [89088, 44545, 1, 384] - - [364, 89.211] - - - [89472, 384, 1, 384] - - [424, 84.73] - - - [89472, 44544, 1, 384] - - [430, 92.095] - - - [89472, 44545, 1, 384] - - [364, 89.406] - - - [89472, 44928, 1, 384] - - [350, 92.031] - - - [89856, 384, 1, 384] - - [424, 85.033] - - - [89856, 44544, 1, 384] - - [430, 91.941] - - - [89856, 44545, 1, 384] - - [364, 89.345] - - - [89856, 44928, 1, 384] - - [350, 91.903] - - - [89856, 44929, 1, 384] - - [364, 89.53] - - - [90240, 384, 1, 384] - - [432, 85.081] - - - [90240, 44928, 1, 384] - - [350, 91.994] - - - [90240, 44929, 1, 384] - - [364, 89.483] - - - [90240, 45312, 1, 384] - - [351, 92.047] - - - [90624, 384, 1, 384] - - [350, 85.687] - - - [90624, 44928, 1, 384] - - [351, 91.981] - - - [90624, 44929, 1, 384] - - [398, 89.581] - - - [90624, 45312, 1, 384] - - [351, 91.989] - - - [90624, 45313, 1, 384] - - [398, 89.344] - - - [91008, 384, 1, 384] - - [350, 85.708] - - - [91008, 45312, 1, 384] - - [351, 91.989] - - - [91008, 45313, 1, 384] - - [364, 89.306] - - - [91008, 45696, 1, 384] - - [350, 91.962] - - - [91392, 384, 1, 384] - - [392, 85.597] - - - [91392, 45312, 1, 384] - - [350, 91.933] - - - [91392, 45313, 1, 384] - - [398, 89.272] - - - [91392, 45696, 1, 384] - - [351, 91.895] - - - [91392, 45697, 1, 384] - - [364, 89.414] - - - [91776, 384, 1, 384] - - [400, 85.667] - - - [91776, 45696, 1, 384] - - [350, 91.863] - - - [91776, 45697, 1, 384] - - [364, 89.305] - - - [91776, 46080, 1, 384] - - [430, 91.64] - - - [92160, 384, 1, 384] - - [350, 85.796] - - - [92160, 45696, 1, 384] - - [350, 91.837] - - - [92160, 45697, 1, 384] - - [364, 89.414] - - - [92160, 46080, 1, 384] - - [430, 91.508] - - - [92160, 46081, 1, 384] - - [368, 86.802] - - - [92544, 384, 1, 384] - - [350, 85.829] - - - [92544, 46080, 1, 384] - - [430, 91.693] - - - [92544, 46081, 1, 384] - - [368, 86.955] - - - [92544, 46464, 1, 384] - - [351, 91.912] - - - [92928, 384, 1, 384] - - [419, 85.695] - - - [92928, 46080, 1, 384] - - [430, 91.599] - - - [92928, 46081, 1, 384] - - [368, 86.867] - - - [92928, 46464, 1, 384] - - [351, 91.873] - - - [92928, 46465, 1, 384] - - [398, 89.235] - - - [93312, 384, 1, 384] - - [350, 86.037] - - - [93312, 46464, 1, 384] - - [351, 91.927] - - - [93312, 46465, 1, 384] - - [364, 89.263] - - - [93312, 46848, 1, 384] - - [430, 91.837] - - - [93696, 384, 1, 384] - - [350, 86.188] - - - [93696, 46464, 1, 384] - - [351, 91.908] - - - [93696, 46465, 1, 384] - - [398, 89.38] - - - [93696, 46848, 1, 384] - - [430, 91.722] - - - [93696, 46849, 1, 384] - - [398, 89.235] - - - [94080, 384, 1, 384] - - [350, 84.976] - - - [94080, 46848, 1, 384] - - [430, 91.844] - - - [94080, 46849, 1, 384] - - [398, 89.099] - - - [94080, 47232, 1, 384] - - [351, 91.861] - - - [94464, 384, 1, 384] - - [419, 85.242] - - - [94464, 46848, 1, 384] - - [430, 91.664] - - - [94464, 46849, 1, 384] - - [398, 89.031] - - - [94464, 47232, 1, 384] - - [351, 91.802] - - - [94464, 47233, 1, 384] - - [398, 89.193] - - - [94848, 384, 1, 384] - - [392, 85.717] - - - [94848, 47232, 1, 384] - - [351, 91.871] - - - [94848, 47233, 1, 384] - - [398, 89.246] - - - [94848, 47616, 1, 384] - - [430, 91.822] - - - [95232, 384, 1, 384] - - [353, 85.621] - - - [95232, 47232, 1, 384] - - [351, 91.601] - - - [95232, 47233, 1, 384] - - [398, 89.04] - - - [95232, 47616, 1, 384] - - [430, 91.499] - - - [95232, 47617, 1, 384] - - [398, 88.921] - - - [95616, 384, 1, 384] - - [427, 85.653] - - - [95616, 47616, 1, 384] - - [430, 91.752] - - - [95616, 47617, 1, 384] - - [398, 89.017] - - - [95616, 48000, 1, 384] - - [351, 91.796] - - - [96000, 384, 1, 384] - - [350, 86.102] - - - [96000, 47616, 1, 384] - - [430, 91.666] - - - [96000, 47617, 1, 384] - - [398, 88.991] - - - [96000, 48000, 1, 384] - - [351, 91.729] - - - [96000, 48001, 1, 384] - - [398, 89.178] - - - [96384, 384, 1, 384] - - [353, 86.005] - - - [96384, 48000, 1, 384] - - [351, 91.638] - - - [96384, 48001, 1, 384] - - [398, 89.079] - - - [96384, 48384, 1, 384] - - [430, 91.67] - - - [96768, 384, 1, 384] - - [350, 86.279] - - - [96768, 48000, 1, 384] - - [351, 91.748] - - - [96768, 48001, 1, 384] - - [398, 89.238] - - - [96768, 48384, 1, 384] - - [351, 91.633] - - - [96768, 48385, 1, 384] - - [398, 89.071] - - - [97152, 384, 1, 384] - - [350, 85.838] - - - [97152, 48384, 1, 384] - - [351, 91.662] - - - [97152, 48385, 1, 384] - - [398, 88.906] - - - [97152, 48768, 1, 384] - - [351, 91.552] - - - [97536, 384, 1, 384] - - [350, 85.994] - - - [97536, 48384, 1, 384] - - [351, 91.615] - - - [97536, 48385, 1, 384] - - [398, 89.002] - - - [97536, 48768, 1, 384] - - [351, 91.602] - - - [97536, 48769, 1, 384] - - [398, 89.136] - - - [97920, 384, 1, 384] - - [350, 85.99] - - - [97920, 48768, 1, 384] - - [351, 91.4] - - - [97920, 48769, 1, 384] - - [364, 88.897] - - - [97920, 49152, 1, 384] - - [398, 90.548] - - - [98304, 384, 1, 384] - - [350, 83.057] - - - [98304, 48768, 1, 384] - - [396, 89.587] - - - [98304, 48769, 1, 384] - - [364, 86.182] - - - [98304, 49152, 1, 384] - - [364, 89.094] - - - [98304, 49153, 1, 384] - - [364, 82.937] - - - [98688, 384, 1, 384] - - [419, 85.323] - - - [98688, 49152, 1, 384] - - [398, 90.484] - - - [98688, 49153, 1, 384] - - [402, 82.498] - - - [98688, 49536, 1, 384] - - [351, 91.75] - - - [99072, 384, 1, 384] - - [419, 85.778] - - - [99072, 49152, 1, 384] - - [398, 90.642] - - - [99072, 49153, 1, 384] - - [402, 82.482] - - - [99072, 49536, 1, 384] - - [351, 91.551] - - - [99072, 49537, 1, 384] - - [398, 88.999] - - - [99456, 384, 1, 384] - - [353, 85.864] - - - [99456, 49536, 1, 384] - - [351, 91.574] - - - [99456, 49537, 1, 384] - - [398, 88.973] - - - [99456, 49920, 1, 384] - - [430, 91.577] - - - [99840, 384, 1, 384] - - [350, 85.883] - - - [99840, 49536, 1, 384] - - [351, 91.536] - - - [99840, 49537, 1, 384] - - [398, 89.039] - - - [99840, 49920, 1, 384] - - [351, 91.479] - - - [99840, 49921, 1, 384] - - [398, 88.911] - - - [100224, 384, 1, 384] - - [353, 86.191] - - - [100224, 49920, 1, 384] - - [351, 91.494] - - - [100224, 49921, 1, 384] - - [398, 88.765] - - - [100224, 50304, 1, 384] - - [351, 91.472] - - - [100608, 384, 1, 384] - - [392, 86.142] - - - [100608, 49920, 1, 384] - - [351, 91.402] - - - [100608, 49921, 1, 384] - - [398, 88.763] - - - [100608, 50304, 1, 384] - - [351, 91.441] - - - [100608, 50305, 1, 384] - - [398, 88.918] - - - [100992, 384, 1, 384] - - [392, 86.271] - - - [100992, 50304, 1, 384] - - [351, 91.422] - - - [100992, 50305, 1, 384] - - [398, 88.828] - - - [100992, 50688, 1, 384] - - [430, 91.441] - - - [101376, 384, 1, 384] - - [350, 86.013] - - - [101376, 50304, 1, 384] - - [351, 91.166] - - - [101376, 50305, 1, 384] - - [398, 88.705] - - - [101376, 50688, 1, 384] - - [430, 91.144] - - - [101376, 50689, 1, 384] - - [398, 88.529] - - - [101760, 384, 1, 384] - - [353, 86.086] - - - [101760, 50688, 1, 384] - - [430, 91.382] - - - [101760, 50689, 1, 384] - - [398, 88.549] - - - [101760, 51072, 1, 384] - - [430, 91.271] - - - [102144, 384, 1, 384] - - [396, 86.082] - - - [102144, 50688, 1, 384] - - [430, 91.293] - - - [102144, 50689, 1, 384] - - [398, 88.623] - - - [102144, 51072, 1, 384] - - [351, 91.312] - - - [102144, 51073, 1, 384] - - [398, 88.831] - - - [102528, 384, 1, 384] - - [392, 85.926] - - - [102528, 51072, 1, 384] - - [351, 91.281] - - - [102528, 51073, 1, 384] - - [398, 88.672] - - - [102528, 51456, 1, 384] - - [430, 91.257] - - - [102912, 384, 1, 384] - - [350, 86.15] - - - [102912, 51072, 1, 384] - - [351, 91.228] - - - [102912, 51073, 1, 384] - - [398, 88.827] - - - [102912, 51456, 1, 384] - - [351, 91.125] - - - [102912, 51457, 1, 384] - - [398, 88.618] - - - [103296, 384, 1, 384] - - [350, 85.513] - - - [103296, 51456, 1, 384] - - [430, 91.175] - - - [103296, 51457, 1, 384] - - [398, 88.489] - - - [103296, 51840, 1, 384] - - [351, 91.143] - - - [103680, 384, 1, 384] - - [424, 85.788] - - - [103680, 51456, 1, 384] - - [430, 91.067] - - - [103680, 51457, 1, 384] - - [398, 88.522] - - - [103680, 51840, 1, 384] - - [351, 91.134] - - - [103680, 51841, 1, 384] - - [398, 88.642] - - - [104064, 384, 1, 384] - - [392, 86.084] - - - [104064, 51840, 1, 384] - - [351, 91.109] - - - [104064, 51841, 1, 384] - - [398, 88.516] - - - [104064, 52224, 1, 384] - - [430, 90.75] - - - [104448, 384, 1, 384] - - [350, 86.027] - - - [104448, 51840, 1, 384] - - [351, 90.838] - - - [104448, 51841, 1, 384] - - [398, 88.47] - - - [104448, 52224, 1, 384] - - [430, 90.639] - - - [104448, 52225, 1, 384] - - [368, 85.828] - - - [104832, 384, 1, 384] - - [353, 86.408] - - - [104832, 52224, 1, 384] - - [430, 90.778] - - - [104832, 52225, 1, 384] - - [402, 85.807] - - - [104832, 52608, 1, 384] - - [430, 90.908] - - - [105216, 384, 1, 384] - - [424, 86.115] - - - [105216, 52224, 1, 384] - - [430, 90.745] - - - [105216, 52225, 1, 384] - - [402, 86.187] - - - [105216, 52608, 1, 384] - - [351, 90.873] - - - [105216, 52609, 1, 384] - - [398, 88.515] - - - [105600, 384, 1, 384] - - [354, 86.027] - - - [105600, 52608, 1, 384] - - [430, 90.842] - - - [105600, 52609, 1, 384] - - [398, 88.233] - - - [105600, 52992, 1, 384] - - [430, 90.931] - - - [105984, 384, 1, 384] - - [346, 86.034] - - - [105984, 52608, 1, 384] - - [430, 90.823] - - - [105984, 52609, 1, 384] - - [398, 88.509] - - - [105984, 52992, 1, 384] - - [430, 90.838] - - - [105984, 52993, 1, 384] - - [398, 88.334] - - - [106368, 384, 1, 384] - - [392, 85.968] - - - [106368, 52992, 1, 384] - - [430, 90.9] - - - [106368, 52993, 1, 384] - - [398, 88.208] - - - [106368, 53376, 1, 384] - - [351, 90.808] - - - [106752, 384, 1, 384] - - [353, 86.194] - - - [106752, 52992, 1, 384] - - [430, 90.773] - - - [106752, 52993, 1, 384] - - [398, 88.173] - - - [106752, 53376, 1, 384] - - [351, 90.731] - - - [106752, 53377, 1, 384] - - [398, 88.314] - - - [107136, 384, 1, 384] - - [427, 85.93] - - - [107136, 53376, 1, 384] - - [351, 90.644] - - - [107136, 53377, 1, 384] - - [398, 88.114] - - - [107136, 53760, 1, 384] - - [430, 90.807] - - - [107520, 384, 1, 384] - - [350, 85.896] - - - [107520, 53376, 1, 384] - - [351, 90.48] - - - [107520, 53377, 1, 384] - - [398, 88.086] - - - [107520, 53760, 1, 384] - - [430, 90.453] - - - [107520, 53761, 1, 384] - - [430, 87.895] - - - [107904, 384, 1, 384] - - [353, 86.037] - - - [107904, 53760, 1, 384] - - [430, 90.813] - - - [107904, 53761, 1, 384] - - [398, 88.061] - - - [107904, 54144, 1, 384] - - [351, 90.677] - - - [108288, 384, 1, 384] - - [400, 85.828] - - - [108288, 53760, 1, 384] - - [430, 90.714] - - - [108288, 53761, 1, 384] - - [398, 88.08] - - - [108288, 54144, 1, 384] - - [351, 90.592] - - - [108288, 54145, 1, 384] - - [398, 88.179] - - - [108672, 384, 1, 384] - - [424, 86.081] - - - [108672, 54144, 1, 384] - - [430, 90.586] - - - [108672, 54145, 1, 384] - - [398, 88.079] - - - [108672, 54528, 1, 384] - - [430, 90.721] - - - [109056, 384, 1, 384] - - [392, 86.201] - - - [109056, 54144, 1, 384] - - [430, 90.531] - - - [109056, 54145, 1, 384] - - [398, 88.18] - - - [109056, 54528, 1, 384] - - [430, 90.493] - - - [109056, 54529, 1, 384] - - [398, 88.025] - - - [109440, 384, 1, 384] - - [350, 86.566] - - - [109440, 54528, 1, 384] - - [430, 90.461] - - - [109440, 54529, 1, 384] - - [398, 87.861] - - - [109440, 54912, 1, 384] - - [351, 90.523] - - - [109824, 384, 1, 384] - - [350, 86.458] - - - [109824, 54528, 1, 384] - - [430, 90.56] - - - [109824, 54529, 1, 384] - - [398, 87.918] - - - [109824, 54912, 1, 384] - - [351, 90.402] - - - [109824, 54913, 1, 384] - - [398, 87.981] - - - [110208, 384, 1, 384] - - [350, 86.586] - - - [110208, 54912, 1, 384] - - [351, 90.404] - - - [110208, 54913, 1, 384] - - [398, 87.882] - - - [110208, 55296, 1, 384] - - [398, 90.055] - - - [110592, 384, 1, 384] - - [353, 86.347] - - - [110592, 54912, 1, 384] - - [351, 90.06] - - - [110592, 54913, 1, 384] - - [398, 87.661] - - - [110592, 55296, 1, 384] - - [430, 89.875] - - - [110592, 55297, 1, 384] - - [402, 85.236] - - - [110976, 384, 1, 384] - - [353, 86.231] - - - [110976, 55296, 1, 384] - - [398, 89.949] - - - [110976, 55297, 1, 384] - - [402, 85.64] - - - [110976, 55680, 1, 384] - - [351, 90.266] - - - [111360, 384, 1, 384] - - [350, 86.338] - - - [111360, 55296, 1, 384] - - [398, 89.944] - - - [111360, 55297, 1, 384] - - [402, 85.665] - - - [111360, 55680, 1, 384] - - [351, 90.22] - - - [111360, 55681, 1, 384] - - [398, 87.766] - - - [111744, 384, 1, 384] - - [432, 85.715] - - - [111744, 55680, 1, 384] - - [351, 90.192] - - - [111744, 55681, 1, 384] - - [398, 87.671] - - - [111744, 56064, 1, 384] - - [351, 90.143] - - - [112128, 384, 1, 384] - - [353, 85.921] - - - [112128, 55680, 1, 384] - - [430, 90.162] - - - [112128, 55681, 1, 384] - - [398, 87.735] - - - [112128, 56064, 1, 384] - - [430, 90.167] - - - [112128, 56065, 1, 384] - - [398, 87.634] - - - [112512, 384, 1, 384] - - [346, 85.839] - - - [112512, 56064, 1, 384] - - [430, 90.184] - - - [112512, 56065, 1, 384] - - [398, 87.481] - - - [112512, 56448, 1, 384] - - [351, 89.931] - - - [112896, 384, 1, 384] - - [350, 85.984] - - - [112896, 56064, 1, 384] - - [430, 90.103] - - - [112896, 56065, 1, 384] - - [398, 87.444] - - - [112896, 56448, 1, 384] - - [430, 89.944] - - - [112896, 56449, 1, 384] - - [398, 87.53] - - - [113280, 384, 1, 384] - - [427, 86.108] - - - [113280, 56448, 1, 384] - - [430, 89.993] - - - [113280, 56449, 1, 384] - - [398, 87.496] - - - [113280, 56832, 1, 384] - - [351, 89.765] - - - [113664, 384, 1, 384] - - [350, 86.115] - - - [113664, 56448, 1, 384] - - [430, 89.691] - - - [113664, 56449, 1, 384] - - [398, 87.315] - - - [113664, 56832, 1, 384] - - [398, 89.384] - - - [113664, 56833, 1, 384] - - [398, 87.134] - - - [114048, 384, 1, 384] - - [350, 86.732] - - - [114048, 56832, 1, 384] - - [430, 89.939] - - - [114048, 56833, 1, 384] - - [398, 87.227] - - - [114048, 57216, 1, 384] - - [430, 89.908] - - - [114432, 384, 1, 384] - - [392, 86.45] - - - [114432, 56832, 1, 384] - - [430, 89.653] - - - [114432, 56833, 1, 384] - - [398, 87.258] - - - [114432, 57216, 1, 384] - - [430, 89.804] - - - [114432, 57217, 1, 384] - - [398, 87.359] - - - [114816, 384, 1, 384] - - [424, 86.445] - - - [114816, 57216, 1, 384] - - [422, 89.56] - - - [114816, 57217, 1, 384] - - [398, 87.133] - - - [114816, 57600, 1, 384] - - [430, 89.76] - - - [115200, 384, 1, 384] - - [353, 86.479] - - - [115200, 57216, 1, 384] - - [430, 89.675] - - - [115200, 57217, 1, 384] - - [398, 87.307] - - - [115200, 57600, 1, 384] - - [430, 89.621] - - - [115200, 57601, 1, 384] - - [398, 87.049] - - - [115584, 384, 1, 384] - - [350, 86.445] - - - [115584, 57600, 1, 384] - - [422, 89.546] - - - [115584, 57601, 1, 384] - - [398, 87.012] - - - [115584, 57984, 1, 384] - - [351, 89.57] - - - [115968, 384, 1, 384] - - [346, 85.815] - - - [115968, 57600, 1, 384] - - [430, 89.545] - - - [115968, 57601, 1, 384] - - [398, 87.0] - - - [115968, 57984, 1, 384] - - [430, 89.52] - - - [115968, 57985, 1, 384] - - [398, 87.122] - - - [116352, 384, 1, 384] - - [363, 85.736] - - - [116352, 57984, 1, 384] - - [422, 89.373] - - - [116352, 57985, 1, 384] - - [398, 87.036] - - - [116352, 58368, 1, 384] - - [430, 88.708] - - - [116736, 384, 1, 384] - - [368, 85.573] - - - [116736, 57984, 1, 384] - - [351, 89.176] - - - [116736, 57985, 1, 384] - - [398, 86.957] - - - [116736, 58368, 1, 384] - - [430, 88.211] - - - [116736, 58369, 1, 384] - - [402, 84.739] - - - [117120, 384, 1, 384] - - [353, 85.921] - - - [117120, 58368, 1, 384] - - [398, 88.393] - - - [117120, 58369, 1, 384] - - [402, 84.93] - - - [117120, 58752, 1, 384] - - [351, 89.287] - - - [117504, 384, 1, 384] - - [346, 85.956] - - - [117504, 58368, 1, 384] - - [398, 88.601] - - - [117504, 58369, 1, 384] - - [402, 85.043] - - - [117504, 58752, 1, 384] - - [351, 89.148] - - - [117504, 58753, 1, 384] - - [398, 86.916] - - - [117888, 384, 1, 384] - - [350, 86.342] - - - [117888, 58752, 1, 384] - - [430, 89.13] - - - [117888, 58753, 1, 384] - - [398, 86.817] - - - [117888, 59136, 1, 384] - - [422, 89.159] - - - [118272, 384, 1, 384] - - [350, 86.461] - - - [118272, 58752, 1, 384] - - [422, 89.072] - - - [118272, 58753, 1, 384] - - [398, 86.911] - - - [118272, 59136, 1, 384] - - [422, 88.914] - - - [118272, 59137, 1, 384] - - [398, 86.676] - - - [118656, 384, 1, 384] - - [353, 86.456] - - - [118656, 59136, 1, 384] - - [422, 89.078] - - - [118656, 59137, 1, 384] - - [398, 86.627] - - - [118656, 59520, 1, 384] - - [351, 89.014] - - - [119040, 384, 1, 384] - - [350, 86.612] - - - [119040, 59136, 1, 384] - - [422, 88.933] - - - [119040, 59137, 1, 384] - - [398, 86.605] - - - [119040, 59520, 1, 384] - - [351, 88.945] - - - [119040, 59521, 1, 384] - - [398, 86.559] - - - [119424, 384, 1, 384] - - [427, 86.523] - - - [119424, 59520, 1, 384] - - [422, 88.955] - - - [119424, 59521, 1, 384] - - [398, 86.576] - - - [119424, 59904, 1, 384] - - [422, 88.787] - - - [119808, 384, 1, 384] - - [350, 86.518] - - - [119808, 59520, 1, 384] - - [422, 88.538] - - - [119808, 59521, 1, 384] - - [398, 86.377] - - - [119808, 59904, 1, 384] - - [403, 88.145] - - - [119808, 59905, 1, 384] - - [398, 86.165] - - - [120192, 384, 1, 384] - - [396, 86.214] - - - [120192, 59904, 1, 384] - - [422, 88.636] - - - [120192, 59905, 1, 384] - - [398, 86.326] - - - [120192, 60288, 1, 384] - - [422, 88.727] - - - [120576, 384, 1, 384] - - [426, 85.979] - - - [120576, 59904, 1, 384] - - [425, 88.559] - - - [120576, 59905, 1, 384] - - [398, 86.328] - - - [120576, 60288, 1, 384] - - [422, 88.564] - - - [120576, 60289, 1, 384] - - [398, 86.416] - - - [120960, 384, 1, 384] - - [350, 85.93] - - - [120960, 60288, 1, 384] - - [422, 88.63] - - - [120960, 60289, 1, 384] - - [398, 86.352] - - - [120960, 60672, 1, 384] - - [422, 88.681] - - - [121344, 384, 1, 384] - - [424, 85.626] - - - [121344, 60288, 1, 384] - - [422, 88.493] - - - [121344, 60289, 1, 384] - - [430, 86.277] - - - [121344, 60672, 1, 384] - - [422, 88.446] - - - [121344, 60673, 1, 384] - - [398, 86.171] - - - [121728, 384, 1, 384] - - [427, 85.861] - - - [121728, 60672, 1, 384] - - [422, 88.488] - - - [121728, 60673, 1, 384] - - [398, 86.026] - - - [121728, 61056, 1, 384] - - [422, 88.464] - - - [122112, 384, 1, 384] - - [353, 85.913] - - - [122112, 60672, 1, 384] - - [422, 88.385] - - - [122112, 60673, 1, 384] - - [398, 86.094] - - - [122112, 61056, 1, 384] - - [422, 88.263] - - - [122112, 61057, 1, 384] - - [398, 86.162] - - - [122496, 384, 1, 384] - - [353, 86.265] - - - [122496, 61056, 1, 384] - - [422, 88.335] - - - [122496, 61057, 1, 384] - - [398, 86.071] - - - [122496, 61440, 1, 384] - - [403, 87.367] - - - [122880, 384, 1, 384] - - [353, 85.646] - - - [122880, 61056, 1, 384] - - [430, 87.448] - - - [122880, 61057, 1, 384] - - [398, 84.557] - - - [122880, 61440, 1, 384] - - [430, 87.227] - - - [122880, 61441, 1, 384] - - [398, 83.738] - - - [123264, 384, 1, 384] - - [350, 86.267] - - - [123264, 61440, 1, 384] - - [403, 87.192] - - - [123264, 61441, 1, 384] - - [402, 83.858] - - - [123264, 61824, 1, 384] - - [422, 88.182] - - - [123648, 384, 1, 384] - - [353, 86.499] - - - [123648, 61440, 1, 384] - - [390, 87.425] - - - [123648, 61441, 1, 384] - - [402, 83.914] - - - [123648, 61824, 1, 384] - - [422, 87.964] - - - [123648, 61825, 1, 384] - - [398, 85.94] - - - [124032, 384, 1, 384] - - [350, 86.695] - - - [124032, 61824, 1, 384] - - [422, 87.983] - - - [124032, 61825, 1, 384] - - [398, 85.809] - - - [124032, 62208, 1, 384] - - [422, 87.936] - - - [124416, 384, 1, 384] - - [350, 86.671] - - - [124416, 61824, 1, 384] - - [422, 87.871] - - - [124416, 61825, 1, 384] - - [398, 85.837] - - - [124416, 62208, 1, 384] - - [390, 87.715] - - - [124416, 62209, 1, 384] - - [398, 85.284] - - - [124800, 384, 1, 384] - - [350, 86.171] - - - [124800, 62208, 1, 384] - - [422, 87.844] - - - [124800, 62209, 1, 384] - - [398, 85.552] - - - [124800, 62592, 1, 384] - - [422, 87.874] - - - [125184, 384, 1, 384] - - [350, 86.132] - - - [125184, 62208, 1, 384] - - [390, 87.641] - - - [125184, 62209, 1, 384] - - [398, 85.518] - - - [125184, 62592, 1, 384] - - [351, 87.72] - - - [125184, 62593, 1, 384] - - [398, 85.662] - - - [125568, 384, 1, 384] - - [392, 85.96] - - - [125568, 62592, 1, 384] - - [422, 87.765] - - - [125568, 62593, 1, 384] - - [398, 85.563] - - - [125568, 62976, 1, 384] - - [422, 87.545] - - - [125952, 384, 1, 384] - - [353, 85.553] - - - [125952, 62592, 1, 384] - - [422, 87.251] - - - [125952, 62593, 1, 384] - - [402, 84.99] - - - [125952, 62976, 1, 384] - - [403, 86.999] - - - [125952, 62977, 1, 384] - - [402, 84.785] - - - [126336, 384, 1, 384] - - [350, 85.989] - - - [126336, 62976, 1, 384] - - [422, 87.457] - - - [126336, 62977, 1, 384] - - [430, 85.313] - - - [126336, 63360, 1, 384] - - [403, 87.528] - - - [126720, 384, 1, 384] - - [350, 86.101] - - - [126720, 62976, 1, 384] - - [403, 87.246] - - - [126720, 62977, 1, 384] - - [398, 85.263] - - - [126720, 63360, 1, 384] - - [422, 87.41] - - - [126720, 63361, 1, 384] - - [398, 85.282] - - - [127104, 384, 1, 384] - - [350, 86.389] - - - [127104, 63360, 1, 384] - - [403, 87.441] - - - [127104, 63361, 1, 384] - - [398, 85.391] - - - [127104, 63744, 1, 384] - - [422, 87.538] - - - [127488, 384, 1, 384] - - [350, 86.42] - - - [127488, 63360, 1, 384] - - [403, 87.22] - - - [127488, 63361, 1, 384] - - [398, 84.888] - - - [127488, 63744, 1, 384] - - [452, 87.376] - - - [127488, 63745, 1, 384] - - [402, 84.753] - - - [127872, 384, 1, 384] - - [426, 86.214] - - - [127872, 63744, 1, 384] - - [403, 87.293] - - - [127872, 63745, 1, 384] - - [398, 85.023] - - - [127872, 64128, 1, 384] - - [403, 87.271] - - - [128256, 384, 1, 384] - - [353, 86.587] - - - [128256, 63744, 1, 384] - - [403, 87.063] - - - [128256, 63745, 1, 384] - - [430, 84.936] - - - [128256, 64128, 1, 384] - - [403, 87.059] - - - [768, 1, 1, 384] - - [453, 0.057] - - - [64128, 127489, 1, 384] - - [403, 85.181] - - - [63744, 126721, 1, 384] - - [403, 85.355] - - - [63744, 127105, 1, 384] - - [403, 85.489] - - - [63744, 127489, 1, 384] - - [403, 85.248] - - - [63360, 125953, 1, 384] - - [403, 84.038] - - - [63360, 126337, 1, 384] - - [403, 85.573] - - - [63360, 126721, 1, 384] - - [403, 85.38] - - - [62976, 125185, 1, 384] - - [403, 85.67] - - - [62976, 125569, 1, 384] - - [403, 85.554] - - - [62976, 125953, 1, 384] - - [403, 84.041] - - - [62592, 124417, 1, 384] - - [403, 85.567] - - - [62592, 124801, 1, 384] - - [403, 85.66] - - - [62592, 125185, 1, 384] - - [403, 85.593] - - - [62208, 123649, 1, 384] - - [403, 85.706] - - - [62208, 124033, 1, 384] - - [403, 85.647] - - - [62208, 124417, 1, 384] - - [403, 85.602] - - - [61824, 122881, 1, 384] - - [402, 81.409] - - - [61824, 123265, 1, 384] - - [403, 85.751] - - - [61824, 123649, 1, 384] - - [403, 85.699] - - - [61440, 122113, 1, 384] - - [403, 85.958] - - - [61440, 122497, 1, 384] - - [403, 85.871] - - - [61440, 122881, 1, 384] - - [402, 81.535] - - - [61056, 121345, 1, 384] - - [403, 85.871] - - - [61056, 121729, 1, 384] - - [403, 85.935] - - - [61056, 122113, 1, 384] - - [403, 85.905] - - - [60672, 120577, 1, 384] - - [403, 86.066] - - - [60672, 120961, 1, 384] - - [403, 86.086] - - - [60672, 121345, 1, 384] - - [403, 85.85] - - - [60288, 119809, 1, 384] - - [403, 84.493] - - - [60288, 120193, 1, 384] - - [403, 86.031] - - - [60288, 120577, 1, 384] - - [403, 86.038] - - - [59904, 119041, 1, 384] - - [403, 86.218] - - - [59904, 119425, 1, 384] - - [403, 86.185] - - - [59904, 119809, 1, 384] - - [403, 84.469] - - - [59520, 118273, 1, 384] - - [403, 86.168] - - - [59520, 118657, 1, 384] - - [403, 86.213] - - - [59520, 119041, 1, 384] - - [403, 86.195] - - - [59136, 117505, 1, 384] - - [403, 86.37] - - - [59136, 117889, 1, 384] - - [403, 86.338] - - - [59136, 118273, 1, 384] - - [403, 86.204] - - - [58752, 116737, 1, 384] - - [403, 84.619] - - - [58752, 117121, 1, 384] - - [403, 86.403] - - - [58752, 117505, 1, 384] - - [403, 86.347] - - - [58368, 115969, 1, 384] - - [403, 86.465] - - - [58368, 116353, 1, 384] - - [403, 86.421] - - - [58368, 116737, 1, 384] - - [403, 84.594] - - - [57984, 115201, 1, 384] - - [403, 86.342] - - - [57984, 115585, 1, 384] - - [403, 86.499] - - - [57984, 115969, 1, 384] - - [403, 86.422] - - - [57600, 114433, 1, 384] - - [403, 86.502] - - - [57600, 114817, 1, 384] - - [403, 86.322] - - - [57600, 115201, 1, 384] - - [403, 86.346] - - - [57216, 113665, 1, 384] - - [403, 84.92] - - - [57216, 114049, 1, 384] - - [403, 86.718] - - - [57216, 114433, 1, 384] - - [403, 86.596] - - - [56832, 112897, 1, 384] - - [403, 86.724] - - - [56832, 113281, 1, 384] - - [400, 86.745] - - - [56832, 113665, 1, 384] - - [400, 84.877] - - - [56448, 112129, 1, 384] - - [403, 86.674] - - - [56448, 112513, 1, 384] - - [403, 86.796] - - - [56448, 112897, 1, 384] - - [403, 86.682] - - - [56064, 111361, 1, 384] - - [400, 86.869] - - - [56064, 111745, 1, 384] - - [403, 86.875] - - - [56064, 112129, 1, 384] - - [403, 86.689] - - - [55680, 110593, 1, 384] - - [402, 84.571] - - - [55680, 110977, 1, 384] - - [403, 86.941] - - - [55680, 111361, 1, 384] - - [403, 86.846] - - - [55296, 109825, 1, 384] - - [403, 87.106] - - - [55296, 110209, 1, 384] - - [400, 87.045] - - - [55296, 110593, 1, 384] - - [402, 84.627] - - - [54912, 109057, 1, 384] - - [400, 86.941] - - - [54912, 109441, 1, 384] - - [400, 87.103] - - - [54912, 109825, 1, 384] - - [400, 87.055] - - - [54528, 108289, 1, 384] - - [400, 87.164] - - - [54528, 108673, 1, 384] - - [400, 87.167] - - - [54528, 109057, 1, 384] - - [400, 86.924] - - - [54144, 107521, 1, 384] - - [402, 85.312] - - - [54144, 107905, 1, 384] - - [400, 87.214] - - - [54144, 108289, 1, 384] - - [400, 87.148] - - - [53760, 106753, 1, 384] - - [400, 87.214] - - - [53760, 107137, 1, 384] - - [400, 87.226] - - - [53760, 107521, 1, 384] - - [402, 85.283] - - - [53376, 105985, 1, 384] - - [400, 87.132] - - - [53376, 106369, 1, 384] - - [400, 87.196] - - - [53376, 106753, 1, 384] - - [400, 87.219] - - - [52992, 105217, 1, 384] - - [400, 87.415] - - - [52992, 105601, 1, 384] - - [400, 87.338] - - - [52992, 105985, 1, 384] - - [400, 87.189] - - - [52608, 104449, 1, 384] - - [368, 85.577] - - - [52608, 104833, 1, 384] - - [400, 87.405] - - - [52608, 105217, 1, 384] - - [400, 87.427] - - - [52224, 103681, 1, 384] - - [400, 87.477] - - - [52224, 104065, 1, 384] - - [400, 87.512] - - - [52224, 104449, 1, 384] - - [368, 85.559] - - - [51840, 102913, 1, 384] - - [400, 87.391] - - - [51840, 103297, 1, 384] - - [400, 87.555] - - - [51840, 103681, 1, 384] - - [400, 87.471] - - - [51456, 102145, 1, 384] - - [400, 87.612] - - - [51456, 102529, 1, 384] - - [400, 87.561] - - - [51456, 102913, 1, 384] - - [400, 87.364] - - - [51072, 101377, 1, 384] - - [368, 85.928] - - - [51072, 101761, 1, 384] - - [400, 87.713] - - - [51072, 102145, 1, 384] - - [400, 87.637] - - - [50688, 100609, 1, 384] - - [400, 87.758] - - - [50688, 100993, 1, 384] - - [400, 87.742] - - - [50688, 101377, 1, 384] - - [368, 85.922] - - - [50304, 99841, 1, 384] - - [400, 87.584] - - - [50304, 100225, 1, 384] - - [400, 87.792] - - - [50304, 100609, 1, 384] - - [400, 87.742] - - - [49920, 99073, 1, 384] - - [400, 87.805] - - - [49920, 99457, 1, 384] - - [400, 87.855] - - - [49920, 99841, 1, 384] - - [400, 87.616] - - - [49536, 98305, 1, 384] - - [401, 78.067] - - - [49536, 98689, 1, 384] - - [400, 88.003] - - - [49536, 99073, 1, 384] - - [400, 87.759] - - - [49152, 97537, 1, 384] - - [400, 87.896] - - - [49152, 97921, 1, 384] - - [368, 87.692] - - - [49152, 98305, 1, 384] - - [401, 78.004] - - - [48768, 96769, 1, 384] - - [400, 87.768] - - - [48768, 97153, 1, 384] - - [400, 87.885] - - - [48768, 97537, 1, 384] - - [400, 87.859] - - - [48384, 96001, 1, 384] - - [400, 87.889] - - - [48384, 96385, 1, 384] - - [400, 88.0] - - - [48384, 96769, 1, 384] - - [400, 87.765] - - - [48000, 95233, 1, 384] - - [368, 86.371] - - - [48000, 95617, 1, 384] - - [400, 88.0] - - - [48000, 96001, 1, 384] - - [400, 87.889] - - - [47616, 94465, 1, 384] - - [400, 88.038] - - - [47616, 94849, 1, 384] - - [400, 88.066] - - - [47616, 95233, 1, 384] - - [368, 86.308] - - - [47232, 93697, 1, 384] - - [400, 87.84] - - - [47232, 94081, 1, 384] - - [400, 88.022] - - - [47232, 94465, 1, 384] - - [400, 87.956] - - - [46848, 92929, 1, 384] - - [400, 88.09] - - - [46848, 93313, 1, 384] - - [400, 88.177] - - - [46848, 93697, 1, 384] - - [400, 87.911] - - - [46464, 92161, 1, 384] - - [368, 86.408] - - - [46464, 92545, 1, 384] - - [400, 88.168] - - - [46464, 92929, 1, 384] - - [400, 88.062] - - - [46080, 91393, 1, 384] - - [400, 88.277] - - - [46080, 91777, 1, 384] - - [351, 88.024] - - - [46080, 92161, 1, 384] - - [368, 86.469] - - - [45696, 90625, 1, 384] - - [350, 88.02] - - - [45696, 91009, 1, 384] - - [400, 88.189] - - - [45696, 91393, 1, 384] - - [400, 88.233] - - - [45312, 89857, 1, 384] - - [400, 88.235] - - - [45312, 90241, 1, 384] - - [400, 88.167] - - - [45312, 90625, 1, 384] - - [400, 88.026] - - - [44928, 89089, 1, 384] - - [368, 86.704] - - - [44928, 89473, 1, 384] - - [400, 88.382] - - - [44928, 89857, 1, 384] - - [400, 88.248] - - - [44544, 88321, 1, 384] - - [400, 88.362] - - - [44544, 88705, 1, 384] - - [400, 88.436] - - - [44544, 89089, 1, 384] - - [368, 86.687] - - - [44160, 87553, 1, 384] - - [400, 88.185] - - - [44160, 87937, 1, 384] - - [400, 88.46] - - - [44160, 88321, 1, 384] - - [400, 88.329] - - - [43776, 86785, 1, 384] - - [400, 88.37] - - - [43776, 87169, 1, 384] - - [368, 88.411] - - - [43776, 87553, 1, 384] - - [350, 88.157] - - - [43392, 86017, 1, 384] - - [368, 86.142] - - - [43392, 86401, 1, 384] - - [350, 88.415] - - - [43392, 86785, 1, 384] - - [400, 88.395] - - - [43008, 85249, 1, 384] - - [350, 88.49] - - - [43008, 85633, 1, 384] - - [350, 88.602] - - - [43008, 86017, 1, 384] - - [368, 86.199] - - - [42624, 84481, 1, 384] - - [350, 88.315] - - - [42624, 84865, 1, 384] - - [368, 88.523] - - - [42624, 85249, 1, 384] - - [368, 88.417] - - - [42240, 83713, 1, 384] - - [368, 88.478] - - - [42240, 84097, 1, 384] - - [368, 88.479] - - - [42240, 84481, 1, 384] - - [350, 88.273] - - - [41856, 82945, 1, 384] - - [368, 86.903] - - - [41856, 83329, 1, 384] - - [368, 88.586] - - - [41856, 83713, 1, 384] - - [400, 88.481] - - - [41472, 82177, 1, 384] - - [350, 88.506] - - - [41472, 82561, 1, 384] - - [368, 88.562] - - - [41472, 82945, 1, 384] - - [368, 86.818] - - - [41088, 81409, 1, 384] - - [350, 88.458] - - - [41088, 81793, 1, 384] - - [350, 88.619] - - - [41088, 82177, 1, 384] - - [350, 88.561] - - - [40704, 80641, 1, 384] - - [350, 88.683] - - - [40704, 81025, 1, 384] - - [350, 88.792] - - - [40704, 81409, 1, 384] - - [350, 88.496] - - - [40320, 79873, 1, 384] - - [368, 86.939] - - - [40320, 80257, 1, 384] - - [396, 88.808] - - - [40320, 80641, 1, 384] - - [368, 88.663] - - - [39936, 79105, 1, 384] - - [350, 88.865] - - - [39936, 79489, 1, 384] - - [396, 88.898] - - - [39936, 79873, 1, 384] - - [368, 86.919] - - - [39552, 78337, 1, 384] - - [350, 88.729] - - - [39552, 78721, 1, 384] - - [396, 88.957] - - - [39552, 79105, 1, 384] - - [350, 88.851] - - - [39168, 77569, 1, 384] - - [350, 88.82] - - - [39168, 77953, 1, 384] - - [398, 88.774] - - - [39168, 78337, 1, 384] - - [350, 88.533] - - - [38784, 76801, 1, 384] - - [368, 87.105] - - - [38784, 77185, 1, 384] - - [396, 89.069] - - - [38784, 77569, 1, 384] - - [350, 88.84] - - - [38400, 76033, 1, 384] - - [350, 88.972] - - - [38400, 76417, 1, 384] - - [350, 88.951] - - - [38400, 76801, 1, 384] - - [368, 87.084] - - - [38016, 75265, 1, 384] - - [350, 88.831] - - - [38016, 75649, 1, 384] - - [364, 89.228] - - - [38016, 76033, 1, 384] - - [350, 89.011] - - - [37632, 74497, 1, 384] - - [350, 88.911] - - - [37632, 74881, 1, 384] - - [364, 89.116] - - - [37632, 75265, 1, 384] - - [350, 88.773] - - - [37248, 73729, 1, 384] - - [394, 84.53] - - - [37248, 74113, 1, 384] - - [364, 89.406] - - - [37248, 74497, 1, 384] - - [350, 89.056] - - - [36864, 72961, 1, 384] - - [368, 88.957] - - - [36864, 73345, 1, 384] - - [364, 89.289] - - - [36864, 73729, 1, 384] - - [394, 84.25] - - - [36480, 72193, 1, 384] - - [364, 89.123] - - - [36480, 72577, 1, 384] - - [364, 89.3] - - - [36480, 72961, 1, 384] - - [350, 88.979] - - - [36096, 71425, 1, 384] - - [350, 89.025] - - - [36096, 71809, 1, 384] - - [350, 89.019] - - - [36096, 72193, 1, 384] - - [350, 88.809] - - - [35712, 70657, 1, 384] - - [368, 87.198] - - - [35712, 71041, 1, 384] - - [364, 89.284] - - - [35712, 71425, 1, 384] - - [350, 89.249] - - - [35328, 69889, 1, 384] - - [364, 89.302] - - - [35328, 70273, 1, 384] - - [364, 89.754] - - - [35328, 70657, 1, 384] - - [368, 87.16] - - - [34944, 69121, 1, 384] - - [364, 89.159] - - - [34944, 69505, 1, 384] - - [364, 89.641] - - - [34944, 69889, 1, 384] - - [364, 89.202] - - - [34560, 68353, 1, 384] - - [350, 89.207] - - - [34560, 68737, 1, 384] - - [364, 89.482] - - - [34560, 69121, 1, 384] - - [364, 89.19] - - - [34176, 67585, 1, 384] - - [368, 87.19] - - - [34176, 67969, 1, 384] - - [364, 89.172] - - - [34176, 68353, 1, 384] - - [350, 89.034] - - - [33792, 66817, 1, 384] - - [364, 89.43] - - - [33792, 67201, 1, 384] - - [364, 89.643] - - - [33792, 67585, 1, 384] - - [368, 87.161] - - - [33408, 66049, 1, 384] - - [364, 89.47] - - - [33408, 66433, 1, 384] - - [364, 89.768] - - - [33408, 66817, 1, 384] - - [364, 89.523] - - - [33024, 65281, 1, 384] - - [364, 89.441] - - - [33024, 65665, 1, 384] - - [364, 89.787] - - - [33024, 66049, 1, 384] - - [364, 89.485] - - - [32640, 64513, 1, 384] - - [368, 87.356] - - - [32640, 64897, 1, 384] - - [350, 89.278] - - - [32640, 65281, 1, 384] - - [350, 89.188] - - - [32256, 63745, 1, 384] - - [350, 89.084] - - - [32256, 64129, 1, 384] - - [364, 89.276] - - - [32256, 64513, 1, 384] - - [368, 87.319] - - - [31872, 62977, 1, 384] - - [364, 89.371] - - - [31872, 63361, 1, 384] - - [364, 89.791] - - - [31872, 63745, 1, 384] - - [350, 89.393] - - - [31488, 62209, 1, 384] - - [364, 89.618] - - - [31488, 62593, 1, 384] - - [364, 90.003] - - - [31488, 62977, 1, 384] - - [364, 89.519] - - - [31104, 61441, 1, 384] - - [368, 86.564] - - - [31104, 61825, 1, 384] - - [357, 89.355] - - - [31104, 62209, 1, 384] - - [350, 89.034] - - - [30720, 60673, 1, 384] - - [350, 89.162] - - - [30720, 61057, 1, 384] - - [364, 89.403] - - - [30720, 61441, 1, 384] - - [368, 86.616] - - - [30336, 59905, 1, 384] - - [350, 89.073] - - - [30336, 60289, 1, 384] - - [350, 89.329] - - - [30336, 60673, 1, 384] - - [350, 89.324] - - - [29952, 59137, 1, 384] - - [364, 89.588] - - - [29952, 59521, 1, 384] - - [364, 89.878] - - - [29952, 59905, 1, 384] - - [364, 89.488] - - - [29568, 58369, 1, 384] - - [368, 87.283] - - - [29568, 58753, 1, 384] - - [364, 89.705] - - - [29568, 59137, 1, 384] - - [364, 89.57] - - - [29184, 57601, 1, 384] - - [364, 89.137] - - - [29184, 57985, 1, 384] - - [364, 89.406] - - - [29184, 58369, 1, 384] - - [368, 87.293] - - - [28800, 56833, 1, 384] - - [350, 89.34] - - - [28800, 57217, 1, 384] - - [350, 89.312] - - - [28800, 57601, 1, 384] - - [350, 89.312] - - - [28416, 56065, 1, 384] - - [364, 89.408] - - - [28416, 56449, 1, 384] - - [364, 89.573] - - - [28416, 56833, 1, 384] - - [364, 89.504] - - - [28032, 55297, 1, 384] - - [368, 87.283] - - - [28032, 55681, 1, 384] - - [364, 89.831] - - - [28032, 56065, 1, 384] - - [364, 89.666] - - - [27648, 54529, 1, 384] - - [350, 89.4] - - - [27648, 54913, 1, 384] - - [350, 89.57] - - - [27648, 55297, 1, 384] - - [368, 87.328] - - - [27264, 53761, 1, 384] - - [364, 89.34] - - - [27264, 54145, 1, 384] - - [350, 89.674] - - - [27264, 54529, 1, 384] - - [350, 89.331] - - - [26880, 52993, 1, 384] - - [350, 89.47] - - - [26880, 53377, 1, 384] - - [350, 89.414] - - - [26880, 53761, 1, 384] - - [350, 89.142] - - - [26496, 52225, 1, 384] - - [368, 87.361] - - - [26496, 52609, 1, 384] - - [364, 89.725] - - - [26496, 52993, 1, 384] - - [364, 89.555] - - - [26112, 51457, 1, 384] - - [364, 89.371] - - - [26112, 51841, 1, 384] - - [364, 89.747] - - - [26112, 52225, 1, 384] - - [380, 87.535] - - - [25728, 50689, 1, 384] - - [350, 89.272] - - - [25728, 51073, 1, 384] - - [350, 89.564] - - - [25728, 51457, 1, 384] - - [350, 89.282] - - - [25344, 49921, 1, 384] - - [350, 89.524] - - - [25344, 50305, 1, 384] - - [350, 89.498] - - - [25344, 50689, 1, 384] - - [350, 89.379] - - - [24960, 49153, 1, 384] - - [364, 84.096] - - - [24960, 49537, 1, 384] - - [350, 89.367] - - - [24960, 49921, 1, 384] - - [350, 89.637] - - - [24576, 48385, 1, 384] - - [364, 89.734] - - - [24576, 48769, 1, 384] - - [364, 89.932] - - - [24576, 49153, 1, 384] - - [384, 84.019] - - - [24192, 47617, 1, 384] - - [350, 89.288] - - - [24192, 48001, 1, 384] - - [350, 89.53] - - - [24192, 48385, 1, 384] - - [350, 89.517] - - - [23808, 46849, 1, 384] - - [350, 89.466] - - - [23808, 47233, 1, 384] - - [350, 89.51] - - - [23808, 47617, 1, 384] - - [376, 89.173] - - - [23424, 46081, 1, 384] - - [368, 87.285] - - - [23424, 46465, 1, 384] - - [350, 89.558] - - - [23424, 46849, 1, 384] - - [350, 89.205] - - - [23040, 45313, 1, 384] - - [350, 89.447] - - - [23040, 45697, 1, 384] - - [364, 89.743] - - - [23040, 46081, 1, 384] - - [380, 87.325] - - - [22656, 44545, 1, 384] - - [364, 89.368] - - - [22656, 44929, 1, 384] - - [364, 89.614] - - - [22656, 45313, 1, 384] - - [364, 89.399] - - - [22272, 43777, 1, 384] - - [350, 89.497] - - - [22272, 44161, 1, 384] - - [350, 89.675] - - - [22272, 44545, 1, 384] - - [350, 89.283] - - - [21888, 43009, 1, 384] - - [350, 87.227] - - - [21888, 43393, 1, 384] - - [364, 89.418] - - - [21888, 43777, 1, 384] - - [350, 89.163] - - - [21504, 42241, 1, 384] - - [350, 89.242] - - - [21504, 42625, 1, 384] - - [350, 89.256] - - - [21504, 43009, 1, 384] - - [368, 87.231] - - - [21120, 41473, 1, 384] - - [350, 89.554] - - - [21120, 41857, 1, 384] - - [350, 89.978] - - - [21120, 42241, 1, 384] - - [350, 89.949] - - - [20736, 40705, 1, 384] - - [376, 89.175] - - - [20736, 41089, 1, 384] - - [376, 89.267] - - - [20736, 41473, 1, 384] - - [376, 89.021] - - - [20352, 39937, 1, 384] - - [350, 87.217] - - - [20352, 40321, 1, 384] - - [364, 89.498] - - - [20352, 40705, 1, 384] - - [350, 89.282] - - - [19968, 39169, 1, 384] - - [364, 89.297] - - - [19968, 39553, 1, 384] - - [350, 89.195] - - - [19968, 39937, 1, 384] - - [350, 87.318] - - - [19584, 38401, 1, 384] - - [350, 89.159] - - - [19584, 38785, 1, 384] - - [350, 89.736] - - - [19584, 39169, 1, 384] - - [350, 89.643] - - - [19200, 37633, 1, 384] - - [350, 89.126] - - - [19200, 38017, 1, 384] - - [350, 89.154] - - - [19200, 38401, 1, 384] - - [359, 89.041] - - - [18816, 36865, 1, 384] - - [350, 86.568] - - - [18816, 37249, 1, 384] - - [350, 89.322] - - - [18816, 37633, 1, 384] - - [350, 89.28] - - - [18432, 36097, 1, 384] - - [350, 89.011] - - - [18432, 36481, 1, 384] - - [364, 89.51] - - - [18432, 36865, 1, 384] - - [368, 86.678] - - - [18048, 35329, 1, 384] - - [364, 88.848] - - - [18048, 35713, 1, 384] - - [350, 89.004] - - - [18048, 36097, 1, 384] - - [350, 88.957] - - - [17664, 34561, 1, 384] - - [350, 88.946] - - - [17664, 34945, 1, 384] - - [350, 88.937] - - - [17664, 35329, 1, 384] - - [350, 88.656] - - - [17280, 33793, 1, 384] - - [368, 86.808] - - - [17280, 34177, 1, 384] - - [357, 88.99] - - - [17280, 34561, 1, 384] - - [350, 88.764] - - - [16896, 33025, 1, 384] - - [350, 89.795] - - - [16896, 33409, 1, 384] - - [350, 89.408] - - - [16896, 33793, 1, 384] - - [368, 86.788] - - - [16512, 32257, 1, 384] - - [362, 88.768] - - - [16512, 32641, 1, 384] - - [364, 88.81] - - - [16512, 33025, 1, 384] - - [350, 88.866] - - - [16128, 31489, 1, 384] - - [350, 88.748] - - - [16128, 31873, 1, 384] - - [350, 89.057] - - - [16128, 32257, 1, 384] - - [350, 88.726] - - - [15744, 30721, 1, 384] - - [368, 86.924] - - - [15744, 31105, 1, 384] - - [350, 88.73] - - - [15744, 31489, 1, 384] - - [350, 88.824] - - - [15360, 29953, 1, 384] - - [350, 88.894] - - - [15360, 30337, 1, 384] - - [362, 89.06] - - - [15360, 30721, 1, 384] - - [368, 86.944] - - - [14976, 29185, 1, 384] - - [350, 88.575] - - - [14976, 29569, 1, 384] - - [353, 88.405] - - - [14976, 29953, 1, 384] - - [350, 88.93] - - - [14592, 28417, 1, 384] - - [366, 88.614] - - - [14592, 28801, 1, 384] - - [364, 88.662] - - - [14592, 29185, 1, 384] - - [350, 88.249] - - - [14208, 27649, 1, 384] - - [350, 86.292] - - - [14208, 28033, 1, 384] - - [350, 88.259] - - - [14208, 28417, 1, 384] - - [350, 88.06] - - - [13824, 26881, 1, 384] - - [350, 88.614] - - - [13824, 27265, 1, 384] - - [350, 88.535] - - - [13824, 27649, 1, 384] - - [368, 86.437] - - - [13440, 26113, 1, 384] - - [362, 88.146] - - - [13440, 26497, 1, 384] - - [350, 88.447] - - - [13440, 26881, 1, 384] - - [366, 88.351] - - - [13056, 25345, 1, 384] - - [350, 88.211] - - - [13056, 25729, 1, 384] - - [350, 88.349] - - - [13056, 26113, 1, 384] - - [350, 88.199] - - - [12672, 24577, 1, 384] - - [350, 86.708] - - - [12672, 24961, 1, 384] - - [350, 88.828] - - - [12672, 25345, 1, 384] - - [350, 88.933] - - - [12288, 23809, 1, 384] - - [353, 88.428] - - - [12288, 24193, 1, 384] - - [353, 88.619] - - - [12288, 24577, 1, 384] - - [368, 85.402] - - - [11904, 23041, 1, 384] - - [362, 88.122] - - - [11904, 23425, 1, 384] - - [362, 88.15] - - - [11904, 23809, 1, 384] - - [366, 88.188] - - - [11520, 22273, 1, 384] - - [362, 88.105] - - - [11520, 22657, 1, 384] - - [350, 87.776] - - - [11520, 23041, 1, 384] - - [364, 87.747] - - - [11136, 21505, 1, 384] - - [363, 85.833] - - - [11136, 21889, 1, 384] - - [362, 88.203] - - - [11136, 22273, 1, 384] - - [362, 88.163] - - - [10752, 20737, 1, 384] - - [350, 88.238] - - - [10752, 21121, 1, 384] - - [350, 88.089] - - - [10752, 21505, 1, 384] - - [350, 86.337] - - - [10368, 19969, 1, 384] - - [350, 88.384] - - - [10368, 20353, 1, 384] - - [350, 88.308] - - - [10368, 20737, 1, 384] - - [350, 88.208] - - - [9984, 19201, 1, 384] - - [350, 88.314] - - - [9984, 19585, 1, 384] - - [350, 88.296] - - - [9984, 19969, 1, 384] - - [362, 88.207] - - - [9600, 18433, 1, 384] - - [351, 86.287] - - - [9600, 18817, 1, 384] - - [350, 88.204] - - - [9600, 19201, 1, 384] - - [362, 88.058] - - - [9216, 17665, 1, 384] - - [350, 88.632] - - - [9216, 18049, 1, 384] - - [350, 88.339] - - - [9216, 18433, 1, 384] - - [350, 86.354] - - - [8832, 16897, 1, 384] - - [357, 88.191] - - - [8832, 17281, 1, 384] - - [357, 88.408] - - - [8832, 17665, 1, 384] - - [357, 88.281] - - - [8448, 16129, 1, 384] - - [350, 88.936] - - - [8448, 16513, 1, 384] - - [350, 88.776] - - - [8448, 16897, 1, 384] - - [359, 88.084] - - - [8064, 15361, 1, 384] - - [350, 86.407] - - - [8064, 15745, 1, 384] - - [350, 88.528] - - - [8064, 16129, 1, 384] - - [353, 88.248] - - - [7680, 14593, 1, 384] - - [350, 87.375] - - - [7680, 14977, 1, 384] - - [350, 88.059] - - - [7680, 15361, 1, 384] - - [350, 85.97] - - - [7296, 13825, 1, 384] - - [350, 87.56] - - - [7296, 14209, 1, 384] - - [350, 87.512] - - - [7296, 14593, 1, 384] - - [350, 87.766] - - - [6912, 13057, 1, 384] - - [357, 87.319] - - - [6912, 13441, 1, 384] - - [350, 87.442] - - - [6912, 13825, 1, 384] - - [353, 87.158] - - - [6528, 12289, 1, 384] - - [354, 84.224] - - - [6528, 12673, 1, 384] - - [357, 87.109] - - - [6528, 13057, 1, 384] - - [353, 86.766] - - - [6144, 11521, 1, 384] - - [353, 86.863] - - - [6144, 11905, 1, 384] - - [350, 86.377] - - - [6144, 12289, 1, 384] - - [350, 84.586] - - - [5760, 10753, 1, 384] - - [353, 86.563] - - - [5760, 11137, 1, 384] - - [350, 86.836] - - - [5760, 11521, 1, 384] - - [350, 86.952] - - - [5376, 9985, 1, 384] - - [350, 85.52] - - - [5376, 10369, 1, 384] - - [350, 86.175] - - - [5376, 10753, 1, 384] - - [350, 86.336] - - - [4992, 9217, 1, 384] - - [350, 83.866] - - - [4992, 9601, 1, 384] - - [350, 86.495] - - - [4992, 9985, 1, 384] - - [350, 85.623] - - - [4608, 8449, 1, 384] - - [353, 85.079] - - - [4608, 8833, 1, 384] - - [354, 85.098] - - - [4608, 9217, 1, 384] - - [350, 83.331] - - - [4224, 7681, 1, 384] - - [350, 83.513] - - - [4224, 8065, 1, 384] - - [350, 83.067] - - - [4224, 8449, 1, 384] - - [350, 83.47] - - - [3840, 6913, 1, 384] - - [350, 82.965] - - - [3840, 7297, 1, 384] - - [350, 83.23] - - - [3840, 7681, 1, 384] - - [350, 83.688] - - - [3456, 6145, 1, 384] - - [350, 75.604] - - - [3456, 6529, 1, 384] - - [351, 81.045] - - - [3456, 6913, 1, 384] - - [350, 81.242] - - - [3072, 5377, 1, 384] - - [350, 76.614] - - - [3072, 5761, 1, 384] - - [350, 76.931] - - - [3072, 6145, 1, 384] - - [350, 78.452] - - - [2688, 4609, 1, 384] - - [346, 72.308] - - - [2688, 4993, 1, 384] - - [347, 75.653] - - - [2688, 5377, 1, 384] - - [348, 76.116] - - - [2304, 3841, 1, 384] - - [341, 71.362] - - - [2304, 4225, 1, 384] - - [344, 71.713] - - - [2304, 4609, 1, 384] - - [337, 72.149] - - - [1920, 3073, 1, 384] - - [341, 63.814] - - - [1920, 3457, 1, 384] - - [342, 70.123] - - - [1920, 3841, 1, 384] - - [336, 68.028] - - - [1536, 2305, 1, 384] - - [336, 60.722] - - - [1536, 2689, 1, 384] - - [338, 64.392] - - - [1536, 3073, 1, 384] - - [339, 64.52] - - - [1152, 1537, 1, 384] - - [335, 47.579] - - - [1152, 1921, 1, 384] - - [336, 54.865] - - - [1152, 2305, 1, 384] - - [337, 54.187] - - - [768, 1153, 1, 384] - - [333, 35.227] - - - [768, 1537, 1, 384] - - [334, 45.293] - - - [384, 769, 1, 384] - - [331, 20.44] - - - [512, 1025, 1, 512] - - [333, 32.551] - - - [1024, 2049, 1, 512] - - [407, 55.903] - - - [1536, 3073, 1, 512] - - [410, 69.281] - - - [2048, 4097, 1, 512] - - [350, 77.876] - - - [2560, 5121, 1, 512] - - [350, 78.958] - - - [3072, 6145, 1, 512] - - [350, 83.924] - - - [3584, 7169, 1, 512] - - [350, 84.559] - - - [1024, 1024, 8, 1024] - - [225, 37.66] - - - [2048, 2048, 4, 2048] - - [226, 55.362] - - - [4096, 4096, 2, 4096] - - [227, 96.942] - - - [8192, 8192, 1, 8192] - - [227, 97.575] - - - [16384, 16384, 1, 16384] - - [227, 99.795] - - - [768, 768, 1, 768] - - [231, 43.067] - - - [1152, 1152, 1, 1152] - - [229, 70.518] - - - [1536, 1536, 1, 1536] - - [232, 76.206] - - - [1920, 1920, 1, 1920] - - [233, 82.25] - - - [2304, 2304, 1, 2304] - - [234, 92.123] - - - [2688, 2688, 1, 2688] - - [235, 87.132] - - - [3072, 3072, 1, 3072] - - [236, 95.558] - - - [3456, 3456, 1, 3456] - - [237, 95.674] - - - [3840, 3840, 1, 3840] - - [238, 97.134] - - - [4224, 4224, 1, 4224] - - [239, 97.539] - - - [4992, 4992, 1, 4992] - - [234, 97.764] - - - [5376, 5376, 1, 5376] - - [233, 96.198] - - - [6144, 6144, 1, 6144] - - [234, 99.046] - - - [6528, 6528, 1, 6528] - - [241, 97.946] - - - [6912, 6912, 1, 6912] - - [239, 97.667] - - - [7296, 7296, 1, 7296] - - [241, 97.983] - - - [7680, 7680, 1, 7680] - - [239, 98.739] - - - [1024, 1024, 1, 2048] - - [242, 64.821] - - - [1024, 1024, 1, 3072] - - [243, 68.93] - - - [1024, 2048, 1, 11264] - - [244, 76.197] - - - [1024, 2048, 1, 15360] - - [245, 76.452] - - - [1024, 2048, 1, 3072] - - [244, 74.102] - - - [1024, 2048, 1, 7168] - - [244, 75.65] - - - [1024, 4096, 1, 13312] - - [246, 91.746] - - - [1024, 4096, 1, 5120] - - [247, 90.65] - - - [1024, 8192, 1, 9216] - - [248, 92.392] - - - [2048, 2048, 1, 4096] - - [249, 83.954] - - - [2048, 2048, 1, 5120] - - [250, 86.155] - - - [2048, 2048, 1, 6144] - - [251, 84.986] - - - [2048, 2048, 1, 7168] - - [252, 83.173] - - - [2048, 4096, 1, 14336] - - [253, 92.588] - - - [2048, 4096, 1, 6144] - - [254, 92.096] - - - [2048, 8192, 1, 10240] - - [242, 94.987] - - - [256, 256, 1, 512] - - [255, 6.691] - - - [3072, 4096, 1, 15360] - - [256, 99.323] - - - [3072, 4096, 1, 7168] - - [256, 98.993] - - - [3072, 8192, 1, 11264] - - [254, 99.396] - - - [4096, 4096, 1, 10240] - - [257, 94.88] - - - [4096, 4096, 1, 11264] - - [258, 94.974] - - - [4096, 4096, 1, 12288] - - [259, 94.911] - - - [4096, 4096, 1, 13312] - - [258, 95.034] - - - [4096, 4096, 1, 14336] - - [259, 95.025] - - - [4096, 4096, 1, 15360] - - [258, 95.07] - - - [4096, 4096, 1, 8192] - - [260, 94.473] - - - [4096, 4096, 1, 9216] - - [258, 94.893] - - - [4096, 8192, 1, 12288] - - [261, 97.713] - - - [512, 512, 1, 1024] - - [262, 31.259] - - - [5120, 8192, 1, 13312] - - [254, 96.746] - - - [6144, 8192, 1, 14336] - - [254, 99.532] - - - [7168, 8192, 1, 15360] - - [254, 98.548] - - - [8192, 8192, 1, 16384] - - [263, 97.775] - - - [1024, 1024, 2, 4096] - - [242, 74.316] - - - [1024, 1024, 2, 5120] - - [244, 75.126] - - - [128, 128, 2, 512] - - [255, 3.259] - - - [2048, 2048, 2, 10240] - - [256, 92.451] - - - [2048, 2048, 2, 11264] - - [264, 92.491] - - - [2048, 2048, 2, 8192] - - [265, 91.95] - - - [2048, 2048, 2, 9216] - - [256, 92.394] - - - [256, 256, 2, 1024] - - [255, 15.589] - - - [4096, 4096, 2, 16384] - - [263, 97.694] - - - [512, 512, 2, 2048] - - [266, 45.997] - - - [1024, 1024, 3, 6144] - - [248, 85.459] - - - [1024, 1024, 3, 7168] - - [248, 85.68] - - - [2048, 2048, 3, 12288] - - [264, 99.208] - - - [2048, 2048, 3, 13312] - - [254, 99.259] - - - [2048, 2048, 3, 14336] - - [267, 99.225] - - - [2048, 2048, 3, 15360] - - [268, 99.235] - - - [512, 512, 3, 3072] - - [269, 55.21] - - - [1024, 1024, 4, 8192] - - [270, 80.106] - - - [1024, 1024, 4, 9216] - - [271, 91.434] - - - [128, 128, 4, 1024] - - [272, 7.695] - - - [2048, 2048, 4, 16384] - - [273, 44.087] - - - [256, 256, 4, 2048] - - [274, 25.766] - - - [512, 512, 4, 4096] - - [275, 67.207] - - - [64, 64, 4, 512] - - [276, 1.575] - - - [1024, 1024, 5, 10240] - - [254, 95.725] - - - [1024, 1024, 5, 11264] - - [277, 95.541] - - - [512, 512, 5, 5120] - - [278, 68.408] - - - [1024, 1024, 6, 12288] - - [279, 90.231] - - - [1024, 1024, 6, 13312] - - [280, 90.527] - - - [256, 256, 6, 3072] - - [281, 31.43] - - - [512, 512, 6, 6144] - - [261, 80.948] - - - [1024, 1024, 7, 14336] - - [260, 88.813] - - - [1024, 1024, 7, 15360] - - [282, 88.875] - - - [512, 512, 7, 7168] - - [283, 52.676] - - - [1024, 1024, 8, 16384] - - [284, 53.337] - - - [128, 128, 8, 2048] - - [285, 8.022] - - - [256, 256, 8, 4096] - - [286, 27.223] - - - [32, 32, 8, 512] - - [262, 0.67] - - - [512, 512, 8, 8192] - - [287, 17.468] - - - [64, 64, 8, 1024] - - [262, 2.24] - - - [512, 512, 9, 9216] - - [288, 61.208] - - - [256, 256, 10, 5120] - - [289, 38.321] - - - [512, 512, 10, 10240] - - [290, 67.597] - - - [512, 512, 11, 11264] - - [291, 71.195] - - - [128, 128, 12, 3072] - - [292, 9.673] - - - [256, 256, 12, 6144] - - [293, 36.782] - - - [512, 512, 12, 12288] - - [294, 54.771] - - - [512, 512, 13, 13312] - - [295, 66.377] - - - [256, 256, 14, 7168] - - [294, 36.06] - - - [512, 512, 14, 14336] - - [296, 47.973] - - - [512, 512, 15, 15360] - - [297, 52.736] - - - [128, 128, 16, 4096] - - [298, 6.562] - - - [256, 256, 16, 8192] - - [299, 9.909] - - - [32, 32, 16, 1024] - - [300, 0.584] - - - [512, 512, 16, 16384] - - [301, 91.972] - - - [64, 64, 16, 2048] - - [302, 2.019] - - - [256, 256, 18, 9216] - - [303, 20.658] - - - [128, 128, 20, 5120] - - [304, 9.686] - - - [256, 256, 20, 10240] - - [305, 18.355] - - - [256, 256, 22, 11264] - - [256, 19.873] - - - [128, 128, 24, 6144] - - [306, 9.309] - - - [256, 256, 24, 12288] - - [307, 13.467] - - - [64, 64, 24, 3072] - - [308, 2.474] - - - [256, 256, 26, 13312] - - [309, 92.025] - - - [128, 128, 28, 7168] - - [310, 9.665] - - - [256, 256, 28, 14336] - - [280, 66.832] - - - [256, 256, 30, 15360] - - [311, 71.605] - - - [128, 128, 32, 8192] - - [312, 4.094] - - - [256, 256, 32, 16384] - - [275, 76.406] - - - [32, 32, 32, 2048] - - [313, 0.535] - - - [64, 64, 32, 4096] - - [314, 1.682] - - - [128, 128, 36, 9216] - - [248, 9.831] - - - [128, 128, 40, 10240] - - [315, 9.459] - - - [64, 64, 40, 5120] - - [316, 2.522] - - - [128, 128, 44, 11264] - - [271, 73.615] - - - [128, 128, 48, 12288] - - [244, 77.025] - - - [32, 32, 48, 3072] - - [317, 0.625] - - - [64, 64, 48, 6144] - - [318, 2.39] - - - [128, 128, 52, 13312] - - [244, 77.699] - - - [128, 128, 56, 14336] - - [319, 49.512] - - - [64, 64, 56, 7168] - - [281, 2.526] - - - [128, 128, 60, 15360] - - [320, 53.128] - - - [128, 128, 64, 16384] - - [321, 56.728] - - - [32, 32, 64, 4096] - - [322, 0.42] - - - [64, 64, 64, 8192] - - [323, 36.465] - - - [64, 64, 72, 9216] - - [262, 37.434] - - - [32, 32, 80, 5120] - - [324, 0.628] - - - [64, 64, 80, 10240] - - [262, 37.946] - - - [64, 64, 88, 11264] - - [262, 38.19] - - - [32, 32, 96, 6144] - - [325, 0.592] - - - [64, 64, 96, 12288] - - [262, 38.343] - - - [64, 64, 104, 13312] - - [262, 38.234] - - - [32, 32, 112, 7168] - - [326, 9.154] - - - [64, 64, 112, 14336] - - [262, 26.486] - - - [64, 64, 120, 15360] - - [327, 28.89] - - - [32, 32, 128, 8192] - - [326, 10.568] - - - [64, 64, 128, 16384] - - [328, 30.662] - - - [32, 32, 144, 9216] - - [329, 11.343] - - - [32, 32, 160, 10240] - - [326, 12.66] - - - [32, 32, 176, 11264] - - [330, 13.317] - - - [32, 32, 192, 12288] - - [255, 14.273] - - - [32, 32, 208, 13312] - - [262, 15.262] - - - [32, 32, 224, 14336] - - [262, 12.166] - - - [32, 32, 240, 15360] - - [327, 13.234] - - - [32, 32, 256, 16384] - - [262, 14.13] - - - [512, 512, 11, 512] - - [489, 53.606] - - - [512, 512, 21, 512] - - [489, 58.098] - - - [512, 512, 31, 512] - - [489, 70.428] - - - [512, 512, 41, 512] - - [489, 78.047] - - - [512, 512, 51, 512] - - [489, 76.725] - - - [512, 512, 61, 512] - - [489, 81.402] - - - [512, 512, 71, 512] - - [489, 79.37] - - - [512, 512, 81, 512] - - [489, 83.325] - - - [512, 512, 91, 512] - - [489, 81.232] -- null diff --git a/library/src/blas3/Tensile/Logic/asm_full/aldebaran_104cu/aldebaran_Cijk_Ailk_Bjlk_DB.yaml b/library/src/blas3/Tensile/Logic/asm_full/aldebaran_104cu/aldebaran_Cijk_Ailk_Bjlk_DB.yaml deleted file mode 100644 index 03a321d..0000000 --- a/library/src/blas3/Tensile/Logic/asm_full/aldebaran_104cu/aldebaran_Cijk_Ailk_Bjlk_DB.yaml +++ /dev/null @@ -1,155832 +0,0 @@ -- {MinimumRequiredVersion: 4.32.1} -- aldebaran -- {Architecture: gfx90a, CUCount: 104} -- [Device 0050, Device 0051, Device 0052, Device 0054, Device 0062, Device 7400, Device - 740c] -- AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] -- - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 512 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MIArchVgpr: false - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 0 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 0 - SourceSwap: false - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 2048 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 2 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MIArchVgpr: false - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 4 - NumGlobalWriteVectorsPerThread: 2 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 4 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 1 - SourceSwap: false - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 8, 2] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 1024 - LdsOffsetA: 0 - LdsOffsetB: 512 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 16 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MIArchVgpr: false - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 2 - SourceSwap: false - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 1024 - LdsOffsetA: 0 - LdsOffsetB: 512 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MIArchVgpr: false - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 4 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 3 - SourceSwap: false - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 1024 - LdsOffsetA: 0 - LdsOffsetB: 512 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MIArchVgpr: false - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 4 - SourceSwap: false - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 512 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MIArchVgpr: false - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 5 - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 512 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MIArchVgpr: false - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 0 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 6 - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MIArchVgpr: false - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 7 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_SE_FL0_WGM11 - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MIArchVgpr: false - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 8 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_SE_FL0_WGM8 - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 1 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MIArchVgpr: false - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 9 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_SE_FL1_WGM8 - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 4 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MIArchVgpr: false - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 10 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x4_SE_FL0_WGM11 - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 4 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 3 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 11 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG16_16_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 12 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG8_32_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 13 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 1 - LSPB: 2 - LVCA: 64 - LVCB: 32 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 14 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_64_WG64_4_1_WGM11 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 15 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 16 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 17 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM10 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 10 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: -1 - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {3: 512} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {0: 128, 1: 128} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: DGEMM_Aldebaran_PKFixedAtomic512Latest - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: Branch - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 1 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 18 - SolutionNameMin: DGEMM_Aldebaran_PKFixedAtomic512Latest - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: -1 - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {3: 512} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {0: 128, 1: 128} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: DGEMM_Aldebaran_PKFixedAtomic512_104 - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: Branch - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 1 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 19 - SolutionNameMin: DGEMM_Aldebaran_PKFixedAtomic512_104 - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 20 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 21 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 22 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 23 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 24 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 25 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 26 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 27 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 28 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 29 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 30 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 31 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 32 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 33 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 34 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 35 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 36 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 37 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 38 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 39 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 40 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 41 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 42 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 43 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 44 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 45 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 46 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 47 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 48 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 49 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 50 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 51 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 52 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 53 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 54 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 55 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 56 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 57 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 58 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 59 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 60 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 61 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 62 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 63 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 64 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 65 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 66 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 67 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 68 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 69 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 70 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 71 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 72 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 73 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 74 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 75 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 76 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 77 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 78 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 79 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 80 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 81 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 82 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 83 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 84 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 85 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 86 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 87 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 88 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 89 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 90 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 91 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 92 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 93 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 94 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 95 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 96 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 97 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 98 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 99 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 100 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 101 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 102 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 103 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 104 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 105 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 106 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 107 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 108 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 109 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 110 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 111 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 112 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 113 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 114 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 115 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 116 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS128_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 117 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 118 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 119 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 120 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 121 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 122 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 123 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 124 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 125 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 126 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 127 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 128 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 129 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 130 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 131 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 132 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 133 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 134 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 135 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 136 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 137 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 138 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU32_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 139 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 140 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 141 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 142 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 143 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 144 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 145 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 146 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 147 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 148 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 149 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 150 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 151 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 152 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 153 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 154 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 155 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 156 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 157 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 158 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 159 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 160 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 161 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 162 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 163 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 164 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 165 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 166 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 167 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 168 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 169 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 170 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 171 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 172 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 173 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 174 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 175 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 176 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 177 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 178 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 179 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT4_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 180 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT4_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 181 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 182 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS128_TT4_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 16 - LSCB: 16 - LSPA: 4 - LSPB: 4 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 256 - LdsNumElementsAlignedA: 128 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 256 - LdsOffsetB: 128 - LdsOffsetB_Blk: 384 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 1] - MIWaveTile: [1, 1] - MIWaveTileA: 1 - MIWaveTileB: 1 - MacroTile0: 16 - MacroTile1: 16 - MacroTileA: 16 - MacroTileB: 16 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 4 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 64 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 183 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT16x16x8_MI16x16x4x1_SN_TT1_16_WG16_4_1 - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 16 - SubGroupA: 4 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [1, 16] - ThreadTile0: 4 - ThreadTile1: 1 - ThreadTileA: 4 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 184 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 185 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 186 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 187 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 188 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 189 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 190 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 191 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 192 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 193 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 194 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_PLR5_SU0_SUS0_SSO8_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 2] - MIWaveTileA: 3 - MIWaveTileB: 2 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 3 - NumLoadsB: 2 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 195 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 32] - ThreadTile0: 12 - ThreadTile1: 2 - ThreadTileA: 12 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 196 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: 1 - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 197 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 198 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO8_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 199 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 200 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 201 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: 1 - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 202 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: 1 - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 203 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_SSO4_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 204 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU0_SUS0_SSO4_TT2_96_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 205 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW2_NLCA3_NLCB1_PLR5_SU32_SUS128_SSO4_TT3_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 206 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 207 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT2_64_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 208 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT4_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 209 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 210 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 1 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 211 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 212 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 1 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 213 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT2_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 214 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 215 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 216 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 217 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 218 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW2_NLCA3_NLCB1_PLR5_SU0_SUS0_SSO4_TT3_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 219 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU0_SUS0_SSO4_TT2_96_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 220 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW2_NLCA3_NLCB1_PLR5_SU0_SUS0_SSO8_TT3_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 221 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO8_TT4_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 222 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 223 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_32_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 224 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO8_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 225 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 226 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO8_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 227 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT2_64_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 228 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT4_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 229 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 230 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT4_32_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 231 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT2_32_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 232 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU32_SUS256_SSO8_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 233 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO4_TT4_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 234 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT2_32_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 235 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS256_SSO4_TT4_48_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 236 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS128_SSO4_TT4_48_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2560 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2560 - LdsOffsetB_Blk: 6656 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [5, 3] - MIWaveTileA: 5 - MIWaveTileB: 3 - MacroTile0: 160 - MacroTile1: 96 - MacroTileA: 160 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 60 - NumGlobalWriteVectorsPerThread: 60 - NumLoadsA: 10 - NumLoadsB: 6 - NumLoadsCoalescedA: 5 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 237 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT160x96x16_MI16x16x4x1_SN_GRVW1_NLCA5_NLCB3_PLR5_SU0_SUS0_SSO4_TT5_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [5, 48] - ThreadTile0: 20 - ThreadTile1: 3 - ThreadTileA: 20 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 238 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT4_64_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 239 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT4_64_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 240 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO8_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 241 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT8_32_WG16_16_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 242 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_64_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 243 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO8_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 244 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT8_32_WG16_16_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 245 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW1_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT8_32_WG16_16_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 246 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT4_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [4, 3] - MIWaveTileA: 4 - MIWaveTileB: 3 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 247 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU32_SUS256_SSO4_TT4_48_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 48] - ThreadTile0: 16 - ThreadTile1: 3 - ThreadTileA: 16 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 248 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT8_32_WG16_16_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 249 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT8_32_WG16_16_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 250 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT8_32_WG16_16_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 251 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT2_32_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 252 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_32_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 253 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO4_TT2_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 254 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS256_SSO4_TT2_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 255 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_ - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 0 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 256 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 257 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 258 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 259 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 260 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_48_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 261 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_48_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 262 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 263 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 264 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_48_VW2_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 265 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 266 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 267 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 268 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 269 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 270 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 271 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_48_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 272 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 273 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 274 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 275 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 276 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 277 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 278 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 279 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 280 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 281 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 282 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 283 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 284 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 285 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 286 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 287 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 288 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 289 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 290 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 291 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 292 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 293 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 294 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 295 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 296 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 297 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 298 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 299 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 300 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 301 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 302 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 303 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 304 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 305 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 306 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 307 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 308 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 309 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 310 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 311 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 312 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 313 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 314 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB2_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 315 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 316 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 317 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 318 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 319 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 320 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 321 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 322 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 323 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 324 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 325 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 326 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 327 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 328 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 329 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 330 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 331 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 332 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 333 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 334 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 335 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 336 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB4_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 337 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB4_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 338 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB4_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 339 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 340 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB4_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 341 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 342 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 343 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_32_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 344 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_48_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 345 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS256_SVW2_TT2_48_VW2_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 346 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 347 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 348 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 349 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 350 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 2] - MIWaveTileA: 3 - MIWaveTileB: 2 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 3 - NumLoadsB: 2 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 351 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU0_SUS0_SVW1_TT3_32_VW1_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 32] - ThreadTile0: 12 - ThreadTile1: 2 - ThreadTileA: 12 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 352 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 353 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 2] - MIWaveTileA: 3 - MIWaveTileB: 2 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 3 - NumLoadsB: 2 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 354 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU0_SUS0_SVW1_TT3_32_VW1_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 32] - ThreadTile0: 12 - ThreadTile1: 2 - ThreadTileA: 12 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 2] - MIWaveTileA: 3 - MIWaveTileB: 2 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 3 - NumLoadsB: 2 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 355 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU32_SUS128_SVW1_TT3_32_VW1_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 32] - ThreadTile0: 12 - ThreadTile1: 2 - ThreadTileA: 12 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 356 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 357 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 8 - LSPB: 2 - LVCA: 32 - LVCB: 128 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 768 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 768 - LdsOffsetB_Blk: 2816 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 358 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x8_MI16x16x4x1_SN_AMAS0_GRVW1_NEPBS2_NLCA3_NLCB1_PLR3_SU0_SUS0_SVW1_TT3_64_VW1_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 359 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU32_SUS256_SVW1_TT3_64_VW1_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 360 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 361 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 362 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU32_SUS128_SVW1_TT3_64_VW1_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 363 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 364 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 365 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 366 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 367 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU32_SUS256_SVW1_TT3_64_VW1_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 368 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 369 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 370 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 371 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 372 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 373 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 374 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 375 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 376 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 377 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB4_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 378 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 379 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 380 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB4_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 381 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB4_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 4 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 382 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB4_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 383 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 384 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_48_VW2_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 385 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_64_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2560 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 3] - MIWaveTileA: 2 - MIWaveTileB: 3 - MacroTile0: 64 - MacroTile1: 96 - MacroTileA: 64 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsA: 2 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 386 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_48_VW2_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 48] - ThreadTile0: 8 - ThreadTile1: 3 - ThreadTileA: 8 - ThreadTileB: 3 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 387 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 388 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 389 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 390 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB2_PLR3_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 391 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 392 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 393 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS1_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 394 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_AMAS0_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU0_SUS0_SVW1_TT3_64_VW1_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 395 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 396 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU32_SUS256_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 397 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 398 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB3_PLR5_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 399 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 400 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 401 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 402 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 403 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 404 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 405 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB1_PLR3_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 406 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 407 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB2_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 408 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 409 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 410 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 411 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 16 - LSPB: 8 - LVCA: 16 - LVCB: 32 - LVPA: 8 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 412 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB2_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 413 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 414 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 415 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_32_VW2_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 416 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 417 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 418 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS0_SVW2_TT2_64_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 419 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 420 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU0_SUS0_SVW2_TT2_96_VW2_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 768 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 421 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x8_MI16x16x4x1_SN_AMAS3_GRVW1_NEPBS2_NLCA1_NLCB3_PLR3_SU32_SUS128_SVW2_TT2_96_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 422 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 423 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 424 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 425 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA2_NLCB1_PLR5_SU32_SUS128_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 426 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA1_NLCB2_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 427 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS1_NLCA2_NLCB2_PLR3_SU0_SUS0_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 4 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 428 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_AMAS3_GRVW2_NEPBS2_NLCA4_NLCB1_PLR5_SU32_SUS256_SVW2_TT2_128_VW2_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 429 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU0_SUS0_SSO4_TT2_96_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 2 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 430 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO8_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 2 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 431 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU0_SUS0_SSO4_TT2_96_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 2 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 432 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 2 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 433 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 2 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 434 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR3_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 2 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 435 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU0_SUS0_SSO4_TT2_96_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 2 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 16 - LSPB: 4 - LVCA: 16 - LVCB: 64 - LVPA: 8 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [3, 4] - MIWaveTileA: 3 - MIWaveTileB: 4 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 436 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW2_NLCA3_NLCB1_PLR5_SU0_SUS0_SSO4_TT3_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 2 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [3, 64] - ThreadTile0: 12 - ThreadTile1: 4 - ThreadTileA: 12 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 4 - LSPB: 16 - LVCA: 64 - LVCB: 16 - LVPA: 2 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1536 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 6] - MIWaveTileA: 2 - MIWaveTileB: 6 - MacroTile0: 128 - MacroTile1: 96 - MacroTileA: 128 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 48 - NumLoadsA: 4 - NumLoadsB: 3 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 3 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 437 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x96x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB3_PLR5_SU0_SUS0_SSO4_TT2_96_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 2 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 96] - ThreadTile0: 8 - ThreadTile1: 6 - ThreadTileA: 8 - ThreadTileB: 6 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 438 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU32_SUS128_SSO8_TT2_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 2 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 8 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: 1 - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 439 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NLCA1_NLCB1_PLR5_SU0_SUS0_SSO4_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 2 - StoreCInUnrollPostLoop: false - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 4 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 32 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: None - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 16 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 16 - LSPB: 16 - LVCA: 16 - LVCB: 16 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 32 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [1, 1] - MIWaveTileA: 1 - MIWaveTileB: 1 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 4 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 440 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x32x32_MI16x16x4x1_SN_AF0EM2_AMAS0_ETN_EPS1_GRVW2_GSU16_PGR2_SUS256_SVW1_TT1_16_VW1_WG32_8_1 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [1, 16] - ThreadTile0: 4 - ThreadTile1: 1 - ThreadTileA: 4 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 32 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 32 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: None - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 8 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 16 - LSPB: 16 - LVCA: 16 - LVCB: 16 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 32 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [2, 2] - MIWaveTile: [1, 1] - MIWaveTileA: 1 - MIWaveTileB: 1 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 4 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 441 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x32x32_MI16x16x4x1_SN_AF0EM2_AMAS0_ETN_EPS1_GRVW2_GSU8_PGR2_SUS256_SVW1_TT1_16_VW1_WG32_8_1 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [1, 16] - ThreadTile0: 4 - ThreadTile1: 1 - ThreadTileA: 4 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 32 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 32 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: Branch - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 16 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 16 - LSPA: 16 - LSPB: 32 - LVCA: 16 - LVCB: 8 - LVPA: 8 - LVPB: 16 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 32 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [4, 4, 4, 4, 4, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 32 - MacroTile1: 16 - MacroTileA: 32 - MacroTileB: 16 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 4 - MatrixInstBM: 4 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 4 - MatrixInstN: 4 - MatrixInstruction: [4, 4, 4, 4] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 2 - NumGlobalWriteVectorsPerThread: 2 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 442 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x16x32_MI4x4x4x4_SN_AF0EM2_AMAS0_ETB_GRVW2_GSU16_PGR2_SUS256_SVW1_VW1 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 1 - ThreadTileA: 2 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 32 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 32 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: false - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 16 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 16 - LSPA: 8 - LSPB: 16 - LVCA: 32 - LVCB: 16 - LVPA: 8 - LVPB: 16 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: false - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 32 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: false - MIBlock: [4, 4, 4, 4, 4, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [1, 4] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 32 - MacroTile1: 16 - MacroTileA: 32 - MacroTileB: 16 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 4 - MatrixInstBM: 4 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 4 - MatrixInstN: 4 - MatrixInstruction: [4, 4, 4, 4] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NonTemporalD: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 2 - NumGlobalWriteVectorsPerThread: 2 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 443 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x16x32_MI4x4x4x4_SN_AF0EM1_AMAS0_ETSP_GRVW1_GSU16_PGR2_SUS256_SVW1_VW1 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 1 - ThreadTileA: 2 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 32 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 0 - LdsOffsetB_Blk: 1024 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 444 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_PLR3_SU32_SUS128_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 0 - LdsOffsetB_Blk: 1024 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 445 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_PLR3_SU8_SUS128_WGM11 - SourceSwap: true - StaggerU: 8 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 446 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_PLR5_SU8_SUS128_WGM11 - SourceSwap: true - StaggerU: 8 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 0 - LdsOffsetB_Blk: 1024 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 447 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_PLR3_SU8_SUS128_WGM8 - SourceSwap: true - StaggerU: 8 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 448 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_PLR5_SU0_SUS0_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 0 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 0 - LdsOffsetB_Blk: 1024 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 449 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_PLR3_SU16_SUS128_WGM11 - SourceSwap: true - StaggerU: 16 - StaggerUMapping: 0 - StaggerUStride: 128 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 450 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM6 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 6 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 451 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM7 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 7 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 452 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM13 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 13 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 453 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM9 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 9 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 454 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM8 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 455 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM10 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 10 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 456 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM5 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 457 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM12 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 12 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 458 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM11 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 459 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM14 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 14 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 32 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DirectToVgprA: true - DirectToVgprB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - Fp16AltImpl: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: SingleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsInitCVgprs: true - LdsNumElements: 4096 - LdsNumElementsAlignedA: 0 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 0 - LdsOffsetB_Blk: 2048 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: FMA - MFMA_BF16_1K: false - MIArchVgpr: true - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoLdsWriteCode: false - NoReject: false - NoTailLoop: true - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 3 - NonTemporalD: 3 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 2 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchAcrossPersistentMode: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - Fp16AltImpl: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 460 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_WGM4 - SourceSwap: true - SplitGlobalRead: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreCInUnroll: false - StoreCInUnrollExact: false - StoreCInUnrollInterval: 1 - StoreCInUnrollPostLoop: false - StorePriorityOpt: true - StoreRemapVectorWidth: 0 - StoreSyncOpt: 2 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - allowLRVWforTLUandMI: false -- [2, 3, 0, 1] -- - - [38144, 38144, 1, 256] - - [20, 74.893] - - - [29568, 128, 1, 384] - - [297, 55.411] - - - [30848, 128, 1, 256] - - [418, 47.977] - - - [25728, 128, 1, 384] - - [38, 57.278] - - - [32256, 32256, 1, 256] - - [22, 75.584] - - - [7680, 7680, 1, 256] - - [55, 73.456] - - - [41984, 41984, 1, 256] - - [23, 75.197] - - - [40448, 40448, 1, 256] - - [22, 74.937] - - - [25728, 128, 1, 256] - - [24, 49.099] - - - [64, 64, 1, 64] - - [183, 0.117] - - - [15104, 15104, 1, 256] - - [25, 75.487] - - - [17280, 17280, 1, 384] - - [25, 90.896] - - - [34688, 128, 1, 384] - - [293, 56.958] - - - [27392, 27392, 1, 256] - - [26, 74.926] - - - [6528, 128, 1, 256] - - [90, 37.791] - - - [35328, 35328, 1, 256] - - [49, 75.386] - - - [18432, 18432, 1, 384] - - [27, 89.766] - - - [31232, 31232, 1, 256] - - [46, 75.543] - - - [7808, 128, 1, 256] - - [91, 35.357] - - - [38400, 38400, 1, 384] - - [23, 90.583] - - - [16128, 16128, 1, 256] - - [35, 75.578] - - - [9472, 9472, 1, 256] - - [25, 74.227] - - - [21888, 21888, 1, 384] - - [26, 89.215] - - - [38656, 38656, 1, 256] - - [28, 74.685] - - - [20224, 20224, 1, 256] - - [35, 75.749] - - - [8960, 8960, 1, 256] - - [29, 73.31] - - - [29952, 29952, 1, 384] - - [25, 90.917] - - - [36864, 36864, 1, 384] - - [30, 87.926] - - - [33408, 33408, 1, 384] - - [27, 90.847] - - - [20608, 128, 1, 384] - - [293, 50.176] - - - [23424, 23424, 1, 384] - - [25, 91.11] - - - [4864, 4864, 1, 256] - - [300, 77.05] - - - [21504, 21504, 1, 384] - - [39, 89.214] - - - [25600, 25600, 1, 256] - - [27, 75.967] - - - [40960, 40960, 1, 256] - - [52, 66.623] - - - [19200, 19200, 1, 384] - - [27, 90.96] - - - [64, 1, 1, 64] - - [183, 0.002] - - - [25088, 25088, 1, 256] - - [40, 75.905] - - - [41728, 41728, 1, 256] - - [26, 73.787] - - - [35840, 35840, 1, 256] - - [35, 75.487] - - - [34560, 34560, 1, 256] - - [25, 75.1] - - - [26368, 26368, 1, 256] - - [23, 75.32] - - - [5888, 5888, 1, 256] - - [299, 81.82] - - - [28032, 28032, 1, 384] - - [25, 91.053] - - - [42496, 42496, 1, 256] - - [26, 75.137] - - - [27008, 128, 1, 256] - - [343, 44.517] - - - [38400, 38400, 1, 256] - - [23, 75.384] - - - [11008, 11008, 1, 256] - - [79, 73.892] - - - [32000, 32000, 1, 256] - - [25, 75.005] - - - [37248, 37248, 1, 384] - - [23, 90.692] - - - [10496, 10496, 1, 256] - - [23, 74.439] - - - [16640, 16640, 1, 256] - - [23, 75.73] - - - [24960, 24960, 1, 384] - - [23, 91.209] - - - [18688, 18688, 1, 256] - - [25, 75.593] - - - [22272, 22272, 1, 384] - - [34, 91.067] - - - [15488, 128, 1, 256] - - [256, 38.387] - - - [28416, 28416, 1, 384] - - [23, 90.867] - - - [3840, 3840, 1, 256] - - [301, 72.23] - - - [19968, 19968, 1, 384] - - [23, 90.738] - - - [43776, 43776, 1, 256] - - [26, 74.28] - - - [35072, 35072, 1, 256] - - [23, 75.071] - - - [20736, 20736, 1, 256] - - [25, 75.611] - - - [7168, 7168, 1, 256] - - [35, 72.544] - - - [18432, 18432, 1, 256] - - [25, 76.049] - - - [38016, 38016, 1, 384] - - [23, 91.017] - - - [35328, 35328, 1, 384] - - [23, 90.526] - - - [38784, 38784, 1, 384] - - [23, 90.979] - - - [26112, 26112, 1, 384] - - [27, 91.076] - - - [27264, 27264, 1, 384] - - [27, 90.776] - - - [44928, 44928, 1, 384] - - [25, 90.696] - - - [41088, 128, 1, 384] - - [297, 57.835] - - - [42368, 128, 1, 256] - - [418, 49.964] - - - [10752, 10752, 1, 256] - - [25, 75.284] - - - [9088, 128, 1, 384] - - [91, 46.368] - - - [17152, 17152, 1, 256] - - [45, 75.562] - - - [44928, 128, 1, 384] - - [392, 61.547] - - - [7808, 128, 1, 384] - - [92, 40.491] - - - [29184, 29184, 1, 256] - - [35, 75.558] - - - [11776, 11776, 1, 256] - - [23, 75.132] - - - [1, 64, 1, 64] - - [183, 0.002] - - - [27136, 27136, 1, 256] - - [40, 75.911] - - - [33408, 128, 1, 256] - - [266, 47.749] - - - [33792, 33792, 1, 384] - - [39, 89.232] - - - [43520, 43520, 1, 256] - - [36, 75.193] - - - [14592, 14592, 1, 384] - - [35, 90.309] - - - [41472, 41472, 1, 256] - - [23, 74.99] - - - [14080, 14080, 1, 256] - - [55, 74.446] - - - [34688, 128, 1, 256] - - [266, 48.232] - - - [16896, 16896, 1, 256] - - [35, 76.106] - - - [15744, 15744, 1, 384] - - [23, 90.68] - - - [28416, 28416, 1, 256] - - [23, 74.869] - - - [23808, 23808, 1, 256] - - [25, 75.555] - - - [27648, 27648, 1, 256] - - [23, 75.552] - - - [1152, 3072, 1, 384] - - [420, 60.318] - - - [21888, 128, 1, 256] - - [357, 43.419] - - - [34816, 34816, 1, 256] - - [27, 75.387] - - - [43776, 43776, 1, 384] - - [39, 89.916] - - - [36096, 36096, 1, 256] - - [39, 74.203] - - - [24320, 24320, 1, 256] - - [22, 75.666] - - - [12544, 12544, 1, 256] - - [25, 74.955] - - - [29184, 29184, 1, 384] - - [27, 90.781] - - - [29568, 29568, 1, 384] - - [22, 90.297] - - - [12928, 128, 1, 384] - - [93, 54.014] - - - [36480, 36480, 1, 384] - - [23, 90.923] - - - [30720, 30720, 1, 256] - - [27, 75.552] - - - [25728, 25728, 1, 384] - - [56, 90.933] - - - [34048, 34048, 1, 256] - - [40, 74.525] - - - [12928, 128, 1, 256] - - [94, 45.988] - - - [9728, 9728, 1, 256] - - [27, 74.669] - - - [128, 128, 1, 256] - - [95, 1.074] - - - [33024, 33024, 1, 256] - - [41, 75.213] - - - [15488, 128, 1, 384] - - [343, 45.315] - - - [39808, 128, 1, 384] - - [272, 57.079] - - - [18176, 18176, 1, 256] - - [25, 75.705] - - - [21504, 21504, 1, 256] - - [25, 76.075] - - - [16384, 16384, 1, 256] - - [52, 62.756] - - - [27008, 128, 1, 384] - - [418, 53.322] - - - [27904, 27904, 1, 256] - - [36, 75.163] - - - [24448, 128, 1, 384] - - [44, 55.216] - - - [35968, 128, 1, 384] - - [418, 58.167] - - - [37632, 37632, 1, 256] - - [27, 74.955] - - - [14848, 14848, 1, 256] - - [26, 75.624] - - - [23552, 23552, 1, 256] - - [27, 76.111] - - - [4608, 4608, 1, 50000] - - [51, 97.942] - - - [13056, 13056, 1, 256] - - [27, 75.313] - - - [38528, 128, 1, 256] - - [290, 50.711] - - - [19584, 19584, 1, 384] - - [35, 91.095] - - - [16768, 128, 1, 384] - - [266, 48.112] - - - [22784, 22784, 1, 256] - - [45, 75.492] - - - [44160, 44160, 1, 384] - - [27, 90.838] - - - [28160, 28160, 1, 256] - - [46, 75.558] - - - [14592, 14592, 1, 256] - - [66, 74.291] - - - [20992, 20992, 1, 256] - - [35, 76.172] - - - [41216, 41216, 1, 256] - - [23, 74.854] - - - [21760, 21760, 1, 256] - - [25, 75.734] - - - [25344, 25344, 1, 256] - - [39, 74.966] - - - [4608, 4608, 1, 256] - - [419, 76.563] - - - [2560, 2048, 1, 256] - - [392, 55.465] - - - [30464, 30464, 1, 256] - - [48, 74.371] - - - [19200, 19200, 1, 256] - - [35, 75.675] - - - [22272, 22272, 1, 256] - - [49, 75.594] - - - [29952, 29952, 1, 256] - - [45, 75.275] - - - [20480, 20480, 1, 256] - - [27, 75.812] - - - [17408, 17408, 1, 256] - - [23, 75.599] - - - [32768, 32768, 1, 256] - - [50, 57.634] - - - [18816, 18816, 1, 384] - - [56, 90.869] - - - [34944, 34944, 1, 384] - - [35, 90.977] - - - [18048, 18048, 1, 384] - - [56, 90.958] - - - [34560, 34560, 1, 384] - - [35, 90.826] - - - [9088, 128, 1, 256] - - [96, 40.41] - - - [24576, 24576, 1, 256] - - [52, 70.192] - - - [32128, 128, 1, 384] - - [293, 58.934] - - - [8448, 8448, 1, 256] - - [55, 73.715] - - - [42752, 42752, 1, 256] - - [25, 74.735] - - - [5376, 5376, 1, 256] - - [390, 78.129] - - - [18048, 128, 1, 256] - - [270, 41.753] - - - [3584, 3584, 1, 256] - - [298, 71.46] - - - [37120, 37120, 1, 256] - - [25, 74.921] - - - [39936, 39936, 1, 384] - - [59, 89.019] - - - [20736, 20736, 1, 384] - - [51, 91.075] - - - [35584, 35584, 1, 256] - - [35, 74.893] - - - [26112, 26112, 1, 256] - - [46, 76.0] - - - [16896, 16896, 1, 384] - - [35, 90.639] - - - [40704, 40704, 1, 384] - - [23, 90.834] - - - [33280, 33280, 1, 256] - - [22, 75.635] - - - [5632, 5632, 1, 256] - - [296, 78.755] - - - [19456, 19456, 1, 256] - - [25, 75.886] - - - [22016, 22016, 1, 256] - - [28, 76.148] - - - [14208, 128, 1, 256] - - [259, 35.232] - - - [13568, 13568, 1, 256] - - [23, 75.309] - - - [30848, 128, 1, 384] - - [418, 57.075] - - - [1408, 128, 1, 384] - - [97, 13.439] - - - [5760, 5760, 1, 5760] - - [368, 96.486] - - - [39936, 39936, 1, 256] - - [23, 75.103] - - - [1920, 3072, 1, 384] - - [420, 67.712] - - - [9984, 9984, 1, 256] - - [33, 74.421] - - - [2816, 2048, 1, 256] - - [291, 58.185] - - - [23168, 128, 1, 256] - - [299, 45.849] - - - [19968, 19968, 1, 256] - - [26, 76.018] - - - [44800, 44800, 1, 256] - - [23, 74.449] - - - [14976, 14976, 1, 384] - - [42, 90.548] - - - [35712, 35712, 1, 384] - - [23, 90.953] - - - [43008, 43008, 1, 384] - - [52, 89.03] - - - [41088, 41088, 1, 384] - - [39, 90.052] - - - [16128, 16128, 1, 384] - - [23, 90.362] - - - [5120, 5120, 1, 256] - - [390, 76.115] - - - [25856, 25856, 1, 256] - - [27, 75.426] - - - [12288, 12288, 1, 256] - - [25, 75.012] - - - [6400, 6400, 1, 256] - - [53, 72.272] - - - [2688, 128, 1, 256] - - [98, 20.246] - - - [11648, 128, 1, 256] - - [99, 42.668] - - - [43264, 43264, 1, 256] - - [23, 74.81] - - - [19712, 19712, 1, 256] - - [30, 75.0] - - - [34176, 34176, 1, 384] - - [25, 90.755] - - - [31104, 31104, 1, 384] - - [25, 90.907] - - - [36608, 36608, 1, 256] - - [36, 74.827] - - - [39808, 128, 1, 256] - - [370, 47.894] - - - [13824, 13824, 1, 384] - - [27, 90.095] - - - [42624, 42624, 1, 384] - - [26, 87.986] - - - [21120, 21120, 1, 384] - - [35, 91.055] - - - [23296, 23296, 1, 256] - - [23, 75.634] - - - [42240, 42240, 1, 256] - - [25, 74.739] - - - [33408, 128, 1, 384] - - [297, 55.777] - - - [43648, 128, 1, 256] - - [291, 52.21] - - - [19328, 128, 1, 384] - - [418, 53.468] - - - [33792, 33792, 1, 256] - - [23, 75.516] - - - [31488, 31488, 1, 256] - - [35, 74.936] - - - [768, 3072, 1, 384] - - [349, 53.344] - - - [6144, 6144, 1, 256] - - [299, 81.999] - - - [20352, 20352, 1, 384] - - [25, 91.117] - - - [23168, 128, 1, 384] - - [418, 53.714] - - - [33536, 33536, 1, 256] - - [48, 75.178] - - - [32640, 32640, 1, 384] - - [27, 90.087] - - - [1536, 3072, 1, 384] - - [267, 65.103] - - - [19328, 128, 1, 256] - - [272, 43.495] - - - [2688, 3072, 1, 384] - - [312, 75.149] - - - [24192, 24192, 1, 384] - - [25, 91.241] - - - [6912, 6912, 1, 256] - - [32, 72.526] - - - [15360, 15360, 1, 256] - - [23, 76.063] - - - [18944, 18944, 1, 256] - - [22, 76.049] - - - [37376, 37376, 1, 256] - - [39, 75.399] - - - [31488, 31488, 1, 384] - - [27, 90.749] - - - [26880, 26880, 1, 256] - - [23, 75.31] - - - [44928, 128, 1, 128] - - [293, 35.219] - - - [24448, 128, 1, 256] - - [55, 46.959] - - - [31872, 31872, 1, 384] - - [27, 90.846] - - - [1408, 128, 1, 256] - - [100, 11.358] - - - [38528, 128, 1, 384] - - [390, 60.3] - - - [15616, 15616, 1, 256] - - [23, 75.505] - - - [39552, 39552, 1, 384] - - [23, 90.918] - - - [4352, 4352, 1, 256] - - [357, 73.426] - - - [28288, 128, 1, 384] - - [392, 54.349] - - - [10368, 128, 1, 256] - - [91, 43.982] - - - [32128, 128, 1, 256] - - [266, 49.566] - - - [4608, 4608, 1, 4608] - - [197, 97.048] - - - [8704, 8704, 1, 256] - - [27, 74.167] - - - [17664, 17664, 1, 256] - - [28, 74.999] - - - [24576, 24576, 1, 384] - - [30, 83.07] - - - [37248, 128, 1, 384] - - [266, 57.238] - - - [34304, 34304, 1, 256] - - [41, 75.525] - - - [42368, 128, 1, 384] - - [418, 59.322] - - - [17664, 17664, 1, 384] - - [25, 90.703] - - - [12800, 12800, 1, 256] - - [25, 75.345] - - - [26624, 26624, 1, 256] - - [35, 75.83] - - - [36864, 36864, 1, 256] - - [27, 74.612] - - - [40704, 40704, 1, 256] - - [25, 74.86] - - - [12032, 12032, 1, 256] - - [25, 74.709] - - - [33024, 33024, 1, 384] - - [23, 90.647] - - - [28800, 28800, 1, 384] - - [51, 91.023] - - - [22656, 22656, 1, 384] - - [23, 91.179] - - - [41472, 41472, 1, 384] - - [25, 90.471] - - - [39680, 39680, 1, 256] - - [27, 74.613] - - - [44032, 44032, 1, 256] - - [35, 75.129] - - - [43392, 43392, 1, 384] - - [23, 90.699] - - - [42240, 42240, 1, 384] - - [25, 90.742] - - - [38912, 38912, 1, 256] - - [27, 75.302] - - - [23040, 23040, 1, 384] - - [25, 91.024] - - - [13312, 13312, 1, 256] - - [23, 75.812] - - - [128, 128, 1, 384] - - [101, 1.211] - - - [39168, 39168, 1, 256] - - [39, 74.682] - - - [25344, 25344, 1, 384] - - [25, 91.121] - - - [5248, 128, 1, 256] - - [91, 31.75] - - - [30208, 30208, 1, 256] - - [40, 75.621] - - - [40192, 40192, 1, 256] - - [27, 74.915] - - - [15872, 15872, 1, 256] - - [23, 76.103] - - - [44544, 44544, 1, 256] - - [20, 74.966] - - - [11520, 11520, 1, 256] - - [27, 74.674] - - - [15360, 15360, 1, 384] - - [45, 89.355] - - - [23040, 23040, 1, 256] - - [25, 76.194] - - - [26496, 26496, 1, 384] - - [27, 91.077] - - - [11264, 11264, 1, 256] - - [23, 75.369] - - - [18048, 128, 1, 384] - - [293, 50.273] - - - [30976, 30976, 1, 256] - - [26, 74.406] - - - [11648, 128, 1, 384] - - [102, 50.278] - - - [2304, 3072, 1, 384] - - [310, 70.811] - - - [28928, 28928, 1, 256] - - [28, 74.911] - - - [43008, 43008, 1, 256] - - [23, 74.922] - - - [29440, 29440, 1, 256] - - [28, 75.408] - - - [36352, 36352, 1, 256] - - [39, 75.406] - - - [32256, 32256, 1, 384] - - [35, 90.64] - - - [23808, 23808, 1, 384] - - [51, 91.162] - - - [37248, 128, 1, 256] - - [392, 50.072] - - - [1, 1, 1, 64] - - [183, 0.0] - - - [37888, 37888, 1, 256] - - [27, 75.212] - - - [35968, 128, 1, 256] - - [418, 49.608] - - - [13824, 13824, 1, 256] - - [49, 75.675] - - - [39168, 39168, 1, 384] - - [23, 90.701] - - - [37632, 37632, 1, 384] - - [23, 90.859] - - - [29568, 128, 1, 256] - - [392, 46.493] - - - [14336, 14336, 1, 256] - - [27, 75.668] - - - [28288, 128, 1, 256] - - [418, 45.207] - - - [16512, 16512, 1, 384] - - [52, 89.482] - - - [30720, 30720, 1, 384] - - [30, 89.477] - - - [21248, 21248, 1, 256] - - [35, 75.578] - - - [29696, 29696, 1, 256] - - [35, 75.577] - - - [384, 3072, 1, 384] - - [91, 48.474] - - - [28672, 28672, 1, 256] - - [25, 75.173] - - - [32512, 32512, 1, 256] - - [22, 75.093] - - - [9216, 9216, 1, 256] - - [25, 74.101] - - - [6656, 6656, 1, 256] - - [32, 72.68] - - - [30336, 30336, 1, 384] - - [23, 90.946] - - - [20608, 128, 1, 256] - - [418, 41.481] - - - [7936, 7936, 1, 256] - - [33, 73.5] - - - [41856, 41856, 1, 384] - - [27, 90.931] - - - [44288, 44288, 1, 256] - - [35, 74.601] - - - [7744, 7744, 1, 7744] - - [23, 97.287] - - - [7424, 7424, 1, 256] - - [47, 73.14] - - - [39424, 39424, 1, 256] - - [23, 75.347] - - - [43648, 128, 1, 384] - - [392, 59.523] - - - [14208, 14208, 1, 384] - - [35, 89.943] - - - [36096, 36096, 1, 384] - - [26, 89.977] - - - [44544, 44544, 1, 384] - - [35, 90.389] - - - [22528, 22528, 1, 256] - - [27, 75.715] - - - [4096, 4096, 1, 256] - - [298, 72.087] - - - [31744, 31744, 1, 256] - - [35, 75.52] - - - [3968, 128, 1, 384] - - [90, 28.344] - - - [17920, 17920, 1, 256] - - [27, 76.073] - - - [5248, 128, 1, 384] - - [90, 36.66] - - - [26880, 26880, 1, 384] - - [35, 91.155] - - - [8192, 8192, 1, 256] - - [23, 72.824] - - - [3968, 128, 1, 256] - - [92, 24.407] - - - [41088, 128, 1, 256] - - [392, 48.356] - - - [21888, 128, 1, 384] - - [291, 51.495] - - - [16768, 128, 1, 256] - - [293, 38.914] - - - [24064, 24064, 1, 256] - - [22, 76.017] - - - [44928, 128, 1, 256] - - [293, 51.805] - - - [27648, 27648, 1, 384] - - [59, 89.139] - - - [24832, 24832, 1, 256] - - [22, 75.662] - - - [10240, 10240, 1, 256] - - [25, 75.096] - - - [40320, 40320, 1, 384] - - [27, 90.989] - - - [18432, 2688, 1, 384] - - [25, 85.949] - - - [43008, 2304, 1, 384] - - [27, 89.185] - - - [3840, 3072, 1, 384] - - [300, 75.263] - - - [33408, 1920, 1, 384] - - [27, 87.863] - - - [33792, 2688, 1, 384] - - [35, 87.542] - - - [8064, 2688, 1, 384] - - [35, 79.14] - - - [33408, 2304, 1, 384] - - [27, 88.628] - - - [31872, 1536, 1, 384] - - [23, 85.612] - - - [41088, 1920, 1, 384] - - [59, 86.524] - - - [41088, 2304, 1, 384] - - [72, 87.329] - - - [5376, 1536, 1, 384] - - [286, 72.717] - - - [16128, 1536, 1, 384] - - [27, 83.019] - - - [36480, 2688, 1, 384] - - [34, 88.977] - - - [15360, 768, 1, 384] - - [393, 75.522] - - - [42624, 768, 1, 384] - - [51, 82.371] - - - [4992, 1536, 1, 384] - - [315, 70.916] - - - [29952, 1536, 1, 384] - - [23, 86.156] - - - [10752, 2688, 1, 384] - - [34, 82.264] - - - [42240, 2688, 1, 384] - - [51, 89.835] - - - [36096, 1536, 1, 384] - - [22, 85.555] - - - [26496, 1536, 1, 384] - - [23, 85.909] - - - [42624, 2688, 1, 384] - - [25, 87.217] - - - [17664, 2688, 1, 384] - - [34, 86.621] - - - [37248, 1536, 1, 384] - - [35, 87.151] - - - [16896, 2304, 1, 384] - - [35, 86.299] - - - [22272, 1920, 1, 384] - - [23, 87.094] - - - [26880, 2688, 1, 384] - - [34, 87.772] - - - [384, 768, 1, 384] - - [92, 21.71] - - - [16896, 1920, 1, 384] - - [35, 86.53] - - - [32640, 2304, 1, 384] - - [23, 85.785] - - - [5760, 2304, 1, 384] - - [306, 80.111] - - - [11904, 2304, 1, 384] - - [25, 82.809] - - - [24576, 2304, 1, 384] - - [35, 85.039] - - - [33024, 1536, 1, 384] - - [35, 84.384] - - - [36096, 2304, 1, 384] - - [22, 86.916] - - - [20352, 2688, 1, 384] - - [42, 87.058] - - - [14592, 2304, 1, 384] - - [35, 85.738] - - - [16128, 1920, 1, 384] - - [25, 82.946] - - - [16512, 1920, 1, 384] - - [39, 79.688] - - - [35712, 1920, 1, 384] - - [35, 87.275] - - - [9216, 2688, 1, 384] - - [25, 83.651] - - - [23808, 2688, 1, 384] - - [51, 88.104] - - - [18048, 768, 1, 384] - - [42, 78.673] - - - [14592, 2688, 1, 384] - - [23, 86.006] - - - [14208, 1920, 1, 384] - - [27, 82.418] - - - [14976, 2688, 1, 384] - - [34, 85.749] - - - [17280, 2304, 1, 384] - - [27, 84.863] - - - [11520, 2304, 1, 384] - - [23, 84.796] - - - [18432, 768, 1, 384] - - [289, 78.339] - - - [4608, 768, 1, 384] - - [294, 59.757] - - - [34944, 1920, 1, 384] - - [35, 87.524] - - - [13824, 2688, 1, 384] - - [51, 86.081] - - - [39936, 2304, 1, 384] - - [23, 88.199] - - - [7680, 2688, 1, 384] - - [25, 81.415] - - - [19968, 2304, 1, 384] - - [25, 86.824] - - - [27648, 2688, 1, 384] - - [35, 86.702] - - - [4224, 768, 1, 384] - - [32, 59.788] - - - [24192, 1920, 1, 384] - - [25, 87.591] - - - [32640, 1920, 1, 384] - - [23, 84.928] - - - [34176, 2688, 1, 384] - - [54, 88.958] - - - [35328, 1536, 1, 384] - - [25, 85.698] - - - [8832, 2688, 1, 384] - - [303, 82.429] - - - [18048, 1920, 1, 384] - - [25, 84.439] - - - [31488, 768, 1, 384] - - [299, 83.234] - - - [21504, 2304, 1, 384] - - [23, 86.898] - - - [11136, 2688, 1, 384] - - [51, 84.703] - - - [768, 1152, 1, 384] - - [271, 31.557] - - - [29184, 2688, 1, 384] - - [42, 88.494] - - - [4608, 2688, 1, 384] - - [302, 77.533] - - - [21504, 2688, 1, 384] - - [25, 85.527] - - - [34176, 768, 1, 384] - - [51, 83.561] - - - [23808, 1536, 1, 384] - - [25, 84.573] - - - [43392, 1536, 1, 384] - - [23, 87.946] - - - [13824, 768, 1, 384] - - [320, 74.018] - - - [38016, 1536, 1, 384] - - [27, 86.692] - - - [20736, 2688, 1, 384] - - [56, 88.463] - - - [15744, 1536, 1, 384] - - [23, 82.237] - - - [16512, 1536, 1, 384] - - [393, 82.43] - - - [3072, 2304, 1, 384] - - [294, 68.389] - - - [5760, 2688, 1, 384] - - [389, 81.364] - - - [38400, 2304, 1, 384] - - [23, 88.2] - - - [15360, 2688, 1, 384] - - [25, 86.326] - - - [29952, 2688, 1, 384] - - [56, 88.86] - - - [43008, 2688, 1, 384] - - [27, 87.686] - - - [13440, 1920, 1, 384] - - [25, 82.655] - - - [6528, 2688, 1, 384] - - [23, 81.588] - - - [2304, 1536, 1, 384] - - [421, 59.413] - - - [40320, 1536, 1, 384] - - [25, 86.967] - - - [13440, 1536, 1, 384] - - [25, 81.341] - - - [40320, 2688, 1, 384] - - [51, 88.955] - - - [30336, 2304, 1, 384] - - [34, 88.619] - - - [24192, 2688, 1, 384] - - [42, 87.204] - - - [35328, 768, 1, 384] - - [51, 81.961] - - - [23040, 768, 1, 384] - - [56, 81.55] - - - [29952, 2304, 1, 384] - - [25, 87.87] - - - [33024, 1920, 1, 384] - - [35, 86.647] - - - [14976, 768, 1, 384] - - [305, 78.431] - - - [42624, 1920, 1, 384] - - [35, 87.484] - - - [32640, 2688, 1, 384] - - [26, 85.433] - - - [11520, 1536, 1, 384] - - [27, 82.185] - - - [6912, 768, 1, 384] - - [297, 64.089] - - - [39552, 1920, 1, 384] - - [25, 87.96] - - - [32256, 1920, 1, 384] - - [25, 87.539] - - - [10752, 1536, 1, 384] - - [23, 77.35] - - - [24576, 2688, 1, 384] - - [35, 84.613] - - - [12672, 2688, 1, 384] - - [51, 86.586] - - - [10752, 1920, 1, 384] - - [25, 81.453] - - - [40704, 1536, 1, 384] - - [23, 87.483] - - - [32256, 768, 1, 384] - - [299, 84.246] - - - [18816, 2688, 1, 384] - - [42, 85.896] - - - [11520, 2688, 1, 384] - - [34, 83.45] - - - [35712, 2688, 1, 384] - - [51, 88.852] - - - [29952, 1920, 1, 384] - - [23, 88.372] - - - [26880, 1920, 1, 384] - - [23, 87.563] - - - [33408, 2688, 1, 384] - - [56, 89.067] - - - [35328, 2688, 1, 384] - - [42, 88.976] - - - [21120, 2688, 1, 384] - - [60, 86.9] - - - [19584, 1920, 1, 384] - - [25, 86.96] - - - [17664, 1536, 1, 384] - - [27, 81.463] - - - [36864, 768, 1, 384] - - [42, 83.653] - - - [14592, 1536, 1, 384] - - [23, 81.728] - - - [11136, 2304, 1, 384] - - [23, 82.333] - - - [9600, 2688, 1, 384] - - [27, 82.609] - - - [9216, 2304, 1, 384] - - [27, 83.156] - - - [21120, 768, 1, 384] - - [299, 80.959] - - - [4992, 2688, 1, 384] - - [389, 82.226] - - - [41472, 768, 1, 384] - - [42, 85.184] - - - [37632, 1536, 1, 384] - - [25, 86.027] - - - [38784, 2304, 1, 384] - - [35, 88.806] - - - [8448, 2688, 1, 384] - - [42, 82.716] - - - [36864, 2304, 1, 384] - - [35, 87.729] - - - [40704, 1920, 1, 384] - - [27, 88.34] - - - [39552, 2688, 1, 384] - - [51, 89.664] - - - [26112, 768, 1, 384] - - [424, 82.316] - - - [29184, 1536, 1, 384] - - [23, 86.987] - - - [32640, 1536, 1, 384] - - [27, 83.298] - - - [5376, 2688, 1, 384] - - [389, 81.429] - - - [13056, 768, 1, 384] - - [391, 77.612] - - - [13824, 2304, 1, 384] - - [25, 85.541] - - - [16896, 768, 1, 384] - - [424, 78.28] - - - [30336, 1920, 1, 384] - - [27, 87.334] - - - [27264, 2304, 1, 384] - - [27, 88.195] - - - [7680, 1536, 1, 384] - - [389, 76.446] - - - [30720, 2688, 1, 384] - - [25, 87.806] - - - [36096, 2688, 1, 384] - - [26, 87.086] - - - [5760, 1920, 1, 384] - - [310, 77.098] - - - [42240, 1536, 1, 384] - - [25, 87.779] - - - [8448, 1920, 1, 384] - - [305, 79.937] - - - [32256, 1536, 1, 384] - - [35, 86.303] - - - [44160, 2304, 1, 384] - - [56, 89.108] - - - [30336, 2688, 1, 384] - - [42, 88.332] - - - [6144, 2688, 1, 384] - - [393, 81.03] - - - [39168, 1536, 1, 384] - - [23, 86.962] - - - [11904, 1920, 1, 384] - - [56, 83.359] - - - [8064, 1536, 1, 384] - - [294, 77.917] - - - [21120, 1920, 1, 384] - - [25, 86.0] - - - [22656, 2304, 1, 384] - - [42, 88.383] - - - [19968, 2688, 1, 384] - - [42, 87.953] - - - [10752, 768, 1, 384] - - [391, 74.532] - - - [18432, 2304, 1, 384] - - [27, 86.351] - - - [14976, 1920, 1, 384] - - [56, 86.02] - - - [33024, 2688, 1, 384] - - [56, 87.96] - - - [1536, 768, 1, 384] - - [344, 46.928] - - - [33024, 2304, 1, 384] - - [42, 87.324] - - - [14208, 2688, 1, 384] - - [42, 84.865] - - - [38016, 2304, 1, 384] - - [51, 88.859] - - - [16896, 2688, 1, 384] - - [34, 85.692] - - - [31104, 768, 1, 384] - - [51, 81.555] - - - [41472, 2304, 1, 384] - - [23, 88.447] - - - [23424, 2688, 1, 384] - - [27, 87.378] - - - [26496, 2688, 1, 384] - - [34, 88.317] - - - [16512, 2304, 1, 384] - - [26, 80.998] - - - [11520, 1920, 1, 384] - - [23, 81.321] - - - [39552, 768, 1, 384] - - [51, 85.633] - - - [6144, 2304, 1, 384] - - [360, 78.902] - - - [14208, 2304, 1, 384] - - [25, 83.895] - - - [19584, 2304, 1, 384] - - [27, 85.331] - - - [36480, 768, 1, 384] - - [56, 83.786] - - - [15744, 2688, 1, 384] - - [51, 86.233] - - - [34560, 1536, 1, 384] - - [27, 86.554] - - - [8448, 2304, 1, 384] - - [23, 82.987] - - - [26112, 2688, 1, 384] - - [51, 88.934] - - - [39936, 768, 1, 384] - - [34, 82.561] - - - [19200, 1920, 1, 384] - - [25, 85.398] - - - [38400, 768, 1, 384] - - [51, 83.34] - - - [8448, 1536, 1, 384] - - [299, 78.822] - - - [13824, 1536, 1, 384] - - [23, 82.987] - - - [9600, 768, 1, 384] - - [295, 72.123] - - - [10368, 768, 1, 384] - - [308, 73.245] - - - [20736, 1536, 1, 384] - - [23, 85.105] - - - [28800, 768, 1, 384] - - [34, 80.331] - - - [10368, 1536, 1, 384] - - [25, 81.171] - - - [21888, 1536, 1, 384] - - [25, 82.833] - - - [38784, 2688, 1, 384] - - [34, 89.453] - - - [27648, 2304, 1, 384] - - [27, 87.437] - - - [11136, 1920, 1, 384] - - [25, 83.769] - - - [37248, 768, 1, 384] - - [42, 85.091] - - - [23040, 2688, 1, 384] - - [34, 87.36] - - - [37632, 1920, 1, 384] - - [23, 87.723] - - - [7680, 768, 1, 384] - - [419, 67.445] - - - [38016, 1920, 1, 384] - - [23, 88.34] - - - [35712, 2304, 1, 384] - - [42, 88.776] - - - [37248, 2688, 1, 384] - - [42, 88.92] - - - [29568, 1920, 1, 384] - - [49, 86.846] - - - [38400, 2688, 1, 384] - - [51, 88.725] - - - [25728, 768, 1, 384] - - [424, 83.385] - - - [8832, 1920, 1, 384] - - [393, 79.39] - - - [43776, 1920, 1, 384] - - [28, 86.47] - - - [15744, 768, 1, 384] - - [393, 77.155] - - - [27264, 1920, 1, 384] - - [27, 85.812] - - - [33792, 2304, 1, 384] - - [25, 87.915] - - - [8832, 2304, 1, 384] - - [25, 80.583] - - - [39168, 2688, 1, 384] - - [34, 89.136] - - - [35328, 1920, 1, 384] - - [25, 88.337] - - - [35328, 2304, 1, 384] - - [25, 88.11] - - - [29184, 768, 1, 384] - - [51, 81.9] - - - [18048, 2688, 1, 384] - - [56, 87.839] - - - [32256, 2688, 1, 384] - - [34, 88.127] - - - [18816, 1536, 1, 384] - - [25, 81.82] - - - [13056, 1536, 1, 384] - - [348, 84.688] - - - [34944, 1536, 1, 384] - - [25, 87.087] - - - [38400, 1920, 1, 384] - - [35, 88.679] - - - [15360, 2304, 1, 384] - - [23, 85.764] - - - [27264, 2688, 1, 384] - - [25, 87.083] - - - [11136, 1536, 1, 384] - - [25, 79.702] - - - [30720, 2304, 1, 384] - - [35, 87.779] - - - [24960, 2688, 1, 384] - - [56, 87.696] - - - [13824, 1920, 1, 384] - - [23, 84.703] - - - [17280, 2688, 1, 384] - - [42, 87.383] - - - [31872, 768, 1, 384] - - [34, 83.012] - - - [11904, 2688, 1, 384] - - [34, 85.826] - - - [7296, 768, 1, 384] - - [291, 67.115] - - - [19200, 1536, 1, 384] - - [27, 83.176] - - - [12288, 768, 1, 384] - - [314, 73.736] - - - [33792, 768, 1, 384] - - [51, 82.575] - - - [21888, 2688, 1, 384] - - [62, 83.957] - - - [2688, 1920, 1, 384] - - [291, 63.361] - - - [19968, 768, 1, 384] - - [42, 78.744] - - - [12288, 2688, 1, 384] - - [25, 83.987] - - - [12288, 2304, 1, 384] - - [23, 84.682] - - - [28416, 768, 1, 384] - - [42, 80.118] - - - [34560, 768, 1, 384] - - [51, 84.329] - - - [39936, 2688, 1, 384] - - [25, 88.07] - - - [8064, 1920, 1, 384] - - [35, 79.126] - - - [26880, 1536, 1, 384] - - [35, 86.713] - - - [28032, 2688, 1, 384] - - [34, 89.022] - - - [41472, 2688, 1, 384] - - [56, 89.01] - - - [29568, 2688, 1, 384] - - [51, 87.448] - - - [31104, 2688, 1, 384] - - [51, 88.587] - - - [5376, 1920, 1, 384] - - [393, 74.146] - - - [41856, 2688, 1, 384] - - [34, 89.376] - - - [9984, 768, 1, 384] - - [309, 70.459] - - - [3456, 2688, 1, 384] - - [360, 73.378] - - - [43392, 2688, 1, 384] - - [42, 89.587] - - - [36480, 1920, 1, 384] - - [42, 88.451] - - - [29568, 1536, 1, 384] - - [45, 84.24] - - - [36864, 2688, 1, 384] - - [35, 87.725] - - - [12672, 768, 1, 384] - - [307, 76.605] - - - [24064, 3072, 1, 256] - - [29, 72.89] - - - [256, 512, 1, 256] - - [256, 4.477] - - - [40960, 27648, 1, 256] - - [20, 67.36] - - - [31744, 3072, 1, 256] - - [27, 73.477] - - - [13056, 1792, 1, 256] - - [299, 76.914] - - - [35328, 22785, 1, 256] - - [71, 73.723] - - - [28160, 15872, 1, 256] - - [63, 75.611] - - - [39168, 1792, 1, 256] - - [57, 71.707] - - - [23808, 11265, 1, 256] - - [30, 71.85] - - - [16640, 4353, 1, 256] - - [71, 69.432] - - - [38912, 26624, 1, 256] - - [35, 75.182] - - - [6912, 3585, 1, 256] - - [390, 73.199] - - - [32768, 1792, 1, 256] - - [20, 58.928] - - - [30976, 18688, 1, 256] - - [26, 74.426] - - - [512, 2048, 1, 256] - - [103, 37.096] - - - [15872, 3584, 1, 256] - - [29, 72.985] - - - [6400, 1792, 1, 256] - - [425, 68.693] - - - [39680, 27393, 1, 256] - - [52, 72.507] - - - [36864, 24577, 1, 256] - - [30, 71.805] - - - [26112, 1536, 1, 256] - - [47, 70.393] - - - [26368, 1536, 1, 256] - - [47, 69.857] - - - [16896, 4353, 1, 256] - - [28, 71.01] - - - [14336, 1793, 1, 256] - - [420, 72.379] - - - [3840, 3072, 1, 256] - - [388, 68.641] - - - [2560, 3072, 1, 256] - - [422, 62.945] - - - [6656, 1536, 1, 256] - - [300, 65.188] - - - [27136, 1792, 1, 256] - - [32, 71.697] - - - [43776, 3072, 1, 256] - - [26, 71.088] - - - [23296, 1792, 1, 256] - - [32, 70.456] - - - [11264, 7937, 1, 256] - - [30, 72.469] - - - [768, 3072, 1, 256] - - [392, 44.03] - - - [6912, 3841, 1, 256] - - [299, 74.373] - - - [40960, 769, 1, 256] - - [350, 61.731] - - - [40448, 9216, 1, 256] - - [40, 74.661] - - - [7680, 4353, 1, 256] - - [299, 77.187] - - - [23296, 3072, 1, 256] - - [23, 72.423] - - - [7936, 4609, 1, 256] - - [299, 78.12] - - - [20736, 8448, 1, 256] - - [27, 74.596] - - - [768, 1024, 1, 256] - - [105, 36.135] - - - [38656, 3072, 1, 256] - - [28, 72.725] - - - [28160, 1792, 1, 256] - - [29, 71.494] - - - [13824, 3072, 1, 256] - - [25, 71.6] - - - [42752, 1792, 1, 256] - - [27, 72.151] - - - [35584, 23041, 1, 256] - - [26, 73.114] - - - [13056, 3072, 1, 256] - - [25, 71.139] - - - [37888, 768, 1, 256] - - [299, 76.452] - - - [19456, 3072, 1, 256] - - [25, 72.97] - - - [15872, 9216, 1, 256] - - [35, 74.823] - - - [30976, 1792, 1, 256] - - [55, 71.133] - - - [26368, 14081, 1, 256] - - [52, 72.876] - - - [35328, 23041, 1, 256] - - [41, 73.686] - - - [27648, 15105, 1, 256] - - [23, 74.161] - - - [25856, 13568, 1, 256] - - [35, 75.047] - - - [23296, 9216, 1, 256] - - [27, 74.259] - - - [2048, 1024, 1, 256] - - [282, 40.635] - - - [12032, 1792, 1, 256] - - [299, 75.969] - - - [11520, 1536, 1, 256] - - [357, 74.16] - - - [16128, 768, 1, 256] - - [421, 70.184] - - - [15360, 3072, 1, 256] - - [25, 72.025] - - - [38912, 26369, 1, 256] - - [52, 73.922] - - - [25344, 13056, 1, 256] - - [39, 74.792] - - - [39168, 26880, 1, 256] - - [36, 74.625] - - - [39424, 768, 1, 256] - - [299, 76.69] - - - [10496, 1792, 1, 256] - - [390, 72.171] - - - [28672, 3072, 1, 256] - - [23, 73.07] - - - [27392, 768, 1, 256] - - [421, 72.829] - - - [39680, 768, 1, 256] - - [390, 76.711] - - - [11520, 8193, 1, 256] - - [52, 70.412] - - - [17408, 4865, 1, 256] - - [23, 71.47] - - - [14080, 1537, 1, 256] - - [424, 70.203] - - - [29184, 768, 1, 256] - - [300, 73.282] - - - [19200, 6913, 1, 256] - - [29, 71.548] - - - [33536, 9216, 1, 256] - - [39, 74.284] - - - [5632, 3072, 1, 256] - - [419, 72.256] - - - [32768, 20480, 1, 256] - - [64, 58.823] - - - [29440, 9216, 1, 256] - - [45, 74.429] - - - [40960, 1792, 1, 256] - - [36, 66.694] - - - [10240, 3072, 1, 256] - - [299, 79.686] - - - [20992, 1792, 1, 256] - - [299, 80.331] - - - [42240, 9216, 1, 256] - - [26, 74.271] - - - [19200, 6912, 1, 256] - - [25, 74.327] - - - [27392, 1792, 1, 256] - - [80, 68.339] - - - [42496, 1536, 1, 256] - - [35, 71.485] - - - [29440, 16897, 1, 256] - - [26, 73.235] - - - [20480, 8192, 1, 256] - - [23, 74.921] - - - [11264, 8193, 1, 256] - - [27, 70.854] - - - [26880, 14337, 1, 256] - - [52, 72.433] - - - [28928, 16641, 1, 256] - - [71, 73.123] - - - [15360, 2817, 1, 256] - - [25, 67.942] - - - [44288, 1536, 1, 256] - - [27, 70.471] - - - [7936, 1536, 1, 256] - - [420, 69.493] - - - [18176, 5633, 1, 256] - - [23, 71.129] - - - [8448, 3072, 1, 256] - - [296, 76.896] - - - [17920, 5632, 1, 256] - - [22, 74.331] - - - [1792, 2048, 1, 256] - - [287, 53.003] - - - [39936, 3072, 1, 256] - - [35, 73.824] - - - [20480, 3072, 1, 256] - - [27, 72.732] - - - [24832, 1792, 1, 256] - - [65, 71.196] - - - [37376, 25088, 1, 256] - - [35, 75.229] - - - [7168, 4097, 1, 256] - - [299, 74.238] - - - [21504, 768, 1, 256] - - [314, 72.919] - - - [13312, 3072, 1, 256] - - [25, 71.791] - - - [40960, 1025, 1, 256] - - [27, 57.587] - - - [12032, 1536, 1, 256] - - [419, 72.741] - - - [9216, 768, 1, 256] - - [282, 61.41] - - - [44288, 27648, 1, 256] - - [26, 74.594] - - - [32512, 1792, 1, 256] - - [32, 71.363] - - - [23808, 11520, 1, 256] - - [23, 75.256] - - - [25600, 13057, 1, 256] - - [52, 73.924] - - - [40448, 1792, 1, 256] - - [27, 72.271] - - - [25088, 12800, 1, 256] - - [45, 75.683] - - - [22784, 10496, 1, 256] - - [65, 74.874] - - - [38400, 26113, 1, 256] - - [26, 73.709] - - - [9728, 3072, 1, 256] - - [289, 78.308] - - - [20736, 1792, 1, 256] - - [390, 80.243] - - - [7680, 3072, 1, 256] - - [299, 75.747] - - - [5376, 2305, 1, 256] - - [348, 67.241] - - - [12800, 3072, 1, 256] - - [29, 71.387] - - - [43520, 3584, 1, 256] - - [35, 74.091] - - - [12288, 3072, 1, 256] - - [357, 81.218] - - - [12800, 1536, 1, 256] - - [419, 72.92] - - - [21504, 8961, 1, 256] - - [23, 73.17] - - - [39680, 9216, 1, 256] - - [25, 74.129] - - - [3584, 513, 1, 256] - - [418, 36.18] - - - [1280, 3072, 1, 256] - - [293, 53.116] - - - [13056, 9216, 1, 256] - - [35, 74.371] - - - [22016, 768, 1, 256] - - [419, 71.825] - - - [33024, 1536, 1, 256] - - [53, 69.126] - - - [26880, 9216, 1, 256] - - [25, 74.278] - - - [44032, 27648, 1, 256] - - [25, 74.877] - - - [7680, 768, 1, 256] - - [295, 58.9] - - - [32000, 19712, 1, 256] - - [35, 75.0] - - - [26880, 14593, 1, 256] - - [52, 72.979] - - - [24064, 9216, 1, 256] - - [41, 74.935] - - - [39424, 26881, 1, 256] - - [52, 73.642] - - - [27392, 3072, 1, 256] - - [41, 70.269] - - - [10752, 1792, 1, 256] - - [417, 72.748] - - - [8960, 5633, 1, 256] - - [27, 70.165] - - - [34560, 3072, 1, 256] - - [23, 72.869] - - - [23808, 9216, 1, 256] - - [23, 74.572] - - - [29696, 17153, 1, 256] - - [27, 74.326] - - - [11776, 1536, 1, 256] - - [289, 72.556] - - - [13568, 1536, 1, 256] - - [300, 74.691] - - - [30208, 9216, 1, 256] - - [40, 75.008] - - - [36608, 1536, 1, 256] - - [53, 70.529] - - - [12800, 513, 1, 256] - - [419, 57.273] - - - [7680, 1792, 1, 256] - - [425, 67.962] - - - [42496, 2305, 1, 256] - - [35, 68.926] - - - [37376, 1536, 1, 256] - - [27, 70.914] - - - [20224, 1792, 1, 256] - - [296, 80.999] - - - [43520, 1536, 1, 256] - - [23, 71.431] - - - [26368, 768, 1, 256] - - [299, 72.307] - - - [18176, 3072, 1, 256] - - [25, 72.178] - - - [24320, 12033, 1, 256] - - [71, 73.218] - - - [17408, 9216, 1, 256] - - [25, 74.837] - - - [36352, 1792, 1, 256] - - [27, 72.37] - - - [20992, 8705, 1, 256] - - [30, 72.643] - - - [19712, 7424, 1, 256] - - [57, 73.379] - - - [38144, 768, 1, 256] - - [289, 76.575] - - - [10752, 1536, 1, 256] - - [360, 73.516] - - - [4096, 3072, 1, 256] - - [298, 70.481] - - - [29696, 17409, 1, 256] - - [52, 72.74] - - - [10240, 6913, 1, 256] - - [30, 71.858] - - - [18944, 1536, 1, 256] - - [299, 77.007] - - - [38656, 26113, 1, 256] - - [39, 72.926] - - - [37376, 25089, 1, 256] - - [39, 73.77] - - - [38400, 1536, 1, 256] - - [23, 70.995] - - - [8448, 1792, 1, 256] - - [299, 70.931] - - - [13056, 769, 1, 256] - - [349, 61.87] - - - [24320, 11777, 1, 256] - - [71, 73.035] - - - [17664, 9216, 1, 256] - - [22, 73.709] - - - [8192, 4865, 1, 256] - - [35, 69.019] - - - [17920, 1792, 1, 256] - - [296, 79.017] - - - [32000, 19713, 1, 256] - - [52, 73.259] - - - [8960, 768, 1, 256] - - [291, 60.567] - - - [31232, 3072, 1, 256] - - [65, 73.089] - - - [12544, 257, 1, 256] - - [419, 46.559] - - - [43776, 3585, 1, 256] - - [26, 68.553] - - - [11008, 1792, 1, 256] - - [287, 74.719] - - - [29696, 17408, 1, 256] - - [25, 75.352] - - - [34560, 22272, 1, 256] - - [23, 74.996] - - - [256, 2048, 1, 256] - - [96, 25.406] - - - [32768, 20481, 1, 256] - - [64, 55.844] - - - [14336, 3072, 1, 256] - - [53, 71.953] - - - [19456, 7168, 1, 256] - - [25, 74.12] - - - [13312, 9216, 1, 256] - - [35, 74.894] - - - [22272, 768, 1, 256] - - [421, 72.955] - - - [24064, 1792, 1, 256] - - [29, 71.538] - - - [16896, 1792, 1, 256] - - [267, 78.07] - - - [27904, 15616, 1, 256] - - [25, 75.078] - - - [37888, 3072, 1, 256] - - [23, 73.795] - - - [13056, 513, 1, 256] - - [419, 55.537] - - - [36608, 24065, 1, 256] - - [26, 72.88] - - - [40704, 3072, 1, 256] - - [25, 73.142] - - - [28928, 16640, 1, 256] - - [45, 75.132] - - - [24576, 12288, 1, 256] - - [52, 69.898] - - - [17152, 3072, 1, 256] - - [47, 71.874] - - - [17152, 4864, 1, 256] - - [33, 73.546] - - - [42496, 9216, 1, 256] - - [26, 74.809] - - - [32256, 768, 1, 256] - - [299, 76.157] - - - [4352, 1792, 1, 256] - - [291, 62.693] - - - [5632, 768, 1, 256] - - [384, 53.845] - - - [40704, 513, 1, 256] - - [300, 64.697] - - - [19712, 768, 1, 256] - - [300, 68.275] - - - [33536, 20993, 1, 256] - - [71, 73.277] - - - [2816, 3072, 1, 256] - - [282, 66.468] - - - [3584, 3072, 1, 256] - - [300, 69.304] - - - [4608, 1537, 1, 256] - - [287, 59.201] - - - [44032, 9216, 1, 256] - - [23, 74.553] - - - [33792, 21249, 1, 256] - - [30, 74.025] - - - [32512, 20225, 1, 256] - - [71, 73.399] - - - [38656, 9216, 1, 256] - - [39, 74.324] - - - [17664, 5377, 1, 256] - - [28, 70.624] - - - [19456, 7169, 1, 256] - - [30, 71.093] - - - [8448, 5121, 1, 256] - - [27, 68.835] - - - [29440, 17152, 1, 256] - - [28, 75.486] - - - [40448, 513, 1, 256] - - [421, 63.05] - - - [41472, 1792, 1, 256] - - [27, 72.408] - - - [17920, 3072, 1, 256] - - [55, 72.264] - - - [35072, 9216, 1, 256] - - [23, 74.214] - - - [34816, 22273, 1, 256] - - [52, 74.005] - - - [35072, 22785, 1, 256] - - [30, 73.107] - - - [39168, 9216, 1, 256] - - [26, 74.256] - - - [42752, 2817, 1, 256] - - [52, 69.038] - - - [11776, 3072, 1, 256] - - [299, 80.578] - - - [24832, 12289, 1, 256] - - [41, 72.414] - - - [24576, 12033, 1, 256] - - [30, 68.22] - - - [6400, 1536, 1, 256] - - [325, 65.726] - - - [32512, 3072, 1, 256] - - [55, 72.723] - - - [30976, 3072, 1, 256] - - [29, 71.129] - - - [22016, 9473, 1, 256] - - [22, 73.241] - - - [19968, 1792, 1, 256] - - [299, 81.293] - - - [29440, 3072, 1, 256] - - [28, 73.013] - - - [43776, 3840, 1, 256] - - [26, 72.616] - - - [41472, 768, 1, 256] - - [390, 77.686] - - - [8192, 1792, 1, 256] - - [420, 70.045] - - - [35840, 3072, 1, 256] - - [25, 73.788] - - - [8704, 3072, 1, 256] - - [296, 79.047] - - - [9728, 1792, 1, 256] - - [348, 73.341] - - - [22272, 9729, 1, 256] - - [68, 72.315] - - - [32768, 3072, 1, 256] - - [20, 59.221] - - - [3072, 2048, 1, 256] - - [289, 60.493] - - - [36864, 24576, 1, 256] - - [23, 74.304] - - - [9984, 1536, 1, 256] - - [300, 71.525] - - - [12032, 8961, 1, 256] - - [45, 71.392] - - - [38400, 25857, 1, 256] - - [26, 73.786] - - - [20224, 7937, 1, 256] - - [52, 72.001] - - - [34304, 21761, 1, 256] - - [41, 73.859] - - - [30720, 18432, 1, 256] - - [25, 75.598] - - - [31744, 9216, 1, 256] - - [25, 74.725] - - - [27136, 14848, 1, 256] - - [45, 75.662] - - - [34048, 9216, 1, 256] - - [41, 73.879] - - - [3584, 257, 1, 256] - - [96, 33.197] - - - [18688, 6145, 1, 256] - - [30, 70.227] - - - [36096, 768, 1, 256] - - [419, 74.999] - - - [36608, 9216, 1, 256] - - [26, 74.237] - - - [35584, 9216, 1, 256] - - [26, 74.453] - - - [29952, 17664, 1, 256] - - [22, 75.301] - - - [34816, 1792, 1, 256] - - [27, 72.664] - - - [24064, 11776, 1, 256] - - [22, 75.569] - - - [40448, 3072, 1, 256] - - [49, 73.246] - - - [18688, 6401, 1, 256] - - [52, 70.993] - - - [20480, 1536, 1, 256] - - [357, 76.715] - - - [18432, 3072, 1, 256] - - [23, 72.829] - - - [20224, 768, 1, 256] - - [314, 69.669] - - - [25344, 768, 1, 256] - - [300, 73.303] - - - [36608, 24320, 1, 256] - - [36, 74.884] - - - [34816, 9216, 1, 256] - - [23, 74.783] - - - [41216, 27648, 1, 256] - - [26, 74.634] - - - [30464, 9216, 1, 256] - - [39, 73.654] - - - [7424, 3072, 1, 256] - - [390, 76.609] - - - [20480, 1792, 1, 256] - - [348, 78.405] - - - [41984, 1793, 1, 256] - - [27, 67.812] - - - [18688, 1792, 1, 256] - - [299, 79.793] - - - [13824, 1792, 1, 256] - - [299, 74.84] - - - [38144, 3072, 1, 256] - - [35, 73.144] - - - [33280, 3072, 1, 256] - - [49, 73.418] - - - [35584, 23296, 1, 256] - - [48, 74.79] - - - [43520, 768, 1, 256] - - [289, 77.031] - - - [40704, 1536, 1, 256] - - [27, 71.038] - - - [29696, 3072, 1, 256] - - [35, 73.642] - - - [32256, 19969, 1, 256] - - [71, 73.669] - - - [40960, 9216, 1, 256] - - [36, 67.099] - - - [37632, 9216, 1, 256] - - [27, 73.983] - - - [42240, 2305, 1, 256] - - [30, 68.068] - - - [17920, 5377, 1, 256] - - [22, 71.454] - - - [27904, 9216, 1, 256] - - [39, 74.405] - - - [34304, 22016, 1, 256] - - [40, 75.492] - - - [11776, 8705, 1, 256] - - [49, 72.055] - - - [22272, 1536, 1, 256] - - [299, 79.112] - - - [25856, 9216, 1, 256] - - [23, 74.358] - - - [19712, 3072, 1, 256] - - [49, 70.534] - - - [41472, 9216, 1, 256] - - [40, 74.605] - - - [42496, 27648, 1, 256] - - [26, 75.01] - - - [44288, 4352, 1, 256] - - [36, 73.828] - - - [42496, 2561, 1, 256] - - [23, 68.824] - - - [9984, 6657, 1, 256] - - [27, 71.101] - - - [43008, 3073, 1, 256] - - [52, 69.489] - - - [36352, 24065, 1, 256] - - [39, 73.734] - - - [24832, 3072, 1, 256] - - [55, 72.622] - - - [29184, 16641, 1, 256] - - [30, 73.633] - - - [1024, 2048, 1, 256] - - [387, 41.031] - - - [42240, 27648, 1, 256] - - [26, 74.587] - - - [9984, 1792, 1, 256] - - [287, 72.573] - - - [44288, 3072, 1, 256] - - [52, 72.705] - - - [11008, 768, 1, 256] - - [421, 65.925] - - - [28672, 16129, 1, 256] - - [52, 73.603] - - - [17920, 9216, 1, 256] - - [72, 74.862] - - - [25088, 12801, 1, 256] - - [41, 73.417] - - - [19712, 9216, 1, 256] - - [39, 74.04] - - - [31744, 19457, 1, 256] - - [23, 72.863] - - - [36864, 1792, 1, 256] - - [25, 72.242] - - - [42496, 1792, 1, 256] - - [23, 72.873] - - - [39936, 9216, 1, 256] - - [23, 74.655] - - - [8960, 1792, 1, 256] - - [348, 70.593] - - - [17664, 5121, 1, 256] - - [49, 69.096] - - - [38144, 25601, 1, 256] - - [26, 72.464] - - - [27136, 14849, 1, 256] - - [41, 73.926] - - - [31744, 19456, 1, 256] - - [25, 75.314] - - - [33024, 3072, 1, 256] - - [22, 72.339] - - - [37888, 9216, 1, 256] - - [25, 74.723] - - - [6912, 1792, 1, 256] - - [267, 69.316] - - - [42240, 2049, 1, 256] - - [56, 66.103] - - - [34048, 3072, 1, 256] - - [29, 72.477] - - - [37120, 9216, 1, 256] - - [39, 74.336] - - - [14080, 9216, 1, 256] - - [49, 73.043] - - - [38400, 1792, 1, 256] - - [35, 72.333] - - - [43776, 9216, 1, 256] - - [39, 74.086] - - - [14336, 2049, 1, 256] - - [296, 73.365] - - - [37120, 24577, 1, 256] - - [39, 72.438] - - - [30976, 18433, 1, 256] - - [26, 72.375] - - - [37632, 3072, 1, 256] - - [23, 72.923] - - - [34560, 1792, 1, 256] - - [53, 71.797] - - - [5120, 3072, 1, 256] - - [419, 72.411] - - - [21760, 9217, 1, 256] - - [30, 71.598] - - - [24064, 11521, 1, 256] - - [28, 73.393] - - - [7936, 3072, 1, 256] - - [390, 76.704] - - - [21760, 9472, 1, 256] - - [27, 75.121] - - - [9216, 6145, 1, 256] - - [25, 69.56] - - - [8192, 1536, 1, 256] - - [294, 68.127] - - - [39936, 27648, 1, 256] - - [23, 74.967] - - - [21248, 9216, 1, 256] - - [25, 74.493] - - - [5376, 2049, 1, 256] - - [312, 65.356] - - - [35072, 22529, 1, 256] - - [52, 72.351] - - - [13312, 769, 1, 256] - - [417, 59.611] - - - [35840, 9216, 1, 256] - - [23, 74.674] - - - [39424, 27136, 1, 256] - - [36, 75.313] - - - [26368, 9216, 1, 256] - - [27, 74.3] - - - [34048, 21505, 1, 256] - - [26, 72.296] - - - [26112, 1792, 1, 256] - - [57, 71.895] - - - [23296, 768, 1, 256] - - [314, 70.36] - - - [43264, 27648, 1, 256] - - [25, 74.283] - - - [18432, 9216, 1, 256] - - [35, 75.027] - - - [38912, 3072, 1, 256] - - [25, 73.845] - - - [30464, 17921, 1, 256] - - [39, 72.452] - - - [37376, 9216, 1, 256] - - [39, 74.853] - - - [256, 3072, 1, 256] - - [105, 36.28] - - - [9472, 3072, 1, 256] - - [390, 79.736] - - - [35840, 23552, 1, 256] - - [27, 75.421] - - - [8960, 3072, 1, 256] - - [296, 79.565] - - - [34816, 3072, 1, 256] - - [23, 73.722] - - - [11008, 3072, 1, 256] - - [295, 78.107] - - - [36864, 1536, 1, 256] - - [25, 70.739] - - - [23552, 9216, 1, 256] - - [23, 74.931] - - - [31232, 18945, 1, 256] - - [39, 73.789] - - - [27136, 9216, 1, 256] - - [41, 74.972] - - - [19968, 7681, 1, 256] - - [39, 72.249] - - - [31488, 18945, 1, 256] - - [30, 73.064] - - - [33280, 1792, 1, 256] - - [29, 72.275] - - - [14592, 3072, 1, 256] - - [57, 70.802] - - - [30976, 18689, 1, 256] - - [39, 72.645] - - - [4096, 769, 1, 256] - - [301, 45.847] - - - [31488, 3072, 1, 256] - - [35, 72.752] - - - [33024, 1792, 1, 256] - - [57, 70.613] - - - [11520, 8449, 1, 256] - - [28, 71.344] - - - [44544, 4353, 1, 256] - - [30, 71.136] - - - [18176, 5889, 1, 256] - - [23, 70.944] - - - [5632, 2305, 1, 256] - - [417, 67.498] - - - [39936, 27393, 1, 256] - - [52, 73.828] - - - [10240, 7169, 1, 256] - - [35, 70.424] - - - [39168, 26625, 1, 256] - - [26, 72.658] - - - [10752, 7681, 1, 256] - - [25, 71.73] - - - [13824, 1536, 1, 256] - - [299, 74.789] - - - [14336, 9216, 1, 256] - - [35, 74.678] - - - [37632, 25345, 1, 256] - - [52, 73.074] - - - [35840, 23553, 1, 256] - - [30, 73.019] - - - [23552, 3072, 1, 256] - - [23, 73.223] - - - [19712, 7169, 1, 256] - - [52, 69.862] - - - [5888, 2561, 1, 256] - - [299, 69.781] - - - [27136, 768, 1, 256] - - [299, 75.15] - - - [22272, 1792, 1, 256] - - [32, 70.322] - - - [15616, 1536, 1, 256] - - [348, 77.084] - - - [3840, 769, 1, 256] - - [418, 46.572] - - - [42240, 2304, 1, 256] - - [35, 72.628] - - - [24576, 3072, 1, 256] - - [20, 68.65] - - - [27136, 1536, 1, 256] - - [29, 70.093] - - - [25344, 12801, 1, 256] - - [26, 72.186] - - - [32512, 20224, 1, 256] - - [45, 75.143] - - - [17664, 3072, 1, 256] - - [55, 71.451] - - - [28160, 15873, 1, 256] - - [30, 73.603] - - - [40960, 3072, 1, 256] - - [36, 66.819] - - - [14592, 9216, 1, 256] - - [69, 72.836] - - - [22784, 10497, 1, 256] - - [39, 72.571] - - - [22272, 3072, 1, 256] - - [55, 72.388] - - - [39680, 27137, 1, 256] - - [30, 72.405] - - - [20992, 8704, 1, 256] - - [35, 75.323] - - - [24320, 1536, 1, 256] - - [357, 79.984] - - - [7936, 4865, 1, 256] - - [390, 78.247] - - - [17664, 5376, 1, 256] - - [55, 73.382] - - - [37888, 25345, 1, 256] - - [52, 73.968] - - - [23296, 10753, 1, 256] - - [30, 72.528] - - - [28416, 15873, 1, 256] - - [68, 73.027] - - - [27648, 15361, 1, 256] - - [30, 72.759] - - - [39424, 1536, 1, 256] - - [29, 70.892] - - - [15104, 2817, 1, 256] - - [27, 68.216] - - - [19456, 9216, 1, 256] - - [23, 74.881] - - - [24064, 11777, 1, 256] - - [41, 73.397] - - - [40448, 1536, 1, 256] - - [28, 71.11] - - - [512, 3072, 1, 256] - - [99, 44.68] - - - [38912, 9216, 1, 256] - - [27, 74.717] - - - [19456, 6913, 1, 256] - - [25, 72.86] - - - [29440, 1792, 1, 256] - - [53, 71.532] - - - [41984, 9216, 1, 256] - - [27, 74.61] - - - [14080, 1793, 1, 256] - - [284, 73.035] - - - [20992, 8449, 1, 256] - - [30, 72.683] - - - [17920, 768, 1, 256] - - [390, 69.974] - - - [10496, 7169, 1, 256] - - [52, 69.631] - - - [40704, 27648, 1, 256] - - [25, 74.406] - - - [13568, 1025, 1, 256] - - [300, 66.483] - - - [38144, 9216, 1, 256] - - [39, 74.19] - - - [27392, 15104, 1, 256] - - [41, 74.726] - - - [2304, 3072, 1, 256] - - [388, 63.009] - - - [9472, 6401, 1, 256] - - [23, 70.416] - - - [39424, 1792, 1, 256] - - [57, 72.285] - - - [41728, 768, 1, 256] - - [299, 76.883] - - - [11264, 3072, 1, 256] - - [299, 81.81] - - - [25344, 3072, 1, 256] - - [55, 72.1] - - - [24576, 1792, 1, 256] - - [23, 67.901] - - - [27392, 14849, 1, 256] - - [39, 72.861] - - - [14848, 2561, 1, 256] - - [299, 77.918] - - - [28160, 3072, 1, 256] - - [25, 73.026] - - - [23552, 11009, 1, 256] - - [52, 73.699] - - - [11776, 8449, 1, 256] - - [52, 72.092] - - - [16640, 1792, 1, 256] - - [299, 77.962] - - - [24576, 12289, 1, 256] - - [30, 67.331] - - - [38656, 26369, 1, 256] - - [26, 72.925] - - - [13824, 9216, 1, 256] - - [59, 74.326] - - - [28928, 1792, 1, 256] - - [25, 71.293] - - - [27904, 15361, 1, 256] - - [39, 72.692] - - - [3840, 1792, 1, 256] - - [301, 61.778] - - - [14848, 3072, 1, 256] - - [35, 71.769] - - - [27904, 1536, 1, 256] - - [65, 70.041] - - - [34816, 1536, 1, 256] - - [25, 71.375] - - - [14592, 2305, 1, 256] - - [299, 74.166] - - - [22528, 9985, 1, 256] - - [27, 73.686] - - - [26368, 13825, 1, 256] - - [71, 72.918] - - - [4096, 1792, 1, 256] - - [295, 62.735] - - - [30720, 18177, 1, 256] - - [52, 74.079] - - - [37120, 24833, 1, 256] - - [39, 73.124] - - - [24320, 3072, 1, 256] - - [23, 72.794] - - - [2560, 1536, 1, 256] - - [392, 53.614] - - - [44032, 4097, 1, 256] - - [52, 70.115] - - - [44544, 27648, 1, 256] - - [26, 74.734] - - - [34048, 21761, 1, 256] - - [26, 72.608] - - - [24064, 1536, 1, 256] - - [299, 79.874] - - - [24832, 12545, 1, 256] - - [71, 73.308] - - - [44032, 3841, 1, 256] - - [23, 70.954] - - - [40448, 257, 1, 256] - - [295, 57.118] - - - [26624, 14337, 1, 256] - - [30, 72.796] - - - [8192, 5121, 1, 256] - - [23, 68.016] - - - [42240, 1536, 1, 256] - - [25, 70.765] - - - [5888, 2817, 1, 256] - - [284, 69.798] - - - [6144, 1792, 1, 256] - - [390, 69.228] - - - [16384, 1792, 1, 256] - - [302, 67.804] - - - [35584, 23297, 1, 256] - - [30, 73.151] - - - [36352, 24064, 1, 256] - - [26, 75.214] - - - [23040, 1536, 1, 256] - - [299, 79.111] - - - [8704, 1536, 1, 256] - - [267, 72.256] - - - [18432, 6145, 1, 256] - - [35, 70.988] - - - [12032, 3072, 1, 256] - - [299, 80.295] - - - [39168, 3072, 1, 256] - - [28, 72.818] - - - [28160, 1536, 1, 256] - - [52, 70.098] - - - [41728, 27648, 1, 256] - - [39, 73.797] - - - [28416, 1792, 1, 256] - - [29, 70.953] - - - [24320, 12032, 1, 256] - - [22, 75.433] - - - [28928, 16385, 1, 256] - - [52, 72.397] - - - [34816, 22528, 1, 256] - - [23, 75.054] - - - [26368, 1792, 1, 256] - - [53, 71.252] - - - [25856, 13569, 1, 256] - - [26, 72.951] - - - [25600, 13312, 1, 256] - - [25, 75.771] - - - [31232, 18689, 1, 256] - - [26, 73.883] - - - [20736, 9216, 1, 256] - - [35, 74.492] - - - [34304, 9216, 1, 256] - - [40, 75.021] - - - [43264, 3073, 1, 256] - - [52, 69.053] - - - [8704, 5633, 1, 256] - - [35, 70.581] - - - [4864, 1793, 1, 256] - - [300, 61.197] - - - [41984, 3072, 1, 256] - - [27, 73.928] - - - [20992, 3072, 1, 256] - - [23, 72.977] - - - [9728, 6401, 1, 256] - - [27, 70.891] - - - [16640, 4097, 1, 256] - - [52, 67.999] - - - [38400, 9216, 1, 256] - - [26, 74.802] - - - [38656, 1536, 1, 256] - - [29, 70.607] - - - [1536, 3072, 1, 256] - - [294, 56.669] - - - [12544, 1792, 1, 256] - - [298, 77.768] - - - [37632, 1792, 1, 256] - - [65, 71.968] - - - [17152, 4609, 1, 256] - - [25, 70.191] - - - [18944, 6656, 1, 256] - - [28, 74.667] - - - [34560, 22017, 1, 256] - - [30, 73.029] - - - [23296, 11008, 1, 256] - - [25, 74.913] - - - [14848, 768, 1, 256] - - [289, 70.288] - - - [38656, 1792, 1, 256] - - [29, 71.937] - - - [8448, 5377, 1, 256] - - [29, 70.346] - - - [29952, 17665, 1, 256] - - [71, 73.321] - - - [33792, 21504, 1, 256] - - [25, 75.406] - - - [24576, 1536, 1, 256] - - [360, 71.81] - - - [37376, 1792, 1, 256] - - [23, 72.422] - - - [42752, 768, 1, 256] - - [390, 76.466] - - - [4096, 1025, 1, 256] - - [388, 49.512] - - - [35840, 768, 1, 256] - - [299, 77.817] - - - [19200, 3072, 1, 256] - - [47, 71.676] - - - [33536, 1792, 1, 256] - - [32, 71.782] - - - [36864, 9216, 1, 256] - - [20, 74.017] - - - [38656, 26368, 1, 256] - - [26, 74.56] - - - [44288, 9216, 1, 256] - - [39, 74.399] - - - [44288, 4097, 1, 256] - - [52, 69.82] - - - [26112, 3072, 1, 256] - - [25, 73.247] - - - [512, 768, 1, 256] - - [90, 22.789] - - - [36096, 3072, 1, 256] - - [22, 71.493] - - - [4864, 1537, 1, 256] - - [300, 58.35] - - - [31232, 18944, 1, 256] - - [28, 75.689] - - - [20224, 7681, 1, 256] - - [26, 71.553] - - - [26112, 9216, 1, 256] - - [39, 74.995] - - - [21504, 3072, 1, 256] - - [23, 73.129] - - - [12544, 3072, 1, 256] - - [296, 81.177] - - - [32256, 19713, 1, 256] - - [68, 73.799] - - - [40704, 1792, 1, 256] - - [55, 72.126] - - - [18176, 5888, 1, 256] - - [32, 73.909] - - - [33792, 9216, 1, 256] - - [25, 74.743] - - - [26624, 14336, 1, 256] - - [35, 75.385] - - - [38912, 1792, 1, 256] - - [27, 72.82] - - - [7936, 1792, 1, 256] - - [301, 69.826] - - - [28672, 16385, 1, 256] - - [52, 72.361] - - - [18944, 3072, 1, 256] - - [57, 72.445] - - - [33280, 20993, 1, 256] - - [41, 73.865] - - - [37120, 24832, 1, 256] - - [25, 74.866] - - - [43520, 1792, 1, 256] - - [25, 72.876] - - - [16896, 4609, 1, 256] - - [28, 70.699] - - - [41472, 1536, 1, 256] - - [35, 71.132] - - - [39936, 768, 1, 256] - - [299, 77.411] - - - [23296, 11009, 1, 256] - - [52, 72.525] - - - [26624, 9216, 1, 256] - - [35, 74.827] - - - [29184, 9216, 1, 256] - - [72, 74.75] - - - [36352, 9216, 1, 256] - - [26, 74.823] - - - [37632, 25344, 1, 256] - - [27, 74.856] - - - [37888, 25600, 1, 256] - - [25, 75.356] - - - [16640, 9216, 1, 256] - - [39, 74.324] - - - [44544, 9216, 1, 256] - - [26, 74.614] - - - [14080, 1792, 1, 256] - - [298, 75.35] - - - [33536, 21249, 1, 256] - - [68, 73.389] - - - [34048, 21760, 1, 256] - - [62, 74.55] - - - [9984, 768, 1, 256] - - [390, 62.485] - - - [40192, 1536, 1, 256] - - [23, 71.09] - - - [41728, 3072, 1, 256] - - [28, 71.64] - - - [35328, 9216, 1, 256] - - [41, 74.937] - - - [32512, 768, 1, 256] - - [390, 75.18] - - - [14592, 2049, 1, 256] - - [299, 72.37] - - - [14848, 9216, 1, 256] - - [39, 74.684] - - - [23808, 3072, 1, 256] - - [23, 72.789] - - - [13568, 9216, 1, 256] - - [27, 74.107] - - - [42496, 2560, 1, 256] - - [23, 73.708] - - - [42752, 3072, 1, 256] - - [23, 73.224] - - - [39680, 27392, 1, 256] - - [35, 74.582] - - - [14592, 1792, 1, 256] - - [299, 76.849] - - - [25600, 13313, 1, 256] - - [30, 72.762] - - - [26624, 1792, 1, 256] - - [25, 72.142] - - - [20480, 8193, 1, 256] - - [52, 71.705] - - - [36096, 23808, 1, 256] - - [39, 74.247] - - - [15104, 2561, 1, 256] - - [299, 77.852] - - - [43520, 3072, 1, 256] - - [25, 73.788] - - - [1280, 2048, 1, 256] - - [299, 43.453] - - - [43008, 1792, 1, 256] - - [23, 72.967] - - - [18688, 3072, 1, 256] - - [23, 72.245] - - - [35328, 23040, 1, 256] - - [49, 75.432] - - - [18944, 6401, 1, 256] - - [71, 71.873] - - - [16128, 3585, 1, 256] - - [23, 69.134] - - - [29952, 1536, 1, 256] - - [35, 70.233] - - - [17408, 5121, 1, 256] - - [52, 70.042] - - - [36608, 1792, 1, 256] - - [32, 71.769] - - - [13056, 768, 1, 256] - - [294, 67.527] - - - [26112, 13824, 1, 256] - - [63, 75.824] - - - [43520, 3585, 1, 256] - - [52, 70.545] - - - [40704, 9216, 1, 256] - - [35, 74.131] - - - [27904, 15617, 1, 256] - - [40, 73.165] - - - [21248, 3072, 1, 256] - - [47, 72.225] - - - [38912, 1536, 1, 256] - - [27, 71.532] - - - [28672, 1792, 1, 256] - - [25, 71.832] - - - [18432, 1792, 1, 256] - - [298, 79.551] - - - [29952, 9216, 1, 256] - - [45, 74.232] - - - [4352, 1025, 1, 256] - - [295, 52.61] - - - [34304, 22017, 1, 256] - - [41, 73.821] - - - [28160, 15617, 1, 256] - - [52, 73.595] - - - [19968, 9216, 1, 256] - - [26, 74.953] - - - [7424, 4353, 1, 256] - - [299, 75.935] - - - [19200, 1792, 1, 256] - - [390, 78.957] - - - [27648, 15360, 1, 256] - - [27, 75.75] - - - [23040, 10497, 1, 256] - - [68, 73.426] - - - [21248, 8961, 1, 256] - - [68, 72.223] - - - [32256, 1792, 1, 256] - - [55, 72.227] - - - [26112, 13569, 1, 256] - - [41, 73.707] - - - [12288, 8961, 1, 256] - - [35, 72.875] - - - [6656, 3585, 1, 256] - - [390, 73.481] - - - [19968, 7425, 1, 256] - - [52, 72.526] - - - [9472, 768, 1, 256] - - [300, 63.494] - - - [33792, 3072, 1, 256] - - [27, 73.595] - - - [15616, 3072, 1, 256] - - [35, 71.795] - - - [8704, 5377, 1, 256] - - [25, 70.194] - - - [11520, 3072, 1, 256] - - [299, 80.145] - - - [25856, 1536, 1, 256] - - [33, 69.702] - - - [28416, 768, 1, 256] - - [289, 74.515] - - - [32256, 3072, 1, 256] - - [23, 73.426] - - - [20736, 1536, 1, 256] - - [390, 78.453] - - - [22784, 10241, 1, 256] - - [39, 71.799] - - - [36608, 24321, 1, 256] - - [39, 72.971] - - - [36096, 9216, 1, 256] - - [26, 73.759] - - - [10752, 768, 1, 256] - - [295, 64.296] - - - [38400, 26112, 1, 256] - - [25, 75.376] - - - [9216, 5889, 1, 256] - - [53, 70.705] - - - [41472, 27648, 1, 256] - - [41, 74.643] - - - [38144, 25856, 1, 256] - - [35, 74.767] - - - [15360, 3073, 1, 256] - - [25, 67.924] - - - [29184, 16896, 1, 256] - - [23, 75.487] - - - [16128, 1792, 1, 256] - - [299, 78.104] - - - [32768, 20225, 1, 256] - - [86, 57.927] - - - [23040, 10752, 1, 256] - - [45, 75.65] - - - [15872, 3585, 1, 256] - - [23, 69.965] - - - [11008, 7681, 1, 256] - - [73, 70.59] - - - [15360, 9216, 1, 256] - - [35, 74.876] - - - [28416, 16128, 1, 256] - - [49, 74.891] - - - [30208, 1792, 1, 256] - - [27, 71.803] - - - [41728, 1792, 1, 256] - - [62, 71.002] - - - [32256, 19968, 1, 256] - - [49, 75.446] - - - [18944, 1792, 1, 256] - - [296, 79.895] - - - [41728, 1793, 1, 256] - - [23, 66.538] - - - [31488, 19201, 1, 256] - - [30, 73.037] - - - [40192, 257, 1, 256] - - [421, 57.031] - - - [42752, 27648, 1, 256] - - [27, 74.278] - - - [40704, 768, 1, 256] - - [336, 75.441] - - - [25088, 12545, 1, 256] - - [22, 73.619] - - - [24576, 9216, 1, 256] - - [36, 70.049] - - - [33024, 20737, 1, 256] - - [71, 73.586] - - - [29696, 9216, 1, 256] - - [25, 74.863] - - - [31232, 1536, 1, 256] - - [23, 70.863] - - - [30208, 17920, 1, 256] - - [22, 75.773] - - - [44544, 4609, 1, 256] - - [52, 71.087] - - - [22016, 9728, 1, 256] - - [35, 75.247] - - - [30208, 17921, 1, 256] - - [40, 73.75] - - - [19200, 6657, 1, 256] - - [68, 71.173] - - - [22016, 9729, 1, 256] - - [22, 73.212] - - - [18176, 768, 1, 256] - - [419, 69.753] - - - [29184, 1792, 1, 256] - - [57, 71.58] - - - [12288, 1792, 1, 256] - - [298, 75.353] - - - [22528, 1536, 1, 256] - - [309, 78.821] - - - [14848, 2305, 1, 256] - - [390, 74.401] - - - [41216, 1025, 1, 256] - - [67, 63.189] - - - [8192, 3072, 1, 256] - - [350, 74.552] - - - [5888, 1792, 1, 256] - - [419, 67.252] - - - [21760, 3072, 1, 256] - - [33, 72.375] - - - [22272, 9985, 1, 256] - - [71, 72.608] - - - [29184, 1536, 1, 256] - - [29, 70.393] - - - [22016, 3072, 1, 256] - - [45, 72.804] - - - [30720, 9216, 1, 256] - - [25, 74.894] - - - [39680, 1792, 1, 256] - - [27, 72.047] - - - [9728, 1536, 1, 256] - - [296, 70.944] - - - [34560, 9216, 1, 256] - - [35, 74.255] - - - [12032, 8705, 1, 256] - - [35, 71.355] - - - [10752, 7425, 1, 256] - - [49, 71.821] - - - [18688, 1536, 1, 256] - - [300, 78.363] - - - [16128, 3840, 1, 256] - - [53, 72.828] - - - [38656, 768, 1, 256] - - [390, 77.663] - - - [21248, 1792, 1, 256] - - [299, 82.048] - - - [36352, 3072, 1, 256] - - [25, 73.493] - - - [19968, 7680, 1, 256] - - [27, 75.125] - - - [3840, 513, 1, 256] - - [349, 38.608] - - - [38400, 3072, 1, 256] - - [35, 73.324] - - - [5376, 768, 1, 256] - - [270, 52.964] - - - [20224, 9216, 1, 256] - - [35, 74.585] - - - [17408, 5120, 1, 256] - - [25, 74.321] - - - [28928, 9216, 1, 256] - - [23, 74.117] - - - [35072, 1792, 1, 256] - - [27, 71.863] - - - [31488, 19200, 1, 256] - - [25, 75.016] - - - [11008, 7937, 1, 256] - - [69, 70.797] - - - [21248, 8705, 1, 256] - - [71, 72.196] - - - [13568, 3072, 1, 256] - - [32, 71.084] - - - [34560, 22273, 1, 256] - - [30, 73.128] - - - [34048, 768, 1, 256] - - [299, 76.64] - - - [40448, 27648, 1, 256] - - [40, 74.661] - - - [28416, 16129, 1, 256] - - [52, 73.072] - - - [34816, 22529, 1, 256] - - [30, 72.704] - - - [22528, 3072, 1, 256] - - [25, 73.326] - - - [27136, 14593, 1, 256] - - [68, 73.833] - - - [35584, 3072, 1, 256] - - [27, 72.72] - - - [43008, 3072, 1, 256] - - [25, 73.831] - - - [30464, 1792, 1, 256] - - [29, 70.013] - - - [16384, 4097, 1, 256] - - [25, 59.138] - - - [20992, 9216, 1, 256] - - [39, 75.069] - - - [31488, 1792, 1, 256] - - [32, 71.727] - - - [31488, 9216, 1, 256] - - [25, 74.117] - - - [22272, 9984, 1, 256] - - [55, 75.013] - - - [41728, 1537, 1, 256] - - [25, 65.171] - - - [26880, 1792, 1, 256] - - [25, 71.327] - - - [30464, 768, 1, 256] - - [390, 74.466] - - - [2816, 1792, 1, 256] - - [425, 56.081] - - - [41472, 1537, 1, 256] - - [35, 66.289] - - - [43008, 27648, 1, 256] - - [23, 74.892] - - - [39424, 27137, 1, 256] - - [39, 73.628] - - - [24320, 1792, 1, 256] - - [29, 70.912] - - - [32000, 3072, 1, 256] - - [23, 72.768] - - - [12800, 1792, 1, 256] - - [296, 77.813] - - - [15872, 3072, 1, 256] - - [23, 72.001] - - - [15872, 1792, 1, 256] - - [296, 78.676] - - - [10496, 7425, 1, 256] - - [52, 71.106] - - - [16896, 4608, 1, 256] - - [23, 73.401] - - - [9984, 6913, 1, 256] - - [23, 71.091] - - - [21248, 8960, 1, 256] - - [25, 74.939] - - - [14336, 1792, 1, 256] - - [289, 76.971] - - - [24832, 12544, 1, 256] - - [29, 75.359] - - - [30464, 18176, 1, 256] - - [26, 74.481] - - - [31744, 19201, 1, 256] - - [25, 74.065] - - - [1792, 768, 1, 256] - - [416, 33.854] - - - [1536, 2048, 1, 256] - - [299, 49.863] - - - [40192, 3072, 1, 256] - - [35, 73.329] - - - [42240, 3072, 1, 256] - - [27, 73.139] - - - [32256, 9216, 1, 256] - - [40, 74.881] - - - [41984, 2049, 1, 256] - - [72, 66.57] - - - [6656, 1792, 1, 256] - - [282, 67.16] - - - [13824, 1537, 1, 256] - - [285, 69.76] - - - [20736, 3072, 1, 256] - - [25, 72.45] - - - [36096, 23809, 1, 256] - - [39, 72.71] - - - [41728, 9216, 1, 256] - - [26, 73.108] - - - [25600, 768, 1, 256] - - [287, 72.574] - - - [37632, 768, 1, 256] - - [299, 76.689] - - - [25600, 9216, 1, 256] - - [27, 75.002] - - - [19968, 3072, 1, 256] - - [27, 72.482] - - - [15616, 9216, 1, 256] - - [27, 74.411] - - - [29184, 16897, 1, 256] - - [30, 73.625] - - - [7168, 3841, 1, 256] - - [296, 74.558] - - - [40704, 769, 1, 256] - - [298, 68.336] - - - [6144, 3073, 1, 256] - - [299, 69.805] - - - [34304, 1792, 1, 256] - - [57, 72.252] - - - [18688, 6400, 1, 256] - - [35, 74.512] - - - [20992, 1536, 1, 256] - - [299, 78.591] - - - [21760, 768, 1, 256] - - [357, 71.08] - - - [43264, 3072, 1, 256] - - [23, 73.263] - - - [21760, 9216, 1, 256] - - [23, 74.502] - - - [11264, 768, 1, 256] - - [300, 65.325] - - - [42496, 3072, 1, 256] - - [35, 73.716] - - - [30208, 17665, 1, 256] - - [71, 73.922] - - - [27392, 15105, 1, 256] - - [26, 72.88] - - - [29952, 17409, 1, 256] - - [71, 72.467] - - - [44032, 3072, 1, 256] - - [35, 73.868] - - - [41216, 9216, 1, 256] - - [39, 74.596] - - - [8448, 1536, 1, 256] - - [298, 69.988] - - - [36352, 768, 1, 256] - - [337, 76.575] - - - [23552, 768, 1, 256] - - [419, 73.196] - - - [7168, 3072, 1, 256] - - [419, 75.672] - - - [44288, 4353, 1, 256] - - [52, 70.362] - - - [36608, 768, 1, 256] - - [299, 77.672] - - - [15616, 3073, 1, 256] - - [25, 68.049] - - - [37376, 24833, 1, 256] - - [26, 73.814] - - - [38144, 25857, 1, 256] - - [26, 72.949] - - - [26880, 14592, 1, 256] - - [27, 75.253] - - - [6144, 2817, 1, 256] - - [304, 70.34] - - - [23808, 768, 1, 256] - - [300, 71.514] - - - [39168, 26881, 1, 256] - - [39, 72.928] - - - [5120, 1793, 1, 256] - - [308, 62.708] - - - [32512, 19969, 1, 256] - - [41, 73.225] - - - [43008, 2817, 1, 256] - - [23, 69.9] - - - [26112, 13825, 1, 256] - - [41, 73.821] - - - [33536, 3072, 1, 256] - - [49, 72.881] - - - [9728, 6657, 1, 256] - - [25, 71.388] - - - [2048, 3072, 1, 256] - - [417, 61.286] - - - [24832, 9216, 1, 256] - - [27, 74.486] - - - [5632, 2561, 1, 256] - - [421, 69.063] - - - [33280, 20992, 1, 256] - - [49, 75.582] - - - [20224, 7936, 1, 256] - - [25, 74.62] - - - [28672, 16384, 1, 256] - - [25, 74.944] - - - [28416, 9216, 1, 256] - - [30, 73.936] - - - [7936, 768, 1, 256] - - [295, 60.517] - - - [23552, 11265, 1, 256] - - [23, 72.46] - - - [25088, 3072, 1, 256] - - [27, 72.818] - - - [32000, 19457, 1, 256] - - [30, 72.725] - - - [44800, 3072, 1, 256] - - [23, 72.719] - - - [37120, 1792, 1, 256] - - [25, 71.806] - - - [30464, 18177, 1, 256] - - [26, 72.706] - - - [44544, 4608, 1, 256] - - [26, 73.899] - - - [7168, 768, 1, 256] - - [293, 57.211] - - - [18944, 9216, 1, 256] - - [40, 74.773] - - - [33280, 20737, 1, 256] - - [68, 73.951] - - - [25856, 3072, 1, 256] - - [47, 72.641] - - - [27648, 9216, 1, 256] - - [35, 74.837] - - - [5120, 2049, 1, 256] - - [289, 63.148] - - - [28160, 9216, 1, 256] - - [46, 74.804] - - - [37632, 25089, 1, 256] - - [30, 73.058] - - - [22016, 1792, 1, 256] - - [27, 71.088] - - - [16384, 9216, 1, 256] - - [20, 63.2] - - - [21504, 9217, 1, 256] - - [30, 72.143] - - - [20480, 7937, 1, 256] - - [30, 72.788] - - - [33536, 21248, 1, 256] - - [49, 75.086] - - - [12800, 768, 1, 256] - - [421, 66.364] - - - [28672, 9216, 1, 256] - - [35, 74.277] - - - [32000, 9216, 1, 256] - - [25, 74.163] - - - [44544, 3072, 1, 256] - - [23, 73.355] - - - [5376, 3072, 1, 256] - - [300, 72.786] - - - [35840, 23297, 1, 256] - - [30, 74.027] - - - [23808, 11521, 1, 256] - - [30, 72.524] - - - [13312, 1025, 1, 256] - - [421, 65.505] - - - [18176, 9216, 1, 256] - - [35, 74.483] - - - [17920, 5633, 1, 256] - - [57, 71.536] - - - [27648, 3072, 1, 256] - - [23, 73.527] - - - [1024, 3072, 1, 256] - - [348, 49.484] - - - [22016, 9216, 1, 256] - - [39, 75.131] - - - [21760, 9473, 1, 256] - - [30, 72.475] - - - [6144, 1536, 1, 256] - - [305, 65.214] - - - [16896, 1536, 1, 256] - - [357, 77.847] - - - [19968, 768, 1, 256] - - [299, 68.668] - - - [23552, 11264, 1, 256] - - [35, 75.562] - - - [27904, 3072, 1, 256] - - [57, 72.351] - - - [19712, 7425, 1, 256] - - [71, 71.126] - - - [26624, 14081, 1, 256] - - [35, 73.889] - - - [3328, 257, 1, 256] - - [92, 31.019] - - - [24320, 9216, 1, 256] - - [45, 74.48] - - - [14080, 3072, 1, 256] - - [57, 70.439] - - - [17408, 3072, 1, 256] - - [35, 72.646] - - - [21504, 9216, 1, 256] - - [25, 74.988] - - - [14848, 2560, 1, 256] - - [296, 82.39] - - - [34304, 3072, 1, 256] - - [45, 73.279] - - - [15104, 9216, 1, 256] - - [23, 74.312] - - - [17152, 4865, 1, 256] - - [23, 70.61] - - - [38912, 26625, 1, 256] - - [52, 72.727] - - - [41216, 1792, 1, 256] - - [23, 71.942] - - - [39424, 3072, 1, 256] - - [27, 73.417] - - - [30720, 18433, 1, 256] - - [30, 72.945] - - - [18944, 6657, 1, 256] - - [22, 72.217] - - - [5632, 1792, 1, 256] - - [291, 65.8] - - - [18176, 1792, 1, 256] - - [390, 79.438] - - - [31232, 9216, 1, 256] - - [26, 74.854] - - - [42752, 2561, 1, 256] - - [30, 68.13] - - - [18688, 9216, 1, 256] - - [35, 74.553] - - - [43776, 1792, 1, 256] - - [39, 69.792] - - - [10240, 1792, 1, 256] - - [299, 72.141] - - - [33792, 21505, 1, 256] - - [30, 72.863] - - - [25856, 13313, 1, 256] - - [26, 72.38] - - - [29952, 3072, 1, 256] - - [28, 72.875] - - - [5888, 768, 1, 256] - - [419, 54.378] - - - [20480, 9216, 1, 256] - - [35, 74.621] - - - [17152, 1792, 1, 256] - - [299, 78.232] - - - [23040, 10753, 1, 256] - - [71, 73.355] - - - [8960, 5889, 1, 256] - - [55, 70.343] - - - [16640, 4352, 1, 256] - - [32, 72.785] - - - [30464, 3072, 1, 256] - - [55, 71.363] - - - [16128, 9216, 1, 256] - - [23, 74.406] - - - [25344, 13057, 1, 256] - - [39, 72.612] - - - [39424, 9216, 1, 256] - - [39, 74.936] - - - [25600, 3072, 1, 256] - - [27, 73.346] - - - [28416, 3072, 1, 256] - - [45, 72.357] - - - [12800, 257, 1, 256] - - [421, 47.209] - - - [43264, 1792, 1, 256] - - [23, 72.181] - - - [20736, 8193, 1, 256] - - [52, 71.591] - - - [30976, 9216, 1, 256] - - [39, 73.289] - - - [40192, 27648, 1, 256] - - [23, 74.477] - - - [31232, 1792, 1, 256] - - [33, 71.89] - - - [36352, 23809, 1, 256] - - [26, 73.868] - - - [9984, 3072, 1, 256] - - [299, 78.695] - - - [11776, 1792, 1, 256] - - [298, 74.637] - - - [37120, 1536, 1, 256] - - [47, 70.444] - - - [14592, 2304, 1, 256] - - [299, 78.538] - - - [7424, 768, 1, 256] - - [270, 58.909] - - - [10240, 1536, 1, 256] - - [390, 72.415] - - - [27392, 9216, 1, 256] - - [39, 74.083] - - - [15104, 3072, 1, 256] - - [32, 71.582] - - - [19200, 9216, 1, 256] - - [26, 74.336] - - - [36096, 23553, 1, 256] - - [26, 72.256] - - - [16128, 3841, 1, 256] - - [25, 69.422] - - - [18432, 5889, 1, 256] - - [23, 72.062] - - - [43776, 3841, 1, 256] - - [26, 69.142] - - - [22528, 10241, 1, 256] - - [52, 72.313] - - - [20224, 3072, 1, 256] - - [27, 72.17] - - - [39680, 3072, 1, 256] - - [35, 72.868] - - - [20736, 8449, 1, 256] - - [30, 72.238] - - - [30720, 1792, 1, 256] - - [53, 72.468] - - - [36864, 24321, 1, 256] - - [30, 73.196] - - - [22784, 1536, 1, 256] - - [289, 78.864] - - - [7424, 4097, 1, 256] - - [299, 75.786] - - - [7680, 4609, 1, 256] - - [299, 79.782] - - - [12032, 768, 1, 256] - - [417, 64.656] - - - [1792, 3072, 1, 256] - - [392, 57.972] - - - [6400, 3073, 1, 256] - - [391, 70.804] - - - [29440, 17153, 1, 256] - - [71, 73.313] - - - [8704, 1792, 1, 256] - - [284, 71.254] - - - [30720, 3072, 1, 256] - - [27, 73.577] - - - [16384, 3841, 1, 256] - - [36, 58.468] - - - [40192, 9216, 1, 256] - - [27, 74.318] - - - [23040, 1792, 1, 256] - - [47, 71.001] - - - [37888, 25601, 1, 256] - - [52, 72.878] - - - [26368, 14080, 1, 256] - - [23, 75.251] - - - [30208, 3072, 1, 256] - - [27, 72.997] - - - [33024, 20736, 1, 256] - - [41, 75.079] - - - [35072, 22784, 1, 256] - - [23, 74.903] - - - [9472, 6145, 1, 256] - - [25, 69.301] - - - [22784, 1792, 1, 256] - - [32, 70.849] - - - [768, 2048, 1, 256] - - [94, 44.352] - - - [1024, 1280, 1, 256] - - [346, 33.09] - - - [41984, 27648, 1, 256] - - [35, 74.925] - - - [33024, 20481, 1, 256] - - [40, 72.827] - - - [33280, 1536, 1, 256] - - [47, 71.098] - - - [9216, 3072, 1, 256] - - [357, 78.086] - - - [22528, 1792, 1, 256] - - [32, 71.567] - - - [25088, 768, 1, 256] - - [289, 72.307] - - - [13825, 128, 1, 128] - - [107, 26.713] - - - [20609, 128, 1, 256] - - [184, 33.58] - - - [6017, 128, 1, 256] - - [108, 30.524] - - - [2305, 128, 1, 128] - - [109, 10.976] - - - [15745, 128, 1, 256] - - [108, 30.656] - - - [8833, 128, 1, 128] - - [110, 23.452] - - - [641, 128, 1, 128] - - [111, 3.197] - - - [9217, 128, 1, 128] - - [112, 23.704] - - - [15361, 128, 1, 256] - - [108, 31.012] - - - [22913, 128, 1, 256] - - [185, 36.177] - - - [2177, 128, 1, 128] - - [100, 10.367] - - - [19073, 128, 1, 256] - - [118, 33.209] - - - [28289, 128, 1, 128] - - [108, 23.868] - - - [13057, 128, 1, 256] - - [102, 37.574] - - - [1793, 128, 1, 128] - - [113, 8.538] - - - [16769, 128, 1, 128] - - [134, 20.301] - - - [23681, 128, 1, 256] - - [186, 36.993] - - - [14593, 128, 1, 256] - - [118, 30.384] - - - [24449, 128, 1, 128] - - [129, 23.36] - - - [4609, 128, 1, 256] - - [114, 24.923] - - - [10625, 128, 1, 128] - - [115, 23.134] - - - [12545, 128, 1, 256] - - [116, 36.242] - - - [5633, 128, 1, 128] - - [117, 18.479] - - - [641, 128, 1, 256] - - [114, 4.731] - - - [18305, 128, 1, 256] - - [118, 31.233] - - - [23297, 128, 1, 256] - - [192, 36.822] - - - [21377, 128, 1, 256] - - [186, 34.302] - - - [9601, 128, 1, 128] - - [118, 24.866] - - - [13697, 128, 1, 256] - - [129, 29.534] - - - [23681, 128, 1, 128] - - [108, 23.301] - - - [24833, 128, 1, 256] - - [187, 38.387] - - - [25985, 128, 1, 128] - - [190, 25.37] - - - [9601, 128, 1, 256] - - [120, 35.612] - - - [17153, 128, 1, 128] - - [118, 20.308] - - - [9985, 128, 1, 128] - - [121, 25.327] - - - [23297, 128, 1, 128] - - [122, 22.878] - - - [19073, 128, 1, 128] - - [130, 21.586] - - - [2689, 128, 1, 256] - - [100, 18.463] - - - [4993, 128, 1, 128] - - [100, 19.814] - - - [6913, 128, 1, 256] - - [122, 33.681] - - - [6785, 128, 1, 128] - - [113, 22.258] - - - [27905, 128, 1, 128] - - [122, 23.693] - - - [7169, 128, 1, 256] - - [123, 27.705] - - - [11905, 128, 1, 256] - - [118, 35.504] - - - [1409, 128, 1, 128] - - [124, 7.027] - - - [12673, 128, 1, 128] - - [120, 26.276] - - - [27521, 128, 1, 256] - - [138, 34.565] - - - [1409, 128, 1, 256] - - [113, 10.4] - - - [25217, 128, 1, 128] - - [190, 24.917] - - - [7297, 128, 1, 128] - - [118, 20.023] - - - [14081, 128, 1, 128] - - [134, 18.804] - - - [22913, 128, 1, 128] - - [120, 23.227] - - - [10753, 128, 1, 256] - - [125, 32.662] - - - [7937, 128, 1, 128] - - [117, 21.46] - - - [11393, 128, 1, 128] - - [126, 24.445] - - - [26369, 128, 1, 128] - - [123, 22.481] - - - [12161, 128, 1, 256] - - [118, 36.415] - - - [8449, 128, 1, 128] - - [110, 22.514] - - - [22145, 128, 1, 256] - - [188, 35.265] - - - [20225, 128, 1, 256] - - [188, 33.174] - - - [10241, 128, 1, 256] - - [127, 35.572] - - - [6913, 128, 1, 128] - - [113, 21.806] - - - [4993, 128, 1, 256] - - [128, 26.999] - - - [6401, 128, 1, 256] - - [129, 32.584] - - - [13057, 128, 1, 128] - - [130, 26.404] - - - [2945, 128, 1, 128] - - [131, 13.846] - - - [3713, 128, 1, 256] - - [100, 20.452] - - - [10753, 128, 1, 128] - - [132, 23.073] - - - [14849, 128, 1, 256] - - [122, 30.049] - - - [3841, 128, 1, 128] - - [100, 15.578] - - - [28289, 128, 1, 256] - - [120, 35.842] - - - [12929, 128, 1, 128] - - [133, 26.806] - - - [14081, 128, 1, 256] - - [112, 29.167] - - - [14977, 128, 1, 256] - - [129, 30.558] - - - [12545, 128, 1, 128] - - [130, 25.865] - - - [16129, 128, 1, 256] - - [108, 32.403] - - - [11777, 128, 1, 256] - - [125, 34.359] - - - [11777, 128, 1, 128] - - [134, 24.695] - - - [17537, 128, 1, 256] - - [108, 31.778] - - - [5377, 128, 1, 128] - - [135, 17.639] - - - [8065, 128, 1, 256] - - [120, 31.167] - - - [6145, 128, 1, 128] - - [129, 19.981] - - - [20993, 128, 1, 128] - - [108, 22.28] - - - [15617, 128, 1, 128] - - [130, 20.563] - - - [5633, 128, 1, 256] - - [136, 28.675] - - - [4865, 128, 1, 128] - - [114, 19.203] - - - [385, 128, 1, 256] - - [109, 2.856] - - - [3841, 128, 1, 256] - - [128, 21.236] - - - [8833, 128, 1, 256] - - [133, 33.694] - - - [4225, 128, 1, 128] - - [109, 17.135] - - - [11009, 128, 1, 256] - - [133, 33.235] - - - [385, 128, 1, 128] - - [109, 1.973] - - - [9473, 128, 1, 256] - - [133, 35.137] - - - [5761, 128, 1, 128] - - [113, 19.155] - - - [11905, 128, 1, 128] - - [130, 25.324] - - - [4097, 128, 1, 256] - - [137, 22.568] - - - [25217, 128, 1, 256] - - [186, 38.617] - - - [9089, 128, 1, 256] - - [118, 25.783] - - - [10369, 128, 1, 256] - - [122, 26.562] - - - [14209, 128, 1, 256] - - [129, 30.171] - - - [6401, 128, 1, 128] - - [139, 20.814] - - - [27137, 128, 1, 256] - - [121, 34.409] - - - [16385, 128, 1, 256] - - [130, 29.276] - - - [24833, 128, 1, 128] - - [129, 23.667] - - - [18689, 128, 1, 128] - - [110, 21.271] - - - [7553, 128, 1, 256] - - [136, 29.265] - - - [8321, 128, 1, 128] - - [132, 22.334] - - - [15361, 128, 1, 128] - - [130, 19.971] - - - [1153, 128, 1, 128] - - [113, 5.75] - - - [1025, 128, 1, 128] - - [109, 5.182] - - - [19841, 128, 1, 256] - - [181, 32.423] - - - [15233, 128, 1, 128] - - [134, 19.827] - - - [21761, 128, 1, 256] - - [185, 34.842] - - - [17153, 128, 1, 256] - - [123, 31.474] - - - [15617, 128, 1, 256] - - [130, 31.685] - - - [4865, 128, 1, 256] - - [113, 26.404] - - - [14209, 128, 1, 128] - - [130, 18.967] - - - [19457, 128, 1, 256] - - [134, 33.461] - - - [9857, 128, 1, 256] - - [140, 36.107] - - - [11521, 128, 1, 128] - - [129, 24.366] - - - [8449, 128, 1, 256] - - [121, 32.146] - - - [4097, 128, 1, 128] - - [100, 16.708] - - - [28673, 128, 1, 256] - - [173, 34.668] - - - [12161, 128, 1, 128] - - [132, 25.868] - - - [1921, 128, 1, 256] - - [109, 13.314] - - - [9985, 128, 1, 256] - - [121, 36.306] - - - [7937, 128, 1, 256] - - [132, 30.434] - - - [9857, 128, 1, 128] - - [112, 25.528] - - - [13825, 128, 1, 256] - - [129, 29.032] - - - [9089, 128, 1, 128] - - [112, 24.22] - - - [6785, 128, 1, 256] - - [122, 33.387] - - - [5249, 128, 1, 256] - - [132, 27.094] - - - [7681, 128, 1, 256] - - [120, 29.376] - - - [3329, 128, 1, 128] - - [109, 15.075] - - - [14465, 128, 1, 128] - - [130, 19.567] - - - [11137, 128, 1, 256] - - [132, 34.039] - - - [1153, 128, 1, 256] - - [113, 8.596] - - - [16001, 128, 1, 128] - - [118, 20.392] - - - [26753, 128, 1, 128] - - [123, 22.997] - - - [13697, 128, 1, 128] - - [130, 27.251] - - - [3073, 128, 1, 128] - - [131, 14.356] - - - [22529, 128, 1, 256] - - [185, 35.8] - - - [18689, 128, 1, 256] - - [121, 32.31] - - - [257, 128, 1, 128] - - [124, 1.282] - - - [15233, 128, 1, 256] - - [118, 31.855] - - - [27521, 128, 1, 128] - - [122, 23.676] - - - [16385, 128, 1, 128] - - [134, 18.806] - - - [4481, 128, 1, 256] - - [141, 24.683] - - - [6017, 128, 1, 128] - - [142, 19.916] - - - [7297, 128, 1, 256] - - [108, 28.273] - - - [7553, 128, 1, 128] - - [143, 20.572] - - - [21761, 128, 1, 128] - - [129, 22.662] - - - [11393, 128, 1, 256] - - [118, 34.324] - - - [11521, 128, 1, 256] - - [144, 34.428] - - - [12929, 128, 1, 256] - - [140, 37.424] - - - [20225, 128, 1, 128] - - [129, 21.928] - - - [13313, 128, 1, 128] - - [145, 25.928] - - - [2561, 128, 1, 128] - - [146, 12.04] - - - [1537, 128, 1, 128] - - [114, 7.718] - - - [24449, 128, 1, 256] - - [193, 37.636] - - - [12289, 128, 1, 256] - - [125, 35.432] - - - [4225, 128, 1, 256] - - [100, 23.101] - - - [26369, 128, 1, 256] - - [191, 40.172] - - - [17921, 128, 1, 256] - - [123, 33.102] - - - [2945, 128, 1, 256] - - [100, 20.127] - - - [24065, 128, 1, 128] - - [108, 22.965] - - - [6529, 128, 1, 128] - - [114, 21.514] - - - [6145, 128, 1, 256] - - [132, 30.96] - - - [25985, 128, 1, 256] - - [188, 39.546] - - - [8705, 128, 1, 256] - - [121, 32.533] - - - [384, 128, 1, 256] - - [147, 3.132] - - - [25601, 128, 1, 256] - - [192, 38.882] - - - [28673, 128, 1, 128] - - [122, 23.306] - - - [20609, 128, 1, 128] - - [129, 21.892] - - - [19457, 128, 1, 128] - - [122, 21.919] - - - [16769, 128, 1, 256] - - [115, 31.99] - - - [12673, 128, 1, 256] - - [93, 36.9] - - - [8321, 128, 1, 256] - - [144, 31.823] - - - [5249, 128, 1, 128] - - [113, 17.854] - - - [16129, 128, 1, 128] - - [134, 21.19] - - - [13441, 128, 1, 256] - - [118, 28.398] - - - [5377, 128, 1, 256] - - [144, 27.466] - - - [21377, 128, 1, 128] - - [121, 22.544] - - - [14465, 128, 1, 256] - - [130, 30.09] - - - [11137, 128, 1, 128] - - [145, 24.107] - - - [7681, 128, 1, 128] - - [112, 20.692] - - - [7169, 128, 1, 128] - - [108, 19.671] - - - [22145, 128, 1, 128] - - [122, 22.962] - - - [11009, 128, 1, 128] - - [148, 23.76] - - - [20993, 128, 1, 256] - - [189, 34.018] - - - [13313, 128, 1, 256] - - [134, 27.763] - - - [25601, 128, 1, 128] - - [190, 24.669] - - - [4609, 128, 1, 128] - - [113, 18.29] - - - [5761, 128, 1, 256] - - [129, 29.529] - - - [17921, 128, 1, 128] - - [121, 21.452] - - - [2689, 128, 1, 128] - - [124, 12.405] - - - [8705, 128, 1, 128] - - [112, 22.785] - - - [10241, 128, 1, 128] - - [134, 25.624] - - - [14977, 128, 1, 128] - - [130, 19.498] - - - [18305, 128, 1, 128] - - [123, 21.751] - - - [3457, 128, 1, 128] - - [109, 15.372] - - - [24065, 128, 1, 256] - - [185, 37.356] - - - [12289, 128, 1, 128] - - [130, 25.127] - - - [14593, 128, 1, 128] - - [110, 19.393] - - - [2177, 128, 1, 256] - - [149, 15.018] - - - [4481, 128, 1, 128] - - [114, 18.173] - - - [8065, 128, 1, 128] - - [108, 21.886] - - - [3457, 128, 1, 256] - - [109, 22.482] - - - [6529, 128, 1, 256] - - [118, 32.895] - - - [26753, 128, 1, 256] - - [129, 33.575] - - - [17537, 128, 1, 128] - - [189, 19.148] - - - [22529, 128, 1, 128] - - [120, 23.032] - - - [10625, 128, 1, 256] - - [144, 32.474] - - - [14849, 128, 1, 128] - - [130, 19.83] - - - [9217, 128, 1, 256] - - [134, 33.105] - - - [19841, 128, 1, 128] - - [118, 22.305] - - - [15745, 128, 1, 128] - - [134, 20.718] - - - [13441, 128, 1, 128] - - [133, 27.181] - - - [3713, 128, 1, 128] - - [100, 15.058] - - - [27137, 128, 1, 128] - - [118, 23.448] - - - [16001, 128, 1, 256] - - [181, 32.047] - - - [10369, 128, 1, 128] - - [112, 26.301] - - - [1921, 128, 1, 128] - - [100, 8.975] - - - [9473, 128, 1, 128] - - [121, 24.62] - - - [27905, 128, 1, 256] - - [115, 35.115] - - - [30976, 1024, 1, 128] - - [348, 48.852] - - - [42240, 26369, 1, 128] - - [20, 45.705] - - - [33024, 17025, 1, 128] - - [25, 46.355] - - - [39168, 512, 1, 128] - - [357, 46.992] - - - [30848, 1024, 1, 128] - - [299, 50.413] - - - [41728, 8192, 1, 128] - - [62, 45.865] - - - [39552, 23553, 1, 128] - - [20, 45.133] - - - [35072, 512, 1, 128] - - [297, 46.21] - - - [29952, 14081, 1, 128] - - [23, 45.934] - - - [33280, 2048, 1, 128] - - [289, 55.226] - - - [40320, 128, 1, 128] - - [392, 33.609] - - - [35456, 1024, 1, 128] - - [390, 51.842] - - - [36096, 1024, 1, 128] - - [421, 50.518] - - - [36992, 20993, 1, 128] - - [20, 45.407] - - - [36096, 20097, 1, 128] - - [62, 44.703] - - - [31488, 15489, 1, 128] - - [20, 45.828] - - - [39552, 23681, 1, 128] - - [20, 45.361] - - - [36864, 128, 1, 128] - - [290, 34.089] - - - [40320, 4096, 1, 128] - - [25, 45.636] - - - [35200, 2048, 1, 128] - - [299, 54.691] - - - [29824, 2048, 1, 128] - - [421, 53.934] - - - [34688, 2048, 1, 128] - - [299, 54.998] - - - [42752, 26753, 1, 128] - - [20, 45.655] - - - [34304, 4096, 1, 128] - - [65, 46.126] - - - [36480, 20481, 1, 128] - - [48, 45.313] - - - [33408, 128, 1, 128] - - [256, 32.437] - - - [38784, 4096, 1, 128] - - [32, 45.248] - - - [43264, 27393, 1, 128] - - [36, 45.698] - - - [34560, 128, 1, 128] - - [293, 33.493] - - - [30336, 4096, 1, 128] - - [65, 45.255] - - - [29056, 2048, 1, 128] - - [289, 54.914] - - - [34816, 512, 1, 128] - - [268, 45.604] - - - [38272, 2048, 1, 128] - - [87, 36.638] - - - [39808, 23937, 1, 128] - - [62, 43.271] - - - [30848, 512, 1, 128] - - [291, 45.104] - - - [40448, 512, 1, 128] - - [293, 46.837] - - - [40448, 24577, 1, 128] - - [62, 45.618] - - - [44544, 28545, 1, 128] - - [48, 45.737] - - - [30208, 14209, 1, 128] - - [28, 46.282] - - - [34688, 18689, 1, 128] - - [27, 45.523] - - - [31360, 512, 1, 128] - - [419, 45.031] - - - [38912, 512, 1, 128] - - [298, 48.489] - - - [39680, 1024, 1, 128] - - [289, 51.163] - - - [34048, 1024, 1, 128] - - [289, 51.866] - - - [39552, 4096, 1, 128] - - [36, 45.488] - - - [40320, 24321, 1, 128] - - [20, 45.408] - - - [40832, 24833, 1, 128] - - [36, 45.035] - - - [36736, 1024, 1, 128] - - [299, 50.994] - - - [44672, 1024, 1, 128] - - [299, 52.375] - - - [32000, 128, 1, 128] - - [418, 33.032] - - - [40704, 4096, 1, 128] - - [25, 45.772] - - - [38144, 1024, 1, 128] - - [421, 51.742] - - - [30720, 14849, 1, 128] - - [36, 46.757] - - - [38144, 8192, 1, 128] - - [35, 46.834] - - - [30208, 1024, 1, 128] - - [390, 50.167] - - - [43136, 1024, 1, 128] - - [299, 52.803] - - - [38528, 1024, 1, 128] - - [348, 51.773] - - - [43264, 2048, 1, 128] - - [56, 44.133] - - - [38400, 22529, 1, 128] - - [62, 45.897] - - - [37120, 128, 1, 128] - - [293, 33.946] - - - [32256, 128, 1, 128] - - [293, 32.926] - - - [29952, 13953, 1, 128] - - [35, 45.983] - - - [34560, 8192, 1, 128] - - [25, 46.641] - - - [37504, 21505, 1, 128] - - [49, 44.546] - - - [33536, 128, 1, 128] - - [346, 32.372] - - - [41856, 2048, 1, 128] - - [58, 44.478] - - - [32896, 4096, 1, 128] - - [72, 43.557] - - - [41856, 8192, 1, 128] - - [35, 46.566] - - - [29440, 4096, 1, 128] - - [36, 45.779] - - - [33664, 8192, 1, 128] - - [45, 46.232] - - - [36992, 512, 1, 128] - - [272, 44.115] - - - [33280, 512, 1, 128] - - [298, 46.696] - - - [41728, 128, 1, 128] - - [418, 34.039] - - - [31744, 128, 1, 128] - - [293, 32.566] - - - [31360, 1024, 1, 128] - - [289, 50.532] - - - [29952, 8192, 1, 128] - - [49, 46.599] - - - [38016, 2048, 1, 128] - - [421, 53.421] - - - [34176, 8192, 1, 128] - - [45, 46.412] - - - [30464, 512, 1, 128] - - [348, 45.767] - - - [41984, 2048, 1, 128] - - [59, 44.722] - - - [40448, 4096, 1, 128] - - [49, 46.116] - - - [33920, 4096, 1, 128] - - [25, 45.581] - - - [41088, 8192, 1, 128] - - [59, 45.778] - - - [39808, 8192, 1, 128] - - [20, 44.228] - - - [40832, 4096, 1, 128] - - [23, 45.606] - - - [30592, 2048, 1, 128] - - [299, 54.008] - - - [36352, 1024, 1, 128] - - [299, 51.621] - - - [30336, 2048, 1, 128] - - [419, 54.023] - - - [30976, 512, 1, 128] - - [268, 46.745] - - - [42368, 1024, 1, 128] - - [299, 51.662] - - - [29056, 1024, 1, 128] - - [299, 50.559] - - - [38784, 22913, 1, 128] - - [20, 45.544] - - - [28928, 512, 1, 128] - - [322, 44.185] - - - [40576, 512, 1, 128] - - [297, 46.772] - - - [34816, 4096, 1, 128] - - [25, 46.696] - - - [41600, 2048, 1, 128] - - [20, 43.878] - - - [29696, 8192, 1, 128] - - [23, 47.304] - - - [41856, 4096, 1, 128] - - [25, 45.762] - - - [35584, 2048, 1, 128] - - [289, 54.461] - - - [30848, 14849, 1, 128] - - [35, 45.765] - - - [33280, 17281, 1, 128] - - [45, 46.514] - - - [43776, 2048, 1, 128] - - [58, 43.542] - - - [42112, 8192, 1, 128] - - [23, 46.462] - - - [37376, 128, 1, 128] - - [270, 33.859] - - - [41600, 4096, 1, 128] - - [36, 45.603] - - - [36224, 20353, 1, 128] - - [36, 45.672] - - - [29952, 1024, 1, 128] - - [390, 49.858] - - - [34176, 1024, 1, 128] - - [289, 52.176] - - - [31744, 512, 1, 128] - - [312, 45.25] - - - [42624, 8192, 1, 128] - - [25, 41.712] - - - [41216, 128, 1, 128] - - [293, 33.084] - - - [42624, 26753, 1, 128] - - [75, 40.481] - - - [32512, 2048, 1, 128] - - [289, 54.428] - - - [40064, 4096, 1, 128] - - [20, 45.152] - - - [32640, 4096, 1, 128] - - [25, 45.254] - - - [42112, 26241, 1, 128] - - [20, 45.551] - - - [32256, 512, 1, 128] - - [298, 45.322] - - - [40960, 1024, 1, 128] - - [310, 46.099] - - - [35968, 128, 1, 128] - - [418, 33.384] - - - [32384, 8192, 1, 128] - - [27, 46.529] - - - [42880, 512, 1, 128] - - [291, 46.649] - - - [33024, 8192, 1, 128] - - [27, 47.007] - - - [43904, 1024, 1, 128] - - [421, 51.544] - - - [33664, 17665, 1, 128] - - [36, 45.702] - - - [41856, 512, 1, 128] - - [291, 47.189] - - - [40704, 128, 1, 128] - - [293, 34.407] - - - [33408, 17537, 1, 128] - - [36, 45.966] - - - [37120, 512, 1, 128] - - [290, 45.963] - - - [41216, 25345, 1, 128] - - [62, 45.806] - - - [39680, 8192, 1, 128] - - [35, 46.567] - - - [40192, 24193, 1, 128] - - [36, 45.525] - - - [33024, 17153, 1, 128] - - [20, 46.343] - - - [38272, 1024, 1, 128] - - [421, 51.257] - - - [35328, 1024, 1, 128] - - [289, 52.833] - - - [31104, 8192, 1, 128] - - [25, 46.434] - - - [40320, 8192, 1, 128] - - [23, 46.483] - - - [29312, 2048, 1, 128] - - [299, 53.627] - - - [36608, 20737, 1, 128] - - [20, 45.83] - - - [42240, 4096, 1, 128] - - [23, 45.827] - - - [43520, 2048, 1, 128] - - [31, 44.598] - - - [29056, 512, 1, 128] - - [293, 45.878] - - - [35328, 19329, 1, 128] - - [48, 46.24] - - - [30464, 128, 1, 128] - - [293, 31.677] - - - [29696, 13697, 1, 128] - - [25, 46.562] - - - [43904, 28033, 1, 128] - - [25, 45.375] - - - [35584, 19713, 1, 128] - - [20, 45.7] - - - [41088, 4096, 1, 128] - - [34, 45.105] - - - [42368, 2048, 1, 128] - - [56, 44.078] - - - [36736, 128, 1, 128] - - [392, 34.257] - - - [30336, 8192, 1, 128] - - [33, 46.227] - - - [43008, 128, 1, 128] - - [291, 35.397] - - - [37120, 1024, 1, 128] - - [289, 52.332] - - - [31104, 2048, 1, 128] - - [299, 54.393] - - - [33152, 4096, 1, 128] - - [32, 45.818] - - - [43392, 27521, 1, 128] - - [36, 45.358] - - - [37248, 21249, 1, 128] - - [36, 45.316] - - - [33920, 17921, 1, 128] - - [36, 45.594] - - - [39680, 4096, 1, 128] - - [23, 45.776] - - - [43264, 512, 1, 128] - - [268, 47.301] - - - [35712, 8192, 1, 128] - - [23, 46.366] - - - [31616, 2048, 1, 128] - - [300, 54.007] - - - [35328, 512, 1, 128] - - [357, 46.404] - - - [43136, 27265, 1, 128] - - [20, 45.393] - - - [30208, 128, 1, 128] - - [392, 31.226] - - - [40320, 24449, 1, 128] - - [36, 45.352] - - - [44288, 2048, 1, 128] - - [34, 44.467] - - - [35072, 1024, 1, 128] - - [289, 51.811] - - - [30464, 14465, 1, 128] - - [72, 43.511] - - - [44160, 8192, 1, 128] - - [23, 46.21] - - - [33792, 17793, 1, 128] - - [27, 46.448] - - - [37632, 1024, 1, 128] - - [390, 51.64] - - - [35968, 2048, 1, 128] - - [300, 53.524] - - - [38400, 8192, 1, 128] - - [28, 46.925] - - - [32512, 4096, 1, 128] - - [25, 45.842] - - - [32512, 16641, 1, 128] - - [49, 46.246] - - - [39424, 128, 1, 128] - - [293, 34.455] - - - [30976, 8192, 1, 128] - - [74, 45.168] - - - [35968, 20097, 1, 128] - - [23, 45.717] - - - [38656, 512, 1, 128] - - [357, 47.742] - - - [34944, 18945, 1, 128] - - [25, 45.624] - - - [33664, 17793, 1, 128] - - [20, 45.754] - - - [38656, 22657, 1, 128] - - [48, 45.798] - - - [34944, 1024, 1, 128] - - [390, 51.457] - - - [31872, 16001, 1, 128] - - [36, 45.704] - - - [43392, 8192, 1, 128] - - [35, 46.144] - - - [38016, 512, 1, 128] - - [291, 46.458] - - - [29440, 8192, 1, 128] - - [20, 46.686] - - - [35200, 1024, 1, 128] - - [348, 51.397] - - - [34304, 18433, 1, 128] - - [48, 46.162] - - - [44672, 28801, 1, 128] - - [36, 45.545] - - - [29184, 4096, 1, 128] - - [55, 45.73] - - - [33408, 8192, 1, 128] - - [25, 46.532] - - - [39040, 128, 1, 128] - - [291, 34.399] - - - [39680, 23681, 1, 128] - - [36, 45.518] - - - [38144, 4096, 1, 128] - - [23, 45.883] - - - [42368, 26497, 1, 128] - - [49, 45.346] - - - [42368, 4096, 1, 128] - - [65, 45.365] - - - [31872, 128, 1, 128] - - [418, 32.462] - - - [41984, 512, 1, 128] - - [268, 48.114] - - - [39296, 2048, 1, 128] - - [36, 43.934] - - - [33920, 2048, 1, 128] - - [299, 54.496] - - - [36736, 20865, 1, 128] - - [20, 45.681] - - - [34432, 8192, 1, 128] - - [28, 44.91] - - - [30848, 14977, 1, 128] - - [35, 45.811] - - - [31744, 15873, 1, 128] - - [27, 46.488] - - - [42880, 27009, 1, 128] - - [62, 44.895] - - - [42240, 26241, 1, 128] - - [36, 45.662] - - - [38400, 4096, 1, 128] - - [35, 45.966] - - - [42624, 26625, 1, 128] - - [25, 40.471] - - - [35072, 4096, 1, 128] - - [35, 45.875] - - - [40576, 4096, 1, 128] - - [23, 45.622] - - - [39296, 8192, 1, 128] - - [27, 46.525] - - - [42624, 512, 1, 128] - - [290, 46.903] - - - [32768, 8192, 1, 128] - - [35, 35.565] - - - [36864, 1024, 1, 128] - - [348, 50.307] - - - [43392, 128, 1, 128] - - [293, 35.086] - - - [41344, 2048, 1, 128] - - [38, 44.187] - - - [35584, 4096, 1, 128] - - [25, 45.526] - - - [40064, 2048, 1, 128] - - [20, 42.994] - - - [40576, 24705, 1, 128] - - [36, 45.497] - - - [39808, 1024, 1, 128] - - [300, 50.871] - - - [36992, 1024, 1, 128] - - [289, 50.145] - - - [42496, 1024, 1, 128] - - [289, 52.509] - - - [43904, 128, 1, 128] - - [290, 35.185] - - - [31232, 512, 1, 128] - - [392, 45.428] - - - [42112, 128, 1, 128] - - [293, 34.193] - - - [37376, 2048, 1, 128] - - [299, 55.376] - - - [38016, 128, 1, 128] - - [266, 34.191] - - - [42368, 8192, 1, 128] - - [28, 46.148] - - - [43392, 512, 1, 128] - - [290, 47.198] - - - [41984, 1024, 1, 128] - - [299, 53.4] - - - [42240, 2048, 1, 128] - - [51, 44.379] - - - [29952, 128, 1, 128] - - [293, 31.675] - - - [36608, 8192, 1, 128] - - [36, 46.612] - - - [32512, 16513, 1, 128] - - [62, 46.298] - - - [29568, 512, 1, 128] - - [291, 43.659] - - - [34304, 1024, 1, 128] - - [289, 51.27] - - - [41984, 4096, 1, 128] - - [23, 46.4] - - - [30464, 4096, 1, 128] - - [74, 43.433] - - - [41216, 2048, 1, 128] - - [34, 44.506] - - - [36480, 20609, 1, 128] - - [36, 45.594] - - - [44800, 4096, 1, 128] - - [27, 45.493] - - - [36864, 512, 1, 128] - - [296, 45.815] - - - [39680, 2048, 1, 128] - - [27, 43.919] - - - [43648, 4096, 1, 128] - - [23, 45.327] - - - [33664, 128, 1, 128] - - [343, 32.371] - - - [41600, 512, 1, 128] - - [293, 46.739] - - - [43776, 1024, 1, 128] - - [421, 50.385] - - - [37632, 512, 1, 128] - - [290, 48.001] - - - [44160, 128, 1, 128] - - [266, 35.066] - - - [37248, 8192, 1, 128] - - [35, 46.313] - - - [34816, 18817, 1, 128] - - [27, 46.801] - - - [38528, 22529, 1, 128] - - [36, 45.314] - - - [40192, 24321, 1, 128] - - [36, 45.492] - - - [40832, 128, 1, 128] - - [418, 33.814] - - - [29312, 8192, 1, 128] - - [25, 46.211] - - - [43776, 27777, 1, 128] - - [59, 43.983] - - - [37632, 21633, 1, 128] - - [36, 45.804] - - - [33792, 4096, 1, 128] - - [23, 46.375] - - - [35968, 1024, 1, 128] - - [298, 50.383] - - - [37888, 512, 1, 128] - - [293, 46.223] - - - [35968, 512, 1, 128] - - [370, 46.624] - - - [30592, 1024, 1, 128] - - [419, 50.081] - - - [38400, 512, 1, 128] - - [293, 46.444] - - - [43264, 1024, 1, 128] - - [299, 51.947] - - - [38528, 4096, 1, 128] - - [47, 45.377] - - - [28928, 1024, 1, 128] - - [348, 50.714] - - - [33152, 1024, 1, 128] - - [421, 50.419] - - - [41344, 1024, 1, 128] - - [299, 51.667] - - - [30848, 8192, 1, 128] - - [35, 46.564] - - - [41344, 4096, 1, 128] - - [27, 45.428] - - - [38912, 2048, 1, 128] - - [27, 44.728] - - - [38272, 128, 1, 128] - - [290, 34.528] - - - [31488, 4096, 1, 128] - - [23, 45.665] - - - [44416, 4096, 1, 128] - - [23, 45.725] - - - [39552, 2048, 1, 128] - - [38, 44.397] - - - [37760, 1024, 1, 128] - - [421, 51.044] - - - [34304, 18305, 1, 128] - - [48, 46.325] - - - [44544, 28673, 1, 128] - - [62, 45.502] - - - [44416, 8192, 1, 128] - - [23, 46.505] - - - [38144, 512, 1, 128] - - [422, 46.43] - - - [30208, 14337, 1, 128] - - [22, 46.144] - - - [38144, 2048, 1, 128] - - [421, 54.208] - - - [40448, 128, 1, 128] - - [293, 34.189] - - - [42240, 8192, 1, 128] - - [23, 46.544] - - - [39424, 2048, 1, 128] - - [34, 44.584] - - - [41088, 512, 1, 128] - - [419, 44.367] - - - [36224, 2048, 1, 128] - - [299, 54.185] - - - [31744, 4096, 1, 128] - - [25, 46.397] - - - [44160, 512, 1, 128] - - [290, 47.418] - - - [32000, 1024, 1, 128] - - [289, 50.509] - - - [42752, 1024, 1, 128] - - [390, 51.965] - - - [42496, 2048, 1, 128] - - [37, 44.786] - - - [32640, 2048, 1, 128] - - [419, 55.131] - - - [42752, 26881, 1, 128] - - [36, 45.625] - - - [32256, 8192, 1, 128] - - [49, 47.12] - - - [44800, 512, 1, 128] - - [293, 48.893] - - - [34816, 128, 1, 128] - - [291, 33.121] - - - [38272, 8192, 1, 128] - - [59, 42.399] - - - [44800, 28929, 1, 128] - - [20, 45.322] - - - [37120, 8192, 1, 128] - - [20, 46.547] - - - [43776, 512, 1, 128] - - [419, 44.901] - - - [43008, 1024, 1, 128] - - [348, 51.898] - - - [34432, 18561, 1, 128] - - [72, 44.536] - - - [36736, 4096, 1, 128] - - [25, 45.56] - - - [36224, 512, 1, 128] - - [357, 46.146] - - - [32768, 512, 1, 128] - - [263, 35.536] - - - [30592, 128, 1, 128] - - [270, 31.095] - - - [43008, 27137, 1, 128] - - [20, 46.575] - - - [34048, 18177, 1, 128] - - [62, 45.732] - - - [43136, 2048, 1, 128] - - [35, 43.944] - - - [29184, 13313, 1, 128] - - [20, 45.766] - - - [40064, 24193, 1, 128] - - [62, 44.519] - - - [40960, 128, 1, 128] - - [293, 33.561] - - - [29184, 2048, 1, 128] - - [390, 54.443] - - - [37248, 128, 1, 128] - - [270, 33.721] - - - [35328, 128, 1, 128] - - [270, 33.003] - - - [43264, 128, 1, 128] - - [297, 34.539] - - - [29952, 4096, 1, 128] - - [55, 45.58] - - - [36736, 20737, 1, 128] - - [36, 45.629] - - - [34176, 4096, 1, 128] - - [29, 45.382] - - - [32768, 1024, 1, 128] - - [302, 39.695] - - - [44160, 4096, 1, 128] - - [23, 45.544] - - - [31104, 1024, 1, 128] - - [390, 50.804] - - - [33792, 512, 1, 128] - - [390, 46.076] - - - [41216, 25217, 1, 128] - - [62, 45.833] - - - [31872, 1024, 1, 128] - - [421, 49.272] - - - [38528, 8192, 1, 128] - - [22, 46.201] - - - [44672, 4096, 1, 128] - - [27, 45.698] - - - [32512, 1024, 1, 128] - - [289, 49.323] - - - [39168, 8192, 1, 128] - - [62, 46.6] - - - [31360, 15361, 1, 128] - - [36, 45.526] - - - [38016, 22145, 1, 128] - - [36, 45.578] - - - [35712, 128, 1, 128] - - [293, 33.752] - - - [30208, 4096, 1, 128] - - [57, 45.872] - - - [33920, 128, 1, 128] - - [392, 32.649] - - - [30336, 128, 1, 128] - - [293, 32.543] - - - [42368, 128, 1, 128] - - [293, 34.986] - - - [38912, 4096, 1, 128] - - [35, 46.69] - - - [34176, 512, 1, 128] - - [297, 46.719] - - - [42752, 8192, 1, 128] - - [25, 46.612] - - - [31488, 1024, 1, 128] - - [299, 50.354] - - - [36608, 1024, 1, 128] - - [299, 52.984] - - - [41856, 128, 1, 128] - - [392, 34.436] - - - [29312, 13441, 1, 128] - - [23, 45.652] - - - [43520, 128, 1, 128] - - [293, 33.71] - - - [31616, 8192, 1, 128] - - [23, 45.923] - - - [40448, 2048, 1, 128] - - [31, 44.652] - - - [35328, 2048, 1, 128] - - [289, 54.708] - - - [36864, 20865, 1, 128] - - [36, 46.566] - - - [32000, 2048, 1, 128] - - [289, 54.942] - - - [34176, 18177, 1, 128] - - [62, 45.62] - - - [37504, 128, 1, 128] - - [266, 34.044] - - - [33792, 1024, 1, 128] - - [421, 50.926] - - - [31872, 8192, 1, 128] - - [23, 46.308] - - - [40704, 512, 1, 128] - - [290, 47.244] - - - [37632, 128, 1, 128] - - [291, 34.168] - - - [32640, 1024, 1, 128] - - [421, 50.334] - - - [44544, 8192, 1, 128] - - [20, 46.449] - - - [39424, 8192, 1, 128] - - [23, 46.96] - - - [39296, 512, 1, 128] - - [293, 48.638] - - - [35840, 128, 1, 128] - - [392, 33.648] - - - [39168, 1024, 1, 128] - - [390, 51.597] - - - [35712, 19841, 1, 128] - - [36, 45.634] - - - [29568, 13569, 1, 128] - - [20, 45.271] - - - [34944, 4096, 1, 128] - - [25, 45.826] - - - [32768, 2048, 1, 128] - - [263, 40.929] - - - [39296, 128, 1, 128] - - [418, 35.124] - - - [29568, 4096, 1, 128] - - [59, 44.664] - - - [39040, 1024, 1, 128] - - [299, 52.286] - - - [37376, 1024, 1, 128] - - [289, 51.383] - - - [33536, 2048, 1, 128] - - [289, 54.565] - - - [31488, 8192, 1, 128] - - [25, 46.506] - - - [37888, 1024, 1, 128] - - [289, 51.478] - - - [41472, 4096, 1, 128] - - [28, 45.992] - - - [30592, 512, 1, 128] - - [293, 46.731] - - - [34560, 18561, 1, 128] - - [27, 45.867] - - - [29184, 512, 1, 128] - - [272, 45.537] - - - [32256, 16257, 1, 128] - - [22, 46.612] - - - [43392, 27393, 1, 128] - - [36, 45.405] - - - [29312, 4096, 1, 128] - - [33, 45.362] - - - [43648, 2048, 1, 128] - - [27, 43.517] - - - [44288, 1024, 1, 128] - - [289, 51.698] - - - [35456, 128, 1, 128] - - [392, 33.171] - - - [44160, 28289, 1, 128] - - [35, 45.17] - - - [40320, 1024, 1, 128] - - [299, 51.96] - - - [37888, 22017, 1, 128] - - [36, 46.388] - - - [29696, 512, 1, 128] - - [268, 45.349] - - - [35840, 2048, 1, 128] - - [300, 53.965] - - - [37504, 2048, 1, 128] - - [390, 54.299] - - - [41728, 4096, 1, 128] - - [65, 45.048] - - - [42752, 4096, 1, 128] - - [35, 45.734] - - - [29824, 4096, 1, 128] - - [59, 44.859] - - - [44800, 1024, 1, 128] - - [390, 52.988] - - - [30592, 4096, 1, 128] - - [23, 45.582] - - - [43904, 4096, 1, 128] - - [25, 45.454] - - - [39552, 8192, 1, 128] - - [36, 46.1] - - - [37632, 2048, 1, 128] - - [421, 54.308] - - - [29312, 128, 1, 128] - - [418, 31.45] - - - [30080, 512, 1, 128] - - [419, 44.767] - - - [33664, 2048, 1, 128] - - [348, 53.431] - - - [43520, 27521, 1, 128] - - [62, 46.018] - - - [36224, 128, 1, 128] - - [291, 33.158] - - - [28928, 12929, 1, 128] - - [35, 45.873] - - - [29440, 1024, 1, 128] - - [299, 50.823] - - - [35840, 19969, 1, 128] - - [36, 46.357] - - - [42880, 4096, 1, 128] - - [55, 44.933] - - - [42496, 8192, 1, 128] - - [23, 46.936] - - - [39936, 24065, 1, 128] - - [36, 46.243] - - - [33408, 1024, 1, 128] - - [289, 50.716] - - - [32256, 2048, 1, 128] - - [289, 54.782] - - - [35712, 19713, 1, 128] - - [36, 45.682] - - - [40192, 4096, 1, 128] - - [35, 45.816] - - - [32000, 16129, 1, 128] - - [25, 46.012] - - - [44032, 512, 1, 128] - - [290, 49.143] - - - [35584, 128, 1, 128] - - [270, 32.849] - - - [35584, 8192, 1, 128] - - [27, 46.454] - - - [37888, 21889, 1, 128] - - [36, 46.38] - - - [37504, 1024, 1, 128] - - [390, 50.955] - - - [33664, 512, 1, 128] - - [291, 45.604] - - - [32384, 1024, 1, 128] - - [299, 50.688] - - - [38400, 1024, 1, 128] - - [390, 50.075] - - - [35200, 128, 1, 128] - - [270, 33.008] - - - [43648, 1024, 1, 128] - - [300, 51.469] - - - [36608, 128, 1, 128] - - [418, 33.255] - - - [32768, 128, 1, 128] - - [370, 30.876] - - - [28928, 4096, 1, 128] - - [27, 45.55] - - - [35200, 19329, 1, 128] - - [20, 45.615] - - - [41216, 8192, 1, 128] - - [23, 46.646] - - - [36864, 8192, 1, 128] - - [27, 47.323] - - - [40064, 128, 1, 128] - - [392, 34.016] - - - [42624, 1024, 1, 128] - - [421, 51.414] - - - [34688, 128, 1, 128] - - [293, 32.606] - - - [43648, 27777, 1, 128] - - [20, 44.988] - - - [37888, 8192, 1, 128] - - [35, 47.162] - - - [41472, 25601, 1, 128] - - [62, 45.782] - - - [38272, 512, 1, 128] - - [421, 44.89] - - - [35456, 4096, 1, 128] - - [47, 45.661] - - - [42496, 26625, 1, 128] - - [48, 45.874] - - - [43136, 4096, 1, 128] - - [35, 45.517] - - - [44800, 8192, 1, 128] - - [27, 46.255] - - - [36480, 8192, 1, 128] - - [27, 46.305] - - - [37504, 4096, 1, 128] - - [61, 44.983] - - - [39040, 8192, 1, 128] - - [25, 46.156] - - - [31104, 512, 1, 128] - - [291, 46.963] - - - [34176, 2048, 1, 128] - - [299, 55.235] - - - [31616, 512, 1, 128] - - [293, 46.888] - - - [35456, 2048, 1, 128] - - [299, 54.055] - - - [43136, 8192, 1, 128] - - [35, 46.264] - - - [33024, 128, 1, 128] - - [257, 30.579] - - - [38656, 4096, 1, 128] - - [49, 45.653] - - - [33408, 17409, 1, 128] - - [36, 45.715] - - - [39424, 1024, 1, 128] - - [390, 51.917] - - - [29312, 13313, 1, 128] - - [25, 45.515] - - - [35840, 4096, 1, 128] - - [27, 46.364] - - - [42496, 512, 1, 128] - - [293, 47.109] - - - [37632, 8192, 1, 128] - - [35, 46.625] - - - [41088, 2048, 1, 128] - - [56, 43.811] - - - [38528, 512, 1, 128] - - [291, 48.218] - - - [35072, 2048, 1, 128] - - [289, 55.181] - - - [31104, 4096, 1, 128] - - [36, 45.51] - - - [33280, 4096, 1, 128] - - [65, 46.23] - - - [43904, 8192, 1, 128] - - [35, 46.33] - - - [34816, 8192, 1, 128] - - [25, 47.493] - - - [38016, 1024, 1, 128] - - [300, 51.149] - - - [33152, 128, 1, 128] - - [259, 31.932] - - - [42496, 128, 1, 128] - - [40, 31.099] - - - [40832, 24961, 1, 128] - - [36, 45.111] - - - [41728, 1024, 1, 128] - - [357, 51.183] - - - [41472, 25473, 1, 128] - - [48, 46.035] - - - [34560, 2048, 1, 128] - - [289, 55.101] - - - [31616, 15617, 1, 128] - - [20, 45.318] - - - [33664, 4096, 1, 128] - - [36, 45.445] - - - [35328, 8192, 1, 128] - - [28, 47.021] - - - [39808, 4096, 1, 128] - - [35, 43.349] - - - [37248, 512, 1, 128] - - [293, 48.329] - - - [31360, 4096, 1, 128] - - [25, 45.616] - - - [41344, 8192, 1, 128] - - [35, 46.237] - - - [32000, 512, 1, 128] - - [348, 47.285] - - - [35968, 19969, 1, 128] - - [25, 45.615] - - - [30080, 14081, 1, 128] - - [45, 43.41] - - - [35840, 8192, 1, 128] - - [35, 47.132] - - - [44672, 2048, 1, 128] - - [61, 44.597] - - - [31872, 2048, 1, 128] - - [289, 53.564] - - - [42496, 4096, 1, 128] - - [20, 46.1] - - - [43776, 128, 1, 128] - - [293, 35.317] - - - [40704, 2048, 1, 128] - - [42, 44.366] - - - [34432, 128, 1, 128] - - [418, 32.611] - - - [44544, 2048, 1, 128] - - [67, 44.537] - - - [32384, 16385, 1, 128] - - [23, 45.9] - - - [43776, 27905, 1, 128] - - [59, 44.016] - - - [44032, 4096, 1, 128] - - [35, 46.373] - - - [36480, 512, 1, 128] - - [290, 46.132] - - - [44160, 1024, 1, 128] - - [299, 51.94] - - - [41216, 4096, 1, 128] - - [35, 45.814] - - - [44032, 2048, 1, 128] - - [56, 44.796] - - - [33152, 2048, 1, 128] - - [390, 53.929] - - - [41984, 25985, 1, 128] - - [36, 46.352] - - - [39552, 512, 1, 128] - - [268, 47.973] - - - [41344, 25473, 1, 128] - - [27, 45.432] - - - [40960, 4096, 1, 128] - - [20, 41.231] - - - [32640, 128, 1, 128] - - [297, 32.983] - - - [35968, 4096, 1, 128] - - [33, 45.611] - - - [33536, 4096, 1, 128] - - [27, 45.86] - - - [30976, 15105, 1, 128] - - [74, 44.642] - - - [35072, 8192, 1, 128] - - [27, 46.675] - - - [39424, 23425, 1, 128] - - [62, 45.929] - - - [43520, 1024, 1, 128] - - [299, 50.785] - - - [44288, 28417, 1, 128] - - [48, 45.373] - - - [30848, 128, 1, 128] - - [272, 31.704] - - - [35712, 512, 1, 128] - - [291, 47.422] - - - [44160, 2048, 1, 128] - - [51, 44.248] - - - [34048, 8192, 1, 128] - - [62, 46.442] - - - [40448, 24449, 1, 128] - - [48, 46.019] - - - [39168, 23297, 1, 128] - - [48, 45.842] - - - [32128, 1024, 1, 128] - - [390, 51.192] - - - [36864, 20993, 1, 128] - - [36, 46.547] - - - [40064, 1024, 1, 128] - - [419, 51.489] - - - [38784, 8192, 1, 128] - - [35, 46.238] - - - [37248, 2048, 1, 128] - - [421, 54.279] - - - [34560, 4096, 1, 128] - - [25, 45.799] - - - [39040, 23041, 1, 128] - - [20, 45.415] - - - [36480, 1024, 1, 128] - - [299, 50.917] - - - [39040, 2048, 1, 128] - - [51, 44.221] - - - [39808, 23809, 1, 128] - - [62, 43.154] - - - [36992, 4096, 1, 128] - - [35, 45.424] - - - [32768, 16897, 1, 128] - - [35, 34.596] - - - [30976, 2048, 1, 128] - - [421, 52.335] - - - [32640, 16769, 1, 128] - - [27, 45.395] - - - [29824, 13953, 1, 128] - - [62, 45.133] - - - [29184, 128, 1, 128] - - [418, 31.788] - - - [30720, 8192, 1, 128] - - [27, 47.523] - - - [30848, 2048, 1, 128] - - [299, 54.183] - - - [38016, 4096, 1, 128] - - [20, 45.516] - - - [35456, 8192, 1, 128] - - [25, 46.552] - - - [36992, 21121, 1, 128] - - [20, 45.411] - - - [36736, 2048, 1, 128] - - [390, 53.318] - - - [37888, 128, 1, 128] - - [392, 34.935] - - - [39808, 2048, 1, 128] - - [25, 42.312] - - - [41856, 25985, 1, 128] - - [20, 45.662] - - - [34688, 4096, 1, 128] - - [47, 45.486] - - - [38784, 1024, 1, 128] - - [299, 51.806] - - - [40960, 25089, 1, 128] - - [20, 40.357] - - - [32000, 4096, 1, 128] - - [23, 45.811] - - - [41600, 25601, 1, 128] - - [20, 45.352] - - - [37504, 512, 1, 128] - - [268, 46.568] - - - [32128, 16129, 1, 128] - - [27, 45.863] - - - [37248, 21377, 1, 128] - - [36, 45.342] - - - [35840, 512, 1, 128] - - [293, 47.524] - - - [36096, 128, 1, 128] - - [293, 34.081] - - - [32512, 8192, 1, 128] - - [23, 46.813] - - - [36736, 8192, 1, 128] - - [27, 46.469] - - - [42880, 1024, 1, 128] - - [299, 52.649] - - - [44288, 8192, 1, 128] - - [25, 46.275] - - - [36224, 1024, 1, 128] - - [390, 52.019] - - - [41344, 25345, 1, 128] - - [27, 45.426] - - - [32384, 512, 1, 128] - - [348, 46.797] - - - [38272, 4096, 1, 128] - - [59, 41.461] - - - [37120, 2048, 1, 128] - - [299, 55.19] - - - [33152, 8192, 1, 128] - - [27, 46.687] - - - [36096, 4096, 1, 128] - - [74, 44.072] - - - [34560, 18689, 1, 128] - - [36, 45.867] - - - [36864, 4096, 1, 128] - - [23, 46.428] - - - [34944, 512, 1, 128] - - [290, 47.785] - - - [37760, 128, 1, 128] - - [418, 33.71] - - - [31616, 128, 1, 128] - - [293, 32.07] - - - [36224, 4096, 1, 128] - - [35, 45.663] - - - [40576, 24577, 1, 128] - - [20, 45.04] - - - [34688, 1024, 1, 128] - - [289, 50.833] - - - [40192, 1024, 1, 128] - - [419, 51.005] - - - [44672, 512, 1, 128] - - [291, 48.022] - - - [33664, 1024, 1, 128] - - [300, 49.961] - - - [39424, 512, 1, 128] - - [291, 45.693] - - - [44416, 1024, 1, 128] - - [299, 51.53] - - - [33408, 2048, 1, 128] - - [390, 54.61] - - - [43648, 8192, 1, 128] - - [35, 46.058] - - - [43520, 27649, 1, 128] - - [62, 45.797] - - - [40448, 1024, 1, 128] - - [390, 52.952] - - - [33152, 17153, 1, 128] - - [35, 46.124] - - - [33024, 512, 1, 128] - - [278, 43.004] - - - [39680, 128, 1, 128] - - [291, 34.309] - - - [29696, 4096, 1, 128] - - [35, 46.382] - - - [42112, 2048, 1, 128] - - [25, 44.11] - - - [38016, 8192, 1, 128] - - [23, 46.349] - - - [30464, 8192, 1, 128] - - [76, 44.92] - - - [43648, 128, 1, 128] - - [392, 35.168] - - - [32896, 16897, 1, 128] - - [72, 44.774] - - - [43008, 8192, 1, 128] - - [25, 47.387] - - - [34304, 512, 1, 128] - - [290, 45.089] - - - [38528, 128, 1, 128] - - [290, 34.46] - - - [41216, 1024, 1, 128] - - [299, 51.568] - - - [38272, 22401, 1, 128] - - [72, 40.701] - - - [34048, 4096, 1, 128] - - [33, 45.503] - - - [30720, 512, 1, 128] - - [297, 45.722] - - - [41728, 512, 1, 128] - - [290, 46.785] - - - [43136, 512, 1, 128] - - [290, 47.353] - - - [41088, 1024, 1, 128] - - [419, 50.562] - - - [33536, 1024, 1, 128] - - [289, 51.065] - - - [41088, 25089, 1, 128] - - [72, 44.401] - - - [36352, 20353, 1, 128] - - [20, 46.128] - - - [29184, 1024, 1, 128] - - [289, 49.854] - - - [44800, 128, 1, 128] - - [418, 35.459] - - - [41600, 8192, 1, 128] - - [20, 46.389] - - - [44416, 28545, 1, 128] - - [20, 45.558] - - - [34048, 512, 1, 128] - - [293, 47.795] - - - [32128, 16257, 1, 128] - - [35, 45.982] - - - [44288, 4096, 1, 128] - - [27, 45.46] - - - [34432, 18433, 1, 128] - - [72, 44.145] - - - [41856, 25857, 1, 128] - - [20, 45.643] - - - [32128, 2048, 1, 128] - - [299, 54.634] - - - [34688, 512, 1, 128] - - [272, 47.274] - - - [39936, 4096, 1, 128] - - [35, 46.381] - - - [38656, 1024, 1, 128] - - [289, 52.188] - - - [37760, 512, 1, 128] - - [296, 47.028] - - - [30336, 512, 1, 128] - - [298, 46.195] - - - [38016, 22017, 1, 128] - - [36, 45.593] - - - [44544, 4096, 1, 128] - - [48, 45.793] - - - [38912, 8192, 1, 128] - - [25, 47.448] - - - [39936, 128, 1, 128] - - [270, 33.783] - - - [36480, 2048, 1, 128] - - [289, 54.787] - - - [35200, 4096, 1, 128] - - [23, 45.526] - - - [30976, 14977, 1, 128] - - [72, 44.608] - - - [31104, 15105, 1, 128] - - [27, 45.786] - - - [40832, 1024, 1, 128] - - [300, 51.355] - - - [32384, 16513, 1, 128] - - [23, 46.085] - - - [43392, 4096, 1, 128] - - [36, 45.466] - - - [32768, 4096, 1, 128] - - [23, 35.663] - - - [38272, 22273, 1, 128] - - [76, 40.665] - - - [32128, 512, 1, 128] - - [297, 46.865] - - - [32896, 2048, 1, 128] - - [419, 54.648] - - - [37376, 21505, 1, 128] - - [35, 45.849] - - - [41856, 1024, 1, 128] - - [299, 52.524] - - - [33536, 8192, 1, 128] - - [62, 46.748] - - - [29568, 1024, 1, 128] - - [300, 48.857] - - - [44032, 28033, 1, 128] - - [36, 46.336] - - - [33280, 8192, 1, 128] - - [49, 47.149] - - - [39296, 4096, 1, 128] - - [23, 45.716] - - - [30592, 14593, 1, 128] - - [27, 45.761] - - - [37504, 8192, 1, 128] - - [59, 45.639] - - - [30336, 14465, 1, 128] - - [27, 45.698] - - - [29952, 2048, 1, 128] - - [299, 53.848] - - - [40832, 512, 1, 128] - - [291, 45.236] - - - [44672, 28673, 1, 128] - - [20, 45.372] - - - [30080, 4096, 1, 128] - - [76, 43.362] - - - [37888, 2048, 1, 128] - - [299, 54.602] - - - [37632, 21761, 1, 128] - - [20, 45.775] - - - [29824, 8192, 1, 128] - - [72, 45.814] - - - [35328, 19457, 1, 128] - - [62, 46.076] - - - [37376, 4096, 1, 128] - - [33, 45.946] - - - [33792, 17921, 1, 128] - - [35, 46.337] - - - [34304, 8192, 1, 128] - - [45, 47.014] - - - [42752, 512, 1, 128] - - [291, 46.81] - - - [36992, 2048, 1, 128] - - [289, 52.65] - - - [39168, 4096, 1, 128] - - [62, 45.615] - - - [31360, 15489, 1, 128] - - [35, 45.835] - - - [43520, 8192, 1, 128] - - [23, 46.897] - - - [30080, 2048, 1, 128] - - [300, 53.678] - - - [30720, 4096, 1, 128] - - [23, 46.713] - - - [34176, 128, 1, 128] - - [293, 32.962] - - - [32768, 16769, 1, 128] - - [35, 34.684] - - - [35072, 128, 1, 128] - - [418, 33.129] - - - [35712, 4096, 1, 128] - - [36, 45.542] - - - [36480, 4096, 1, 128] - - [35, 45.435] - - - [39424, 4096, 1, 128] - - [25, 45.949] - - - [38400, 128, 1, 128] - - [418, 34.664] - - - [34432, 2048, 1, 128] - - [300, 53.791] - - - [41344, 512, 1, 128] - - [272, 46.949] - - - [35200, 512, 1, 128] - - [293, 47.51] - - - [39936, 8192, 1, 128] - - [27, 47.064] - - - [31488, 128, 1, 128] - - [418, 32.696] - - - [43008, 512, 1, 128] - - [291, 48.494] - - - [33024, 4096, 1, 128] - - [23, 46.064] - - - [36608, 512, 1, 128] - - [268, 47.207] - - - [37376, 8192, 1, 128] - - [35, 46.86] - - - [29824, 13825, 1, 128] - - [48, 45.018] - - - [36352, 2048, 1, 128] - - [299, 54.587] - - - [30336, 1024, 1, 128] - - [299, 49.573] - - - [44416, 28417, 1, 128] - - [36, 45.544] - - - [38144, 22273, 1, 128] - - [36, 45.787] - - - [28928, 2048, 1, 128] - - [289, 54.875] - - - [29568, 13697, 1, 128] - - [36, 45.321] - - - [43136, 27137, 1, 128] - - [20, 45.344] - - - [42112, 4096, 1, 128] - - [35, 45.686] - - - [40960, 512, 1, 128] - - [263, 41.578] - - - [35584, 1024, 1, 128] - - [390, 49.929] - - - [31232, 15361, 1, 128] - - [23, 45.928] - - - [40960, 8192, 1, 128] - - [23, 41.571] - - - [31232, 1024, 1, 128] - - [299, 50.601] - - - [29312, 512, 1, 128] - - [272, 46.76] - - - [44416, 512, 1, 128] - - [290, 48.15] - - - [42240, 512, 1, 128] - - [296, 46.665] - - - [31232, 8192, 1, 128] - - [25, 46.787] - - - [35072, 19201, 1, 128] - - [35, 45.897] - - - [29568, 128, 1, 128] - - [418, 31.18] - - - [33792, 2048, 1, 128] - - [419, 53.719] - - - [35712, 2048, 1, 128] - - [300, 53.813] - - - [40576, 128, 1, 128] - - [297, 33.304] - - - [40704, 1024, 1, 128] - - [299, 51.575] - - - [29824, 1024, 1, 128] - - [348, 49.431] - - - [33536, 17665, 1, 128] - - [62, 46.186] - - - [43008, 27009, 1, 128] - - [20, 46.557] - - - [34304, 2048, 1, 128] - - [289, 54.898] - - - [37120, 21249, 1, 128] - - [27, 45.742] - - - [41600, 1024, 1, 128] - - [421, 51.695] - - - [33024, 1024, 1, 128] - - [300, 49.418] - - - [42368, 512, 1, 128] - - [290, 48.364] - - - [30592, 14721, 1, 128] - - [36, 45.745] - - - [29696, 2048, 1, 128] - - [300, 53.442] - - - [31232, 128, 1, 128] - - [293, 32.484] - - - [38784, 22785, 1, 128] - - [20, 45.55] - - - [32896, 1024, 1, 128] - - [419, 49.586] - - - [32128, 128, 1, 128] - - [293, 32.637] - - - [35968, 8192, 1, 128] - - [35, 46.444] - - - [38400, 2048, 1, 128] - - [54, 44.604] - - - [36864, 2048, 1, 128] - - [304, 52.657] - - - [31616, 4096, 1, 128] - - [62, 45.071] - - - [34688, 18817, 1, 128] - - [20, 45.573] - - - [42624, 4096, 1, 128] - - [27, 41.862] - - - [29312, 1024, 1, 128] - - [299, 50.337] - - - [37760, 2048, 1, 128] - - [300, 53.333] - - - [39808, 512, 1, 128] - - [290, 46.602] - - - [41472, 128, 1, 128] - - [290, 34.821] - - - [32128, 4096, 1, 128] - - [35, 45.8] - - - [43520, 4096, 1, 128] - - [25, 46.119] - - - [41472, 512, 1, 128] - - [290, 49.526] - - - [38912, 22913, 1, 128] - - [36, 46.708] - - - [30464, 1024, 1, 128] - - [421, 49.336] - - - [33280, 128, 1, 128] - - [257, 31.854] - - - [31872, 15873, 1, 128] - - [36, 45.676] - - - [36352, 4096, 1, 128] - - [23, 45.942] - - - [30720, 2048, 1, 128] - - [289, 54.014] - - - [33792, 128, 1, 128] - - [392, 32.602] - - - [36096, 8192, 1, 128] - - [57, 45.087] - - - [38784, 128, 1, 128] - - [293, 35.165] - - - [30208, 2048, 1, 128] - - [299, 54.579] - - - [34432, 4096, 1, 128] - - [48, 44.362] - - - [42880, 128, 1, 128] - - [418, 34.937] - - - [31616, 15745, 1, 128] - - [20, 45.341] - - - [40960, 2048, 1, 128] - - [20, 39.906] - - - [41344, 128, 1, 128] - - [418, 33.539] - - - [41728, 25857, 1, 128] - - [62, 45.145] - - - [32896, 512, 1, 128] - - [300, 43.691] - - - [41728, 2048, 1, 128] - - [57, 43.251] - - - [42368, 26369, 1, 128] - - [45, 45.288] - - - [30720, 14721, 1, 128] - - [27, 46.809] - - - [37376, 512, 1, 128] - - [272, 47.547] - - - [35456, 19457, 1, 128] - - [25, 45.548] - - - [29184, 13185, 1, 128] - - [20, 45.997] - - - [34944, 128, 1, 128] - - [418, 33.436] - - - [36608, 20609, 1, 128] - - [25, 45.82] - - - [35584, 19585, 1, 128] - - [36, 45.685] - - - [42880, 8192, 1, 128] - - [45, 45.728] - - - [39936, 1024, 1, 128] - - [289, 53.418] - - - [34944, 19073, 1, 128] - - [27, 45.75] - - - [32512, 128, 1, 128] - - [392, 32.168] - - - [40064, 512, 1, 128] - - [291, 47.544] - - - [30464, 2048, 1, 128] - - [419, 53.32] - - - [30592, 8192, 1, 128] - - [25, 46.473] - - - [39040, 512, 1, 128] - - [290, 46.48] - - - [41088, 128, 1, 128] - - [392, 33.187] - - - [29824, 128, 1, 128] - - [293, 30.86] - - - [32384, 128, 1, 128] - - [291, 32.804] - - - [41728, 25729, 1, 128] - - [48, 45.1] - - - [30976, 4096, 1, 128] - - [49, 44.37] - - - [42624, 128, 1, 128] - - [297, 33.868] - - - [42112, 512, 1, 128] - - [290, 47.191] - - - [38784, 2048, 1, 128] - - [58, 44.143] - - - [35200, 8192, 1, 128] - - [25, 46.357] - - - [30976, 128, 1, 128] - - [266, 31.922] - - - [32640, 16641, 1, 128] - - [25, 45.367] - - - [41984, 8192, 1, 128] - - [25, 47.088] - - - [30080, 128, 1, 128] - - [293, 31.135] - - - [35584, 512, 1, 128] - - [290, 45.213] - - - [44800, 2048, 1, 128] - - [36, 44.034] - - - [34048, 128, 1, 128] - - [418, 32.757] - - - [35712, 1024, 1, 128] - - [300, 49.769] - - - [43136, 128, 1, 128] - - [290, 35.517] - - - [33280, 1024, 1, 128] - - [289, 50.937] - - - [34816, 18945, 1, 128] - - [20, 46.764] - - - [40704, 8192, 1, 128] - - [35, 46.688] - - - [34304, 128, 1, 128] - - [293, 33.16] - - - [39936, 512, 1, 128] - - [290, 48.841] - - - [36096, 2048, 1, 128] - - [300, 53.406] - - - [40832, 8192, 1, 128] - - [27, 46.421] - - - [37760, 4096, 1, 128] - - [35, 45.537] - - - [36736, 512, 1, 128] - - [268, 46.971] - - - [31744, 8192, 1, 128] - - [25, 47.209] - - - [33920, 1024, 1, 128] - - [421, 51.027] - - - [39808, 128, 1, 128] - - [293, 34.327] - - - [36608, 2048, 1, 128] - - [299, 55.186] - - - [30464, 14593, 1, 128] - - [76, 42.867] - - - [35200, 19201, 1, 128] - - [36, 45.538] - - - [41472, 1024, 1, 128] - - [289, 53.315] - - - [30720, 128, 1, 128] - - [266, 32.217] - - - [41600, 128, 1, 128] - - [291, 33.981] - - - [38144, 22145, 1, 128] - - [20, 45.754] - - - [37120, 4096, 1, 128] - - [36, 45.686] - - - [40704, 24705, 1, 128] - - [20, 45.747] - - - [41088, 25217, 1, 128] - - [36, 44.43] - - - [43776, 8192, 1, 128] - - [59, 45.549] - - - [38912, 1024, 1, 128] - - [289, 51.605] - - - [43008, 2048, 1, 128] - - [23, 44.887] - - - [42496, 26497, 1, 128] - - [62, 46.082] - - - [33536, 512, 1, 128] - - [300, 46.345] - - - [43520, 512, 1, 128] - - [348, 46.079] - - - [39040, 23169, 1, 128] - - [20, 45.446] - - - [29568, 2048, 1, 128] - - [300, 53.798] - - - [44672, 8192, 1, 128] - - [25, 46.483] - - - [29824, 512, 1, 128] - - [419, 45.799] - - - [34944, 2048, 1, 128] - - [289, 55.213] - - - [33408, 4096, 1, 128] - - [23, 45.857] - - - [41600, 25729, 1, 128] - - [36, 45.529] - - - [40832, 2048, 1, 128] - - [25, 44.139] - - - [38912, 128, 1, 128] - - [418, 35.187] - - - [34048, 2048, 1, 128] - - [299, 53.968] - - - [43904, 2048, 1, 128] - - [51, 44.341] - - - [39296, 23297, 1, 128] - - [20, 45.524] - - - [31232, 4096, 1, 128] - - [35, 45.845] - - - [35840, 1024, 1, 128] - - [300, 51.076] - - - [28928, 128, 1, 128] - - [256, 30.875] - - - [42752, 2048, 1, 128] - - [36, 44.089] - - - [44032, 1024, 1, 128] - - [390, 53.051] - - - [29440, 13569, 1, 128] - - [23, 45.99] - - - [35456, 19585, 1, 128] - - [27, 45.703] - - - [35840, 19841, 1, 128] - - [35, 46.395] - - - [31360, 128, 1, 128] - - [256, 32.087] - - - [40192, 2048, 1, 128] - - [58, 44.606] - - - [33920, 8192, 1, 128] - - [25, 46.43] - - - [43648, 512, 1, 128] - - [291, 46.693] - - - [30080, 14209, 1, 128] - - [43, 43.283] - - - [39680, 23809, 1, 128] - - [20, 45.422] - - - [32512, 512, 1, 128] - - [421, 45.837] - - - [34816, 2048, 1, 128] - - [390, 54.9] - - - [43392, 1024, 1, 128] - - [289, 51.67] - - - [39040, 4096, 1, 128] - - [25, 45.454] - - - [43264, 4096, 1, 128] - - [27, 45.807] - - - [44416, 2048, 1, 128] - - [34, 44.36] - - - [31488, 512, 1, 128] - - [289, 46.661] - - - [31616, 1024, 1, 128] - - [300, 50.43] - - - [44032, 8192, 1, 128] - - [25, 47.089] - - - [39424, 23553, 1, 128] - - [36, 45.76] - - - [31360, 8192, 1, 128] - - [25, 46.446] - - - [42752, 128, 1, 128] - - [418, 34.6] - - - [40192, 512, 1, 128] - - [291, 48.033] - - - [36096, 20225, 1, 128] - - [45, 44.722] - - - [41984, 26113, 1, 128] - - [20, 46.343] - - - [39936, 2048, 1, 128] - - [56, 44.806] - - - [42880, 2048, 1, 128] - - [31, 43.617] - - - [29440, 128, 1, 128] - - [290, 30.651] - - - [40192, 128, 1, 128] - - [291, 33.491] - - - [36608, 4096, 1, 128] - - [35, 45.643] - - - [37760, 21761, 1, 128] - - [20, 45.57] - - - [44160, 28161, 1, 128] - - [27, 45.034] - - - [44288, 512, 1, 128] - - [290, 48.322] - - - [29056, 13185, 1, 128] - - [27, 45.782] - - - [43904, 512, 1, 128] - - [290, 48.982] - - - [29696, 128, 1, 128] - - [418, 30.959] - - - [36224, 8192, 1, 128] - - [25, 46.49] - - - [33024, 2048, 1, 128] - - [390, 54.032] - - - [44032, 28161, 1, 128] - - [36, 46.258] - - - [44032, 128, 1, 128] - - [418, 35.101] - - - [38784, 512, 1, 128] - - [291, 47.097] - - - [29056, 8192, 1, 128] - - [35, 46.411] - - - [33920, 18049, 1, 128] - - [36, 45.703] - - - [34816, 1024, 1, 128] - - [289, 51.641] - - - [29056, 128, 1, 128] - - [293, 30.941] - - - [39552, 1024, 1, 128] - - [299, 52.403] - - - [36992, 8192, 1, 128] - - [35, 46.344] - - - [44544, 1024, 1, 128] - - [390, 51.778] - - - [43904, 27905, 1, 128] - - [23, 45.391] - - - [29440, 512, 1, 128] - - [290, 46.958] - - - [29568, 8192, 1, 128] - - [27, 46.022] - - - [41472, 2048, 1, 128] - - [70, 44.493] - - - [29184, 8192, 1, 128] - - [33, 46.616] - - - [33408, 512, 1, 128] - - [300, 46.109] - - - [38656, 22785, 1, 128] - - [62, 45.782] - - - [31744, 15745, 1, 128] - - [35, 46.536] - - - [38656, 2048, 1, 128] - - [61, 44.176] - - - [30080, 8192, 1, 128] - - [72, 44.644] - - - [44672, 128, 1, 128] - - [392, 36.159] - - - [40704, 24833, 1, 128] - - [48, 45.754] - - - [33792, 8192, 1, 128] - - [25, 47.181] - - - [33920, 512, 1, 128] - - [293, 47.638] - - - [40576, 1024, 1, 128] - - [289, 52.164] - - - [36224, 20225, 1, 128] - - [36, 45.701] - - - [34432, 1024, 1, 128] - - [419, 49.862] - - - [31488, 15617, 1, 128] - - [25, 45.941] - - - [40576, 2048, 1, 128] - - [51, 44.334] - - - [30208, 512, 1, 128] - - [268, 44.712] - - - [36480, 128, 1, 128] - - [291, 33.734] - - - [37504, 21633, 1, 128] - - [28, 44.759] - - - [32896, 17025, 1, 128] - - [72, 44.944] - - - [39168, 2048, 1, 128] - - [21, 44.037] - - - [29440, 2048, 1, 128] - - [390, 54.05] - - - [29440, 13441, 1, 128] - - [25, 45.952] - - - [32640, 8192, 1, 128] - - [35, 46.225] - - - [35072, 19073, 1, 128] - - [36, 45.943] - - - [33152, 512, 1, 128] - - [297, 45.951] - - - [40576, 8192, 1, 128] - - [27, 46.372] - - - [34944, 8192, 1, 128] - - [25, 46.479] - - - [38656, 128, 1, 128] - - [291, 34.859] - - - [33536, 17537, 1, 128] - - [48, 46.122] - - - [29952, 512, 1, 128] - - [289, 45.523] - - - [31488, 2048, 1, 128] - - [289, 54.879] - - - [31872, 4096, 1, 128] - - [36, 45.435] - - - [31232, 15233, 1, 128] - - [48, 46.235] - - - [38912, 23041, 1, 128] - - [20, 46.582] - - - [31232, 2048, 1, 128] - - [289, 55.375] - - - [40448, 8192, 1, 128] - - [45, 46.84] - - - [36352, 128, 1, 128] - - [291, 34.037] - - - [43776, 4096, 1, 128] - - [59, 44.709] - - - [32000, 8192, 1, 128] - - [27, 46.687] - - - [37760, 8192, 1, 128] - - [23, 46.459] - - - [30080, 1024, 1, 128] - - [421, 49.187] - - - [44544, 128, 1, 128] - - [266, 34.9] - - - [29696, 1024, 1, 128] - - [300, 49.526] - - - [32640, 512, 1, 128] - - [290, 45.514] - - - [44416, 128, 1, 128] - - [418, 34.998] - - - [41216, 512, 1, 128] - - [421, 46.274] - - - [31872, 512, 1, 128] - - [297, 45.527] - - - [34432, 512, 1, 128] - - [357, 45.968] - - - [34560, 1024, 1, 128] - - [299, 52.567] - - - [42240, 128, 1, 128] - - [266, 34.747] - - - [44288, 28289, 1, 128] - - [20, 45.366] - - - [30336, 14337, 1, 128] - - [45, 45.421] - - - [32384, 2048, 1, 128] - - [299, 53.886] - - - [38400, 22401, 1, 128] - - [62, 46.128] - - - [39296, 1024, 1, 128] - - [421, 51.438] - - - [28928, 8192, 1, 128] - - [25, 46.561] - - - [40320, 2048, 1, 128] - - [61, 44.54] - - - [31104, 15233, 1, 128] - - [23, 45.866] - - - [39680, 512, 1, 128] - - [299, 46.943] - - - [34048, 18049, 1, 128] - - [62, 45.789] - - - [30720, 1024, 1, 128] - - [419, 50.708] - - - [42880, 26881, 1, 128] - - [62, 44.892] - - - [32896, 8192, 1, 128] - - [72, 45.437] - - - [43264, 8192, 1, 128] - - [35, 46.531] - - - [37632, 4096, 1, 128] - - [20, 45.643] - - - [32256, 4096, 1, 128] - - [55, 46.088] - - - [37248, 4096, 1, 128] - - [25, 45.415] - - - [33280, 17409, 1, 128] - - [62, 46.304] - - - [36096, 512, 1, 128] - - [419, 45.537] - - - [37120, 21121, 1, 128] - - [25, 45.711] - - - [32896, 128, 1, 128] - - [418, 31.145] - - - [36352, 20481, 1, 128] - - [36, 45.734] - - - [43392, 2048, 1, 128] - - [35, 43.776] - - - [36352, 512, 1, 128] - - [291, 46.401] - - - [29056, 13057, 1, 128] - - [20, 45.668] - - - [29056, 4096, 1, 128] - - [53, 45.545] - - - [37888, 4096, 1, 128] - - [35, 46.418] - - - [40320, 512, 1, 128] - - [290, 46.72] - - - [39168, 128, 1, 128] - - [293, 34.765] - - - [41472, 8192, 1, 128] - - [49, 46.812] - - - [34560, 512, 1, 128] - - [293, 48.139] - - - [34176, 18305, 1, 128] - - [49, 45.623] - - - [34688, 8192, 1, 128] - - [27, 46.367] - - - [29696, 13825, 1, 128] - - [35, 46.514] - - - [33152, 17281, 1, 128] - - [35, 46.072] - - - [30208, 8192, 1, 128] - - [49, 46.932] - - - [43648, 27649, 1, 128] - - [36, 44.631] - - - [31360, 2048, 1, 128] - - [289, 54.384] - - - [41984, 128, 1, 128] - - [392, 34.455] - - - [38528, 2048, 1, 128] - - [21, 43.961] - - - [32256, 16385, 1, 128] - - [28, 46.375] - - - [42240, 1024, 1, 128] - - [289, 51.81] - - - [32000, 16001, 1, 128] - - [20, 46.01] - - - [37248, 1024, 1, 128] - - [289, 51.288] - - - [32256, 1024, 1, 128] - - [299, 51.245] - - - [39296, 23425, 1, 128] - - [20, 45.506] - - - [43008, 4096, 1, 128] - - [23, 46.736] - - - [31104, 128, 1, 128] - - [418, 31.857] - - - [38656, 8192, 1, 128] - - [62, 46.561] - - - [44288, 128, 1, 128] - - [418, 35.255] - - - [38528, 22657, 1, 128] - - [20, 45.514] - - - [39552, 128, 1, 128] - - [293, 34.022] - - - [37376, 21377, 1, 128] - - [36, 46.05] - - - [28928, 13057, 1, 128] - - [25, 45.96] - - - [43264, 27265, 1, 128] - - [20, 45.734] - - - [35328, 4096, 1, 128] - - [62, 46.012] - - - [30848, 4096, 1, 128] - - [33, 45.587] - - - [44800, 28801, 1, 128] - - [36, 45.329] - - - [35456, 512, 1, 128] - - [291, 48.424] - - - [40960, 24961, 1, 128] - - [20, 40.336] - - - [39936, 23937, 1, 128] - - [20, 46.264] - - - [31744, 1024, 1, 128] - - [300, 49.42] - - - [32128, 8192, 1, 128] - - [20, 46.465] - - - [42112, 26113, 1, 128] - - [20, 45.455] - - - [31744, 2048, 1, 128] - - [419, 53.911] - - - [42112, 1024, 1, 128] - - [299, 52.106] - - - [40064, 8192, 1, 128] - - [20, 45.948] - - - [38144, 128, 1, 128] - - [272, 34.22] - - - [42624, 2048, 1, 128] - - [25, 40.989] - - - [36992, 128, 1, 128] - - [272, 33.291] - - - [40192, 8192, 1, 128] - - [27, 46.619] - - - [40064, 24065, 1, 128] - - [48, 44.533] - - - [37760, 21889, 1, 128] - - [20, 45.545] - - - [36352, 8192, 1, 128] - - [23, 46.787] - - - [44544, 512, 1, 128] - - [298, 47.185] - - - [32384, 4096, 1, 128] - - [25, 45.721] - - - [39168, 23169, 1, 128] - - [48, 45.761] - - - [1408, 897, 1, 128] - - [105, 29.971] - - - [16512, 512, 1, 128] - - [300, 39.434] - - - [20480, 12673, 1, 128] - - [27, 47.23] - - - [20992, 512, 1, 128] - - [293, 44.717] - - - [9344, 512, 1, 128] - - [418, 36.788] - - - [18048, 2048, 1, 128] - - [419, 52.973] - - - [20352, 12673, 1, 128] - - [27, 45.916] - - - [640, 128, 1, 128] - - [109, 3.662] - - - [28160, 512, 1, 128] - - [268, 45.015] - - - [20608, 4096, 1, 128] - - [25, 44.987] - - - [19328, 1024, 1, 128] - - [357, 48.029] - - - [26496, 4096, 1, 128] - - [65, 44.965] - - - [10624, 512, 1, 128] - - [293, 38.164] - - - [20352, 1024, 1, 128] - - [296, 48.547] - - - [10240, 6529, 1, 128] - - [421, 54.745] - - - [22144, 14465, 1, 128] - - [25, 45.898] - - - [13184, 2048, 1, 128] - - [300, 50.562] - - - [14720, 6913, 1, 128] - - [23, 44.874] - - - [21248, 512, 1, 128] - - [290, 41.725] - - - [10496, 128, 1, 128] - - [96, 31.751] - - - [13056, 5377, 1, 128] - - [289, 54.304] - - - [10880, 128, 1, 128] - - [96, 29.524] - - - [18688, 512, 1, 128] - - [293, 43.679] - - - [22656, 4096, 1, 128] - - [53, 45.081] - - - [15232, 1024, 1, 128] - - [421, 47.74] - - - [20224, 4096, 1, 128] - - [33, 45.205] - - - [6016, 2305, 1, 128] - - [420, 45.883] - - - [13184, 4096, 1, 128] - - [419, 54.674] - - - [256, 129, 1, 128] - - [150, 1.465] - - - [11264, 7553, 1, 128] - - [35, 45.632] - - - [18176, 128, 1, 128] - - [293, 27.695] - - - [15872, 8193, 1, 128] - - [35, 45.629] - - - [26112, 4096, 1, 128] - - [55, 45.885] - - - [22784, 2048, 1, 128] - - [419, 53.805] - - - [10880, 7297, 1, 128] - - [35, 44.655] - - - [14720, 2048, 1, 128] - - [421, 50.308] - - - [9216, 5633, 1, 128] - - [421, 53.171] - - - [23040, 15233, 1, 128] - - [35, 46.577] - - - [8832, 5121, 1, 128] - - [419, 53.012] - - - [18816, 1024, 1, 128] - - [300, 48.449] - - - [128, 129, 1, 128] - - [101, 0.738] - - - [15488, 512, 1, 128] - - [290, 42.143] - - - [18176, 1024, 1, 128] - - [421, 48.229] - - - [16128, 8449, 1, 128] - - [25, 45.56] - - - [16000, 2048, 1, 128] - - [421, 52.236] - - - [24960, 9089, 1, 128] - - [35, 45.317] - - - [14336, 1024, 1, 128] - - [348, 46.097] - - - [25472, 8192, 1, 128] - - [23, 46.584] - - - [23040, 128, 1, 128] - - [21, 31.621] - - - [9472, 512, 1, 128] - - [418, 36.413] - - - [19072, 128, 1, 128] - - [290, 27.559] - - - [10624, 6913, 1, 128] - - [289, 53.711] - - - [7808, 1024, 1, 128] - - [300, 41.537] - - - [27008, 11137, 1, 128] - - [48, 45.459] - - - [21504, 4096, 1, 128] - - [27, 46.162] - - - [7936, 1024, 1, 128] - - [421, 41.041] - - - [12928, 5121, 1, 128] - - [421, 54.18] - - - [26240, 8192, 1, 128] - - [27, 42.362] - - - [18304, 2048, 1, 128] - - [299, 53.549] - - - [24576, 1024, 1, 128] - - [302, 44.83] - - - [10624, 128, 1, 128] - - [91, 28.619] - - - [24576, 128, 1, 128] - - [32, 33.175] - - - [25600, 9601, 1, 128] - - [36, 46.181] - - - [5248, 128, 1, 128] - - [146, 22.135] - - - [24448, 4096, 1, 128] - - [33, 45.432] - - - [19328, 128, 1, 128] - - [290, 27.624] - - - [24064, 512, 1, 128] - - [293, 45.491] - - - [11136, 512, 1, 128] - - [418, 38.518] - - - [14592, 1024, 1, 128] - - [300, 45.672] - - - [12544, 4737, 1, 128] - - [289, 53.311] - - - [17280, 128, 1, 128] - - [266, 26.918] - - - [25344, 8192, 1, 128] - - [62, 46.021] - - - [4608, 512, 1, 128] - - [392, 28.144] - - - [4608, 128, 1, 128] - - [98, 20.246] - - - [21760, 512, 1, 128] - - [298, 43.191] - - - [7936, 128, 1, 128] - - [90, 26.151] - - - [11008, 7425, 1, 128] - - [79, 44.734] - - - [13824, 2048, 1, 128] - - [300, 50.009] - - - [18048, 512, 1, 128] - - [370, 42.155] - - - [19584, 11905, 1, 128] - - [23, 45.771] - - - [22656, 512, 1, 128] - - [290, 44.496] - - - [4608, 3073, 1, 128] - - [291, 45.21] - - - [5504, 128, 1, 128] - - [105, 22.823] - - - [4864, 1024, 1, 128] - - [291, 36.173] - - - [17664, 1024, 1, 128] - - [419, 46.24] - - - [18176, 2048, 1, 128] - - [421, 52.941] - - - [2048, 1537, 1, 128] - - [418, 30.331] - - - [22528, 128, 1, 128] - - [348, 29.227] - - - [21760, 13953, 1, 128] - - [20, 46.306] - - - [7040, 128, 1, 128] - - [92, 27.349] - - - [3328, 1665, 1, 128] - - [294, 38.65] - - - [768, 512, 1, 128] - - [101, 16.196] - - - [21504, 13697, 1, 128] - - [20, 46.635] - - - [18560, 10881, 1, 128] - - [27, 45.473] - - - [2560, 128, 1, 128] - - [98, 13.997] - - - [15616, 1024, 1, 128] - - [348, 46.849] - - - [19456, 4096, 1, 128] - - [25, 45.916] - - - [25600, 2048, 1, 128] - - [289, 54.114] - - - [2304, 128, 1, 128] - - [131, 12.504] - - - [1664, 1025, 1, 128] - - [344, 22.494] - - - [23168, 15361, 1, 128] - - [23, 45.518] - - - [9856, 128, 1, 128] - - [90, 30.957] - - - [13312, 2048, 1, 128] - - [299, 50.798] - - - [19200, 512, 1, 128] - - [272, 41.737] - - - [19200, 2048, 1, 128] - - [289, 51.895] - - - [23168, 2048, 1, 128] - - [300, 53.617] - - - [18688, 128, 1, 128] - - [290, 27.388] - - - [13568, 1024, 1, 128] - - [390, 45.357] - - - [17792, 9985, 1, 128] - - [48, 45.306] - - - [20608, 1024, 1, 128] - - [314, 46.71] - - - [11648, 8065, 1, 128] - - [35, 44.991] - - - [1280, 128, 1, 128] - - [95, 7.324] - - - [16256, 4096, 1, 128] - - [421, 57.409] - - - [17024, 1024, 1, 128] - - [348, 47.259] - - - [19456, 128, 1, 128] - - [392, 28.839] - - - [20736, 512, 1, 128] - - [290, 44.957] - - - [14464, 6785, 1, 128] - - [23, 44.794] - - - [20736, 13057, 1, 128] - - [35, 46.071] - - - [8704, 2048, 1, 128] - - [300, 48.855] - - - [640, 512, 1, 128] - - [95, 13.792] - - - [768, 129, 1, 128] - - [150, 4.395] - - - [27776, 1024, 1, 128] - - [390, 49.211] - - - [19200, 11521, 1, 128] - - [25, 45.806] - - - [6400, 2048, 1, 128] - - [294, 47.4] - - - [14976, 7297, 1, 128] - - [25, 44.964] - - - [7040, 2048, 1, 128] - - [300, 47.265] - - - [25984, 128, 1, 128] - - [32, 34.207] - - - [13696, 128, 1, 128] - - [259, 22.203] - - - [2688, 1153, 1, 128] - - [270, 31.129] - - - [15232, 2048, 1, 128] - - [299, 51.36] - - - [11776, 128, 1, 128] - - [151, 29.869] - - - [3328, 512, 1, 128] - - [105, 32.644] - - - [11648, 7937, 1, 128] - - [23, 44.948] - - - [19456, 2048, 1, 128] - - [299, 53.82] - - - [11008, 128, 1, 128] - - [91, 29.228] - - - [9984, 6401, 1, 128] - - [421, 53.535] - - - [25856, 9857, 1, 128] - - [35, 45.618] - - - [4224, 512, 1, 128] - - [418, 26.826] - - - [13568, 5761, 1, 128] - - [23, 44.662] - - - [5632, 2049, 1, 128] - - [287, 44.156] - - - [8832, 2048, 1, 128] - - [348, 48.641] - - - [5632, 3969, 1, 128] - - [287, 49.058] - - - [25856, 2048, 1, 128] - - [421, 53.949] - - - [25472, 2048, 1, 128] - - [419, 53.722] - - - [20736, 12929, 1, 128] - - [25, 46.039] - - - [14592, 128, 1, 128] - - [293, 23.452] - - - [1792, 512, 1, 128] - - [91, 24.27] - - - [14208, 2048, 1, 128] - - [299, 50.766] - - - [15360, 7681, 1, 128] - - [35, 45.649] - - - [5760, 2048, 1, 128] - - [287, 45.723] - - - [6400, 512, 1, 128] - - [418, 32.672] - - - [5248, 3713, 1, 128] - - [300, 49.769] - - - [16768, 1024, 1, 128] - - [300, 45.64] - - - [10752, 512, 1, 128] - - [291, 38.122] - - - [26624, 2048, 1, 128] - - [299, 54.67] - - - [384, 128, 1, 128] - - [98, 2.232] - - - [27392, 8192, 1, 128] - - [72, 45.808] - - - [24448, 512, 1, 128] - - [390, 43.739] - - - [11136, 7553, 1, 128] - - [27, 44.898] - - - [17024, 9345, 1, 128] - - [25, 45.64] - - - [16000, 8193, 1, 128] - - [25, 45.091] - - - [5888, 2048, 1, 128] - - [420, 45.596] - - - [18304, 10497, 1, 128] - - [23, 45.496] - - - [3968, 128, 1, 128] - - [95, 17.751] - - - [14336, 6529, 1, 128] - - [27, 45.676] - - - [19840, 128, 1, 128] - - [297, 28.609] - - - [25600, 8192, 1, 128] - - [27, 47.314] - - - [18688, 11009, 1, 128] - - [23, 45.921] - - - [7680, 1024, 1, 128] - - [294, 40.948] - - - [7168, 128, 1, 128] - - [152, 24.049] - - - [1664, 512, 1, 128] - - [90, 26.701] - - - [12544, 1024, 1, 128] - - [348, 44.839] - - - [6528, 2048, 1, 128] - - [421, 48.57] - - - [19072, 4096, 1, 128] - - [421, 55.239] - - - [2048, 512, 1, 128] - - [153, 27.115] - - - [13568, 5889, 1, 128] - - [25, 44.448] - - - [23680, 16001, 1, 128] - - [62, 45.198] - - - [26112, 10113, 1, 128] - - [48, 46.101] - - - [15872, 128, 1, 128] - - [418, 24.978] - - - [16384, 512, 1, 128] - - [418, 35.323] - - - [9856, 6273, 1, 128] - - [419, 53.917] - - - [26368, 1024, 1, 128] - - [390, 49.113] - - - [16256, 2048, 1, 128] - - [300, 52.607] - - - [3968, 2305, 1, 128] - - [301, 43.744] - - - [28672, 8192, 1, 128] - - [23, 47.54] - - - [10368, 1024, 1, 128] - - [294, 43.855] - - - [11008, 1024, 1, 128] - - [295, 43.831] - - - [11776, 4097, 1, 128] - - [421, 52.955] - - - [26496, 2048, 1, 128] - - [300, 53.866] - - - [17792, 4096, 1, 128] - - [289, 54.3] - - - [2304, 512, 1, 128] - - [91, 29.196] - - - [9216, 2048, 1, 128] - - [421, 47.915] - - - [12416, 512, 1, 128] - - [291, 38.03] - - - [18048, 128, 1, 128] - - [418, 26.691] - - - [21888, 14209, 1, 128] - - [74, 41.982] - - - [9344, 5761, 1, 128] - - [299, 53.299] - - - [19712, 2048, 1, 128] - - [300, 52.819] - - - [12288, 1024, 1, 128] - - [296, 43.752] - - - [3584, 1921, 1, 128] - - [294, 40.907] - - - [22784, 128, 1, 128] - - [308, 29.434] - - - [26880, 128, 1, 128] - - [343, 29.521] - - - [17408, 1024, 1, 128] - - [296, 48.138] - - - [15488, 4096, 1, 128] - - [299, 55.362] - - - [13312, 5633, 1, 128] - - [299, 54.253] - - - [22016, 14337, 1, 128] - - [28, 46.472] - - - [19328, 2048, 1, 128] - - [299, 53.431] - - - [25600, 128, 1, 128] - - [58, 33.998] - - - [22784, 15105, 1, 128] - - [62, 46.187] - - - [5376, 3713, 1, 128] - - [421, 48.796] - - - [14208, 512, 1, 128] - - [290, 40.195] - - - [12928, 4096, 1, 128] - - [300, 54.968] - - - [768, 257, 1, 128] - - [156, 3.819] - - - [27776, 11777, 1, 128] - - [35, 45.416] - - - [12032, 1024, 1, 128] - - [301, 43.329] - - - [14208, 4096, 1, 128] - - [299, 55.301] - - - [19840, 12161, 1, 128] - - [35, 45.78] - - - [17536, 512, 1, 128] - - [268, 42.669] - - - [19840, 4096, 1, 128] - - [32, 45.206] - - - [26624, 512, 1, 128] - - [392, 44.708] - - - [27136, 11137, 1, 128] - - [28, 46.29] - - - [11008, 512, 1, 128] - - [345, 37.608] - - - [1024, 513, 1, 128] - - [147, 18.031] - - - [15744, 512, 1, 128] - - [290, 41.164] - - - [22016, 128, 1, 128] - - [259, 28.887] - - - [9344, 1024, 1, 128] - - [421, 43.949] - - - [28544, 1024, 1, 128] - - [289, 49.116] - - - [13440, 5633, 1, 128] - - [300, 53.569] - - - [21632, 13825, 1, 128] - - [36, 45.844] - - - [24064, 4096, 1, 128] - - [28, 45.78] - - - [24192, 512, 1, 128] - - [418, 43.49] - - - [22912, 15233, 1, 128] - - [25, 46.056] - - - [20864, 13185, 1, 128] - - [27, 45.554] - - - [8064, 4353, 1, 128] - - [287, 51.705] - - - [8704, 5121, 1, 128] - - [419, 53.387] - - - [19840, 1024, 1, 128] - - [421, 48.157] - - - [15616, 128, 1, 128] - - [290, 24.845] - - - [21632, 512, 1, 128] - - [268, 44.051] - - - [13440, 512, 1, 128] - - [290, 39.255] - - - [23936, 128, 1, 128] - - [392, 30.586] - - - [8960, 5377, 1, 128] - - [421, 53.068] - - - [27008, 512, 1, 128] - - [293, 46.138] - - - [13440, 5761, 1, 128] - - [419, 54.22] - - - [3072, 512, 1, 128] - - [154, 30.642] - - - [4096, 1024, 1, 128] - - [418, 34.487] - - - [7296, 3585, 1, 128] - - [420, 49.692] - - - [12416, 4737, 1, 128] - - [289, 53.227] - - - [6912, 512, 1, 128] - - [418, 31.786] - - - [11136, 2048, 1, 128] - - [390, 49.837] - - - [18176, 10369, 1, 128] - - [23, 45.848] - - - [14976, 4096, 1, 128] - - [421, 55.01] - - - [19712, 4096, 1, 128] - - [38, 43.38] - - - [8064, 1024, 1, 128] - - [294, 42.69] - - - [9600, 128, 1, 128] - - [105, 30.281] - - - [26240, 1024, 1, 128] - - [300, 48.83] - - - [5248, 3585, 1, 128] - - [300, 49.02] - - - [16768, 2048, 1, 128] - - [300, 52.223] - - - [13184, 128, 1, 128] - - [416, 23.851] - - - [19328, 11521, 1, 128] - - [35, 45.479] - - - [4864, 512, 1, 128] - - [418, 29.015] - - - [3584, 2049, 1, 128] - - [300, 41.721] - - - [18560, 128, 1, 128] - - [293, 27.51] - - - [27392, 11393, 1, 128] - - [23, 44.508] - - - [27520, 512, 1, 128] - - [300, 44.51] - - - [18176, 4096, 1, 128] - - [289, 55.776] - - - [7808, 4225, 1, 128] - - [421, 50.554] - - - [15232, 128, 1, 128] - - [266, 24.374] - - - [25728, 1024, 1, 128] - - [419, 48.1] - - - [23936, 512, 1, 128] - - [290, 44.79] - - - [23424, 2048, 1, 128] - - [419, 54.142] - - - [28032, 12161, 1, 128] - - [23, 45.603] - - - [27136, 512, 1, 128] - - [290, 43.429] - - - [14336, 6657, 1, 128] - - [20, 45.789] - - - [15616, 4096, 1, 128] - - [390, 55.401] - - - [3328, 1793, 1, 128] - - [419, 39.844] - - - [28416, 512, 1, 128] - - [390, 44.651] - - - [16384, 8705, 1, 128] - - [27, 37.424] - - - [3200, 1537, 1, 128] - - [392, 36.062] - - - [26368, 128, 1, 128] - - [61, 35.08] - - - [16000, 512, 1, 128] - - [268, 41.241] - - - [25216, 9345, 1, 128] - - [25, 45.479] - - - [28288, 4096, 1, 128] - - [20, 45.457] - - - [24832, 512, 1, 128] - - [309, 42.656] - - - [18048, 10369, 1, 128] - - [22, 44.334] - - - [20480, 4096, 1, 128] - - [32, 46.324] - - - [17792, 10113, 1, 128] - - [22, 45.432] - - - [13312, 5505, 1, 128] - - [299, 54.01] - - - [17024, 2048, 1, 128] - - [421, 51.601] - - - [20608, 12929, 1, 128] - - [27, 45.702] - - - [16896, 4096, 1, 128] - - [300, 55.017] - - - [27776, 2048, 1, 128] - - [300, 53.865] - - - [6912, 3201, 1, 128] - - [417, 49.42] - - - [15744, 2048, 1, 128] - - [421, 51.29] - - - [24448, 128, 1, 128] - - [298, 30.666] - - - [2688, 128, 1, 128] - - [98, 14.172] - - - [7808, 2048, 1, 128] - - [289, 47.973] - - - [1408, 512, 1, 128] - - [105, 23.224] - - - [12032, 512, 1, 128] - - [298, 38.7] - - - [26752, 512, 1, 128] - - [268, 45.517] - - - [16128, 8321, 1, 128] - - [27, 45.623] - - - [25856, 128, 1, 128] - - [57, 34.214] - - - [24064, 8192, 1, 128] - - [55, 47.018] - - - [28160, 4096, 1, 128] - - [55, 45.418] - - - [13312, 128, 1, 128] - - [96, 31.595] - - - [10112, 6401, 1, 128] - - [421, 53.899] - - - [16384, 4096, 1, 128] - - [310, 47.528] - - - [16512, 2048, 1, 128] - - [421, 51.98] - - - [27520, 11521, 1, 128] - - [25, 45.627] - - - [8192, 4481, 1, 128] - - [294, 51.551] - - - [16768, 512, 1, 128] - - [291, 42.626] - - - [6144, 128, 1, 128] - - [90, 25.055] - - - [13568, 512, 1, 128] - - [290, 39.924] - - - [9344, 5633, 1, 128] - - [299, 53.059] - - - [13440, 4096, 1, 128] - - [421, 54.12] - - - [2176, 1665, 1, 128] - - [301, 33.293] - - - [28288, 128, 1, 128] - - [256, 30.564] - - - [11776, 4096, 1, 128] - - [419, 53.897] - - - [17280, 512, 1, 128] - - [293, 43.256] - - - [5504, 3841, 1, 128] - - [287, 49.825] - - - [14848, 7041, 1, 128] - - [23, 45.163] - - - [3584, 128, 1, 128] - - [95, 16.958] - - - [26880, 8192, 1, 128] - - [35, 46.482] - - - [2944, 1409, 1, 128] - - [384, 33.188] - - - [26368, 10369, 1, 128] - - [35, 45.792] - - - [21888, 512, 1, 128] - - [272, 40.618] - - - [15872, 2048, 1, 128] - - [419, 51.353] - - - [20224, 512, 1, 128] - - [300, 42.372] - - - [24320, 8449, 1, 128] - - [35, 45.716] - - - [5632, 1024, 1, 128] - - [291, 37.929] - - - [17152, 9473, 1, 128] - - [25, 45.713] - - - [4096, 128, 1, 128] - - [147, 18.547] - - - [8832, 128, 1, 128] - - [151, 28.344] - - - [2048, 1409, 1, 128] - - [256, 29.1] - - - [28160, 12289, 1, 128] - - [49, 45.864] - - - [9088, 5505, 1, 128] - - [289, 53.461] - - - [19200, 1024, 1, 128] - - [299, 45.813] - - - [18048, 4096, 1, 128] - - [421, 54.757] - - - [12928, 512, 1, 128] - - [297, 39.441] - - - [20864, 4096, 1, 128] - - [53, 45.018] - - - [27008, 2048, 1, 128] - - [390, 53.958] - - - [16640, 128, 1, 128] - - [293, 24.227] - - - [24960, 8192, 1, 128] - - [25, 46.265] - - - [24320, 1024, 1, 128] - - [348, 47.796] - - - [23552, 15873, 1, 128] - - [23, 46.69] - - - [26240, 4096, 1, 128] - - [25, 42.717] - - - [24320, 128, 1, 128] - - [55, 32.949] - - - [26240, 128, 1, 128] - - [21, 34.73] - - - [3200, 1665, 1, 128] - - [295, 37.437] - - - [11776, 2048, 1, 128] - - [300, 48.835] - - - [6144, 512, 1, 128] - - [392, 32.383] - - - [24960, 128, 1, 128] - - [21, 33.334] - - - [23424, 128, 1, 128] - - [422, 30.377] - - - [11776, 8065, 1, 128] - - [65, 45.272] - - - [19072, 11265, 1, 128] - - [23, 45.403] - - - [8192, 4609, 1, 128] - - [420, 51.195] - - - [21888, 4096, 1, 128] - - [72, 40.382] - - - [14976, 2048, 1, 128] - - [289, 50.585] - - - [23680, 4096, 1, 128] - - [48, 44.576] - - - [14080, 1024, 1, 128] - - [294, 44.971] - - - [19968, 4096, 1, 128] - - [53, 45.484] - - - [8704, 128, 1, 128] - - [92, 28.178] - - - [23424, 15745, 1, 128] - - [27, 43.83] - - - [8320, 2048, 1, 128] - - [295, 47.239] - - - [6144, 2433, 1, 128] - - [420, 46.0] - - - [19200, 11393, 1, 128] - - [35, 45.778] - - - [28416, 128, 1, 128] - - [418, 30.745] - - - [14080, 2048, 1, 128] - - [419, 52.099] - - - [12544, 4096, 1, 128] - - [299, 54.715] - - - [17024, 128, 1, 128] - - [290, 26.848] - - - [23936, 16257, 1, 128] - - [25, 46.159] - - - [12288, 128, 1, 128] - - [105, 30.85] - - - [28800, 1024, 1, 128] - - [299, 47.944] - - - [13824, 6017, 1, 128] - - [35, 45.055] - - - [23040, 2048, 1, 128] - - [421, 52.5] - - - [9984, 6273, 1, 128] - - [419, 52.8] - - - [23680, 512, 1, 128] - - [268, 44.844] - - - [7936, 4353, 1, 128] - - [421, 50.061] - - - [24192, 2048, 1, 128] - - [299, 53.594] - - - [8448, 512, 1, 128] - - [256, 34.153] - - - [5760, 2177, 1, 128] - - [417, 45.458] - - - [22656, 14977, 1, 128] - - [20, 45.731] - - - [17024, 4096, 1, 128] - - [299, 55.287] - - - [24960, 8961, 1, 128] - - [23, 45.169] - - - [5888, 1024, 1, 128] - - [294, 40.29] - - - [9344, 2048, 1, 128] - - [287, 48.467] - - - [11520, 1024, 1, 128] - - [419, 44.428] - - - [17024, 9217, 1, 128] - - [27, 45.351] - - - [10368, 6657, 1, 128] - - [300, 54.298] - - - [21632, 2048, 1, 128] - - [299, 53.754] - - - [26880, 2048, 1, 128] - - [289, 54.234] - - - [20736, 4096, 1, 128] - - [25, 45.481] - - - [26624, 8192, 1, 128] - - [23, 47.587] - - - [26752, 2048, 1, 128] - - [300, 53.795] - - - [24192, 8321, 1, 128] - - [35, 45.195] - - - [4736, 1024, 1, 128] - - [301, 36.939] - - - [27648, 8192, 1, 128] - - [25, 47.306] - - - [27392, 11521, 1, 128] - - [72, 44.51] - - - [27776, 4096, 1, 128] - - [33, 45.284] - - - [28672, 12801, 1, 128] - - [23, 46.648] - - - [13056, 512, 1, 128] - - [391, 39.481] - - - [25088, 2048, 1, 128] - - [289, 54.494] - - - [17408, 9601, 1, 128] - - [35, 46.144] - - - [5120, 3585, 1, 128] - - [417, 49.524] - - - [13824, 512, 1, 128] - - [291, 39.765] - - - [8576, 1024, 1, 128] - - [300, 42.205] - - - [16768, 4096, 1, 128] - - [419, 56.018] - - - [25728, 9729, 1, 128] - - [25, 45.32] - - - [27392, 512, 1, 128] - - [421, 43.933] - - - [13824, 128, 1, 128] - - [96, 32.393] - - - [27264, 1024, 1, 128] - - [33, 41.537] - - - [22272, 14465, 1, 128] - - [28, 46.223] - - - [19840, 2048, 1, 128] - - [300, 52.551] - - - [18176, 10497, 1, 128] - - [23, 45.819] - - - [4992, 3329, 1, 128] - - [291, 46.928] - - - [14976, 7169, 1, 128] - - [27, 44.407] - - - [10112, 512, 1, 128] - - [293, 37.671] - - - [24704, 128, 1, 128] - - [25, 33.286] - - - [16896, 128, 1, 128] - - [268, 25.928] - - - [10880, 7169, 1, 128] - - [419, 56.759] - - - [9600, 512, 1, 128] - - [291, 36.913] - - - [22528, 1024, 1, 128] - - [296, 48.758] - - - [27008, 128, 1, 128] - - [256, 30.056] - - - [4480, 2945, 1, 128] - - [420, 47.872] - - - [15872, 8065, 1, 128] - - [23, 45.84] - - - [28672, 128, 1, 128] - - [343, 30.697] - - - [9344, 128, 1, 128] - - [92, 29.6] - - - [15360, 2048, 1, 128] - - [299, 51.481] - - - [11392, 512, 1, 128] - - [290, 38.856] - - - [9216, 128, 1, 128] - - [152, 28.702] - - - [8192, 2048, 1, 128] - - [294, 45.162] - - - [14464, 1024, 1, 128] - - [290, 45.934] - - - [4096, 2433, 1, 128] - - [57, 39.249] - - - [6528, 2945, 1, 128] - - [419, 50.392] - - - [12672, 512, 1, 128] - - [298, 39.978] - - - [26624, 128, 1, 128] - - [315, 31.426] - - - [19712, 1024, 1, 128] - - [421, 46.679] - - - [4480, 2817, 1, 128] - - [295, 46.027] - - - [13440, 2048, 1, 128] - - [300, 50.073] - - - [256, 257, 1, 128] - - [147, 2.987] - - - [16000, 128, 1, 128] - - [293, 25.173] - - - [7552, 3969, 1, 128] - - [289, 49.954] - - - [12416, 2048, 1, 128] - - [289, 48.645] - - - [18432, 512, 1, 128] - - [348, 42.921] - - - [14464, 512, 1, 128] - - [291, 40.968] - - - [1280, 769, 1, 128] - - [105, 25.116] - - - [14976, 512, 1, 128] - - [300, 40.323] - - - [28032, 4096, 1, 128] - - [47, 45.397] - - - [27904, 128, 1, 128] - - [290, 30.141] - - - [20224, 12545, 1, 128] - - [25, 45.854] - - - [15872, 4096, 1, 128] - - [300, 54.551] - - - [3456, 1793, 1, 128] - - [301, 39.924] - - - [14336, 128, 1, 128] - - [256, 23.187] - - - [21248, 2048, 1, 128] - - [419, 53.28] - - - [23040, 1024, 1, 128] - - [390, 48.834] - - - [15232, 7425, 1, 128] - - [20, 44.631] - - - [14592, 512, 1, 128] - - [419, 40.608] - - - [22912, 15105, 1, 128] - - [27, 45.856] - - - [22528, 2048, 1, 128] - - [289, 54.224] - - - [3072, 1024, 1, 128] - - [313, 31.621] - - - [17536, 4096, 1, 128] - - [289, 55.835] - - - [384, 257, 1, 128] - - [104, 4.378] - - - [14464, 6657, 1, 128] - - [27, 44.75] - - - [20096, 1024, 1, 128] - - [299, 47.777] - - - [26880, 4096, 1, 128] - - [23, 45.686] - - - [18816, 2048, 1, 128] - - [300, 53.333] - - - [17152, 512, 1, 128] - - [419, 42.23] - - - [18432, 4096, 1, 128] - - [299, 56.697] - - - [10368, 2048, 1, 128] - - [300, 48.08] - - - [1408, 769, 1, 128] - - [96, 27.266] - - - [7168, 2048, 1, 128] - - [290, 47.029] - - - [17664, 128, 1, 128] - - [57, 26.138] - - - [1152, 513, 1, 128] - - [97, 19.814] - - - [7296, 3713, 1, 128] - - [301, 50.165] - - - [24064, 2048, 1, 128] - - [289, 53.783] - - - [8576, 2048, 1, 128] - - [419, 47.297] - - - [23168, 15489, 1, 128] - - [27, 45.702] - - - [14848, 7169, 1, 128] - - [25, 44.814] - - - [2432, 512, 1, 128] - - [155, 30.686] - - - [19712, 12033, 1, 128] - - [48, 44.208] - - - [25856, 4096, 1, 128] - - [23, 45.649] - - - [17152, 9345, 1, 128] - - [28, 45.62] - - - [3712, 128, 1, 128] - - [146, 16.913] - - - [22272, 128, 1, 128] - - [259, 29.31] - - - [25600, 9729, 1, 128] - - [25, 46.199] - - - [6016, 2433, 1, 128] - - [295, 47.366] - - - [12928, 128, 1, 128] - - [96, 32.348] - - - [25088, 8192, 1, 128] - - [62, 46.92] - - - [7040, 1024, 1, 128] - - [294, 39.855] - - - [4736, 3201, 1, 128] - - [421, 46.411] - - - [16000, 1024, 1, 128] - - [289, 46.75] - - - [1920, 512, 1, 128] - - [105, 25.535] - - - [8192, 1024, 1, 128] - - [295, 41.701] - - - [8448, 4865, 1, 128] - - [421, 52.116] - - - [11136, 7425, 1, 128] - - [35, 45.079] - - - [23296, 4096, 1, 128] - - [25, 45.463] - - - [27904, 2048, 1, 128] - - [390, 53.557] - - - [23552, 4096, 1, 128] - - [35, 46.138] - - - [24960, 2048, 1, 128] - - [289, 54.204] - - - [2816, 128, 1, 128] - - [109, 14.953] - - - [7424, 3841, 1, 128] - - [294, 50.11] - - - [20480, 128, 1, 128] - - [348, 27.718] - - - [18816, 11137, 1, 128] - - [23, 45.576] - - - [26496, 128, 1, 128] - - [390, 31.037] - - - [16896, 9217, 1, 128] - - [23, 45.874] - - - [23296, 512, 1, 128] - - [291, 44.724] - - - [8064, 2048, 1, 128] - - [357, 46.918] - - - [19968, 128, 1, 128] - - [392, 27.276] - - - [8320, 4737, 1, 128] - - [419, 52.283] - - - [27648, 1024, 1, 128] - - [390, 49.409] - - - [3712, 512, 1, 128] - - [418, 24.324] - - - [256, 128, 1, 128] - - [147, 1.5] - - - [3072, 1537, 1, 128] - - [300, 37.101] - - - [5504, 1024, 1, 128] - - [421, 38.741] - - - [20992, 2048, 1, 128] - - [421, 52.578] - - - [20480, 1024, 1, 128] - - [420, 45.739] - - - [20864, 128, 1, 128] - - [392, 28.413] - - - [28544, 12545, 1, 128] - - [35, 45.641] - - - [1152, 512, 1, 128] - - [147, 20.49] - - - [24320, 8321, 1, 128] - - [25, 45.636] - - - [2688, 512, 1, 128] - - [106, 29.178] - - - [27904, 8192, 1, 128] - - [48, 46.305] - - - [3840, 2177, 1, 128] - - [295, 43.787] - - - [25344, 128, 1, 128] - - [418, 31.292] - - - [13184, 512, 1, 128] - - [290, 39.676] - - - [7680, 512, 1, 128] - - [291, 33.382] - - - [11904, 2048, 1, 128] - - [300, 49.484] - - - [12544, 512, 1, 128] - - [291, 39.249] - - - [8448, 4737, 1, 128] - - [419, 51.681] - - - [28544, 128, 1, 128] - - [343, 30.479] - - - [21760, 14081, 1, 128] - - [25, 46.293] - - - [12800, 128, 1, 128] - - [103, 31.388] - - - [17664, 4096, 1, 128] - - [300, 55.788] - - - [2432, 1793, 1, 128] - - [293, 35.712] - - - [16384, 8577, 1, 128] - - [27, 37.574] - - - [28544, 512, 1, 128] - - [268, 44.82] - - - [28032, 12033, 1, 128] - - [35, 45.692] - - - [4864, 3329, 1, 128] - - [301, 48.742] - - - [12928, 5249, 1, 128] - - [419, 54.523] - - - [4736, 512, 1, 128] - - [291, 28.645] - - - [27264, 2048, 1, 128] - - [421, 53.774] - - - [19840, 12033, 1, 128] - - [25, 45.797] - - - [19584, 4096, 1, 128] - - [25, 45.273] - - - [21376, 4096, 1, 128] - - [35, 45.323] - - - [20352, 4096, 1, 128] - - [53, 45.354] - - - [6400, 2689, 1, 128] - - [300, 46.205] - - - [24704, 8192, 1, 128] - - [59, 45.87] - - - [22528, 14849, 1, 128] - - [25, 46.927] - - - [18304, 512, 1, 128] - - [290, 43.445] - - - [6656, 1024, 1, 128] - - [420, 40.141] - - - [13568, 4096, 1, 128] - - [289, 54.883] - - - [6016, 512, 1, 128] - - [392, 31.89] - - - [17664, 2048, 1, 128] - - [289, 52.614] - - - [17408, 512, 1, 128] - - [290, 43.151] - - - [24960, 4096, 1, 128] - - [35, 45.456] - - - [20608, 12801, 1, 128] - - [35, 45.592] - - - [27648, 11649, 1, 128] - - [23, 46.394] - - - [5760, 128, 1, 128] - - [91, 23.233] - - - [17792, 512, 1, 128] - - [418, 41.765] - - - [17664, 512, 1, 128] - - [290, 42.386] - - - [19968, 12161, 1, 128] - - [23, 46.15] - - - [19840, 512, 1, 128] - - [290, 43.613] - - - [12032, 4353, 1, 128] - - [300, 53.274] - - - [25984, 512, 1, 128] - - [421, 45.476] - - - [27648, 4096, 1, 128] - - [35, 46.345] - - - [10752, 7041, 1, 128] - - [289, 54.715] - - - [28544, 2048, 1, 128] - - [419, 52.391] - - - [7680, 2048, 1, 128] - - [421, 48.937] - - - [13184, 5377, 1, 128] - - [300, 53.603] - - - [6784, 3201, 1, 128] - - [417, 49.204] - - - [16384, 2048, 1, 128] - - [310, 44.152] - - - [22656, 1024, 1, 128] - - [357, 48.83] - - - [12800, 512, 1, 128] - - [291, 39.962] - - - [23936, 1024, 1, 128] - - [299, 49.414] - - - [15360, 1024, 1, 128] - - [419, 47.941] - - - [15488, 2048, 1, 128] - - [300, 51.296] - - - [11392, 1024, 1, 128] - - [390, 44.323] - - - [15744, 1024, 1, 128] - - [314, 46.034] - - - [9856, 2048, 1, 128] - - [421, 48.566] - - - [5888, 2305, 1, 128] - - [421, 45.387] - - - [10496, 512, 1, 128] - - [272, 37.101] - - - [1664, 1153, 1, 128] - - [392, 23.994] - - - [3456, 1024, 1, 128] - - [301, 33.032] - - - [20992, 13313, 1, 128] - - [27, 46.058] - - - [11904, 4096, 1, 128] - - [419, 53.88] - - - [13056, 1024, 1, 128] - - [298, 46.291] - - - [12800, 2048, 1, 128] - - [289, 50.352] - - - [12160, 512, 1, 128] - - [291, 38.703] - - - [5760, 2049, 1, 128] - - [287, 45.291] - - - [11392, 128, 1, 128] - - [96, 30.138] - - - [5632, 128, 1, 128] - - [91, 22.716] - - - [11520, 2048, 1, 128] - - [348, 49.716] - - - [11648, 2048, 1, 128] - - [289, 50.277] - - - [28544, 8192, 1, 128] - - [35, 46.393] - - - [22912, 1024, 1, 128] - - [419, 47.99] - - - [10752, 7169, 1, 128] - - [299, 54.618] - - - [8320, 128, 1, 128] - - [91, 26.935] - - - [23808, 1024, 1, 128] - - [390, 48.631] - - - [25984, 8192, 1, 128] - - [22, 46.384] - - - [22656, 2048, 1, 128] - - [299, 53.898] - - - [7296, 1024, 1, 128] - - [294, 39.746] - - - [28032, 512, 1, 128] - - [291, 47.082] - - - [22400, 2048, 1, 128] - - [289, 52.737] - - - [22144, 512, 1, 128] - - [293, 45.065] - - - [13312, 4096, 1, 128] - - [289, 55.602] - - - [10240, 2048, 1, 128] - - [300, 49.823] - - - [12672, 128, 1, 128] - - [96, 32.477] - - - [10752, 2048, 1, 128] - - [357, 48.851] - - - [1152, 128, 1, 128] - - [101, 6.643] - - - [13696, 5889, 1, 128] - - [47, 42.592] - - - [9216, 1024, 1, 128] - - [298, 43.271] - - - [17152, 128, 1, 128] - - [392, 26.61] - - - [24320, 2048, 1, 128] - - [298, 52.72] - - - [16512, 8705, 1, 128] - - [25, 44.284] - - - [3072, 1409, 1, 128] - - [420, 34.33] - - - [1024, 128, 1, 128] - - [98, 5.905] - - - [22400, 14593, 1, 128] - - [20, 45.652] - - - [4096, 512, 1, 128] - - [388, 25.77] - - - [4992, 128, 1, 128] - - [95, 21.548] - - - [9472, 5889, 1, 128] - - [300, 53.662] - - - [9472, 5761, 1, 128] - - [300, 53.673] - - - [27136, 1024, 1, 128] - - [390, 50.181] - - - [6528, 1024, 1, 128] - - [421, 40.894] - - - [25472, 1024, 1, 128] - - [300, 49.036] - - - [5120, 512, 1, 128] - - [392, 28.709] - - - [5504, 512, 1, 128] - - [293, 29.114] - - - [21120, 13441, 1, 128] - - [48, 44.906] - - - [4352, 128, 1, 128] - - [95, 19.236] - - - [8832, 5249, 1, 128] - - [419, 53.42] - - - [1536, 1025, 1, 128] - - [277, 21.25] - - - [11520, 512, 1, 128] - - [293, 38.624] - - - [5632, 2048, 1, 128] - - [384, 43.663] - - - [7424, 128, 1, 128] - - [92, 24.684] - - - [18432, 128, 1, 128] - - [291, 27.214] - - - [12672, 2048, 1, 128] - - [300, 50.265] - - - [14208, 128, 1, 128] - - [418, 23.737] - - - [15360, 7553, 1, 128] - - [25, 45.889] - - - [26496, 1024, 1, 128] - - [348, 49.645] - - - [27136, 128, 1, 128] - - [343, 30.147] - - - [12032, 2048, 1, 128] - - [421, 49.481] - - - [11648, 1024, 1, 128] - - [421, 45.223] - - - [11776, 512, 1, 128] - - [291, 37.753] - - - [1024, 512, 1, 128] - - [415, 9.915] - - - [11264, 7681, 1, 128] - - [25, 45.447] - - - [19456, 11777, 1, 128] - - [35, 46.512] - - - [14080, 4096, 1, 128] - - [421, 54.196] - - - [7040, 3329, 1, 128] - - [287, 50.672] - - - [27392, 4096, 1, 128] - - [74, 44.661] - - - [14720, 7041, 1, 128] - - [27, 44.828] - - - [19584, 1024, 1, 128] - - [300, 48.593] - - - [21376, 13569, 1, 128] - - [27, 45.75] - - - [20480, 12801, 1, 128] - - [27, 47.065] - - - [21248, 128, 1, 128] - - [422, 28.485] - - - [9728, 1024, 1, 128] - - [299, 43.513] - - - [18688, 10881, 1, 128] - - [27, 45.714] - - - [21120, 13313, 1, 128] - - [49, 44.673] - - - [20096, 2048, 1, 128] - - [419, 53.016] - - - [16640, 4096, 1, 128] - - [300, 55.612] - - - [28160, 12161, 1, 128] - - [48, 45.981] - - - [640, 129, 1, 128] - - [147, 3.662] - - - [28672, 512, 1, 128] - - [418, 44.181] - - - [12416, 4096, 1, 128] - - [289, 53.69] - - - [25344, 9473, 1, 128] - - [55, 44.899] - - - [18304, 1024, 1, 128] - - [289, 46.511] - - - [25600, 4096, 1, 128] - - [35, 46.394] - - - [22272, 512, 1, 128] - - [268, 44.963] - - - [21504, 13825, 1, 128] - - [36, 46.611] - - - [4736, 128, 1, 128] - - [343, 10.888] - - - [26496, 10625, 1, 128] - - [62, 45.111] - - - [7040, 512, 1, 128] - - [345, 32.309] - - - [14336, 4096, 1, 128] - - [289, 55.544] - - - [9216, 512, 1, 128] - - [293, 36.451] - - - [1280, 641, 1, 128] - - [90, 24.901] - - - [16768, 8961, 1, 128] - - [27, 45.569] - - - [18944, 11137, 1, 128] - - [48, 46.171] - - - [21504, 2048, 1, 128] - - [299, 54.423] - - - [21888, 1024, 1, 128] - - [419, 48.193] - - - [11264, 512, 1, 128] - - [268, 38.548] - - - [27776, 8192, 1, 128] - - [23, 46.27] - - - [10368, 6785, 1, 128] - - [300, 54.417] - - - [18432, 10753, 1, 128] - - [35, 46.711] - - - [19968, 2048, 1, 128] - - [390, 53.078] - - - [16640, 512, 1, 128] - - [264, 39.76] - - - [24576, 8577, 1, 128] - - [36, 42.526] - - - [28672, 2048, 1, 128] - - [357, 51.778] - - - [11136, 128, 1, 128] - - [90, 29.674] - - - [12288, 4609, 1, 128] - - [287, 52.392] - - - [14848, 1024, 1, 128] - - [419, 44.415] - - - [14848, 128, 1, 128] - - [418, 23.494] - - - [7424, 1024, 1, 128] - - [290, 40.154] - - - [2560, 1024, 1, 128] - - [291, 28.336] - - - [6400, 128, 1, 128] - - [105, 25.674] - - - [15488, 7809, 1, 128] - - [27, 45.064] - - - [17920, 2048, 1, 128] - - [348, 51.598] - - - [5760, 512, 1, 128] - - [348, 31.007] - - - [16640, 1024, 1, 128] - - [421, 46.005] - - - [28160, 2048, 1, 128] - - [390, 54.242] - - - [5504, 3969, 1, 128] - - [419, 49.587] - - - [11776, 1024, 1, 128] - - [391, 42.472] - - - [18816, 128, 1, 128] - - [392, 27.693] - - - [27904, 12033, 1, 128] - - [48, 45.586] - - - [11520, 7937, 1, 128] - - [25, 44.746] - - - [18944, 11265, 1, 128] - - [62, 46.055] - - - [5376, 1024, 1, 128] - - [295, 38.415] - - - [12032, 4225, 1, 128] - - [421, 53.132] - - - [5376, 128, 1, 128] - - [90, 22.045] - - - [9856, 1024, 1, 128] - - [300, 43.608] - - - [26752, 10881, 1, 128] - - [25, 45.393] - - - [20352, 128, 1, 128] - - [309, 27.369] - - - [14464, 128, 1, 128] - - [256, 23.428] - - - [1024, 385, 1, 128] - - [104, 15.36] - - - [3840, 128, 1, 128] - - [97, 17.389] - - - [24192, 128, 1, 128] - - [58, 32.783] - - - [28544, 12673, 1, 128] - - [25, 45.789] - - - [1664, 128, 1, 128] - - [97, 9.305] - - - [26752, 8192, 1, 128] - - [20, 46.339] - - - [16896, 1024, 1, 128] - - [298, 46.454] - - - [9728, 128, 1, 128] - - [91, 30.297] - - - [11264, 2048, 1, 128] - - [300, 48.948] - - - [11392, 2048, 1, 128] - - [289, 50.785] - - - [20224, 2048, 1, 128] - - [289, 53.57] - - - [26880, 1024, 1, 128] - - [299, 50.239] - - - [15104, 512, 1, 128] - - [293, 41.305] - - - [26368, 2048, 1, 128] - - [421, 53.796] - - - [6784, 3073, 1, 128] - - [300, 47.411] - - - [23168, 128, 1, 128] - - [307, 30.032] - - - [8448, 1024, 1, 128] - - [293, 40.097] - - - [16896, 9089, 1, 128] - - [22, 46.11] - - - [17536, 128, 1, 128] - - [290, 26.848] - - - [22912, 512, 1, 128] - - [299, 43.864] - - - [28032, 128, 1, 128] - - [290, 29.876] - - - [19584, 512, 1, 128] - - [290, 44.485] - - - [27136, 11265, 1, 128] - - [22, 46.115] - - - [4992, 512, 1, 128] - - [256, 28.539] - - - [8448, 128, 1, 128] - - [151, 27.47] - - - [27648, 128, 1, 128] - - [418, 30.544] - - - [16640, 2048, 1, 128] - - [300, 51.631] - - - [26752, 10753, 1, 128] - - [36, 45.477] - - - [2944, 1281, 1, 128] - - [418, 33.347] - - - [5376, 3841, 1, 128] - - [417, 48.686] - - - [10496, 6913, 1, 128] - - [419, 53.876] - - - [17024, 512, 1, 128] - - [290, 42.615] - - - [11008, 7297, 1, 128] - - [88, 44.623] - - - [14080, 128, 1, 128] - - [272, 23.005] - - - [5888, 512, 1, 128] - - [291, 31.076] - - - [19200, 128, 1, 128] - - [291, 27.619] - - - [14208, 6529, 1, 128] - - [25, 44.534] - - - [22912, 4096, 1, 128] - - [32, 45.339] - - - [14336, 2048, 1, 128] - - [300, 51.453] - - - [17792, 128, 1, 128] - - [266, 27.004] - - - [22656, 14849, 1, 128] - - [20, 45.629] - - - [19712, 512, 1, 128] - - [299, 41.691] - - - [5248, 1024, 1, 128] - - [295, 37.721] - - - [3712, 2049, 1, 128] - - [291, 41.231] - - - [24448, 8449, 1, 128] - - [25, 45.255] - - - [8192, 512, 1, 128] - - [392, 33.366] - - - [25472, 4096, 1, 128] - - [23, 45.651] - - - [25088, 512, 1, 128] - - [390, 45.309] - - - [23168, 1024, 1, 128] - - [300, 49.417] - - - [24320, 8192, 1, 128] - - [55, 46.734] - - - [24192, 8192, 1, 128] - - [23, 46.474] - - - [2176, 512, 1, 128] - - [103, 28.055] - - - [4992, 3457, 1, 128] - - [287, 48.707] - - - [896, 257, 1, 128] - - [156, 9.692] - - - [28288, 1024, 1, 128] - - [299, 49.598] - - - [20864, 1024, 1, 128] - - [289, 48.463] - - - [18432, 2048, 1, 128] - - [289, 53.385] - - - [17280, 9601, 1, 128] - - [27, 45.382] - - - [18944, 4096, 1, 128] - - [300, 55.19] - - - [13440, 128, 1, 128] - - [152, 32.958] - - - [7424, 2048, 1, 128] - - [290, 47.289] - - - [768, 128, 1, 128] - - [98, 4.429] - - - [16128, 512, 1, 128] - - [264, 40.311] - - - [28288, 12289, 1, 128] - - [27, 45.397] - - - [23552, 128, 1, 128] - - [51, 31.909] - - - [24832, 8192, 1, 128] - - [57, 46.933] - - - [10240, 1024, 1, 128] - - [291, 42.863] - - - [8960, 2048, 1, 128] - - [390, 47.301] - - - [17664, 9985, 1, 128] - - [27, 45.395] - - - [25088, 4096, 1, 128] - - [65, 45.937] - - - [7552, 2048, 1, 128] - - [300, 48.022] - - - [15104, 7297, 1, 128] - - [23, 45.225] - - - [7168, 1024, 1, 128] - - [420, 41.527] - - - [26112, 8192, 1, 128] - - [55, 47.055] - - - [24192, 1024, 1, 128] - - [419, 48.481] - - - [22912, 2048, 1, 128] - - [421, 53.137] - - - [10368, 512, 1, 128] - - [293, 38.034] - - - [22528, 4096, 1, 128] - - [27, 46.425] - - - [6528, 128, 1, 128] - - [91, 26.046] - - - [26752, 4096, 1, 128] - - [36, 45.179] - - - [2816, 512, 1, 128] - - [90, 30.344] - - - [22016, 14209, 1, 128] - - [22, 46.54] - - - [8832, 1024, 1, 128] - - [419, 42.667] - - - [16384, 128, 1, 128] - - [293, 24.871] - - - [5120, 1024, 1, 128] - - [392, 36.945] - - - [24832, 8833, 1, 128] - - [48, 45.868] - - - [11520, 128, 1, 128] - - [103, 30.26] - - - [24960, 512, 1, 128] - - [289, 45.76] - - - [27520, 2048, 1, 128] - - [419, 53.844] - - - [22272, 14593, 1, 128] - - [48, 45.954] - - - [2048, 128, 1, 128] - - [98, 11.281] - - - [2176, 1537, 1, 128] - - [419, 31.989] - - - [10496, 1024, 1, 128] - - [291, 42.845] - - - [12160, 4353, 1, 128] - - [421, 53.159] - - - [6144, 1024, 1, 128] - - [289, 38.808] - - - [26752, 1024, 1, 128] - - [289, 48.969] - - - [17280, 4096, 1, 128] - - [299, 55.614] - - - [16896, 512, 1, 128] - - [291, 41.476] - - - [4480, 128, 1, 128] - - [98, 19.801] - - - [18944, 128, 1, 128] - - [293, 27.456] - - - [9600, 2048, 1, 128] - - [306, 47.391] - - - [19456, 1024, 1, 128] - - [35, 41.595] - - - [9984, 2048, 1, 128] - - [23, 42.255] - - - [25216, 9217, 1, 128] - - [20, 45.153] - - - [19968, 1024, 1, 128] - - [390, 48.69] - - - [13952, 2048, 1, 128] - - [421, 49.255] - - - [10496, 2048, 1, 128] - - [417, 48.899] - - - [12672, 1024, 1, 128] - - [357, 45.114] - - - [19072, 11393, 1, 128] - - [36, 45.652] - - - [11008, 2048, 1, 128] - - [295, 48.437] - - - [27520, 11649, 1, 128] - - [25, 45.664] - - - [10880, 512, 1, 128] - - [290, 39.149] - - - [14592, 6785, 1, 128] - - [57, 44.037] - - - [7424, 512, 1, 128] - - [418, 33.623] - - - [13056, 5249, 1, 128] - - [300, 53.815] - - - [23296, 15489, 1, 128] - - [23, 46.155] - - - [28416, 8192, 1, 128] - - [25, 46.394] - - - [11392, 7681, 1, 128] - - [25, 44.48] - - - [18048, 1024, 1, 128] - - [421, 47.706] - - - [15616, 7809, 1, 128] - - [27, 45.364] - - - [128, 128, 1, 128] - - [95, 0.75] - - - [24704, 512, 1, 128] - - [300, 41.874] - - - [7680, 4097, 1, 128] - - [421, 49.7] - - - [16640, 8961, 1, 128] - - [23, 45.941] - - - [18944, 1024, 1, 128] - - [421, 48.784] - - - [12928, 2048, 1, 128] - - [299, 50.04] - - - [22272, 2048, 1, 128] - - [421, 53.445] - - - [27904, 11905, 1, 128] - - [49, 45.645] - - - [26240, 2048, 1, 128] - - [290, 52.67] - - - [9728, 6017, 1, 128] - - [299, 53.768] - - - [20736, 1024, 1, 128] - - [300, 46.478] - - - [3456, 1921, 1, 128] - - [287, 40.106] - - - [8064, 512, 1, 128] - - [300, 33.329] - - - [4224, 1024, 1, 128] - - [392, 34.247] - - - [25984, 10113, 1, 128] - - [28, 45.563] - - - [13696, 6017, 1, 128] - - [79, 42.514] - - - [27520, 8192, 1, 128] - - [35, 46.441] - - - [18944, 512, 1, 128] - - [293, 44.207] - - - [6272, 128, 1, 128] - - [92, 25.437] - - - [27264, 4096, 1, 128] - - [47, 45.21] - - - [1792, 1153, 1, 128] - - [284, 25.496] - - - [17536, 9729, 1, 128] - - [35, 45.307] - - - [13184, 5505, 1, 128] - - [417, 53.155] - - - [2944, 128, 1, 128] - - [97, 15.411] - - - [25344, 512, 1, 128] - - [293, 45.588] - - - [23040, 15361, 1, 128] - - [35, 46.101] - - - [8704, 512, 1, 128] - - [418, 35.53] - - - [20864, 13057, 1, 128] - - [20, 45.465] - - - [19328, 4096, 1, 128] - - [35, 45.15] - - - [28288, 8192, 1, 128] - - [36, 46.469] - - - [10112, 1024, 1, 128] - - [300, 43.778] - - - [17536, 2048, 1, 128] - - [300, 52.868] - - - [7552, 128, 1, 128] - - [90, 24.997] - - - [15616, 7937, 1, 128] - - [27, 45.393] - - - [23040, 512, 1, 128] - - [300, 42.981] - - - [25984, 2048, 1, 128] - - [421, 53.693] - - - [14720, 128, 1, 128] - - [392, 23.464] - - - [23424, 1024, 1, 128] - - [421, 47.18] - - - [1920, 1281, 1, 128] - - [256, 26.965] - - - [27136, 2048, 1, 128] - - [299, 53.897] - - - [28800, 8192, 1, 128] - - [27, 46.396] - - - [15488, 128, 1, 128] - - [259, 24.417] - - - [28800, 12929, 1, 128] - - [35, 45.572] - - - [21888, 14081, 1, 128] - - [74, 41.962] - - - [25600, 1024, 1, 128] - - [289, 49.147] - - - [21632, 1024, 1, 128] - - [291, 47.781] - - - [24448, 1024, 1, 128] - - [357, 49.673] - - - [4352, 2689, 1, 128] - - [420, 46.742] - - - [20480, 512, 1, 128] - - [300, 41.797] - - - [7296, 128, 1, 128] - - [96, 24.703] - - - [4992, 1024, 1, 128] - - [291, 37.418] - - - [27264, 11393, 1, 128] - - [27, 45.662] - - - [26752, 128, 1, 128] - - [259, 29.375] - - - [24960, 1024, 1, 128] - - [420, 48.573] - - - [21504, 512, 1, 128] - - [300, 43.872] - - - [6272, 2561, 1, 128] - - [294, 46.917] - - - [25088, 9089, 1, 128] - - [45, 45.869] - - - [20864, 512, 1, 128] - - [272, 42.939] - - - [4224, 2561, 1, 128] - - [294, 44.953] - - - [15744, 8065, 1, 128] - - [23, 45.33] - - - [21632, 128, 1, 128] - - [348, 28.123] - - - [15104, 4096, 1, 128] - - [289, 55.822] - - - [20352, 512, 1, 128] - - [290, 43.151] - - - [25472, 9601, 1, 128] - - [27, 45.504] - - - [27904, 512, 1, 128] - - [291, 45.575] - - - [19968, 512, 1, 128] - - [291, 43.505] - - - [5760, 1024, 1, 128] - - [294, 38.883] - - - [28416, 12545, 1, 128] - - [35, 45.589] - - - [16512, 8833, 1, 128] - - [25, 44.373] - - - [6016, 128, 1, 128] - - [103, 24.399] - - - [13056, 4096, 1, 128] - - [421, 55.215] - - - [19968, 12289, 1, 128] - - [25, 45.801] - - - [7424, 3713, 1, 128] - - [300, 48.671] - - - [28800, 128, 1, 128] - - [257, 30.38] - - - [512, 512, 1, 128] - - [97, 10.955] - - - [24832, 2048, 1, 128] - - [421, 53.162] - - - [20736, 128, 1, 128] - - [348, 28.276] - - - [26368, 512, 1, 128] - - [268, 45.956] - - - [26496, 8192, 1, 128] - - [55, 45.86] - - - [13824, 4096, 1, 128] - - [421, 54.35] - - - [27264, 128, 1, 128] - - [418, 30.027] - - - [21760, 1024, 1, 128] - - [419, 47.729] - - - [2432, 1921, 1, 128] - - [418, 36.475] - - - [27136, 8192, 1, 128] - - [45, 47.092] - - - [6784, 2048, 1, 128] - - [300, 45.319] - - - [11264, 128, 1, 128] - - [157, 28.769] - - - [7552, 512, 1, 128] - - [291, 33.662] - - - [19328, 11649, 1, 128] - - [25, 45.711] - - - [17152, 2048, 1, 128] - - [421, 53.212] - - - [23808, 16129, 1, 128] - - [25, 46.207] - - - [20224, 12417, 1, 128] - - [20, 45.968] - - - [27904, 1024, 1, 128] - - [357, 48.848] - - - [3456, 512, 1, 128] - - [259, 23.086] - - - [13312, 512, 1, 128] - - [291, 40.039] - - - [26368, 4096, 1, 128] - - [23, 45.563] - - - [23296, 15617, 1, 128] - - [27, 46.156] - - - [26112, 10241, 1, 128] - - [22, 45.807] - - - [26240, 512, 1, 128] - - [293, 44.828] - - - [4352, 1024, 1, 128] - - [290, 35.452] - - - [10624, 2048, 1, 128] - - [291, 50.014] - - - [23808, 16001, 1, 128] - - [23, 46.117] - - - [17536, 9857, 1, 128] - - [23, 45.598] - - - [23936, 4096, 1, 128] - - [27, 45.442] - - - [1408, 128, 1, 128] - - [147, 7.933] - - - [14848, 512, 1, 128] - - [290, 41.071] - - - [8704, 4993, 1, 128] - - [289, 53.749] - - - [15104, 2048, 1, 128] - - [421, 51.309] - - - [2560, 512, 1, 128] - - [103, 30.724] - - - [27264, 8192, 1, 128] - - [27, 46.344] - - - [23808, 4096, 1, 128] - - [36, 45.544] - - - [14080, 6273, 1, 128] - - [33, 43.219] - - - [10112, 6529, 1, 128] - - [299, 53.992] - - - [27648, 512, 1, 128] - - [421, 45.529] - - - [20992, 128, 1, 128] - - [259, 27.974] - - - [15104, 128, 1, 128] - - [418, 23.632] - - - [7808, 128, 1, 128] - - [105, 25.729] - - - [3584, 1024, 1, 128] - - [295, 34.519] - - - [15232, 512, 1, 128] - - [421, 40.624] - - - [21376, 13697, 1, 128] - - [25, 45.838] - - - [11392, 7809, 1, 128] - - [52, 44.52] - - - [11904, 1024, 1, 128] - - [421, 45.301] - - - [28800, 2048, 1, 128] - - [299, 52.811] - - - [8960, 512, 1, 128] - - [264, 34.586] - - - [19456, 11649, 1, 128] - - [27, 46.451] - - - [11904, 128, 1, 128] - - [96, 31.381] - - - [18560, 512, 1, 128] - - [290, 43.635] - - - [6656, 128, 1, 128] - - [92, 26.414] - - - [17792, 2048, 1, 128] - - [300, 52.506] - - - [21632, 4096, 1, 128] - - [35, 45.217] - - - [25728, 4096, 1, 128] - - [47, 45.444] - - - [18048, 10241, 1, 128] - - [49, 44.043] - - - [1792, 1281, 1, 128] - - [294, 27.052] - - - [512, 385, 1, 128] - - [97, 8.547] - - - [26112, 512, 1, 128] - - [291, 45.537] - - - [16128, 1024, 1, 128] - - [421, 47.485] - - - [4480, 1024, 1, 128] - - [420, 35.885] - - - [14720, 4096, 1, 128] - - [300, 54.349] - - - [23552, 2048, 1, 128] - - [289, 54.612] - - - [22528, 512, 1, 128] - - [289, 44.519] - - - [22912, 128, 1, 128] - - [348, 29.984] - - - [25344, 1024, 1, 128] - - [422, 47.791] - - - [24064, 16257, 1, 128] - - [48, 46.571] - - - [9088, 5377, 1, 128] - - [421, 53.106] - - - [27776, 128, 1, 128] - - [392, 30.529] - - - [15616, 512, 1, 128] - - [293, 41.809] - - - [13568, 128, 1, 128] - - [152, 32.943] - - - [15488, 7681, 1, 128] - - [23, 44.725] - - - [20096, 512, 1, 128] - - [293, 43.407] - - - [24832, 4096, 1, 128] - - [25, 45.784] - - - [28800, 4096, 1, 128] - - [32, 45.168] - - - [11904, 4225, 1, 128] - - [300, 51.9] - - - [3968, 1024, 1, 128] - - [291, 33.737] - - - [6400, 2817, 1, 128] - - [301, 46.794] - - - [24576, 4096, 1, 128] - - [27, 42.846] - - - [9088, 128, 1, 128] - - [96, 29.292] - - - [17152, 4096, 1, 128] - - [421, 55.441] - - - [22528, 14721, 1, 128] - - [35, 46.947] - - - [27392, 2048, 1, 128] - - [421, 54.187] - - - [8832, 512, 1, 128] - - [418, 35.938] - - - [8960, 5249, 1, 128] - - [419, 53.972] - - - [3200, 1024, 1, 128] - - [298, 32.081] - - - [4736, 3073, 1, 128] - - [419, 47.595] - - - [28032, 2048, 1, 128] - - [299, 54.453] - - - [14592, 2048, 1, 128] - - [300, 50.398] - - - [13440, 1024, 1, 128] - - [421, 45.492] - - - [14464, 2048, 1, 128] - - [390, 51.099] - - - [6912, 2048, 1, 128] - - [420, 46.963] - - - [19584, 2048, 1, 128] - - [289, 52.995] - - - [17920, 128, 1, 128] - - [272, 27.449] - - - [19584, 11777, 1, 128] - - [35, 45.594] - - - [23936, 16129, 1, 128] - - [27, 46.1] - - - [10496, 6785, 1, 128] - - [300, 54.676] - - - [27648, 2048, 1, 128] - - [299, 54.641] - - - [23808, 128, 1, 128] - - [58, 32.621] - - - [20864, 2048, 1, 128] - - [419, 52.346] - - - [9088, 512, 1, 128] - - [392, 35.775] - - - [3584, 512, 1, 128] - - [418, 24.309] - - - [8576, 4993, 1, 128] - - [299, 52.708] - - - [3328, 1024, 1, 128] - - [294, 32.873] - - - [20608, 2048, 1, 128] - - [289, 52.689] - - - [23552, 15745, 1, 128] - - [35, 46.73] - - - [23424, 15617, 1, 128] - - [43, 43.609] - - - [21120, 512, 1, 128] - - [272, 42.934] - - - [6656, 512, 1, 128] - - [264, 31.387] - - - [12544, 128, 1, 128] - - [96, 31.709] - - - [24448, 8577, 1, 128] - - [35, 45.437] - - - [9984, 512, 1, 128] - - [392, 35.863] - - - [18304, 4096, 1, 128] - - [289, 56.337] - - - [17920, 512, 1, 128] - - [293, 42.602] - - - [12160, 4096, 1, 128] - - [300, 54.042] - - - [3968, 2433, 1, 128] - - [301, 44.633] - - - [27008, 4096, 1, 128] - - [62, 45.118] - - - [22272, 1024, 1, 128] - - [289, 49.5] - - - [14336, 512, 1, 128] - - [290, 40.961] - - - [18560, 10753, 1, 128] - - [35, 45.476] - - - [6272, 2048, 1, 128] - - [421, 47.381] - - - [12800, 1024, 1, 128] - - [421, 46.273] - - - [9600, 5889, 1, 128] - - [390, 53.627] - - - [13056, 128, 1, 128] - - [152, 32.667] - - - [7296, 2048, 1, 128] - - [290, 45.13] - - - [21376, 512, 1, 128] - - [300, 44.412] - - - [11904, 512, 1, 128] - - [291, 37.938] - - - [6400, 1024, 1, 128] - - [390, 39.795] - - - [27008, 1024, 1, 128] - - [289, 50.493] - - - [22400, 14721, 1, 128] - - [20, 45.543] - - - [6272, 1024, 1, 128] - - [419, 38.681] - - - [17408, 128, 1, 128] - - [392, 26.506] - - - [26624, 10625, 1, 128] - - [27, 46.721] - - - [22400, 1024, 1, 128] - - [289, 48.647] - - - [18304, 10625, 1, 128] - - [35, 45.509] - - - [15872, 1024, 1, 128] - - [300, 46.737] - - - [21120, 128, 1, 128] - - [392, 27.855] - - - [22784, 4096, 1, 128] - - [55, 45.163] - - - [25728, 9857, 1, 128] - - [27, 45.466] - - - [16256, 1024, 1, 128] - - [300, 46.468] - - - [18560, 4096, 1, 128] - - [421, 55.401] - - - [7936, 4225, 1, 128] - - [301, 50.567] - - - [7680, 3969, 1, 128] - - [298, 51.002] - - - [9472, 2048, 1, 128] - - [421, 50.417] - - - [28160, 128, 1, 128] - - [266, 30.452] - - - [18816, 512, 1, 128] - - [293, 43.924] - - - [9856, 512, 1, 128] - - [290, 37.565] - - - [17664, 9857, 1, 128] - - [36, 45.257] - - - [27392, 128, 1, 128] - - [129, 29.082] - - - [24448, 2048, 1, 128] - - [419, 53.885] - - - [7808, 512, 1, 128] - - [290, 34.279] - - - [13952, 512, 1, 128] - - [290, 40.502] - - - [24576, 512, 1, 128] - - [23, 39.435] - - - [27520, 128, 1, 128] - - [293, 30.163] - - - [26496, 512, 1, 128] - - [291, 45.707] - - - [8576, 512, 1, 128] - - [421, 34.358] - - - [11648, 512, 1, 128] - - [293, 38.332] - - - [17408, 2048, 1, 128] - - [299, 53.089] - - - [17920, 10241, 1, 128] - - [45, 45.751] - - - [16384, 1024, 1, 128] - - [302, 40.303] - - - [6016, 2048, 1, 128] - - [345, 44.904] - - - [9728, 512, 1, 128] - - [293, 37.319] - - - [19712, 128, 1, 128] - - [272, 27.867] - - - [26112, 1024, 1, 128] - - [421, 49.07] - - - [16768, 128, 1, 128] - - [418, 26.095] - - - [8960, 1024, 1, 128] - - [296, 42.348] - - - [6784, 128, 1, 128] - - [105, 26.778] - - - [12800, 4993, 1, 128] - - [289, 53.518] - - - [6144, 2561, 1, 128] - - [420, 46.693] - - - [26880, 10881, 1, 128] - - [36, 45.591] - - - [12928, 1024, 1, 128] - - [300, 45.611] - - - [7040, 3457, 1, 128] - - [300, 49.225] - - - [15744, 4096, 1, 128] - - [300, 55.578] - - - [20096, 4096, 1, 128] - - [53, 44.898] - - - [21760, 128, 1, 128] - - [418, 28.794] - - - [7936, 2048, 1, 128] - - [301, 46.346] - - - [24448, 8192, 1, 128] - - [27, 46.582] - - - [21120, 2048, 1, 128] - - [421, 52.781] - - - [12160, 1024, 1, 128] - - [421, 45.569] - - - [7168, 3457, 1, 128] - - [300, 48.544] - - - [15232, 7553, 1, 128] - - [36, 44.833] - - - [26624, 1024, 1, 128] - - [299, 49.844] - - - [25344, 2048, 1, 128] - - [421, 53.205] - - - [12544, 4865, 1, 128] - - [419, 52.992] - - - [21120, 4096, 1, 128] - - [57, 44.469] - - - [20224, 128, 1, 128] - - [298, 27.244] - - - [14592, 4096, 1, 128] - - [300, 54.595] - - - [16256, 8577, 1, 128] - - [27, 45.045] - - - [24192, 4096, 1, 128] - - [27, 45.486] - - - [21248, 1024, 1, 128] - - [390, 48.848] - - - [25216, 1024, 1, 128] - - [421, 48.29] - - - [5888, 2177, 1, 128] - - [294, 45.71] - - - [21504, 1024, 1, 128] - - [289, 48.179] - - - [17536, 1024, 1, 128] - - [421, 48.694] - - - [9728, 2048, 1, 128] - - [390, 49.6] - - - [13952, 6273, 1, 128] - - [27, 44.621] - - - [28800, 512, 1, 128] - - [291, 44.263] - - - [2304, 1793, 1, 128] - - [418, 34.085] - - - [12416, 128, 1, 128] - - [92, 31.601] - - - [20224, 1024, 1, 128] - - [419, 48.17] - - - [22144, 128, 1, 128] - - [391, 28.741] - - - [22784, 1024, 1, 128] - - [299, 49.832] - - - [27136, 4096, 1, 128] - - [47, 46.044] - - - [27264, 512, 1, 128] - - [290, 44.503] - - - [26240, 10241, 1, 128] - - [23, 41.87] - - - [27904, 4096, 1, 128] - - [62, 45.283] - - - [21504, 128, 1, 128] - - [298, 29.247] - - - [3712, 2177, 1, 128] - - [294, 42.492] - - - [18432, 1024, 1, 128] - - [298, 48.132] - - - [28672, 4096, 1, 128] - - [25, 46.523] - - - [25344, 4096, 1, 128] - - [65, 45.027] - - - [26880, 512, 1, 128] - - [293, 46.558] - - - [21888, 2048, 1, 128] - - [419, 52.801] - - - [1792, 128, 1, 128] - - [95, 10.02] - - - [6016, 1024, 1, 128] - - [421, 38.356] - - - [15104, 7425, 1, 128] - - [25, 45.481] - - - [22016, 2048, 1, 128] - - [299, 53.377] - - - [13952, 4096, 1, 128] - - [421, 54.675] - - - [20992, 4096, 1, 128] - - [47, 45.563] - - - [8064, 4481, 1, 128] - - [295, 52.133] - - - [12672, 4096, 1, 128] - - [287, 53.125] - - - [20096, 12289, 1, 128] - - [23, 45.138] - - - [14848, 2048, 1, 128] - - [419, 50.311] - - - [23168, 512, 1, 128] - - [277, 42.669] - - - [7680, 128, 1, 128] - - [90, 25.42] - - - [13312, 1024, 1, 128] - - [348, 45.292] - - - [10624, 1024, 1, 128] - - [419, 44.997] - - - [3840, 512, 1, 128] - - [256, 25.259] - - - [22144, 14337, 1, 128] - - [35, 45.549] - - - [3200, 128, 1, 128] - - [156, 16.517] - - - [25472, 9473, 1, 128] - - [35, 45.643] - - - [16768, 9089, 1, 128] - - [23, 45.453] - - - [12288, 2048, 1, 128] - - [392, 48.378] - - - [20608, 512, 1, 128] - - [300, 41.57] - - - [2816, 1024, 1, 128] - - [290, 30.563] - - - [7552, 1024, 1, 128] - - [419, 40.716] - - - [5120, 3457, 1, 128] - - [301, 47.908] - - - [25216, 2048, 1, 128] - - [300, 53.796] - - - [12672, 4865, 1, 128] - - [300, 52.332] - - - [10880, 2048, 1, 128] - - [300, 50.236] - - - [18176, 512, 1, 128] - - [348, 42.69] - - - [8320, 4609, 1, 128] - - [419, 52.236] - - - [16000, 4096, 1, 128] - - [419, 55.979] - - - [22144, 2048, 1, 128] - - [58, 43.669] - - - [22784, 512, 1, 128] - - [272, 45.034] - - - [4096, 2561, 1, 128] - - [294, 45.275] - - - [24576, 2048, 1, 128] - - [310, 48.56] - - - [26624, 4096, 1, 128] - - [23, 46.684] - - - [18560, 2048, 1, 128] - - [390, 53.16] - - - [19584, 128, 1, 128] - - [290, 28.129] - - - [23936, 2048, 1, 128] - - [299, 53.21] - - - [23552, 512, 1, 128] - - [291, 45.422] - - - [12032, 4096, 1, 128] - - [299, 54.747] - - - [3840, 2305, 1, 128] - - [290, 41.743] - - - [25088, 128, 1, 128] - - [290, 31.662] - - - [16640, 8833, 1, 128] - - [27, 46.213] - - - [896, 128, 1, 128] - - [95, 5.167] - - - [17280, 2048, 1, 128] - - [421, 52.894] - - - [16896, 2048, 1, 128] - - [300, 51.206] - - - [22656, 128, 1, 128] - - [293, 29.812] - - - [25728, 8192, 1, 128] - - [27, 46.455] - - - [16128, 128, 1, 128] - - [256, 24.278] - - - [3840, 1024, 1, 128] - - [290, 34.074] - - - [2944, 512, 1, 128] - - [105, 31.266] - - - [24064, 1024, 1, 128] - - [390, 49.251] - - - [896, 385, 1, 128] - - [158, 13.814] - - - [8064, 128, 1, 128] - - [96, 26.572] - - - [12416, 1024, 1, 128] - - [348, 44.511] - - - [20608, 128, 1, 128] - - [298, 27.909] - - - [2944, 1024, 1, 128] - - [418, 30.916] - - - [6656, 2048, 1, 128] - - [294, 45.691] - - - [24064, 128, 1, 128] - - [428, 30.327] - - - [15744, 7937, 1, 128] - - [23, 45.436] - - - [2688, 1024, 1, 128] - - [392, 28.994] - - - [24192, 8193, 1, 128] - - [27, 44.945] - - - [24320, 4096, 1, 128] - - [35, 45.703] - - - [24576, 8705, 1, 128] - - [20, 42.522] - - - [13824, 1024, 1, 128] - - [298, 45.228] - - - [27776, 512, 1, 128] - - [272, 44.826] - - - [10240, 128, 1, 128] - - [103, 30.977] - - - [26240, 10369, 1, 128] - - [35, 41.875] - - - [16512, 4096, 1, 128] - - [300, 56.358] - - - [9856, 6145, 1, 128] - - [421, 54.241] - - - [27392, 1024, 1, 128] - - [421, 47.931] - - - [14976, 1024, 1, 128] - - [420, 45.815] - - - [1280, 512, 1, 128] - - [147, 22.362] - - - [6528, 2817, 1, 128] - - [287, 49.356] - - - [12288, 512, 1, 128] - - [370, 38.025] - - - [5248, 512, 1, 128] - - [291, 28.916] - - - [28544, 4096, 1, 128] - - [53, 45.444] - - - [21248, 13569, 1, 128] - - [49, 45.898] - - - [26112, 2048, 1, 128] - - [421, 52.907] - - - [14208, 6401, 1, 128] - - [25, 44.543] - - - [13952, 128, 1, 128] - - [96, 33.545] - - - [2304, 1665, 1, 128] - - [293, 32.587] - - - [6912, 1024, 1, 128] - - [419, 40.889] - - - [28672, 1024, 1, 128] - - [357, 48.238] - - - [14592, 6913, 1, 128] - - [55, 43.881] - - - [24704, 1024, 1, 128] - - [390, 47.638] - - - [22400, 512, 1, 128] - - [370, 44.307] - - - [23424, 4096, 1, 128] - - [33, 43.442] - - - [24832, 128, 1, 128] - - [38, 33.52] - - - [23680, 2048, 1, 128] - - [421, 53.387] - - - [25984, 9985, 1, 128] - - [28, 45.49] - - - [15360, 512, 1, 128] - - [293, 41.341] - - - [21376, 2048, 1, 128] - - [289, 53.446] - - - [16128, 2048, 1, 128] - - [421, 51.304] - - - [15872, 512, 1, 128] - - [291, 40.882] - - - [3072, 128, 1, 128] - - [98, 16.313] - - - [27520, 4096, 1, 128] - - [25, 45.446] - - - [25216, 4096, 1, 128] - - [33, 45.281] - - - [28672, 12673, 1, 128] - - [35, 46.81] - - - [28288, 512, 1, 128] - - [272, 45.464] - - - [22400, 4096, 1, 128] - - [36, 45.103] - - - [25344, 9345, 1, 128] - - [45, 45.119] - - - [9984, 128, 1, 128] - - [96, 30.964] - - - [28416, 1024, 1, 128] - - [421, 49.699] - - - [27008, 8192, 1, 128] - - [48, 46.31] - - - [13184, 1024, 1, 128] - - [290, 44.786] - - - [10240, 512, 1, 128] - - [291, 37.482] - - - [3456, 128, 1, 128] - - [147, 17.354] - - - [16000, 8321, 1, 128] - - [27, 45.088] - - - [27520, 1024, 1, 128] - - [421, 48.903] - - - [25088, 1024, 1, 128] - - [419, 49.867] - - - [6784, 512, 1, 128] - - [418, 31.994] - - - [18432, 10625, 1, 128] - - [35, 46.727] - - - [16128, 4096, 1, 128] - - [419, 55.283] - - - [26880, 11009, 1, 128] - - [27, 45.732] - - - [28800, 12801, 1, 128] - - [27, 45.526] - - - [12288, 4096, 1, 128] - - [348, 54.183] - - - [20096, 12417, 1, 128] - - [27, 45.622] - - - [1920, 128, 1, 128] - - [98, 10.498] - - - [13056, 2048, 1, 128] - - [421, 51.34] - - - [384, 385, 1, 128] - - [159, 6.362] - - - [9088, 1024, 1, 128] - - [291, 42.512] - - - [6784, 1024, 1, 128] - - [421, 40.415] - - - [21760, 4096, 1, 128] - - [32, 45.548] - - - [27008, 11009, 1, 128] - - [48, 45.472] - - - [14208, 1024, 1, 128] - - [295, 45.848] - - - [25600, 512, 1, 128] - - [312, 44.801] - - - [23680, 1024, 1, 128] - - [421, 48.327] - - - [28160, 8192, 1, 128] - - [49, 46.763] - - - [22016, 4096, 1, 128] - - [28, 45.788] - - - [18688, 4096, 1, 128] - - [289, 55.26] - - - [10752, 1024, 1, 128] - - [300, 43.744] - - - [2432, 128, 1, 128] - - [160, 13.103] - - - [7296, 512, 1, 128] - - [392, 33.34] - - - [19200, 4096, 1, 128] - - [25, 45.06] - - - [4608, 2945, 1, 128] - - [295, 47.64] - - - [18816, 11009, 1, 128] - - [25, 45.335] - - - [9600, 1024, 1, 128] - - [299, 43.317] - - - [7168, 512, 1, 128] - - [293, 32.653] - - - [11904, 4097, 1, 128] - - [419, 52.74] - - - [17920, 1024, 1, 128] - - [300, 47.25] - - - [11520, 7809, 1, 128] - - [47, 44.699] - - - [22784, 14977, 1, 128] - - [45, 46.074] - - - [13696, 1024, 1, 128] - - [420, 45.641] - - - [15104, 1024, 1, 128] - - [419, 46.537] - - - [25216, 512, 1, 128] - - [293, 45.41] - - - [5376, 512, 1, 128] - - [345, 29.375] - - - [17408, 4096, 1, 128] - - [289, 56.231] - - - [25728, 512, 1, 128] - - [293, 45.225] - - - [896, 512, 1, 128] - - [100, 16.743] - - - [6912, 3329, 1, 128] - - [300, 48.663] - - - [22016, 512, 1, 128] - - [268, 42.321] - - - [22144, 4096, 1, 128] - - [53, 45.107] - - - [10368, 128, 1, 128] - - [90, 31.493] - - - [23296, 2048, 1, 128] - - [300, 53.89] - - - [17920, 10113, 1, 128] - - [62, 46.032] - - - [14848, 4096, 1, 128] - - [419, 54.645] - - - [26112, 128, 1, 128] - - [21, 34.81] - - - [28032, 8192, 1, 128] - - [35, 46.536] - - - [20096, 128, 1, 128] - - [312, 27.145] - - - [15360, 4096, 1, 128] - - [299, 56.263] - - - [3328, 128, 1, 128] - - [100, 17.059] - - - [25472, 512, 1, 128] - - [370, 44.43] - - - [18304, 128, 1, 128] - - [418, 27.086] - - - [20352, 12545, 1, 128] - - [27, 45.911] - - - [26624, 10753, 1, 128] - - [35, 46.719] - - - [20480, 2048, 1, 128] - - [301, 51.568] - - - [26496, 10497, 1, 128] - - [45, 45.036] - - - [22400, 128, 1, 128] - - [425, 29.198] - - - [9216, 5505, 1, 128] - - [300, 53.504] - - - [24064, 8193, 1, 128] - - [28, 45.758] - - - [4224, 128, 1, 128] - - [100, 18.782] - - - [6656, 3073, 1, 128] - - [287, 48.384] - - - [10880, 1024, 1, 128] - - [290, 45.131] - - - [23808, 512, 1, 128] - - [290, 45.296] - - - [15488, 1024, 1, 128] - - [419, 46.626] - - - [24704, 8705, 1, 128] - - [23, 44.615] - - - [12416, 4609, 1, 128] - - [419, 52.201] - - - [3712, 1024, 1, 128] - - [293, 33.035] - - - [25856, 8192, 1, 128] - - [36, 46.514] - - - [8320, 1024, 1, 128] - - [420, 40.83] - - - [16256, 512, 1, 128] - - [300, 41.358] - - - [18944, 2048, 1, 128] - - [300, 53.042] - - - [23168, 4096, 1, 128] - - [25, 45.243] - - - [15616, 2048, 1, 128] - - [421, 51.848] - - - [24320, 512, 1, 128] - - [298, 44.588] - - - [2688, 1025, 1, 128] - - [357, 29.009] - - - [12800, 5121, 1, 128] - - [419, 53.797] - - - [5120, 128, 1, 128] - - [259, 11.845] - - - [4352, 512, 1, 128] - - [256, 26.87] - - - [24576, 8192, 1, 128] - - [36, 43.71] - - - [8320, 512, 1, 128] - - [264, 33.816] - - - [12160, 4481, 1, 128] - - [300, 53.398] - - - [2560, 1025, 1, 128] - - [291, 28.015] - - - [19072, 1024, 1, 128] - - [419, 46.547] - - - [2816, 1153, 1, 128] - - [295, 31.034] - - - [6912, 128, 1, 128] - - [96, 27.43] - - - [9088, 2048, 1, 128] - - [301, 47.731] - - - [26368, 8192, 1, 128] - - [35, 46.613] - - - [17408, 9729, 1, 128] - - [35, 46.24] - - - [18816, 4096, 1, 128] - - [419, 55.01] - - - [4480, 512, 1, 128] - - [392, 27.78] - - - [11648, 128, 1, 128] - - [105, 30.816] - - - [1536, 897, 1, 128] - - [152, 27.683] - - - [11136, 1024, 1, 128] - - [421, 44.282] - - - [8704, 1024, 1, 128] - - [289, 41.854] - - - [19072, 2048, 1, 128] - - [421, 52.84] - - - [25856, 1024, 1, 128] - - [421, 48.448] - - - [7552, 3841, 1, 128] - - [390, 50.515] - - - [23296, 128, 1, 128] - - [320, 29.729] - - - [23424, 512, 1, 128] - - [297, 44.63] - - - [26368, 10497, 1, 128] - - [25, 45.865] - - - [18560, 1024, 1, 128] - - [289, 47.589] - - - [8192, 128, 1, 128] - - [91, 26.405] - - - [27776, 11905, 1, 128] - - [35, 45.554] - - - [18688, 1024, 1, 128] - - [419, 46.431] - - - [21248, 4096, 1, 128] - - [33, 45.353] - - - [16256, 8449, 1, 128] - - [27, 44.905] - - - [1920, 1409, 1, 128] - - [344, 28.228] - - - [24704, 4096, 1, 128] - - [59, 44.774] - - - [13824, 6145, 1, 128] - - [25, 44.647] - - - [6528, 512, 1, 128] - - [418, 33.838] - - - [21376, 128, 1, 128] - - [390, 28.422] - - - [11264, 1024, 1, 128] - - [421, 44.622] - - - [4352, 2817, 1, 128] - - [420, 47.509] - - - [22272, 4096, 1, 128] - - [20, 45.315] - - - [27264, 11265, 1, 128] - - [23, 45.335] - - - [28160, 1024, 1, 128] - - [348, 49.274] - - - [16256, 128, 1, 128] - - [418, 26.309] - - - [18688, 2048, 1, 128] - - [419, 52.991] - - - [9600, 6017, 1, 128] - - [421, 53.267] - - - [23552, 1024, 1, 128] - - [299, 49.065] - - - [8576, 128, 1, 128] - - [96, 27.764] - - - [20992, 13185, 1, 128] - - [27, 46.298] - - - [20992, 1024, 1, 128] - - [296, 48.326] - - - [14720, 512, 1, 128] - - [290, 41.511] - - - [28032, 1024, 1, 128] - - [299, 49.774] - - - [20352, 2048, 1, 128] - - [300, 52.962] - - - [15360, 128, 1, 128] - - [418, 24.801] - - - [8448, 2048, 1, 128] - - [348, 47.344] - - - [6272, 2689, 1, 128] - - [284, 47.108] - - - [7808, 4097, 1, 128] - - [421, 50.181] - - - [25472, 128, 1, 128] - - [58, 33.95] - - - [12288, 4481, 1, 128] - - [417, 52.441] - - - [28416, 4096, 1, 128] - - [32, 45.681] - - - [2176, 128, 1, 128] - - [98, 11.724] - - - [21760, 2048, 1, 128] - - [300, 53.745] - - - [21376, 1024, 1, 128] - - [295, 47.866] - - - [13696, 2048, 1, 128] - - [421, 50.463] - - - [28288, 12417, 1, 128] - - [23, 45.754] - - - [5632, 512, 1, 128] - - [418, 30.008] - - - [22016, 1024, 1, 128] - - [357, 47.25] - - - [25216, 128, 1, 128] - - [293, 31.42] - - - [25216, 8192, 1, 128] - - [20, 46.249] - - - [12032, 128, 1, 128] - - [152, 30.73] - - - [6144, 2048, 1, 128] - - [291, 46.555] - - - [23680, 128, 1, 128] - - [25, 32.26] - - - [15744, 128, 1, 128] - - [270, 24.377] - - - [3968, 512, 1, 128] - - [343, 25.162] - - - [16512, 1024, 1, 128] - - [300, 48.192] - - - [1536, 128, 1, 128] - - [101, 8.857] - - - [25984, 4096, 1, 128] - - [49, 45.164] - - - [19456, 512, 1, 128] - - [348, 42.791] - - - [9984, 1024, 1, 128] - - [291, 43.044] - - - [14080, 6401, 1, 128] - - [57, 43.172] - - - [20736, 2048, 1, 128] - - [289, 53.415] - - - [4224, 2689, 1, 128] - - [301, 44.317] - - - [13696, 512, 1, 128] - - [300, 38.896] - - - [17280, 1024, 1, 128] - - [289, 47.822] - - - [10752, 128, 1, 128] - - [96, 29.285] - - - [1536, 512, 1, 128] - - [105, 24.918] - - - [25728, 2048, 1, 128] - - [300, 53.712] - - - [9472, 128, 1, 128] - - [92, 29.878] - - - [7168, 3585, 1, 128] - - [300, 49.222] - - - [14720, 1024, 1, 128] - - [298, 45.869] - - - [25728, 128, 1, 128] - - [56, 34.113] - - - [14976, 128, 1, 128] - - [290, 24.222] - - - [24832, 1024, 1, 128] - - [300, 48.957] - - - [14080, 512, 1, 128] - - [290, 39.541] - - - [17152, 1024, 1, 128] - - [290, 46.193] - - - [19072, 512, 1, 128] - - [297, 42.426] - - - [21120, 1024, 1, 128] - - [348, 47.241] - - - [4864, 128, 1, 128] - - [95, 20.996] - - - [7936, 512, 1, 128] - - [290, 32.781] - - - [21248, 13441, 1, 128] - - [27, 45.979] - - - [12160, 2048, 1, 128] - - [390, 49.17] - - - [19712, 11905, 1, 128] - - [62, 44.366] - - - [23296, 1024, 1, 128] - - [300, 47.694] - - - [24832, 8961, 1, 128] - - [49, 45.757] - - - [13568, 2048, 1, 128] - - [421, 50.568] - - - [13696, 4096, 1, 128] - - [300, 55.096] - - - [5888, 128, 1, 128] - - [90, 24.145] - - - [10112, 2048, 1, 128] - - [299, 48.96] - - - [21632, 13953, 1, 128] - - [20, 45.908] - - - [19328, 512, 1, 128] - - [290, 43.911] - - - [6272, 512, 1, 128] - - [392, 32.383] - - - [4864, 3201, 1, 128] - - [295, 49.001] - - - [15232, 4096, 1, 128] - - [289, 54.861] - - - [23040, 4096, 1, 128] - - [33, 45.59] - - - [2816, 1281, 1, 128] - - [417, 33.382] - - - [8960, 128, 1, 128] - - [92, 28.755] - - - [9472, 1024, 1, 128] - - [289, 43.108] - - - [27648, 11777, 1, 128] - - [35, 46.371] - - - [28416, 2048, 1, 128] - - [289, 54.572] - - - [13952, 6145, 1, 128] - - [32, 44.191] - - - [13952, 1024, 1, 128] - - [295, 45.357] - - - [12544, 2048, 1, 128] - - [300, 50.529] - - - [10624, 7041, 1, 128] - - [421, 53.985] - - - [24704, 2048, 1, 128] - - [300, 53.136] - - - [17280, 9473, 1, 128] - - [36, 45.407] - - - [25088, 9217, 1, 128] - - [45, 45.765] - - - [10240, 6657, 1, 128] - - [300, 54.669] - - - [12800, 4096, 1, 128] - - [300, 53.771] - - - [17792, 1024, 1, 128] - - [419, 46.176] - - - [12160, 128, 1, 128] - - [152, 31.828] - - - [16512, 128, 1, 128] - - [418, 25.658] - - - [25856, 512, 1, 128] - - [289, 45.543] - - - [8576, 4865, 1, 128] - - [421, 53.16] - - - [25984, 1024, 1, 128] - - [289, 48.89] - - - [512, 128, 1, 128] - - [158, 2.93] - - - [10112, 128, 1, 128] - - [92, 31.099] - - - [28288, 2048, 1, 128] - - [299, 54.449] - - - [1152, 641, 1, 128] - - [91, 23.269] - - - [17920, 4096, 1, 128] - - [421, 54.473] - - - [2560, 1921, 1, 128] - - [392, 37.225] - - - [24704, 8833, 1, 128] - - [27, 44.724] - - - [3200, 512, 1, 128] - - [105, 32.245] - - - [6656, 2945, 1, 128] - - [289, 47.751] - - - [12672, 4993, 1, 128] - - [419, 52.9] - - - [4608, 1024, 1, 128] - - [294, 37.546] - - - [25856, 9985, 1, 128] - - [20, 45.586] - - - [23808, 2048, 1, 128] - - [421, 53.552] - - - [9728, 6145, 1, 128] - - [299, 54.041] - - - [28416, 12417, 1, 128] - - [27, 45.628] - - - [14464, 4096, 1, 128] - - [289, 54.336] - - - [21888, 128, 1, 128] - - [370, 28.4] - - - [23680, 15873, 1, 128] - - [48, 45.366] - - - [22144, 1024, 1, 128] - - [390, 49.046] - - - [17664, 512, 1, 256] - - [290, 64.19] - - - [25600, 1024, 1, 256] - - [340, 75.317] - - - [28928, 512, 1, 256] - - [289, 72.493] - - - [15104, 512, 1, 256] - - [422, 61.706] - - - [38912, 1024, 1, 256] - - [21, 70.324] - - - [34304, 8192, 1, 256] - - [40, 74.829] - - - [23552, 1024, 1, 256] - - [289, 74.314] - - - [39424, 23552, 1, 256] - - [39, 75.106] - - - [9472, 1024, 1, 256] - - [424, 66.576] - - - [28928, 13056, 1, 256] - - [49, 74.869] - - - [42496, 1024, 1, 256] - - [67, 70.355] - - - [18432, 1024, 1, 256] - - [299, 74.874] - - - [40192, 24320, 1, 256] - - [25, 74.909] - - - [33280, 17152, 1, 256] - - [22, 75.787] - - - [27904, 512, 1, 256] - - [390, 70.61] - - - [39680, 8192, 1, 256] - - [23, 74.356] - - - [28160, 8192, 1, 256] - - [27, 74.637] - - - [25088, 8192, 1, 256] - - [22, 74.701] - - - [23040, 15360, 1, 256] - - [23, 75.671] - - - [19712, 11776, 1, 256] - - [39, 74.182] - - - [43520, 27648, 1, 256] - - [39, 74.947] - - - [44544, 4096, 1, 256] - - [39, 73.451] - - - [20224, 4096, 1, 256] - - [25, 72.535] - - - [31744, 4096, 1, 256] - - [26, 73.158] - - - [33024, 16896, 1, 256] - - [40, 75.252] - - - [32768, 8192, 1, 256] - - [81, 58.919] - - - [42752, 4096, 1, 256] - - [25, 72.964] - - - [19968, 512, 1, 256] - - [268, 63.814] - - - [10496, 512, 1, 256] - - [291, 57.527] - - - [36864, 4096, 1, 256] - - [52, 72.634] - - - [12288, 1024, 1, 256] - - [296, 66.771] - - - [22784, 14848, 1, 256] - - [26, 75.154] - - - [17152, 9472, 1, 256] - - [29, 74.859] - - - [31488, 1024, 1, 256] - - [299, 78.031] - - - [25344, 1024, 1, 256] - - [309, 75.675] - - - [33536, 512, 1, 256] - - [309, 70.579] - - - [28672, 8192, 1, 256] - - [27, 74.718] - - - [15104, 7168, 1, 256] - - [25, 73.585] - - - [38144, 22272, 1, 256] - - [36, 74.833] - - - [25344, 4096, 1, 256] - - [41, 72.024] - - - [6400, 2560, 1, 256] - - [360, 72.622] - - - [21248, 13568, 1, 256] - - [35, 75.179] - - - [2304, 1536, 1, 256] - - [419, 51.168] - - - [20992, 512, 1, 256] - - [290, 65.944] - - - [3072, 1024, 1, 256] - - [357, 50.612] - - - [36864, 20736, 1, 256] - - [23, 75.147] - - - [39936, 24064, 1, 256] - - [35, 75.501] - - - [2816, 512, 1, 256] - - [346, 35.059] - - - [37888, 512, 1, 256] - - [267, 71.591] - - - [39680, 1024, 1, 256] - - [35, 69.234] - - - [35584, 19712, 1, 256] - - [20, 74.82] - - - [25600, 9728, 1, 256] - - [23, 75.44] - - - [2816, 1024, 1, 256] - - [290, 46.506] - - - [13056, 1024, 1, 256] - - [425, 71.566] - - - [39680, 4096, 1, 256] - - [35, 72.728] - - - [4864, 3072, 1, 256] - - [294, 72.427] - - - [27648, 11776, 1, 256] - - [27, 75.73] - - - [13056, 4096, 1, 256] - - [35, 72.06] - - - [4096, 2304, 1, 256] - - [308, 65.813] - - - [34048, 1024, 1, 256] - - [299, 81.141] - - - [6400, 512, 1, 256] - - [38, 51.306] - - - [15872, 4096, 1, 256] - - [35, 72.281] - - - [29440, 1024, 1, 256] - - [299, 80.3] - - - [7424, 512, 1, 256] - - [291, 50.348] - - - [19200, 4096, 1, 256] - - [35, 72.099] - - - [37376, 21504, 1, 256] - - [39, 75.215] - - - [37888, 1024, 1, 256] - - [299, 81.195] - - - [40704, 24832, 1, 256] - - [25, 74.84] - - - [26112, 1024, 1, 256] - - [289, 76.34] - - - [25088, 8960, 1, 256] - - [57, 75.437] - - - [27136, 512, 1, 256] - - [289, 69.218] - - - [4608, 512, 1, 256] - - [270, 44.165] - - - [31232, 8192, 1, 256] - - [39, 74.803] - - - [33024, 512, 1, 256] - - [291, 69.154] - - - [27648, 512, 1, 256] - - [304, 68.877] - - - [28928, 4096, 1, 256] - - [25, 72.464] - - - [44544, 2048, 1, 256] - - [39, 72.039] - - - [43776, 27648, 1, 256] - - [39, 74.437] - - - [19456, 4096, 1, 256] - - [23, 72.834] - - - [33536, 17664, 1, 256] - - [22, 75.192] - - - [35328, 4096, 1, 256] - - [41, 73.657] - - - [13312, 5376, 1, 256] - - [32, 74.004] - - - [32768, 1024, 1, 256] - - [302, 60.483] - - - [39168, 4096, 1, 256] - - [39, 73.18] - - - [15616, 7936, 1, 256] - - [35, 74.429] - - - [41472, 25600, 1, 256] - - [36, 74.75] - - - [14592, 4096, 1, 256] - - [29, 70.479] - - - [37632, 21760, 1, 256] - - [23, 74.813] - - - [37376, 21248, 1, 256] - - [35, 75.33] - - - [14336, 6656, 1, 256] - - [27, 74.692] - - - [36608, 20480, 1, 256] - - [39, 74.602] - - - [32256, 16384, 1, 256] - - [41, 75.223] - - - [44544, 28416, 1, 256] - - [36, 75.158] - - - [26112, 512, 1, 256] - - [357, 71.374] - - - [41216, 25344, 1, 256] - - [25, 74.919] - - - [16640, 512, 1, 256] - - [291, 61.406] - - - [30464, 14336, 1, 256] - - [39, 73.987] - - - [13312, 4096, 1, 256] - - [27, 72.101] - - - [22528, 1024, 1, 256] - - [299, 76.98] - - - [5632, 1024, 1, 256] - - [291, 58.191] - - - [27392, 1024, 1, 256] - - [299, 74.117] - - - [27648, 8192, 1, 256] - - [23, 75.078] - - - [26368, 1024, 1, 256] - - [390, 77.839] - - - [43776, 4096, 1, 256] - - [26, 73.189] - - - [23552, 15872, 1, 256] - - [23, 76.241] - - - [26624, 10496, 1, 256] - - [25, 75.778] - - - [27392, 8192, 1, 256] - - [26, 74.062] - - - [17408, 9728, 1, 256] - - [23, 75.456] - - - [16896, 9216, 1, 256] - - [25, 74.885] - - - [26880, 11008, 1, 256] - - [25, 74.818] - - - [31488, 512, 1, 256] - - [314, 69.717] - - - [14336, 6400, 1, 256] - - [47, 74.846] - - - [17152, 512, 1, 256] - - [290, 63.984] - - - [7168, 512, 1, 256] - - [290, 50.22] - - - [41984, 26112, 1, 256] - - [20, 75.545] - - - [11776, 512, 1, 256] - - [291, 58.434] - - - [16128, 8448, 1, 256] - - [32, 74.56] - - - [11520, 1024, 1, 256] - - [421, 66.093] - - - [27904, 1024, 1, 256] - - [299, 76.772] - - - [37888, 8192, 1, 256] - - [27, 74.991] - - - [20480, 12544, 1, 256] - - [20, 75.828] - - - [23552, 15616, 1, 256] - - [27, 76.268] - - - [21504, 13824, 1, 256] - - [25, 76.158] - - - [27136, 11008, 1, 256] - - [22, 75.587] - - - [32000, 512, 1, 256] - - [309, 69.662] - - - [26624, 1024, 1, 256] - - [299, 76.934] - - - [34816, 8192, 1, 256] - - [35, 75.041] - - - [23040, 512, 1, 256] - - [291, 67.275] - - - [36608, 1024, 1, 256] - - [299, 81.349] - - - [43264, 8192, 1, 256] - - [27, 74.358] - - - [30208, 14336, 1, 256] - - [40, 75.063] - - - [43520, 512, 1, 256] - - [289, 72.271] - - - [32256, 4096, 1, 256] - - [41, 73.355] - - - [33792, 17664, 1, 256] - - [25, 75.868] - - - [10752, 6912, 1, 256] - - [25, 74.308] - - - [29696, 8192, 1, 256] - - [27, 74.986] - - - [41472, 512, 1, 256] - - [299, 75.122] - - - [44544, 8192, 1, 256] - - [25, 74.517] - - - [41472, 8192, 1, 256] - - [41, 74.446] - - - [38656, 4096, 1, 256] - - [39, 73.31] - - - [44800, 512, 1, 256] - - [357, 74.267] - - - [37376, 4096, 1, 256] - - [39, 73.81] - - - [19200, 1024, 1, 256] - - [289, 74.01] - - - [39680, 23552, 1, 256] - - [35, 74.227] - - - [30976, 8192, 1, 256] - - [39, 73.668] - - - [25856, 1024, 1, 256] - - [296, 76.076] - - - [22016, 14336, 1, 256] - - [35, 75.309] - - - [17152, 9216, 1, 256] - - [23, 74.262] - - - [18432, 10752, 1, 256] - - [23, 75.858] - - - [5376, 1024, 1, 256] - - [421, 55.394] - - - [21760, 13824, 1, 256] - - [27, 75.487] - - - [15360, 512, 1, 256] - - [422, 61.52] - - - [2560, 512, 1, 256] - - [347, 33.004] - - - [36096, 8192, 1, 256] - - [26, 73.872] - - - [42752, 26624, 1, 256] - - [25, 74.446] - - - [35584, 19456, 1, 256] - - [39, 74.674] - - - [6144, 2304, 1, 256] - - [287, 69.505] - - - [42240, 1024, 1, 256] - - [42, 69.828] - - - [26880, 4096, 1, 256] - - [23, 72.673] - - - [28160, 12032, 1, 256] - - [25, 75.503] - - - [18688, 10752, 1, 256] - - [25, 75.22] - - - [43520, 8192, 1, 256] - - [27, 74.662] - - - [8192, 4352, 1, 256] - - [304, 77.625] - - - [6912, 3072, 1, 256] - - [296, 78.097] - - - [31744, 15616, 1, 256] - - [35, 75.832] - - - [36352, 20224, 1, 256] - - [39, 75.358] - - - [41216, 25088, 1, 256] - - [20, 74.747] - - - [37632, 1024, 1, 256] - - [390, 79.689] - - - [18944, 512, 1, 256] - - [390, 66.947] - - - [15616, 1024, 1, 256] - - [306, 70.905] - - - [44288, 512, 1, 256] - - [308, 74.282] - - - [24832, 8704, 1, 256] - - [49, 74.982] - - - [21504, 13568, 1, 256] - - [35, 76.086] - - - [18176, 10496, 1, 256] - - [25, 75.104] - - - [21248, 1024, 1, 256] - - [313, 75.451] - - - [16384, 1024, 1, 256] - - [302, 63.94] - - - [25600, 8192, 1, 256] - - [35, 75.08] - - - [28672, 12544, 1, 256] - - [20, 75.492] - - - [16128, 1024, 1, 256] - - [348, 71.533] - - - [22272, 14592, 1, 256] - - [45, 75.412] - - - [1280, 512, 1, 256] - - [103, 30.85] - - - [36864, 20992, 1, 256] - - [35, 75.008] - - - [3584, 1792, 1, 256] - - [293, 61.511] - - - [35072, 19200, 1, 256] - - [23, 74.991] - - - [32000, 4096, 1, 256] - - [52, 72.71] - - - [28416, 1024, 1, 256] - - [299, 77.456] - - - [20480, 12800, 1, 256] - - [23, 75.76] - - - [21760, 4096, 1, 256] - - [84, 72.584] - - - [44288, 8192, 1, 256] - - [39, 74.353] - - - [33280, 4096, 1, 256] - - [26, 73.587] - - - [32512, 1024, 1, 256] - - [390, 77.366] - - - [38400, 22528, 1, 256] - - [26, 75.187] - - - [40448, 1024, 1, 256] - - [58, 70.029] - - - [5120, 512, 1, 256] - - [289, 43.639] - - - [29952, 8192, 1, 256] - - [49, 74.456] - - - [40448, 24576, 1, 256] - - [41, 74.509] - - - [29696, 4096, 1, 256] - - [23, 73.234] - - - [21504, 1024, 1, 256] - - [299, 73.97] - - - [19968, 1024, 1, 256] - - [299, 73.04] - - - [16896, 512, 1, 256] - - [268, 63.87] - - - [33536, 17408, 1, 256] - - [26, 74.787] - - - [19712, 512, 1, 256] - - [348, 62.611] - - - [16384, 8704, 1, 256] - - [36, 63.818] - - - [29952, 13824, 1, 256] - - [28, 75.112] - - - [14592, 6656, 1, 256] - - [65, 72.415] - - - [36864, 1024, 1, 256] - - [306, 78.36] - - - [31744, 15872, 1, 256] - - [25, 75.897] - - - [24832, 8960, 1, 256] - - [29, 75.119] - - - [23808, 1024, 1, 256] - - [289, 74.659] - - - [19200, 11264, 1, 256] - - [26, 74.727] - - - [23296, 15360, 1, 256] - - [27, 75.187] - - - [34304, 18432, 1, 256] - - [40, 75.048] - - - [22016, 1024, 1, 256] - - [390, 74.404] - - - [40704, 4096, 1, 256] - - [26, 73.066] - - - [25600, 4096, 1, 256] - - [26, 73.152] - - - [3328, 1024, 1, 256] - - [300, 50.927] - - - [30464, 8192, 1, 256] - - [39, 73.421] - - - [39424, 8192, 1, 256] - - [35, 74.587] - - - [23808, 15872, 1, 256] - - [35, 75.524] - - - [8960, 1024, 1, 256] - - [268, 64.081] - - - [44032, 4096, 1, 256] - - [35, 73.194] - - - [35584, 8192, 1, 256] - - [40, 74.268] - - - [29184, 8192, 1, 256] - - [28, 74.597] - - - [13824, 1024, 1, 256] - - [296, 70.822] - - - [36608, 8192, 1, 256] - - [23, 74.225] - - - [30976, 512, 1, 256] - - [423, 70.181] - - - [33024, 4096, 1, 256] - - [40, 73.199] - - - [11776, 7936, 1, 256] - - [25, 74.374] - - - [23808, 16128, 1, 256] - - [25, 75.37] - - - [22272, 14336, 1, 256] - - [39, 74.944] - - - [27392, 11520, 1, 256] - - [40, 74.289] - - - [30464, 4096, 1, 256] - - [39, 71.791] - - - [20992, 13312, 1, 256] - - [26, 75.507] - - - [44800, 1024, 1, 256] - - [27, 69.404] - - - [32512, 4096, 1, 256] - - [82, 72.937] - - - [23296, 15616, 1, 256] - - [25, 75.408] - - - [9216, 1024, 1, 256] - - [289, 67.319] - - - [20224, 12544, 1, 256] - - [23, 75.301] - - - [32256, 1024, 1, 256] - - [289, 77.752] - - - [38400, 512, 1, 256] - - [424, 69.875] - - - [29952, 1024, 1, 256] - - [299, 77.929] - - - [36352, 512, 1, 256] - - [298, 71.381] - - - [41728, 25600, 1, 256] - - [39, 73.877] - - - [32000, 1024, 1, 256] - - [289, 78.645] - - - [38144, 22016, 1, 256] - - [20, 74.81] - - - [27136, 11264, 1, 256] - - [45, 75.389] - - - [34048, 18176, 1, 256] - - [63, 74.655] - - - [22016, 14080, 1, 256] - - [57, 76.113] - - - [19712, 12032, 1, 256] - - [39, 74.357] - - - [23552, 4096, 1, 256] - - [39, 72.963] - - - [15872, 1024, 1, 256] - - [325, 71.921] - - - [37120, 512, 1, 256] - - [306, 70.893] - - - [9984, 1024, 1, 256] - - [268, 64.724] - - - [32512, 8192, 1, 256] - - [22, 74.44] - - - [15360, 4096, 1, 256] - - [25, 72.39] - - - [13056, 512, 1, 256] - - [426, 61.842] - - - [44032, 8192, 1, 256] - - [35, 74.818] - - - [24576, 8192, 1, 256] - - [52, 70.1] - - - [36352, 8192, 1, 256] - - [26, 74.844] - - - [26368, 8192, 1, 256] - - [35, 74.449] - - - [20480, 1024, 1, 256] - - [357, 74.04] - - - [35072, 8192, 1, 256] - - [35, 74.34] - - - [32000, 15872, 1, 256] - - [25, 75.055] - - - [40704, 24576, 1, 256] - - [26, 74.2] - - - [15104, 7424, 1, 256] - - [27, 74.433] - - - [25856, 4096, 1, 256] - - [26, 72.553] - - - [14848, 512, 1, 256] - - [293, 62.463] - - - [39424, 4096, 1, 256] - - [63, 73.507] - - - [24832, 512, 1, 256] - - [293, 67.623] - - - [44288, 28416, 1, 256] - - [20, 74.674] - - - [12544, 4608, 1, 256] - - [33, 72.008] - - - [12800, 4864, 1, 256] - - [57, 73.386] - - - [29440, 512, 1, 256] - - [348, 70.338] - - - [40192, 24064, 1, 256] - - [35, 74.898] - - - [18176, 4096, 1, 256] - - [27, 72.486] - - - [40960, 8192, 1, 256] - - [52, 67.262] - - - [42240, 512, 1, 256] - - [289, 75.361] - - - [9728, 512, 1, 256] - - [289, 57.808] - - - [14848, 7168, 1, 256] - - [23, 73.661] - - - [44800, 28672, 1, 256] - - [52, 74.049] - - - [15616, 7680, 1, 256] - - [27, 74.594] - - - [33280, 17408, 1, 256] - - [40, 75.374] - - - [42752, 1024, 1, 256] - - [34, 69.863] - - - [35328, 8192, 1, 256] - - [45, 74.689] - - - [36352, 1024, 1, 256] - - [299, 81.063] - - - [35840, 1024, 1, 256] - - [299, 80.244] - - - [41472, 4096, 1, 256] - - [26, 73.656] - - - [3584, 1024, 1, 256] - - [295, 52.432] - - - [22528, 14592, 1, 256] - - [23, 76.245] - - - [44032, 512, 1, 256] - - [360, 74.061] - - - [30720, 1024, 1, 256] - - [299, 78.893] - - - [39680, 512, 1, 256] - - [289, 71.628] - - - [22272, 1024, 1, 256] - - [340, 74.062] - - - [42240, 26368, 1, 256] - - [27, 74.63] - - - [10240, 6400, 1, 256] - - [53, 74.368] - - - [30976, 14848, 1, 256] - - [26, 74.145] - - - [41728, 25856, 1, 256] - - [39, 74.256] - - - [28928, 12800, 1, 256] - - [45, 74.82] - - - [21760, 14080, 1, 256] - - [23, 75.486] - - - [5888, 1024, 1, 256] - - [421, 60.454] - - - [24576, 8704, 1, 256] - - [36, 70.872] - - - [38912, 4096, 1, 256] - - [23, 73.118] - - - [15360, 1024, 1, 256] - - [300, 70.958] - - - [18688, 512, 1, 256] - - [298, 65.318] - - - [27392, 512, 1, 256] - - [337, 66.116] - - - [22784, 512, 1, 256] - - [348, 69.533] - - - [40448, 4096, 1, 256] - - [59, 73.641] - - - [19200, 512, 1, 256] - - [315, 66.393] - - - [26368, 10496, 1, 256] - - [25, 74.947] - - - [25088, 9216, 1, 256] - - [40, 74.874] - - - [33536, 1024, 1, 256] - - [299, 79.345] - - - [25600, 9472, 1, 256] - - [25, 75.887] - - - [13824, 4096, 1, 256] - - [25, 71.984] - - - [5632, 3840, 1, 256] - - [299, 75.584] - - - [9216, 5376, 1, 256] - - [47, 72.915] - - - [8960, 5120, 1, 256] - - [29, 71.62] - - - [19456, 512, 1, 256] - - [424, 67.758] - - - [24576, 4096, 1, 256] - - [30, 68.369] - - - [27392, 11264, 1, 256] - - [39, 74.537] - - - [35072, 4096, 1, 256] - - [39, 72.993] - - - [44288, 4096, 1, 256] - - [39, 73.28] - - - [40448, 8192, 1, 256] - - [49, 74.465] - - - [33280, 512, 1, 256] - - [267, 70.6] - - - [22272, 4096, 1, 256] - - [39, 72.461] - - - [35584, 512, 1, 256] - - [289, 69.945] - - - [10752, 512, 1, 256] - - [293, 57.924] - - - [19968, 4096, 1, 256] - - [59, 72.937] - - - [34304, 1024, 1, 256] - - [299, 79.467] - - - [41216, 8192, 1, 256] - - [39, 74.477] - - - [35840, 19712, 1, 256] - - [25, 75.823] - - - [43520, 27392, 1, 256] - - [36, 75.44] - - - [30720, 14848, 1, 256] - - [25, 75.633] - - - [38400, 22272, 1, 256] - - [36, 75.441] - - - [1536, 1024, 1, 256] - - [347, 36.389] - - - [40192, 1024, 1, 256] - - [59, 69.757] - - - [44800, 256, 1, 256] - - [291, 64.53] - - - [1536, 512, 1, 256] - - [105, 35.992] - - - [34560, 18432, 1, 256] - - [23, 74.758] - - - [1792, 1024, 1, 256] - - [345, 39.728] - - - [5376, 3584, 1, 256] - - [348, 73.951] - - - [30208, 1024, 1, 256] - - [299, 76.992] - - - [31232, 512, 1, 256] - - [312, 70.96] - - - [23040, 4096, 1, 256] - - [82, 73.309] - - - [35840, 4096, 1, 256] - - [25, 73.139] - - - [38144, 512, 1, 256] - - [299, 71.773] - - - [31744, 512, 1, 256] - - [304, 70.665] - - - [14592, 6912, 1, 256] - - [66, 73.121] - - - [19456, 11520, 1, 256] - - [25, 76.138] - - - [7168, 1024, 1, 256] - - [299, 61.009] - - - [18944, 11264, 1, 256] - - [40, 75.342] - - - [19712, 1024, 1, 256] - - [421, 71.985] - - - [26112, 9984, 1, 256] - - [35, 75.563] - - - [38656, 22784, 1, 256] - - [45, 74.655] - - - [24320, 8192, 1, 256] - - [45, 74.704] - - - [4864, 1024, 1, 256] - - [308, 56.013] - - - [20480, 4096, 1, 256] - - [35, 72.506] - - - [10240, 1024, 1, 256] - - [299, 65.3] - - - [31232, 15360, 1, 256] - - [22, 75.415] - - - [24320, 4096, 1, 256] - - [45, 72.813] - - - [33792, 1024, 1, 256] - - [299, 79.06] - - - [12032, 1024, 1, 256] - - [299, 68.736] - - - [39168, 512, 1, 256] - - [298, 72.75] - - - [16896, 4096, 1, 256] - - [25, 72.514] - - - [36096, 1024, 1, 256] - - [299, 77.927] - - - [28416, 12544, 1, 256] - - [49, 74.959] - - - [30720, 4096, 1, 256] - - [23, 73.127] - - - [19712, 4096, 1, 256] - - [39, 71.466] - - - [37120, 21248, 1, 256] - - [35, 74.681] - - - [16384, 4096, 1, 256] - - [25, 61.887] - - - [18688, 11008, 1, 256] - - [27, 75.091] - - - [38400, 8192, 1, 256] - - [35, 74.678] - - - [11264, 7424, 1, 256] - - [25, 74.924] - - - [23296, 512, 1, 256] - - [293, 66.977] - - - [25344, 512, 1, 256] - - [348, 70.579] - - - [44544, 256, 1, 256] - - [424, 65.676] - - - [43264, 4096, 1, 256] - - [52, 72.828] - - - [32512, 16640, 1, 256] - - [28, 75.357] - - - [39936, 8192, 1, 256] - - [27, 74.974] - - - [43264, 512, 1, 256] - - [298, 73.773] - - - [16640, 8704, 1, 256] - - [35, 74.811] - - - [26624, 8192, 1, 256] - - [35, 75.099] - - - [35328, 19456, 1, 256] - - [40, 75.247] - - - [42752, 26880, 1, 256] - - [23, 74.838] - - - [25344, 9216, 1, 256] - - [41, 73.445] - - - [34048, 8192, 1, 256] - - [41, 73.799] - - - [18688, 4096, 1, 256] - - [27, 72.394] - - - [37632, 8192, 1, 256] - - [52, 74.221] - - - [19968, 12032, 1, 256] - - [25, 75.67] - - - [8448, 4608, 1, 256] - - [299, 81.788] - - - [2048, 1536, 1, 256] - - [293, 48.655] - - - [31488, 15616, 1, 256] - - [35, 74.939] - - - [35328, 512, 1, 256] - - [309, 71.218] - - - [37376, 8192, 1, 256] - - [39, 74.732] - - - [33792, 8192, 1, 256] - - [35, 74.959] - - - [36608, 4096, 1, 256] - - [39, 72.898] - - - [28416, 8192, 1, 256] - - [45, 74.104] - - - [5632, 512, 1, 256] - - [290, 46.222] - - - [13568, 4096, 1, 256] - - [25, 71.849] - - - [17664, 9728, 1, 256] - - [55, 74.224] - - - [13568, 1024, 1, 256] - - [289, 69.657] - - - [8448, 512, 1, 256] - - [291, 52.262] - - - [22528, 4096, 1, 256] - - [27, 72.967] - - - [33536, 8192, 1, 256] - - [45, 74.444] - - - [23296, 1024, 1, 256] - - [299, 74.139] - - - [43520, 4096, 1, 256] - - [39, 73.78] - - - [39936, 23808, 1, 256] - - [35, 75.482] - - - [12544, 4096, 1, 256] - - [32, 71.348] - - - [22016, 4096, 1, 256] - - [72, 73.077] - - - [14592, 512, 1, 256] - - [298, 60.746] - - - [39936, 4096, 1, 256] - - [25, 73.242] - - - [18176, 1024, 1, 256] - - [289, 72.46] - - - [44800, 2048, 1, 256] - - [41, 70.237] - - - [14848, 4096, 1, 256] - - [25, 72.5] - - - [20224, 12288, 1, 256] - - [39, 74.421] - - - [16896, 8960, 1, 256] - - [35, 75.55] - - - [43264, 27392, 1, 256] - - [35, 74.802] - - - [24064, 16128, 1, 256] - - [45, 75.907] - - - [1024, 512, 1, 256] - - [91, 24.986] - - - [24576, 8448, 1, 256] - - [20, 70.765] - - - [25344, 9472, 1, 256] - - [26, 74.117] - - - [3328, 1536, 1, 256] - - [293, 55.032] - - - [31488, 4096, 1, 256] - - [25, 72.728] - - - [43008, 8192, 1, 256] - - [35, 74.904] - - - [28672, 12800, 1, 256] - - [23, 75.357] - - - [20736, 13056, 1, 256] - - [25, 75.359] - - - [17664, 9984, 1, 256] - - [55, 74.504] - - - [17920, 1024, 1, 256] - - [299, 71.163] - - - [11008, 1024, 1, 256] - - [300, 66.317] - - - [44800, 4096, 1, 256] - - [30, 72.511] - - - [29952, 14080, 1, 256] - - [22, 75.218] - - - [39168, 23296, 1, 256] - - [39, 74.594] - - - [9472, 512, 1, 256] - - [291, 56.468] - - - [27904, 8192, 1, 256] - - [26, 74.273] - - - [5120, 1024, 1, 256] - - [291, 55.472] - - - [15872, 7936, 1, 256] - - [23, 75.157] - - - [13568, 5632, 1, 256] - - [32, 73.703] - - - [17920, 9984, 1, 256] - - [35, 75.469] - - - [16640, 8960, 1, 256] - - [49, 74.67] - - - [41984, 4096, 1, 256] - - [25, 73.1] - - - [6912, 512, 1, 256] - - [293, 48.297] - - - [28416, 4096, 1, 256] - - [30, 72.453] - - - [27648, 11520, 1, 256] - - [27, 75.978] - - - [7680, 3840, 1, 256] - - [421, 79.051] - - - [34048, 4096, 1, 256] - - [40, 72.436] - - - [11264, 512, 1, 256] - - [277, 57.829] - - - [26368, 4096, 1, 256] - - [35, 72.565] - - - [21248, 13312, 1, 256] - - [25, 75.106] - - - [15104, 1024, 1, 256] - - [289, 71.008] - - - [35072, 18944, 1, 256] - - [25, 74.936] - - - [6144, 1024, 1, 256] - - [298, 60.175] - - - [44800, 8192, 1, 256] - - [35, 74.033] - - - [25088, 512, 1, 256] - - [348, 70.787] - - - [27904, 12032, 1, 256] - - [41, 74.855] - - - [27648, 1024, 1, 256] - - [289, 77.008] - - - [28928, 8192, 1, 256] - - [45, 74.128] - - - [29440, 13312, 1, 256] - - [28, 75.126] - - - [43264, 27136, 1, 256] - - [36, 74.753] - - - [23552, 512, 1, 256] - - [291, 67.252] - - - [26880, 10752, 1, 256] - - [27, 75.05] - - - [44032, 28160, 1, 256] - - [35, 75.553] - - - [36096, 512, 1, 256] - - [312, 69.982] - - - [4352, 2560, 1, 256] - - [287, 69.01] - - - [38912, 8192, 1, 256] - - [23, 74.954] - - - [12032, 4096, 1, 256] - - [23, 71.167] - - - [37632, 512, 1, 256] - - [426, 70.677] - - - [30208, 512, 1, 256] - - [289, 69.714] - - - [2304, 512, 1, 256] - - [103, 41.352] - - - [24320, 8448, 1, 256] - - [55, 74.743] - - - [39424, 512, 1, 256] - - [307, 71.008] - - - [37632, 21504, 1, 256] - - [36, 74.577] - - - [17152, 1024, 1, 256] - - [306, 72.043] - - - [22784, 15104, 1, 256] - - [49, 75.335] - - - [27904, 11776, 1, 256] - - [40, 74.855] - - - [43008, 26880, 1, 256] - - [36, 75.604] - - - [41728, 4096, 1, 256] - - [39, 72.671] - - - [25344, 8192, 1, 256] - - [39, 73.596] - - - [44800, 28928, 1, 256] - - [23, 74.343] - - - [38912, 22784, 1, 256] - - [36, 75.639] - - - [44032, 1024, 1, 256] - - [42, 70.885] - - - [30976, 4096, 1, 256] - - [26, 71.622] - - - [15872, 8192, 1, 256] - - [23, 75.001] - - - [40960, 4096, 1, 256] - - [30, 66.076] - - - [35584, 1024, 1, 256] - - [299, 81.001] - - - [18944, 4096, 1, 256] - - [74, 72.836] - - - [36096, 20224, 1, 256] - - [26, 74.266] - - - [11008, 7168, 1, 256] - - [79, 71.456] - - - [7936, 1024, 1, 256] - - [294, 62.6] - - - [44288, 1024, 1, 256] - - [37, 69.726] - - - [38656, 8192, 1, 256] - - [39, 74.287] - - - [38144, 1024, 1, 256] - - [299, 80.817] - - - [41984, 1024, 1, 256] - - [56, 70.499] - - - [20736, 512, 1, 256] - - [289, 65.991] - - - [32768, 16640, 1, 256] - - [81, 59.277] - - - [40960, 1024, 1, 256] - - [23, 64.607] - - - [25856, 9984, 1, 256] - - [23, 74.885] - - - [29696, 13824, 1, 256] - - [27, 75.882] - - - [37120, 4096, 1, 256] - - [39, 73.15] - - - [37120, 20992, 1, 256] - - [23, 74.854] - - - [35072, 512, 1, 256] - - [313, 70.433] - - - [38656, 1024, 1, 256] - - [34, 69.426] - - - [37376, 512, 1, 256] - - [306, 72.652] - - - [32000, 16128, 1, 256] - - [27, 74.91] - - - [41984, 25856, 1, 256] - - [36, 75.626] - - - [23040, 15104, 1, 256] - - [28, 76.021] - - - [31232, 15104, 1, 256] - - [49, 75.761] - - - [25088, 4096, 1, 256] - - [59, 73.174] - - - [15360, 7424, 1, 256] - - [23, 75.274] - - - [16384, 8448, 1, 256] - - [20, 63.356] - - - [26624, 4096, 1, 256] - - [25, 73.175] - - - [14080, 6400, 1, 256] - - [79, 73.029] - - - [16128, 4096, 1, 256] - - [25, 72.085] - - - [43776, 27904, 1, 256] - - [39, 74.364] - - - [15872, 512, 1, 256] - - [291, 61.811] - - - [43776, 8192, 1, 256] - - [39, 74.087] - - - [10496, 6656, 1, 256] - - [33, 73.241] - - - [13312, 512, 1, 256] - - [268, 60.02] - - - [29184, 512, 1, 256] - - [320, 69.77] - - - [15360, 7680, 1, 256] - - [23, 75.288] - - - [40192, 8192, 1, 256] - - [23, 74.473] - - - [34560, 8192, 1, 256] - - [27, 74.307] - - - [25856, 8192, 1, 256] - - [23, 74.455] - - - [32512, 16384, 1, 256] - - [40, 75.174] - - - [12288, 4352, 1, 256] - - [23, 73.324] - - - [29440, 13568, 1, 256] - - [49, 75.133] - - - [28160, 1024, 1, 256] - - [390, 76.933] - - - [32768, 4096, 1, 256] - - [64, 58.306] - - - [24832, 4096, 1, 256] - - [39, 72.684] - - - [39680, 23808, 1, 256] - - [35, 74.405] - - - [22784, 4096, 1, 256] - - [59, 72.548] - - - [7936, 4096, 1, 256] - - [299, 79.37] - - - [8704, 4864, 1, 256] - - [53, 72.098] - - - [29696, 512, 1, 256] - - [298, 69.432] - - - [39424, 23296, 1, 256] - - [20, 75.413] - - - [17408, 9472, 1, 256] - - [23, 75.849] - - - [33792, 4096, 1, 256] - - [52, 72.911] - - - [17920, 512, 1, 256] - - [291, 64.532] - - - [25856, 512, 1, 256] - - [348, 71.005] - - - [44288, 28160, 1, 256] - - [20, 74.632] - - - [40192, 4096, 1, 256] - - [72, 73.012] - - - [21248, 512, 1, 256] - - [305, 65.875] - - - [3072, 512, 1, 256] - - [163, 44.68] - - - [29184, 13312, 1, 256] - - [23, 75.113] - - - [44544, 1024, 1, 256] - - [58, 70.345] - - - [37888, 21760, 1, 256] - - [25, 75.792] - - - [33792, 17920, 1, 256] - - [25, 75.738] - - - [6912, 1024, 1, 256] - - [295, 60.989] - - - [41216, 512, 1, 256] - - [298, 73.667] - - - [42240, 26112, 1, 256] - - [27, 74.753] - - - [30720, 8192, 1, 256] - - [35, 75.059] - - - [11776, 1024, 1, 256] - - [299, 68.264] - - - [43008, 4096, 1, 256] - - [36, 73.053] - - - [34560, 18688, 1, 256] - - [27, 74.923] - - - [41984, 512, 1, 256] - - [390, 75.617] - - - [41728, 512, 1, 256] - - [360, 75.142] - - - [2560, 1792, 1, 256] - - [392, 56.281] - - - [36864, 8192, 1, 256] - - [25, 74.544] - - - [40704, 8192, 1, 256] - - [25, 74.341] - - - [30720, 14592, 1, 256] - - [23, 75.998] - - - [32256, 512, 1, 256] - - [325, 69.713] - - - [40192, 512, 1, 256] - - [426, 72.759] - - - [8960, 512, 1, 256] - - [418, 53.98] - - - [16640, 4096, 1, 256] - - [46, 71.565] - - - [30976, 15104, 1, 256] - - [39, 74.383] - - - [27136, 8192, 1, 256] - - [28, 74.986] - - - [30208, 8192, 1, 256] - - [45, 74.616] - - - [21504, 512, 1, 256] - - [289, 67.646] - - - [9728, 5888, 1, 256] - - [32, 73.206] - - - [38912, 23040, 1, 256] - - [36, 75.609] - - - [7424, 1024, 1, 256] - - [348, 62.46] - - - [38656, 22528, 1, 256] - - [26, 74.572] - - - [26880, 512, 1, 256] - - [290, 69.26] - - - [29184, 13056, 1, 256] - - [45, 75.458] - - - [44032, 27904, 1, 256] - - [23, 75.577] - - - [38144, 8192, 1, 256] - - [39, 74.149] - - - [29952, 512, 1, 256] - - [296, 70.253] - - - [18432, 4096, 1, 256] - - [25, 72.431] - - - [28160, 12288, 1, 256] - - [26, 74.924] - - - [29696, 1024, 1, 256] - - [390, 77.395] - - - [39936, 1024, 1, 256] - - [56, 70.178] - - - [25600, 512, 1, 256] - - [428, 69.092] - - - [40448, 24320, 1, 256] - - [22, 75.043] - - - [40448, 512, 1, 256] - - [289, 73.834] - - - [7424, 3584, 1, 256] - - [390, 78.96] - - - [5376, 512, 1, 256] - - [299, 44.241] - - - [27136, 4096, 1, 256] - - [41, 73.399] - - - [35840, 19968, 1, 256] - - [25, 75.549] - - - [18944, 11008, 1, 256] - - [57, 75.428] - - - [34816, 18688, 1, 256] - - [25, 75.739] - - - [38400, 1024, 1, 256] - - [56, 70.053] - - - [36352, 20480, 1, 256] - - [26, 75.176] - - - [36608, 20736, 1, 256] - - [20, 74.849] - - - [28672, 1024, 1, 256] - - [296, 76.847] - - - [42496, 26624, 1, 256] - - [39, 75.087] - - - [31488, 15360, 1, 256] - - [25, 74.816] - - - [20992, 4096, 1, 256] - - [26, 73.044] - - - [12544, 512, 1, 256] - - [289, 60.652] - - - [24064, 8192, 1, 256] - - [45, 74.93] - - - [26880, 8192, 1, 256] - - [25, 74.524] - - - [4352, 512, 1, 256] - - [285, 42.087] - - - [7680, 1024, 1, 256] - - [350, 62.457] - - - [16128, 8192, 1, 256] - - [23, 74.553] - - - [39168, 8192, 1, 256] - - [26, 74.205] - - - [29440, 4096, 1, 256] - - [45, 72.927] - - - [33536, 4096, 1, 256] - - [39, 73.191] - - - [33024, 17152, 1, 256] - - [45, 75.207] - - - [34816, 18944, 1, 256] - - [23, 75.801] - - - [22016, 512, 1, 256] - - [290, 67.391] - - - [14848, 6912, 1, 256] - - [23, 74.685] - - - [20736, 12800, 1, 256] - - [23, 75.324] - - - [32256, 16128, 1, 256] - - [28, 75.679] - - - [7680, 512, 1, 256] - - [293, 52.405] - - - [19968, 12288, 1, 256] - - [39, 75.413] - - - [29184, 4096, 1, 256] - - [72, 73.343] - - - [15616, 4096, 1, 256] - - [27, 72.242] - - - [44544, 28672, 1, 256] - - [20, 74.533] - - - [26112, 4096, 1, 256] - - [63, 73.758] - - - [26624, 10752, 1, 256] - - [27, 75.829] - - - [15104, 4096, 1, 256] - - [25, 71.987] - - - [23296, 4096, 1, 256] - - [27, 72.489] - - - [37888, 22016, 1, 256] - - [27, 75.62] - - - [11520, 7680, 1, 256] - - [57, 73.86] - - - [41728, 1024, 1, 256] - - [25, 69.281] - - - [2304, 1792, 1, 256] - - [418, 54.815] - - - [34048, 17920, 1, 256] - - [22, 74.538] - - - [1536, 768, 1, 256] - - [103, 40.855] - - - [33280, 8192, 1, 256] - - [22, 74.883] - - - [11264, 1024, 1, 256] - - [357, 70.476] - - - [21760, 1024, 1, 256] - - [348, 73.617] - - - [18432, 10496, 1, 256] - - [25, 75.833] - - - [41216, 4096, 1, 256] - - [41, 73.138] - - - [41472, 25344, 1, 256] - - [35, 75.153] - - - [17408, 1024, 1, 256] - - [296, 72.212] - - - [19456, 1024, 1, 256] - - [299, 75.373] - - - [36096, 19968, 1, 256] - - [39, 74.202] - - - [8704, 512, 1, 256] - - [299, 53.56] - - - [30464, 1024, 1, 256] - - [269, 77.416] - - - [8192, 1024, 1, 256] - - [295, 61.48] - - - [11520, 512, 1, 256] - - [298, 57.237] - - - [44544, 512, 1, 256] - - [307, 73.821] - - - [20736, 4096, 1, 256] - - [23, 72.521] - - - [42752, 8192, 1, 256] - - [27, 74.408] - - - [39936, 512, 1, 256] - - [289, 73.199] - - - [42496, 26368, 1, 256] - - [25, 75.25] - - - [28672, 4096, 1, 256] - - [35, 72.526] - - - [35840, 8192, 1, 256] - - [27, 74.949] - - - [17664, 1024, 1, 256] - - [300, 71.901] - - - [21248, 4096, 1, 256] - - [35, 72.544] - - - [1280, 768, 1, 256] - - [103, 35.43] - - - [28160, 512, 1, 256] - - [315, 69.999] - - - [34304, 18176, 1, 256] - - [40, 75.541] - - - [19200, 11520, 1, 256] - - [25, 75.066] - - - [25856, 9728, 1, 256] - - [25, 74.614] - - - [35328, 19200, 1, 256] - - [49, 75.548] - - - [29440, 8192, 1, 256] - - [22, 74.577] - - - [20992, 13056, 1, 256] - - [23, 75.803] - - - [21760, 512, 1, 256] - - [357, 68.067] - - - [12800, 512, 1, 256] - - [422, 60.175] - - - [28416, 12288, 1, 256] - - [28, 74.006] - - - [29696, 13568, 1, 256] - - [25, 75.794] - - - [21504, 4096, 1, 256] - - [26, 72.907] - - - [30464, 14592, 1, 256] - - [45, 74.281] - - - [13056, 5120, 1, 256] - - [47, 73.164] - - - [34560, 4096, 1, 256] - - [27, 72.865] - - - [32768, 16896, 1, 256] - - [81, 59.383] - - - [13824, 5888, 1, 256] - - [20, 73.544] - - - [33024, 8192, 1, 256] - - [41, 74.541] - - - [14080, 4096, 1, 256] - - [45, 70.417] - - - [43008, 1024, 1, 256] - - [34, 70.968] - - - [31744, 1024, 1, 256] - - [390, 77.827] - - - [11008, 512, 1, 256] - - [291, 55.805] - - - [24832, 8192, 1, 256] - - [29, 74.582] - - - [43776, 512, 1, 256] - - [272, 69.54] - - - [24064, 1024, 1, 256] - - [289, 75.64] - - - [12800, 4096, 1, 256] - - [72, 71.929] - - - [19456, 11776, 1, 256] - - [23, 75.91] - - - [22528, 14848, 1, 256] - - [25, 75.994] - - - [30208, 14080, 1, 256] - - [22, 75.797] - - - [40704, 1024, 1, 256] - - [70, 69.656] - - - [35584, 4096, 1, 256] - - [41, 73.157] - - - [26112, 8192, 1, 256] - - [27, 74.852] - - - [9472, 5632, 1, 256] - - [32, 72.733] - - - [15616, 512, 1, 256] - - [293, 61.163] - - - [34816, 4096, 1, 256] - - [27, 73.248] - - - [31232, 4096, 1, 256] - - [39, 73.472] - - - [9728, 1024, 1, 256] - - [315, 67.139] - - - [13312, 1024, 1, 256] - - [390, 69.46] - - - [20224, 1024, 1, 256] - - [299, 73.687] - - - [4864, 512, 1, 256] - - [385, 45.023] - - - [34304, 4096, 1, 256] - - [40, 73.894] - - - [43776, 1024, 1, 256] - - [74, 67.204] - - - [37120, 8192, 1, 256] - - [23, 74.285] - - - [33792, 512, 1, 256] - - [320, 71.663] - - - [42496, 512, 1, 256] - - [357, 73.4] - - - [9216, 512, 1, 256] - - [291, 56.305] - - - [14336, 4096, 1, 256] - - [25, 72.202] - - - [43008, 27136, 1, 256] - - [20, 75.512] - - - [35840, 512, 1, 256] - - [348, 70.791] - - - [40960, 25088, 1, 256] - - [36, 67.979] - - - [17408, 512, 1, 256] - - [291, 63.161] - - - [12288, 4096, 1, 256] - - [25, 71.721] - - - [6656, 512, 1, 256] - - [392, 46.767] - - - [40960, 24832, 1, 256] - - [36, 68.288] - - - [39168, 23040, 1, 256] - - [26, 74.625] - - - [512, 1, 1, 128] - - [183, 0.025] - - - [384, 1, 1, 384] - - [95, 0.018] - - - [256, 1, 1, 256] - - [183, 0.017] - - - [128, 1, 1, 128] - - [183, 0.006] - - - [640, 1, 1, 128] - - [183, 0.032] - - - [1, 128, 1, 256] - - [183, 0.009] - - - [512, 128, 1, 256] - - [95, 4.27] - - - [2049, 128, 1, 256] - - [160, 14.336] - - - [49, 128, 1, 256] - - [183, 0.202] - - - [1537, 128, 1, 256] - - [164, 11.345] - - - [257, 128, 1, 256] - - [139, 1.888] - - - [9728, 128, 1, 256] - - [103, 42.487] - - - [3840, 128, 1, 256] - - [96, 23.522] - - - [1280, 128, 1, 256] - - [95, 10.616] - - - [7168, 128, 1, 256] - - [90, 32.965] - - - [6656, 128, 1, 256] - - [346, 24.673] - - - [2561, 128, 1, 256] - - [100, 17.834] - - - [6912, 128, 1, 256] - - [90, 38.946] - - - [2048, 128, 1, 256] - - [98, 15.996] - - - [2304, 128, 1, 256] - - [98, 17.807] - - - [1536, 128, 1, 256] - - [147, 12.739] - - - [4864, 128, 1, 256] - - [90, 29.428] - - - [8448, 128, 1, 256] - - [90, 38.138] - - - [3072, 128, 1, 256] - - [98, 23.02] - - - [3329, 128, 1, 256] - - [100, 22.543] - - - [3328, 128, 1, 256] - - [92, 24.565] - - - [8960, 128, 1, 256] - - [91, 39.721] - - - [9216, 128, 1, 256] - - [165, 38.761] - - - [2817, 128, 1, 256] - - [131, 19.524] - - - [6400, 128, 1, 256] - - [151, 37.641] - - - [561, 128, 1, 256] - - [100, 3.87] - - - [2816, 128, 1, 256] - - [96, 21.21] - - - [3073, 128, 1, 256] - - [141, 20.073] - - - [2097, 128, 1, 256] - - [100, 14.071] - - - [768, 128, 1, 256] - - [147, 6.405] - - - [9984, 128, 1, 256] - - [96, 42.97] - - - [3584, 128, 1, 256] - - [91, 22.61] - - - [817, 128, 1, 256] - - [109, 5.636] - - - [5632, 128, 1, 256] - - [91, 33.525] - - - [9472, 128, 1, 256] - - [96, 41.248] - - - [2305, 128, 1, 256] - - [109, 15.9] - - - [1329, 128, 1, 256] - - [100, 9.168] - - - [5888, 128, 1, 256] - - [92, 34.907] - - - [7680, 128, 1, 256] - - [256, 26.499] - - - [4608, 128, 1, 256] - - [96, 28.462] - - - [2353, 128, 1, 256] - - [146, 15.789] - - - [5120, 128, 1, 256] - - [90, 30.601] - - - [769, 128, 1, 256] - - [164, 5.733] - - - [1792, 128, 1, 256] - - [256, 8.2] - - - [1073, 128, 1, 256] - - [100, 7.437] - - - [513, 128, 1, 256] - - [113, 3.806] - - - [4096, 128, 1, 256] - - [92, 25.3] - - - [7424, 128, 1, 256] - - [105, 34.356] - - - [4352, 128, 1, 256] - - [92, 26.77] - - - [1793, 128, 1, 256] - - [100, 12.725] - - - [8192, 128, 1, 256] - - [166, 35.992] - - - [1281, 128, 1, 256] - - [111, 9.551] - - - [305, 128, 1, 256] - - [183, 1.365] - - - [2560, 128, 1, 256] - - [98, 19.481] - - - [2609, 128, 1, 256] - - [100, 17.831] - - - [1585, 128, 1, 256] - - [131, 10.783] - - - [8704, 128, 1, 256] - - [91, 37.244] - - - [10240, 128, 1, 256] - - [347, 32.667] - - - [256, 128, 1, 256] - - [97, 2.135] - - - [1025, 128, 1, 256] - - [164, 7.604] - - - [2865, 128, 1, 256] - - [124, 19.051] - - - [5376, 128, 1, 256] - - [92, 32.001] - - - [1841, 128, 1, 256] - - [146, 12.41] - - - [7936, 128, 1, 256] - - [90, 36.158] - - - [6144, 128, 1, 256] - - [92, 35.292] - - - [1024, 128, 1, 256] - - [97, 8.54] - - - [36096, 1281, 1, 256] - - [83, 63.047] - - - [38656, 2816, 1, 256] - - [72, 72.503] - - - [35072, 2048, 1, 256] - - [72, 71.133] - - - [39424, 2865, 1, 256] - - [52, 70.508] - - - [39168, 3328, 1, 256] - - [26, 72.947] - - - [36096, 2865, 1, 256] - - [49, 68.618] - - - [39216, 5632, 1, 256] - - [56, 62.951] - - - [38144, 6144, 1, 256] - - [25, 74.027] - - - [35328, 3072, 1, 256] - - [45, 73.434] - - - [39936, 256, 1, 256] - - [268, 62.202] - - - [36864, 3328, 1, 256] - - [27, 72.842] - - - [39168, 6144, 1, 256] - - [39, 73.921] - - - [36352, 4352, 1, 256] - - [27, 74.346] - - - [37680, 10240, 1, 256] - - [39, 62.962] - - - [38144, 256, 1, 256] - - [309, 63.217] - - - [37632, 1281, 1, 256] - - [72, 63.847] - - - [35632, 1792, 1, 256] - - [56, 62.655] - - - [36096, 4096, 1, 256] - - [39, 72.242] - - - [36144, 2816, 1, 256] - - [34, 62.834] - - - [36352, 256, 1, 256] - - [309, 62.698] - - - [35888, 2865, 1, 256] - - [51, 61.552] - - - [38912, 1280, 1, 256] - - [27, 71.695] - - - [37120, 3072, 1, 256] - - [23, 73.092] - - - [38448, 10240, 1, 256] - - [59, 62.59] - - - [39936, 3328, 1, 256] - - [25, 73.571] - - - [39168, 10240, 1, 256] - - [26, 74.409] - - - [39680, 3329, 1, 256] - - [52, 69.552] - - - [37168, 2865, 1, 256] - - [51, 61.603] - - - [38144, 5888, 1, 256] - - [35, 73.803] - - - [37120, 1281, 1, 256] - - [72, 64.097] - - - [37376, 10240, 1, 256] - - [23, 74.986] - - - [38704, 5120, 1, 256] - - [42, 62.977] - - - [39168, 5376, 1, 256] - - [26, 73.875] - - - [38656, 2865, 1, 256] - - [52, 69.906] - - - [37376, 3584, 1, 256] - - [23, 74.017] - - - [35072, 6144, 1, 256] - - [23, 74.189] - - - [39936, 6144, 1, 256] - - [35, 74.671] - - - [37632, 5376, 1, 256] - - [23, 73.89] - - - [36352, 2304, 1, 256] - - [23, 73.025] - - - [35840, 2048, 1, 256] - - [80, 71.487] - - - [36608, 1280, 1, 256] - - [23, 70.589] - - - [39936, 1792, 1, 256] - - [25, 73.01] - - - [36608, 3329, 1, 256] - - [52, 69.48] - - - [35072, 3329, 1, 256] - - [52, 69.553] - - - [37168, 3584, 1, 256] - - [42, 62.89] - - - [36096, 1792, 1, 256] - - [29, 70.106] - - - [39424, 3329, 1, 256] - - [39, 70.143] - - - [39424, 2048, 1, 256] - - [39, 71.915] - - - [39984, 2865, 1, 256] - - [34, 61.483] - - - [38448, 256, 1, 256] - - [422, 58.498] - - - [35584, 256, 1, 256] - - [266, 61.322] - - - [36608, 10240, 1, 256] - - [27, 74.583] - - - [38960, 5376, 1, 256] - - [26, 62.631] - - - [36352, 2048, 1, 256] - - [39, 71.757] - - - [39680, 1281, 1, 256] - - [25, 64.293] - - - [36608, 2304, 1, 256] - - [27, 72.296] - - - [39936, 1280, 1, 256] - - [25, 71.595] - - - [39680, 5376, 1, 256] - - [35, 74.025] - - - [35584, 10240, 1, 256] - - [22, 74.399] - - - [36864, 512, 1, 256] - - [357, 70.145] - - - [39424, 2816, 1, 256] - - [49, 73.142] - - - [35840, 2816, 1, 256] - - [23, 73.659] - - - [38192, 2816, 1, 256] - - [34, 62.921] - - - [35584, 2048, 1, 256] - - [39, 70.786] - - - [37936, 2865, 1, 256] - - [27, 61.41] - - - [39936, 2865, 1, 256] - - [27, 70.99] - - - [38656, 10240, 1, 256] - - [39, 74.41] - - - [36608, 2048, 1, 256] - - [59, 70.999] - - - [35120, 2816, 1, 256] - - [51, 63.338] - - - [39424, 5888, 1, 256] - - [23, 74.33] - - - [37680, 2816, 1, 256] - - [34, 63.798] - - - [36096, 6144, 1, 256] - - [39, 73.136] - - - [38144, 1281, 1, 256] - - [84, 64.27] - - - [37632, 2048, 1, 256] - - [59, 70.581] - - - [39680, 256, 1, 256] - - [270, 62.288] - - - [37680, 3840, 1, 256] - - [51, 63.791] - - - [39168, 2816, 1, 256] - - [39, 72.679] - - - [38192, 2865, 1, 256] - - [51, 60.856] - - - [38912, 4608, 1, 256] - - [25, 74.158] - - - [37120, 2048, 1, 256] - - [39, 71.342] - - - [35376, 1536, 1, 256] - - [34, 61.119] - - - [38448, 4864, 1, 256] - - [34, 63.358] - - - [38192, 10240, 1, 256] - - [72, 62.37] - - - [37632, 2816, 1, 256] - - [27, 72.628] - - - [39424, 1024, 1, 256] - - [74, 70.073] - - - [39168, 256, 1, 256] - - [270, 62.262] - - - [39984, 6144, 1, 256] - - [39, 63.37] - - - [38144, 4608, 1, 256] - - [35, 73.45] - - - [35840, 2865, 1, 256] - - [25, 71.032] - - - [36352, 6144, 1, 256] - - [39, 74.483] - - - [36864, 768, 1, 256] - - [390, 74.766] - - - [37888, 3328, 1, 256] - - [23, 73.568] - - - [36912, 3328, 1, 256] - - [39, 61.472] - - - [37120, 3584, 1, 256] - - [23, 73.471] - - - [38912, 1281, 1, 256] - - [25, 64.419] - - - [39472, 256, 1, 256] - - [293, 57.324] - - - [39936, 1281, 1, 256] - - [26, 64.808] - - - [37376, 5120, 1, 256] - - [35, 74.278] - - - [37888, 2048, 1, 256] - - [26, 71.546] - - - [37632, 1280, 1, 256] - - [27, 70.731] - - - [35376, 2816, 1, 256] - - [56, 63.641] - - - [38656, 3329, 1, 256] - - [28, 69.446] - - - [36912, 256, 1, 256] - - [308, 57.683] - - - [39168, 768, 1, 256] - - [299, 78.242] - - - [37424, 256, 1, 256] - - [392, 57.831] - - - [38448, 2816, 1, 256] - - [51, 63.43] - - - [35840, 3840, 1, 256] - - [23, 74.545] - - - [38912, 2865, 1, 256] - - [20, 71.049] - - - [36096, 1280, 1, 256] - - [55, 68.901] - - - [35328, 1024, 1, 256] - - [299, 79.772] - - - [39680, 3328, 1, 256] - - [27, 72.712] - - - [36352, 2816, 1, 256] - - [23, 73.066] - - - [38912, 256, 1, 256] - - [392, 64.297] - - - [39424, 3328, 1, 256] - - [26, 73.378] - - - [35888, 2816, 1, 256] - - [56, 63.218] - - - [36096, 2816, 1, 256] - - [62, 71.255] - - - [38960, 10240, 1, 256] - - [39, 62.954] - - - [35840, 3584, 1, 256] - - [25, 74.213] - - - [39424, 5120, 1, 256] - - [35, 74.236] - - - [37376, 1024, 1, 256] - - [299, 80.634] - - - [37632, 4096, 1, 256] - - [25, 72.873] - - - [36400, 2865, 1, 256] - - [51, 61.927] - - - [36144, 2560, 1, 256] - - [51, 63.37] - - - [36864, 1281, 1, 256] - - [25, 63.521] - - - [39424, 5376, 1, 256] - - [35, 74.471] - - - [36400, 2816, 1, 256] - - [56, 63.396] - - - [38656, 6144, 1, 256] - - [28, 73.826] - - - [37888, 5632, 1, 256] - - [25, 75.071] - - - [36912, 2865, 1, 256] - - [25, 59.852] - - - [38656, 4352, 1, 256] - - [28, 73.699] - - - [37632, 1536, 1, 256] - - [35, 70.658] - - - [35072, 2865, 1, 256] - - [27, 70.225] - - - [35888, 2304, 1, 256] - - [34, 63.831] - - - [38912, 3329, 1, 256] - - [52, 70.513] - - - [37680, 4096, 1, 256] - - [59, 62.597] - - - [38400, 6144, 1, 256] - - [26, 74.375] - - - [37888, 3840, 1, 256] - - [25, 74.613] - - - [36608, 3328, 1, 256] - - [26, 72.821] - - - [35328, 256, 1, 256] - - [308, 60.999] - - - [36096, 3329, 1, 256] - - [26, 68.384] - - - [37888, 5888, 1, 256] - - [35, 74.629] - - - [36864, 3329, 1, 256] - - [52, 69.763] - - - [35632, 256, 1, 256] - - [291, 57.841] - - - [38656, 4864, 1, 256] - - [45, 73.807] - - - [37888, 2816, 1, 256] - - [27, 73.767] - - - [37120, 3328, 1, 256] - - [26, 72.974] - - - [35328, 1536, 1, 256] - - [29, 70.794] - - - [35328, 1280, 1, 256] - - [55, 70.894] - - - [35888, 10240, 1, 256] - - [26, 63.793] - - - [36400, 10240, 1, 256] - - [39, 62.71] - - - [35072, 10240, 1, 256] - - [23, 74.629] - - - [39680, 2816, 1, 256] - - [35, 72.727] - - - [35584, 3329, 1, 256] - - [71, 69.36] - - - [36656, 256, 1, 256] - - [325, 58.723] - - - [38144, 4096, 1, 256] - - [72, 73.033] - - - [39936, 2816, 1, 256] - - [23, 73.724] - - - [36864, 3072, 1, 256] - - [23, 73.23] - - - [37936, 2816, 1, 256] - - [51, 63.534] - - - [37632, 3584, 1, 256] - - [23, 73.496] - - - [39984, 10240, 1, 256] - - [39, 63.649] - - - [38656, 512, 1, 256] - - [312, 71.803] - - - [35328, 10240, 1, 256] - - [49, 74.983] - - - [36096, 2048, 1, 256] - - [26, 69.762] - - - [37120, 4864, 1, 256] - - [23, 73.988] - - - [35840, 10240, 1, 256] - - [27, 75.284] - - - [39680, 5632, 1, 256] - - [35, 74.307] - - - [38144, 4352, 1, 256] - - [27, 73.886] - - - [36400, 2560, 1, 256] - - [34, 63.434] - - - [35840, 3329, 1, 256] - - [52, 70.438] - - - [37424, 10240, 1, 256] - - [39, 62.738] - - - [38912, 10240, 1, 256] - - [23, 75.232] - - - [35072, 768, 1, 256] - - [289, 75.31] - - - [36096, 3840, 1, 256] - - [52, 72.331] - - - [36656, 3072, 1, 256] - - [25, 62.236] - - - [39680, 1536, 1, 256] - - [30, 70.578] - - - [36656, 2865, 1, 256] - - [34, 61.222] - - - [38912, 512, 1, 256] - - [314, 72.382] - - - [38400, 256, 1, 256] - - [338, 63.156] - - - [38704, 10240, 1, 256] - - [26, 62.954] - - - [38912, 5376, 1, 256] - - [23, 74.867] - - - [35120, 256, 1, 256] - - [392, 58.523] - - - [38656, 3328, 1, 256] - - [26, 72.917] - - - [37888, 1536, 1, 256] - - [35, 71.305] - - - [39216, 5376, 1, 256] - - [34, 63.737] - - - [37376, 3329, 1, 256] - - [30, 70.049] - - - [37680, 256, 1, 256] - - [298, 59.226] - - - [39680, 6144, 1, 256] - - [27, 74.112] - - - [38400, 2865, 1, 256] - - [52, 70.515] - - - [36608, 2865, 1, 256] - - [52, 69.97] - - - [38912, 768, 1, 256] - - [390, 76.675] - - - [35584, 1792, 1, 256] - - [32, 71.502] - - - [39424, 256, 1, 256] - - [423, 61.065] - - - [36352, 1281, 1, 256] - - [72, 64.906] - - - [38400, 2048, 1, 256] - - [26, 71.748] - - - [38144, 3329, 1, 256] - - [52, 69.524] - - - [39680, 2048, 1, 256] - - [51, 69.64] - - - [38656, 256, 1, 256] - - [418, 64.056] - - - [39728, 2816, 1, 256] - - [56, 63.085] - - - [36352, 3329, 1, 256] - - [27, 70.011] - - - [38400, 10240, 1, 256] - - [23, 74.925] - - - [39984, 6400, 1, 256] - - [39, 63.356] - - - [37888, 4352, 1, 256] - - [35, 74.699] - - - [37888, 4096, 1, 256] - - [35, 73.073] - - - [35584, 1536, 1, 256] - - [33, 70.284] - - - [36096, 256, 1, 256] - - [293, 61.916] - - - [36864, 2048, 1, 256] - - [25, 68.862] - - - [36144, 2865, 1, 256] - - [51, 61.056] - - - [35584, 3584, 1, 256] - - [35, 73.116] - - - [35072, 1024, 1, 256] - - [299, 79.066] - - - [36352, 3328, 1, 256] - - [39, 73.516] - - - [39424, 1281, 1, 256] - - [84, 65.011] - - - [39728, 10240, 1, 256] - - [39, 62.94] - - - [37632, 2865, 1, 256] - - [52, 70.226] - - - [37168, 3328, 1, 256] - - [34, 63.032] - - - [37376, 5376, 1, 256] - - [35, 74.538] - - - [35328, 2865, 1, 256] - - [68, 70.435] - - - [35584, 6144, 1, 256] - - [23, 73.895] - - - [38704, 2816, 1, 256] - - [51, 63.386] - - - [36608, 3072, 1, 256] - - [35, 73.008] - - - [39680, 1280, 1, 256] - - [23, 70.822] - - - [35328, 1281, 1, 256] - - [72, 64.856] - - - [36608, 512, 1, 256] - - [298, 71.918] - - - [39936, 1536, 1, 256] - - [30, 71.413] - - - [39728, 5888, 1, 256] - - [56, 63.022] - - - [39168, 1281, 1, 256] - - [76, 64.281] - - - [37120, 256, 1, 256] - - [290, 62.834] - - - [38960, 2865, 1, 256] - - [25, 61.03] - - - [39168, 5120, 1, 256] - - [22, 73.651] - - - [36864, 256, 1, 256] - - [325, 62.318] - - - [36912, 2816, 1, 256] - - [26, 61.151] - - - [36096, 2304, 1, 256] - - [55, 70.752] - - - [35840, 3328, 1, 256] - - [25, 73.553] - - - [38704, 2865, 1, 256] - - [56, 61.61] - - - [38144, 1792, 1, 256] - - [35, 71.784] - - - [36608, 2560, 1, 256] - - [27, 72.736] - - - [35376, 10240, 1, 256] - - [39, 63.016] - - - [35840, 2304, 1, 256] - - [35, 73.298] - - - [35840, 1280, 1, 256] - - [35, 71.379] - - - [37376, 1280, 1, 256] - - [35, 71.316] - - - [35584, 3328, 1, 256] - - [40, 72.682] - - - [35584, 2865, 1, 256] - - [30, 69.888] - - - [39936, 10240, 1, 256] - - [25, 75.215] - - - [38912, 5120, 1, 256] - - [35, 74.698] - - - [37632, 3329, 1, 256] - - [30, 69.618] - - - [37888, 1792, 1, 256] - - [25, 72.793] - - - [36608, 1281, 1, 256] - - [26, 64.424] - - - [38192, 4352, 1, 256] - - [56, 62.726] - - - [39936, 2048, 1, 256] - - [39, 72.048] - - - [35072, 1281, 1, 256] - - [76, 64.274] - - - [39472, 2816, 1, 256] - - [51, 63.357] - - - [39728, 2865, 1, 256] - - [34, 61.192] - - - [38400, 2816, 1, 256] - - [23, 73.171] - - - [38400, 4608, 1, 256] - - [40, 73.988] - - - [39216, 10240, 1, 256] - - [59, 62.289] - - - [35072, 3072, 1, 256] - - [23, 72.885] - - - [38400, 4352, 1, 256] - - [27, 74.322] - - - [39216, 2816, 1, 256] - - [56, 63.182] - - - [35840, 1792, 1, 256] - - [27, 72.46] - - - [35632, 2048, 1, 256] - - [34, 62.969] - - - [38704, 256, 1, 256] - - [308, 57.79] - - - [37888, 3329, 1, 256] - - [27, 70.547] - - - [37888, 6144, 1, 256] - - [25, 74.7] - - - [37376, 6144, 1, 256] - - [39, 74.51] - - - [37376, 256, 1, 256] - - [360, 63.84] - - - [36400, 256, 1, 256] - - [308, 57.886] - - - [37936, 4096, 1, 256] - - [39, 62.768] - - - [38144, 10240, 1, 256] - - [35, 74.524] - - - [35376, 1792, 1, 256] - - [34, 62.315] - - - [37168, 10240, 1, 256] - - [59, 62.52] - - - [39984, 2816, 1, 256] - - [42, 63.532] - - - [37168, 2816, 1, 256] - - [51, 62.931] - - - [39424, 5632, 1, 256] - - [25, 74.685] - - - [36352, 1280, 1, 256] - - [36, 70.957] - - - [39680, 10240, 1, 256] - - [27, 74.481] - - - [38144, 3328, 1, 256] - - [63, 72.575] - - - [39168, 2048, 1, 256] - - [39, 71.015] - - - [35328, 6144, 1, 256] - - [41, 74.535] - - - [35632, 2865, 1, 256] - - [51, 61.238] - - - [36656, 10240, 1, 256] - - [39, 63.069] - - - [36608, 4352, 1, 256] - - [35, 73.899] - - - [35120, 2865, 1, 256] - - [42, 61.404] - - - [36608, 6144, 1, 256] - - [25, 74.021] - - - [37888, 2865, 1, 256] - - [35, 70.976] - - - [39168, 1024, 1, 256] - - [67, 69.668] - - - [38704, 4864, 1, 256] - - [42, 63.525] - - - [39168, 2865, 1, 256] - - [27, 69.967] - - - [38960, 5120, 1, 256] - - [26, 62.521] - - - [36864, 2816, 1, 256] - - [23, 73.037] - - - [38656, 1280, 1, 256] - - [29, 70.603] - - - [35584, 1281, 1, 256] - - [37, 64.143] - - - [39216, 2865, 1, 256] - - [34, 61.369] - - - [35120, 1280, 1, 256] - - [56, 62.736] - - - [36096, 3328, 1, 256] - - [40, 71.742] - - - [38912, 6144, 1, 256] - - [25, 74.651] - - - [37376, 3840, 1, 256] - - [27, 74.191] - - - [37424, 2816, 1, 256] - - [51, 63.705] - - - [36864, 10240, 1, 256] - - [25, 74.71] - - - [35328, 3328, 1, 256] - - [41, 73.449] - - - [37632, 5632, 1, 256] - - [35, 74.204] - - - [35072, 1536, 1, 256] - - [35, 70.539] - - - [36864, 2865, 1, 256] - - [27, 70.301] - - - [36864, 4608, 1, 256] - - [27, 73.415] - - - [37888, 1280, 1, 256] - - [32, 71.371] - - - [36864, 4864, 1, 256] - - [27, 74.336] - - - [37632, 256, 1, 256] - - [392, 62.695] - - - [38912, 2816, 1, 256] - - [25, 73.697] - - - [38656, 5120, 1, 256] - - [22, 73.766] - - - [35072, 1280, 1, 256] - - [47, 70.409] - - - [38400, 3329, 1, 256] - - [52, 70.087] - - - [35840, 1281, 1, 256] - - [78, 64.22] - - - [39680, 2865, 1, 256] - - [52, 70.163] - - - [38192, 256, 1, 256] - - [308, 57.994] - - - [37632, 10240, 1, 256] - - [35, 74.479] - - - [39984, 256, 1, 256] - - [293, 57.661] - - - [37424, 2865, 1, 256] - - [34, 61.628] - - - [37888, 256, 1, 256] - - [428, 63.503] - - - [36864, 6144, 1, 256] - - [35, 74.004] - - - [38656, 1281, 1, 256] - - [82, 64.467] - - - [37936, 256, 1, 256] - - [423, 58.431] - - - [39168, 4864, 1, 256] - - [28, 73.766] - - - [35840, 256, 1, 256] - - [418, 62.347] - - - [37888, 10240, 1, 256] - - [27, 75.274] - - - [39728, 6144, 1, 256] - - [26, 62.67] - - - [39680, 5888, 1, 256] - - [35, 73.797] - - - [38144, 2816, 1, 256] - - [25, 72.629] - - - [39728, 256, 1, 256] - - [357, 56.776] - - - [37376, 2816, 1, 256] - - [23, 73.31] - - - [36352, 2865, 1, 256] - - [25, 70.628] - - - [39216, 256, 1, 256] - - [332, 57.779] - - - [37888, 1281, 1, 256] - - [72, 64.407] - - - [39472, 10240, 1, 256] - - [26, 63.129] - - - [37376, 2048, 1, 256] - - [41, 71.636] - - - [36096, 10240, 1, 256] - - [52, 73.83] - - - [35584, 1280, 1, 256] - - [47, 70.281] - - - [39168, 5632, 1, 256] - - [45, 73.916] - - - [39936, 5632, 1, 256] - - [25, 74.951] - - - [35072, 256, 1, 256] - - [266, 63.027] - - - [35376, 2865, 1, 256] - - [51, 61.863] - - - [38400, 4864, 1, 256] - - [35, 74.544] - - - [35888, 256, 1, 256] - - [297, 59.191] - - - [35072, 3328, 1, 256] - - [39, 72.749] - - - [37936, 10240, 1, 256] - - [39, 63.616] - - - [36352, 10240, 1, 256] - - [39, 75.006] - - - [38656, 2048, 1, 256] - - [39, 71.141] - - - [35632, 2816, 1, 256] - - [34, 63.573] - - - [36912, 10240, 1, 256] - - [26, 62.221] - - - [39936, 5888, 1, 256] - - [27, 74.727] - - - [38448, 2865, 1, 256] - - [42, 61.645] - - - [38144, 3840, 1, 256] - - [25, 73.821] - - - [37632, 6144, 1, 256] - - [35, 74.0] - - - [37376, 3328, 1, 256] - - [39, 73.555] - - - [36608, 2816, 1, 256] - - [25, 72.815] - - - [36912, 3072, 1, 256] - - [26, 60.786] - - - [37120, 2816, 1, 256] - - [26, 72.663] - - - [38144, 2865, 1, 256] - - [52, 70.012] - - - [38912, 2048, 1, 256] - - [39, 70.594] - - - [38192, 4608, 1, 256] - - [56, 63.111] - - - [37120, 5120, 1, 256] - - [25, 74.028] - - - [38400, 3328, 1, 256] - - [26, 73.36] - - - [35632, 10240, 1, 256] - - [26, 62.964] - - - [38912, 4864, 1, 256] - - [25, 74.806] - - - [37120, 10240, 1, 256] - - [25, 74.59] - - - [37120, 3329, 1, 256] - - [30, 69.606] - - - [35840, 6144, 1, 256] - - [25, 74.653] - - - [38400, 1281, 1, 256] - - [40, 64.99] - - - [36144, 10240, 1, 256] - - [39, 62.723] - - - [38144, 1280, 1, 256] - - [33, 70.558] - - - [39424, 10240, 1, 256] - - [25, 74.889] - - - [39424, 6144, 1, 256] - - [39, 74.434] - - - [39424, 1280, 1, 256] - - [32, 70.931] - - - [35328, 3329, 1, 256] - - [45, 70.11] - - - [39472, 5888, 1, 256] - - [51, 63.366] - - - [36352, 4096, 1, 256] - - [26, 73.815] - - - [38656, 4608, 1, 256] - - [26, 73.336] - - - [37168, 256, 1, 256] - - [390, 57.969] - - - [38144, 2048, 1, 256] - - [39, 71.056] - - - [35840, 1536, 1, 256] - - [35, 71.264] - - - [37120, 1280, 1, 256] - - [33, 70.842] - - - [37424, 3840, 1, 256] - - [34, 63.367] - - - [37424, 3584, 1, 256] - - [51, 63.462] - - - [36864, 2560, 1, 256] - - [20, 72.877] - - - [39936, 6400, 1, 256] - - [27, 75.294] - - - [36096, 2560, 1, 256] - - [55, 71.482] - - - [37120, 768, 1, 256] - - [299, 77.374] - - - [35328, 2048, 1, 256] - - [39, 71.615] - - - [36608, 4608, 1, 256] - - [26, 73.42] - - - [38400, 4096, 1, 256] - - [39, 73.807] - - - [35328, 2816, 1, 256] - - [49, 73.251] - - - [36144, 256, 1, 256] - - [304, 57.963] - - - [36608, 256, 1, 256] - - [313, 63.035] - - - [39168, 3329, 1, 256] - - [22, 69.417] - - - [38448, 4608, 1, 256] - - [56, 62.818] - - - [37632, 3328, 1, 256] - - [35, 72.662] - - - [37680, 2865, 1, 256] - - [51, 61.765] - - - [35120, 10240, 1, 256] - - [42, 62.831] - - - [37120, 6144, 1, 256] - - [39, 73.953] - - - [36656, 2816, 1, 256] - - [56, 63.274] - - - [39936, 3329, 1, 256] - - [52, 70.508] - - - [35328, 1792, 1, 256] - - [29, 72.387] - - - [35120, 1536, 1, 256] - - [34, 61.566] - - - [39472, 2865, 1, 256] - - [34, 61.639] - - - [37936, 4352, 1, 256] - - [23, 63.579] - - - [35888, 2048, 1, 256] - - [51, 62.81] - - - [37888, 3584, 1, 256] - - [25, 74.221] - - - [37376, 2865, 1, 256] - - [35, 70.609] - - - [36864, 1280, 1, 256] - - [53, 70.756] - - - [39472, 5632, 1, 256] - - [34, 63.752] - - - [37120, 1024, 1, 256] - - [299, 78.903] - - - [37120, 2865, 1, 256] - - [52, 70.15] - - - [38400, 1280, 1, 256] - - [32, 71.196] - - - [35584, 2816, 1, 256] - - [33, 72.373] - - - [37376, 1281, 1, 256] - - [74, 64.736] - - - [36352, 2560, 1, 256] - - [25, 73.332] - - - [36144, 2304, 1, 256] - - [42, 62.853] - - - [37632, 3840, 1, 256] - - [25, 73.728] - - - [38960, 2816, 1, 256] - - [27, 62.02] - - - [37376, 3072, 1, 256] - - [36, 73.241] - - - [35072, 2816, 1, 256] - - [25, 72.753] - - - [38912, 3328, 1, 256] - - [35, 73.576] - - - [38960, 256, 1, 256] - - [299, 57.433] - - - [35376, 256, 1, 256] - - [315, 57.47] - - - [39168, 1280, 1, 256] - - [25, 70.66] - - - [44032, 5888, 1, 256] - - [35, 74.602] - - - [40192, 2865, 1, 256] - - [23, 70.132] - - - [43312, 256, 1, 256] - - [360, 58.871] - - - [43520, 1280, 1, 256] - - [23, 71.545] - - - [41216, 2816, 1, 256] - - [23, 72.943] - - - [41520, 7936, 1, 256] - - [34, 63.037] - - - [43008, 2048, 1, 256] - - [26, 70.834] - - - [42496, 2048, 1, 256] - - [39, 72.118] - - - [40704, 3328, 1, 256] - - [39, 72.847] - - - [41776, 7936, 1, 256] - - [42, 63.282] - - - [40192, 1792, 1, 256] - - [25, 72.033] - - - [43520, 6144, 1, 256] - - [35, 74.517] - - - [42032, 2865, 1, 256] - - [34, 61.445] - - - [41472, 3329, 1, 256] - - [52, 70.049] - - - [41008, 7424, 1, 256] - - [26, 61.948] - - - [40448, 2865, 1, 256] - - [25, 70.49] - - - [41264, 2865, 1, 256] - - [42, 62.119] - - - [43312, 9728, 1, 256] - - [56, 62.838] - - - [40704, 2816, 1, 256] - - [25, 72.846] - - - [42544, 8704, 1, 256] - - [51, 62.483] - - - [40960, 7168, 1, 256] - - [36, 66.483] - - - [41216, 3329, 1, 256] - - [71, 69.564] - - - [41984, 6144, 1, 256] - - [25, 74.663] - - - [42240, 10240, 1, 256] - - [25, 74.534] - - - [42752, 2865, 1, 256] - - [52, 70.315] - - - [41216, 1280, 1, 256] - - [29, 71.322] - - - [40704, 7168, 1, 256] - - [23, 73.285] - - - [41216, 10240, 1, 256] - - [26, 74.599] - - - [40960, 256, 1, 256] - - [360, 59.816] - - - [40704, 2560, 1, 256] - - [23, 73.113] - - - [42752, 3329, 1, 256] - - [30, 69.627] - - - [43264, 3329, 1, 256] - - [52, 69.794] - - - [40192, 6144, 1, 256] - - [23, 74.232] - - - [43008, 10240, 1, 256] - - [23, 75.153] - - - [43520, 1281, 1, 256] - - [39, 65.119] - - - [42496, 8960, 1, 256] - - [27, 75.285] - - - [43312, 10240, 1, 256] - - [42, 62.292] - - - [44032, 6144, 1, 256] - - [23, 74.572] - - - [40192, 256, 1, 256] - - [293, 63.269] - - - [41984, 1536, 1, 256] - - [25, 71.539] - - - [41216, 768, 1, 256] - - [299, 78.305] - - - [40752, 256, 1, 256] - - [293, 58.784] - - - [44288, 1280, 1, 256] - - [55, 70.78] - - - [43520, 9216, 1, 256] - - [39, 74.767] - - - [42032, 8192, 1, 256] - - [39, 63.425] - - - [41728, 3584, 1, 256] - - [48, 71.854] - - - [40448, 1280, 1, 256] - - [25, 71.25] - - - [41216, 7168, 1, 256] - - [25, 73.415] - - - [42496, 1280, 1, 256] - - [23, 71.711] - - - [40448, 6656, 1, 256] - - [22, 74.479] - - - [40240, 256, 1, 256] - - [290, 58.373] - - - [41264, 2816, 1, 256] - - [34, 63.512] - - - [43264, 3328, 1, 256] - - [25, 72.954] - - - [43008, 9216, 1, 256] - - [23, 74.511] - - - [42240, 1281, 1, 256] - - [39, 64.663] - - - [42288, 2865, 1, 256] - - [51, 61.45] - - - [43008, 3328, 1, 256] - - [23, 73.593] - - - [40496, 256, 1, 256] - - [297, 58.686] - - - [43264, 8960, 1, 256] - - [27, 74.747] - - - [43056, 9472, 1, 256] - - [39, 62.988] - - - [40448, 3328, 1, 256] - - [41, 73.139] - - - [41776, 8192, 1, 256] - - [39, 62.693] - - - [40704, 6400, 1, 256] - - [23, 74.452] - - - [41984, 7680, 1, 256] - - [35, 75.275] - - - [43312, 9472, 1, 256] - - [42, 62.742] - - - [40192, 1280, 1, 256] - - [35, 71.066] - - - [43776, 5632, 1, 256] - - [39, 73.586] - - - [41984, 2865, 1, 256] - - [30, 71.09] - - - [40448, 2816, 1, 256] - - [25, 73.248] - - - [42240, 3328, 1, 256] - - [39, 73.03] - - - [42752, 2048, 1, 256] - - [72, 70.609] - - - [42240, 256, 1, 256] - - [293, 65.316] - - - [43008, 3329, 1, 256] - - [30, 70.532] - - - [44032, 5632, 1, 256] - - [23, 75.031] - - - [40192, 2048, 1, 256] - - [72, 71.401] - - - [41216, 256, 1, 256] - - [270, 62.343] - - - [44288, 9984, 1, 256] - - [25, 74.699] - - - [43008, 1280, 1, 256] - - [27, 71.961] - - - [41984, 2816, 1, 256] - - [23, 73.74] - - - [42752, 6144, 1, 256] - - [35, 74.06] - - - [43776, 3329, 1, 256] - - [26, 68.689] - - - [43008, 2865, 1, 256] - - [20, 71.08] - - - [43776, 9728, 1, 256] - - [26, 74.27] - - - [42240, 7936, 1, 256] - - [25, 74.473] - - - [41472, 7424, 1, 256] - - [23, 74.806] - - - [43776, 5376, 1, 256] - - [26, 73.395] - - - [43008, 6144, 1, 256] - - [25, 74.574] - - - [41216, 3072, 1, 256] - - [35, 73.182] - - - [42496, 8192, 1, 256] - - [35, 74.717] - - - [40704, 6144, 1, 256] - - [23, 74.159] - - - [44032, 3329, 1, 256] - - [52, 70.575] - - - [43520, 2048, 1, 256] - - [59, 71.975] - - - [43264, 2048, 1, 256] - - [72, 71.216] - - - [40448, 1281, 1, 256] - - [59, 64.998] - - - [40496, 2865, 1, 256] - - [42, 61.745] - - - [40448, 6144, 1, 256] - - [23, 74.257] - - - [41008, 10240, 1, 256] - - [26, 62.391] - - - [43056, 2865, 1, 256] - - [25, 60.62] - - - [43264, 1280, 1, 256] - - [27, 71.169] - - - [40192, 10240, 1, 256] - - [25, 74.615] - - - [41216, 7680, 1, 256] - - [25, 74.657] - - - [41008, 7168, 1, 256] - - [39, 62.179] - - - [44288, 2048, 1, 256] - - [26, 71.786] - - - [41472, 6144, 1, 256] - - [41, 74.195] - - - [43264, 2865, 1, 256] - - [52, 70.362] - - - [40448, 6912, 1, 256] - - [28, 74.755] - - - [41216, 6912, 1, 256] - - [35, 74.568] - - - [41984, 1792, 1, 256] - - [25, 72.966] - - - [40192, 1281, 1, 256] - - [72, 64.585] - - - [40960, 3329, 1, 256] - - [52, 62.664] - - - [41520, 10240, 1, 256] - - [72, 62.859] - - - [44032, 10240, 1, 256] - - [23, 75.084] - - - [43264, 2816, 1, 256] - - [35, 72.884] - - - [43008, 4608, 1, 256] - - [35, 74.106] - - - [43776, 1281, 1, 256] - - [59, 63.973] - - - [40240, 6656, 1, 256] - - [42, 63.007] - - - [43264, 9216, 1, 256] - - [27, 74.059] - - - [40704, 3329, 1, 256] - - [30, 69.77] - - - [42752, 3328, 1, 256] - - [35, 72.653] - - - [41984, 2048, 1, 256] - - [39, 71.786] - - - [44288, 3329, 1, 256] - - [26, 69.201] - - - [40192, 3328, 1, 256] - - [39, 72.906] - - - [40960, 10240, 1, 256] - - [36, 67.284] - - - [42496, 256, 1, 256] - - [272, 63.874] - - - [40496, 10240, 1, 256] - - [72, 62.619] - - - [40496, 6912, 1, 256] - - [51, 63.78] - - - [43776, 6144, 1, 256] - - [26, 73.627] - - - [40960, 1280, 1, 256] - - [23, 64.486] - - - [42288, 8704, 1, 256] - - [26, 62.698] - - - [42496, 3328, 1, 256] - - [26, 73.671] - - - [41216, 2865, 1, 256] - - [52, 70.173] - - - [42496, 3329, 1, 256] - - [30, 70.164] - - - [41984, 7936, 1, 256] - - [25, 75.243] - - - [41472, 1281, 1, 256] - - [59, 65.032] - - - [41776, 256, 1, 256] - - [418, 59.015] - - - [42752, 8960, 1, 256] - - [20, 74.694] - - - [41472, 7168, 1, 256] - - [23, 73.532] - - - [40240, 10240, 1, 256] - - [39, 62.723] - - - [41728, 1280, 1, 256] - - [57, 69.927] - - - [40752, 2865, 1, 256] - - [34, 61.872] - - - [40960, 2048, 1, 256] - - [25, 62.991] - - - [41472, 7680, 1, 256] - - [35, 74.81] - - - [41472, 10240, 1, 256] - - [23, 74.705] - - - [41264, 7680, 1, 256] - - [51, 63.108] - - - [42800, 8960, 1, 256] - - [51, 63.579] - - - [41728, 10240, 1, 256] - - [39, 74.303] - - - [44032, 3328, 1, 256] - - [27, 73.632] - - - [40704, 6912, 1, 256] - - [25, 74.49] - - - [41472, 2048, 1, 256] - - [46, 71.859] - - - [40960, 6144, 1, 256] - - [36, 67.029] - - - [43776, 3328, 1, 256] - - [26, 72.192] - - - [42496, 2865, 1, 256] - - [25, 70.925] - - - [40960, 3328, 1, 256] - - [36, 66.416] - - - [41728, 7936, 1, 256] - - [39, 73.453] - - - [41984, 3329, 1, 256] - - [23, 70.588] - - - [43008, 256, 1, 256] - - [304, 64.878] - - - [42240, 1280, 1, 256] - - [20, 70.817] - - - [43776, 10240, 1, 256] - - [26, 74.271] - - - [42752, 8448, 1, 256] - - [25, 74.38] - - - [42496, 1281, 1, 256] - - [39, 65.226] - - - [44032, 1536, 1, 256] - - [23, 71.765] - - - [40960, 2816, 1, 256] - - [20, 66.747] - - - [44288, 1792, 1, 256] - - [23, 71.649] - - - [43264, 1281, 1, 256] - - [72, 64.133] - - - [43008, 8704, 1, 256] - - [23, 75.246] - - - [41728, 1536, 1, 256] - - [55, 69.755] - - - [41728, 2048, 1, 256] - - [82, 69.934] - - - [43520, 9728, 1, 256] - - [25, 74.909] - - - [42032, 256, 1, 256] - - [370, 59.298] - - - [43776, 256, 1, 256] - - [307, 64.142] - - - [43008, 9472, 1, 256] - - [23, 75.471] - - - [44032, 1792, 1, 256] - - [35, 72.939] - - - [40704, 2865, 1, 256] - - [35, 70.238] - - - [42240, 1792, 1, 256] - - [25, 71.959] - - - [40704, 2304, 1, 256] - - [25, 72.693] - - - [42800, 9216, 1, 256] - - [26, 62.893] - - - [42240, 8704, 1, 256] - - [35, 74.597] - - - [42496, 6144, 1, 256] - - [23, 74.501] - - - [43568, 9728, 1, 256] - - [39, 63.01] - - - [40704, 2048, 1, 256] - - [26, 71.158] - - - [41472, 7936, 1, 256] - - [23, 74.77] - - - [42752, 2816, 1, 256] - - [25, 72.973] - - - [41008, 2865, 1, 256] - - [26, 58.776] - - - [40960, 6912, 1, 256] - - [36, 68.041] - - - [44032, 256, 1, 256] - - [313, 64.942] - - - [42496, 4352, 1, 256] - - [23, 74.491] - - - [42032, 8448, 1, 256] - - [26, 63.624] - - - [42752, 4608, 1, 256] - - [27, 73.474] - - - [44032, 1280, 1, 256] - - [25, 72.047] - - - [44288, 6144, 1, 256] - - [39, 74.003] - - - [42800, 2865, 1, 256] - - [34, 61.011] - - - [41008, 2816, 1, 256] - - [39, 61.755] - - - [41984, 8192, 1, 256] - - [23, 74.9] - - - [43264, 256, 1, 256] - - [418, 64.237] - - - [41728, 2865, 1, 256] - - [49, 68.836] - - - [43520, 5120, 1, 256] - - [23, 74.521] - - - [41984, 3584, 1, 256] - - [35, 74.444] - - - [41216, 3328, 1, 256] - - [26, 73.012] - - - [43520, 9472, 1, 256] - - [25, 75.262] - - - [43264, 9728, 1, 256] - - [23, 74.48] - - - [41728, 1281, 1, 256] - - [84, 64.039] - - - [40704, 1281, 1, 256] - - [39, 64.284] - - - [42288, 256, 1, 256] - - [298, 59.285] - - - [40960, 512, 1, 256] - - [25, 64.162] - - - [42752, 4352, 1, 256] - - [23, 74.11] - - - [40752, 10240, 1, 256] - - [26, 62.91] - - - [41728, 3328, 1, 256] - - [39, 72.265] - - - [43568, 2816, 1, 256] - - [34, 62.54] - - - [43008, 512, 1, 256] - - [350, 72.581] - - - [41216, 2048, 1, 256] - - [39, 71.246] - - - [42800, 256, 1, 256] - - [426, 58.162] - - - [43312, 2816, 1, 256] - - [34, 63.107] - - - [40192, 6400, 1, 256] - - [25, 74.557] - - - [41264, 7424, 1, 256] - - [34, 63.342] - - - [42544, 8960, 1, 256] - - [42, 63.039] - - - [41472, 256, 1, 256] - - [290, 64.647] - - - [42288, 10240, 1, 256] - - [39, 62.646] - - - [43520, 1024, 1, 256] - - [51, 70.335] - - - [42288, 8448, 1, 256] - - [42, 62.928] - - - [43776, 9472, 1, 256] - - [72, 74.186] - - - [43008, 1281, 1, 256] - - [23, 64.667] - - - [43008, 8960, 1, 256] - - [36, 75.419] - - - [41728, 256, 1, 256] - - [268, 63.634] - - - [41520, 7680, 1, 256] - - [26, 62.77] - - - [42240, 3329, 1, 256] - - [30, 69.539] - - - [41472, 2816, 1, 256] - - [45, 72.969] - - - [41216, 6144, 1, 256] - - [25, 74.096] - - - [40752, 2816, 1, 256] - - [42, 62.707] - - - [42496, 8704, 1, 256] - - [27, 75.149] - - - [40448, 6400, 1, 256] - - [35, 74.798] - - - [44032, 1281, 1, 256] - - [23, 64.951] - - - [41472, 1024, 1, 256] - - [61, 70.004] - - - [41216, 7424, 1, 256] - - [23, 74.529] - - - [43312, 2865, 1, 256] - - [34, 61.386] - - - [40960, 768, 1, 256] - - [311, 68.396] - - - [40240, 2865, 1, 256] - - [42, 61.619] - - - [43264, 768, 1, 256] - - [299, 77.248] - - - [40192, 3329, 1, 256] - - [35, 69.591] - - - [42800, 10240, 1, 256] - - [39, 62.897] - - - [42752, 512, 1, 256] - - [291, 71.607] - - - [40752, 6912, 1, 256] - - [34, 63.427] - - - [42240, 8192, 1, 256] - - [26, 74.303] - - - [42288, 2816, 1, 256] - - [34, 62.835] - - - [40960, 2865, 1, 256] - - [52, 63.581] - - - [42800, 2816, 1, 256] - - [56, 62.968] - - - [42496, 2816, 1, 256] - - [27, 73.572] - - - [41728, 7680, 1, 256] - - [26, 73.673] - - - [42240, 8448, 1, 256] - - [39, 74.494] - - - [41984, 1281, 1, 256] - - [72, 64.815] - - - [41984, 3328, 1, 256] - - [25, 73.564] - - - [40240, 6400, 1, 256] - - [42, 63.269] - - - [44288, 256, 1, 256] - - [425, 65.693] - - - [42496, 4096, 1, 256] - - [39, 73.815] - - - [43520, 3329, 1, 256] - - [52, 70.231] - - - [44288, 5888, 1, 256] - - [72, 73.788] - - - [42752, 1281, 1, 256] - - [35, 64.042] - - - [43776, 9984, 1, 256] - - [39, 73.921] - - - [41008, 256, 1, 256] - - [291, 56.583] - - - [40960, 1281, 1, 256] - - [36, 57.818] - - - [40704, 6656, 1, 256] - - [27, 74.314] - - - [40192, 2816, 1, 256] - - [25, 72.942] - - - [43264, 10240, 1, 256] - - [25, 74.641] - - - [44032, 9984, 1, 256] - - [25, 75.513] - - - [43520, 2865, 1, 256] - - [25, 70.933] - - - [42240, 3840, 1, 256] - - [23, 73.925] - - - [43056, 9216, 1, 256] - - [26, 62.923] - - - [43520, 10240, 1, 256] - - [23, 74.931] - - - [42544, 10240, 1, 256] - - [39, 62.521] - - - [40448, 2304, 1, 256] - - [27, 72.902] - - - [40704, 1280, 1, 256] - - [55, 70.86] - - - [43520, 2816, 1, 256] - - [23, 73.456] - - - [43520, 5376, 1, 256] - - [25, 74.646] - - - [41984, 256, 1, 256] - - [424, 63.793] - - - [43776, 1280, 1, 256] - - [59, 68.773] - - - [43568, 2865, 1, 256] - - [51, 61.386] - - - [41520, 256, 1, 256] - - [297, 59.04] - - - [41472, 3328, 1, 256] - - [41, 73.35] - - - [40192, 6656, 1, 256] - - [23, 74.325] - - - [40448, 2048, 1, 256] - - [26, 71.758] - - - [41520, 2816, 1, 256] - - [56, 63.286] - - - [43520, 9984, 1, 256] - - [23, 75.254] - - - [42544, 2865, 1, 256] - - [56, 61.581] - - - [42240, 2048, 1, 256] - - [39, 71.349] - - - [41472, 1280, 1, 256] - - [65, 71.256] - - - [40192, 5888, 1, 256] - - [25, 73.973] - - - [42240, 2865, 1, 256] - - [25, 70.024] - - - [41984, 10240, 1, 256] - - [23, 75.156] - - - [41264, 10240, 1, 256] - - [25, 62.503] - - - [42752, 10240, 1, 256] - - [25, 74.606] - - - [41216, 1024, 1, 256] - - [70, 69.891] - - - [41776, 10240, 1, 256] - - [39, 62.884] - - - [40960, 7424, 1, 256] - - [36, 67.903] - - - [40960, 2560, 1, 256] - - [20, 66.523] - - - [41216, 1281, 1, 256] - - [82, 64.548] - - - [41984, 1280, 1, 256] - - [25, 71.781] - - - [40448, 3329, 1, 256] - - [30, 69.924] - - - [41776, 2816, 1, 256] - - [51, 63.494] - - - [40704, 256, 1, 256] - - [270, 62.686] - - - [43264, 4864, 1, 256] - - [27, 74.253] - - - [42240, 6144, 1, 256] - - [25, 74.187] - - - [43520, 3328, 1, 256] - - [39, 73.577] - - - [42752, 256, 1, 256] - - [350, 63.935] - - - [40752, 7168, 1, 256] - - [26, 62.197] - - - [43776, 1536, 1, 256] - - [84, 67.546] - - - [42032, 10240, 1, 256] - - [26, 63.661] - - - [43008, 4864, 1, 256] - - [23, 74.853] - - - [40704, 10240, 1, 256] - - [25, 74.63] - - - [44288, 1281, 1, 256] - - [72, 64.844] - - - [41520, 2865, 1, 256] - - [42, 61.695] - - - [41264, 256, 1, 256] - - [290, 58.707] - - - [40496, 6656, 1, 256] - - [42, 63.334] - - - [42240, 4096, 1, 256] - - [39, 73.206] - - - [43568, 9984, 1, 256] - - [51, 63.212] - - - [43264, 9472, 1, 256] - - [25, 74.68] - - - [43008, 768, 1, 256] - - [390, 77.023] - - - [43776, 2816, 1, 256] - - [39, 72.397] - - - [43008, 2816, 1, 256] - - [20, 73.792] - - - [41984, 8448, 1, 256] - - [27, 75.147] - - - [43520, 256, 1, 256] - - [425, 62.346] - - - [43776, 2865, 1, 256] - - [26, 68.682] - - - [41984, 3840, 1, 256] - - [25, 74.662] - - - [42544, 256, 1, 256] - - [426, 58.09] - - - [43056, 2816, 1, 256] - - [39, 61.699] - - - [41472, 3072, 1, 256] - - [35, 73.12] - - - [41776, 2865, 1, 256] - - [51, 61.177] - - - [43056, 256, 1, 256] - - [289, 58.887] - - - [41728, 6144, 1, 256] - - [26, 73.428] - - - [42496, 8448, 1, 256] - - [26, 74.955] - - - [43568, 256, 1, 256] - - [370, 58.736] - - - [42752, 8704, 1, 256] - - [25, 74.639] - - - [42544, 2816, 1, 256] - - [51, 63.276] - - - [40448, 10240, 1, 256] - - [45, 74.672] - - - [41728, 2816, 1, 256] - - [84, 71.895] - - - [43568, 10240, 1, 256] - - [39, 62.816] - - - [44032, 2048, 1, 256] - - [39, 71.938] - - - [41472, 2865, 1, 256] - - [23, 70.643] - - - [40448, 256, 1, 256] - - [266, 63.033] - - - [41728, 3329, 1, 256] - - [26, 68.274] - - - [43264, 6144, 1, 256] - - [23, 74.082] - - - [40960, 6656, 1, 256] - - [20, 67.684] - - - [42752, 9216, 1, 256] - - [27, 74.194] - - - [40496, 2816, 1, 256] - - [42, 63.636] - - - [40704, 512, 1, 256] - - [269, 72.582] - - - [43056, 10240, 1, 256] - - [39, 63.048] - - - [44032, 9728, 1, 256] - - [25, 75.072] - - - [41728, 8192, 1, 256] - - [39, 74.078] - - - [43264, 1024, 1, 256] - - [51, 69.973] - - - [43776, 2048, 1, 256] - - [26, 70.652] - - - [40240, 2816, 1, 256] - - [56, 62.853] - - - [42752, 1280, 1, 256] - - [25, 71.138] - - - [44288, 10240, 1, 256] - - [39, 74.537] - - - [42240, 2816, 1, 256] - - [25, 72.834] - - - [41728, 7424, 1, 256] - - [26, 73.837] - - - [44288, 3328, 1, 256] - - [26, 72.991] - - - [43264, 5120, 1, 256] - - [25, 74.085] - - - [42032, 2816, 1, 256] - - [51, 63.145] - - - [11776, 6144, 1, 256] - - [35, 73.095] - - - [11264, 1792, 1, 256] - - [357, 73.071] - - - [4352, 2865, 1, 256] - - [419, 68.33] - - - [14640, 1536, 1, 256] - - [314, 72.224] - - - [4096, 2865, 1, 256] - - [282, 65.917] - - - [5168, 256, 1, 256] - - [103, 41.239] - - - [19968, 3328, 1, 256] - - [25, 72.493] - - - [12544, 3328, 1, 256] - - [33, 70.717] - - - [15408, 2816, 1, 256] - - [23, 62.707] - - - [16640, 3329, 1, 256] - - [23, 67.757] - - - [768, 768, 1, 256] - - [271, 19.275] - - - [3840, 512, 1, 256] - - [284, 38.293] - - - [7424, 5888, 1, 256] - - [25, 71.721] - - - [48, 49, 1, 256] - - [183, 0.08] - - - [16384, 768, 1, 256] - - [302, 61.438] - - - [15664, 2865, 1, 256] - - [23, 61.915] - - - [12544, 2048, 1, 256] - - [299, 77.196] - - - [7680, 4096, 1, 256] - - [390, 78.186] - - - [8240, 5376, 1, 256] - - [35, 65.489] - - - [11520, 256, 1, 256] - - [293, 47.226] - - - [12800, 256, 1, 256] - - [424, 50.715] - - - [10544, 2865, 1, 256] - - [306, 72.676] - - - [10032, 6912, 1, 256] - - [34, 63.949] - - - [3072, 3072, 1, 256] - - [357, 65.865] - - - [5888, 2865, 1, 256] - - [348, 69.693] - - - [8448, 3328, 1, 256] - - [298, 77.962] - - - [17920, 4096, 1, 256] - - [72, 72.746] - - - [19200, 5376, 1, 256] - - [23, 73.642] - - - [16432, 2865, 1, 256] - - [36, 64.164] - - - [12032, 3329, 1, 256] - - [33, 67.426] - - - [11776, 8704, 1, 256] - - [65, 74.604] - - - [11520, 1281, 1, 256] - - [294, 65.937] - - - [19760, 10240, 1, 256] - - [42, 62.731] - - - [15360, 1281, 1, 256] - - [300, 70.3] - - - [19712, 2865, 1, 256] - - [28, 67.478] - - - [9216, 6400, 1, 256] - - [23, 73.86] - - - [18944, 3329, 1, 256] - - [28, 69.341] - - - [5632, 2816, 1, 256] - - [298, 72.597] - - - [13872, 256, 1, 256] - - [293, 45.442] - - - [9984, 1280, 1, 256] - - [426, 69.386] - - - [19248, 10240, 1, 256] - - [34, 63.303] - - - [14128, 256, 1, 256] - - [293, 46.685] - - - [12080, 9216, 1, 256] - - [51, 63.728] - - - [18224, 5120, 1, 256] - - [56, 63.714] - - - [2352, 256, 1, 256] - - [105, 28.229] - - - [17712, 4608, 1, 256] - - [56, 63.487] - - - [8192, 5376, 1, 256] - - [33, 71.577] - - - [8752, 5888, 1, 256] - - [34, 63.141] - - - [11264, 3584, 1, 256] - - [47, 72.09] - - - [816, 256, 1, 256] - - [97, 12.953] - - - [5376, 3328, 1, 256] - - [349, 72.391] - - - [6144, 2560, 1, 256] - - [417, 71.131] - - - [9264, 256, 1, 256] - - [418, 42.01] - - - [8960, 5376, 1, 256] - - [55, 72.299] - - - [2608, 2353, 1, 256] - - [290, 55.412] - - - [2096, 256, 1, 256] - - [105, 25.259] - - - [9984, 7168, 1, 256] - - [32, 72.618] - - - [7424, 3329, 1, 256] - - [419, 74.598] - - - [2352, 2304, 1, 256] - - [418, 56.857] - - - [9984, 512, 1, 256] - - [291, 54.422] - - - [6656, 3840, 1, 256] - - [296, 76.92] - - - [17408, 3329, 1, 256] - - [35, 69.421] - - - [8496, 5376, 1, 256] - - [27, 64.353] - - - [11264, 3840, 1, 256] - - [32, 72.724] - - - [13312, 2865, 1, 256] - - [390, 78.96] - - - [3584, 768, 1, 256] - - [345, 45.1] - - - [11520, 6144, 1, 256] - - [57, 72.709] - - - [15360, 2048, 1, 256] - - [357, 79.23] - - - [7936, 3328, 1, 256] - - [390, 77.449] - - - [6144, 1281, 1, 256] - - [301, 60.16] - - - [19968, 6656, 1, 256] - - [35, 74.609] - - - [15152, 256, 1, 256] - - [418, 48.256] - - - [18432, 4608, 1, 256] - - [33, 73.713] - - - [1072, 256, 1, 256] - - [98, 16.399] - - - [6400, 4864, 1, 256] - - [390, 79.685] - - - [19712, 1281, 1, 256] - - [300, 71.331] - - - [1792, 1280, 1, 256] - - [387, 43.557] - - - [8192, 2865, 1, 256] - - [388, 70.807] - - - [3376, 256, 1, 256] - - [90, 37.471] - - - [10544, 2816, 1, 256] - - [391, 75.103] - - - [14336, 2816, 1, 256] - - [25, 71.657] - - - [16384, 1280, 1, 256] - - [309, 63.915] - - - [1280, 256, 1, 256] - - [98, 19.581] - - - [12544, 8960, 1, 256] - - [25, 74.392] - - - [13824, 1281, 1, 256] - - [390, 68.265] - - - [3072, 256, 1, 256] - - [91, 35.708] - - - [19760, 2816, 1, 256] - - [56, 63.795] - - - [8448, 5376, 1, 256] - - [47, 71.936] - - - [11824, 2865, 1, 256] - - [299, 72.795] - - - [6656, 3584, 1, 256] - - [299, 75.325] - - - [12288, 8704, 1, 256] - - [35, 75.086] - - - [11312, 256, 1, 256] - - [348, 44.515] - - - [15920, 2816, 1, 256] - - [51, 64.786] - - - [12032, 8448, 1, 256] - - [32, 73.765] - - - [14080, 2048, 1, 256] - - [350, 76.414] - - - [6400, 5120, 1, 256] - - [390, 79.936] - - - [7216, 2865, 1, 256] - - [312, 69.077] - - - [4400, 1280, 1, 256] - - [418, 56.058] - - - [5376, 3840, 1, 256] - - [421, 76.221] - - - [7168, 2816, 1, 256] - - [299, 73.11] - - - [19200, 5632, 1, 256] - - [25, 73.869] - - - [4144, 1024, 1, 256] - - [290, 50.869] - - - [12800, 3329, 1, 256] - - [65, 68.406] - - - [6400, 2865, 1, 256] - - [300, 71.825] - - - [12800, 5376, 1, 256] - - [55, 73.266] - - - [7168, 1536, 1, 256] - - [417, 67.716] - - - [19968, 1281, 1, 256] - - [421, 71.842] - - - [17664, 1281, 1, 256] - - [294, 69.974] - - - [11264, 3329, 1, 256] - - [296, 77.269] - - - [17712, 256, 1, 256] - - [278, 49.465] - - - [6656, 5376, 1, 256] - - [357, 80.057] - - - [13056, 5376, 1, 256] - - [32, 73.45] - - - [11568, 2865, 1, 256] - - [391, 73.176] - - - [3328, 1281, 1, 256] - - [287, 51.335] - - - [19968, 2048, 1, 256] - - [56, 70.791] - - - [2304, 2048, 1, 256] - - [294, 56.36] - - - [7728, 256, 1, 256] - - [343, 37.109] - - - [7424, 4352, 1, 256] - - [299, 79.485] - - - [5376, 2048, 1, 256] - - [360, 66.476] - - - [19456, 2816, 1, 256] - - [23, 72.622] - - - [7216, 2816, 1, 256] - - [315, 70.644] - - - [18688, 5376, 1, 256] - - [23, 73.843] - - - [4656, 1792, 1, 256] - - [284, 62.388] - - - [10240, 768, 1, 256] - - [290, 61.862] - - - [19456, 1280, 1, 256] - - [391, 76.031] - - - [18432, 3329, 1, 256] - - [27, 69.63] - - - [17920, 6144, 1, 256] - - [27, 74.033] - - - [1536, 1280, 1, 256] - - [259, 39.215] - - - [19456, 6400, 1, 256] - - [35, 75.291] - - - [15360, 6144, 1, 256] - - [23, 74.341] - - - [15664, 10240, 1, 256] - - [23, 63.248] - - - [3840, 256, 1, 256] - - [168, 34.149] - - - [4864, 3328, 1, 256] - - [348, 73.904] - - - [18224, 2865, 1, 256] - - [35, 61.253] - - - [13056, 9984, 1, 256] - - [23, 75.033] - - - [12288, 256, 1, 256] - - [308, 49.492] - - - [7168, 3840, 1, 256] - - [348, 77.334] - - - [17712, 4352, 1, 256] - - [56, 63.205] - - - [14592, 10240, 1, 256] - - [66, 73.199] - - - [8704, 5376, 1, 256] - - [47, 72.78] - - - [16128, 2816, 1, 256] - - [25, 71.212] - - - [4352, 3329, 1, 256] - - [305, 70.167] - - - [13568, 512, 1, 256] - - [293, 61.16] - - - [15872, 2865, 1, 256] - - [23, 69.974] - - - [12032, 1281, 1, 256] - - [300, 68.452] - - - [11520, 2048, 1, 256] - - [299, 75.482] - - - [12032, 2048, 1, 256] - - [299, 76.946] - - - [5632, 1281, 1, 256] - - [301, 60.464] - - - [13312, 9984, 1, 256] - - [27, 75.719] - - - [4912, 2865, 1, 256] - - [325, 66.907] - - - [15408, 2304, 1, 256] - - [426, 76.048] - - - [7472, 2816, 1, 256] - - [320, 70.873] - - - [18688, 10240, 1, 256] - - [27, 74.915] - - - [10752, 7936, 1, 256] - - [23, 74.542] - - - [2048, 1793, 1, 256] - - [419, 51.981] - - - [11776, 1280, 1, 256] - - [360, 72.587] - - - [10032, 256, 1, 256] - - [291, 40.04] - - - [17408, 1536, 1, 256] - - [308, 78.251] - - - [14080, 2865, 1, 256] - - [65, 68.381] - - - [16688, 3328, 1, 256] - - [42, 65.797] - - - [18944, 1024, 1, 256] - - [390, 72.647] - - - [2352, 2097, 1, 256] - - [344, 51.094] - - - [11008, 2048, 1, 256] - - [390, 73.598] - - - [10240, 6912, 1, 256] - - [47, 74.534] - - - [8448, 768, 1, 256] - - [270, 60.06] - - - [16640, 1024, 1, 256] - - [300, 69.67] - - - [11824, 8960, 1, 256] - - [51, 64.102] - - - [7936, 1280, 1, 256] - - [290, 66.226] - - - [6960, 3840, 1, 256] - - [391, 74.409] - - - [3328, 2048, 1, 256] - - [282, 60.906] - - - [16944, 2865, 1, 256] - - [35, 63.601] - - - [1024, 256, 1, 256] - - [98, 15.829] - - - [16944, 3840, 1, 256] - - [51, 64.63] - - - [3376, 2816, 1, 256] - - [420, 64.137] - - - [12288, 768, 1, 256] - - [417, 64.882] - - - [17152, 3329, 1, 256] - - [53, 69.147] - - - [6192, 2865, 1, 256] - - [349, 67.219] - - - [5888, 1281, 1, 256] - - [417, 59.031] - - - [11824, 256, 1, 256] - - [315, 45.574] - - - [18688, 1280, 1, 256] - - [299, 74.278] - - - [11520, 7936, 1, 256] - - [57, 73.686] - - - [15616, 1281, 1, 256] - - [421, 71.355] - - - [16944, 10240, 1, 256] - - [23, 64.198] - - - [12032, 4352, 1, 256] - - [33, 71.886] - - - [9984, 6656, 1, 256] - - [33, 73.374] - - - [17408, 1281, 1, 256] - - [300, 71.277] - - - [6912, 3329, 1, 256] - - [390, 73.715] - - - [16176, 2865, 1, 256] - - [27, 63.843] - - - [7936, 4864, 1, 256] - - [299, 82.374] - - - [7168, 256, 1, 256] - - [343, 36.907] - - - [9728, 6144, 1, 256] - - [25, 72.968] - - - [10752, 7680, 1, 256] - - [23, 74.394] - - - [13056, 5632, 1, 256] - - [35, 73.854] - - - [17152, 2865, 1, 256] - - [25, 68.888] - - - [4096, 512, 1, 256] - - [282, 39.785] - - - [3584, 2304, 1, 256] - - [417, 65.602] - - - [11264, 2048, 1, 256] - - [320, 74.815] - - - [18944, 5376, 1, 256] - - [65, 74.399] - - - [8960, 3329, 1, 256] - - [419, 75.18] - - - [7936, 1281, 1, 256] - - [419, 63.127] - - - [12848, 2816, 1, 256] - - [428, 76.876] - - - [9472, 3328, 1, 256] - - [299, 80.397] - - - [2816, 2816, 1, 256] - - [299, 64.243] - - - [15616, 10240, 1, 256] - - [27, 74.822] - - - [2816, 256, 1, 256] - - [91, 33.257] - - - [48, 256, 1, 256] - - [95, 0.779] - - - [17408, 1792, 1, 256] - - [357, 77.59] - - - [10032, 2865, 1, 256] - - [306, 72.736] - - - [3584, 2865, 1, 256] - - [284, 62.902] - - - [9472, 2816, 1, 256] - - [390, 78.964] - - - [2096, 2048, 1, 256] - - [290, 50.752] - - - [9216, 1536, 1, 256] - - [299, 72.507] - - - [5936, 256, 1, 256] - - [162, 41.044] - - - [11520, 1280, 1, 256] - - [289, 72.223] - - - [16896, 3328, 1, 256] - - [35, 72.404] - - - [7984, 4864, 1, 256] - - [299, 78.151] - - - [11008, 1280, 1, 256] - - [419, 68.318] - - - [18432, 6144, 1, 256] - - [23, 74.435] - - - [2096, 1841, 1, 256] - - [293, 48.8] - - - [8448, 1024, 1, 256] - - [421, 61.041] - - - [17968, 10240, 1, 256] - - [34, 63.221] - - - [1536, 1536, 1, 256] - - [284, 44.775] - - - [7728, 4864, 1, 256] - - [298, 77.575] - - - [18944, 3328, 1, 256] - - [29, 72.475] - - - [4608, 1792, 1, 256] - - [284, 64.449] - - - [8960, 6144, 1, 256] - - [33, 72.367] - - - [18736, 2816, 1, 256] - - [42, 62.4] - - - [8704, 5120, 1, 256] - - [33, 72.429] - - - [19456, 6144, 1, 256] - - [23, 74.371] - - - [19456, 1281, 1, 256] - - [421, 70.639] - - - [17200, 3840, 1, 256] - - [25, 64.105] - - - [2352, 2353, 1, 256] - - [290, 54.388] - - - [17408, 2816, 1, 256] - - [53, 72.23] - - - [13312, 2816, 1, 256] - - [390, 81.853] - - - [8960, 2816, 1, 256] - - [357, 76.353] - - - [2048, 1792, 1, 256] - - [419, 52.107] - - - [17152, 10240, 1, 256] - - [55, 74.66] - - - [16176, 10240, 1, 256] - - [25, 65.472] - - - [10288, 2865, 1, 256] - - [325, 71.515] - - - [8704, 2816, 1, 256] - - [267, 78.006] - - - [7424, 4096, 1, 256] - - [390, 78.605] - - - [6656, 1024, 1, 256] - - [293, 59.157] - - - [2304, 256, 1, 256] - - [90, 28.11] - - - [16384, 2865, 1, 256] - - [20, 59.821] - - - [7680, 2816, 1, 256] - - [289, 76.206] - - - [11520, 3329, 1, 256] - - [299, 77.673] - - - [10752, 1280, 1, 256] - - [357, 69.636] - - - [3120, 2816, 1, 256] - - [293, 61.432] - - - [15872, 1281, 1, 256] - - [294, 70.391] - - - [13824, 6144, 1, 256] - - [27, 73.534] - - - [6912, 3584, 1, 256] - - [296, 75.77] - - - [12032, 3328, 1, 256] - - [53, 70.445] - - - [11264, 1281, 1, 256] - - [419, 67.039] - - - [19456, 5632, 1, 256] - - [23, 74.96] - - - [17200, 2816, 1, 256] - - [56, 63.832] - - - [11520, 3840, 1, 256] - - [65, 71.601] - - - [11520, 2865, 1, 256] - - [285, 76.33] - - - [14848, 1280, 1, 256] - - [296, 72.098] - - - [16176, 256, 1, 256] - - [293, 52.53] - - - [16384, 256, 1, 256] - - [370, 47.465] - - - [4096, 768, 1, 256] - - [418, 49.705] - - - [4864, 2816, 1, 256] - - [298, 69.86] - - - [13568, 256, 1, 256] - - [290, 47.798] - - - [4608, 2048, 1, 256] - - [299, 66.917] - - - [9984, 6144, 1, 256] - - [33, 72.711] - - - [3632, 768, 1, 256] - - [271, 44.01] - - - [19200, 5888, 1, 256] - - [32, 73.564] - - - [5632, 2865, 1, 256] - - [420, 70.971] - - - [15360, 1280, 1, 256] - - [299, 72.783] - - - [12800, 1280, 1, 256] - - [298, 71.76] - - - [7168, 3328, 1, 256] - - [289, 75.708] - - - [11264, 8448, 1, 256] - - [35, 74.919] - - - [18176, 3328, 1, 256] - - [47, 72.027] - - - [4096, 2560, 1, 256] - - [419, 66.227] - - - [12544, 768, 1, 256] - - [325, 65.0] - - - [11568, 8448, 1, 256] - - [42, 63.978] - - - [8704, 1280, 1, 256] - - [272, 66.706] - - - [13056, 1536, 1, 256] - - [390, 75.233] - - - [2304, 1024, 1, 256] - - [282, 44.64] - - - [3072, 1281, 1, 256] - - [293, 51.032] - - - [6912, 1280, 1, 256] - - [392, 64.009] - - - [9216, 2816, 1, 256] - - [390, 77.111] - - - [17152, 6144, 1, 256] - - [25, 73.807] - - - [18992, 2865, 1, 256] - - [23, 60.345] - - - [10240, 2560, 1, 256] - - [296, 77.621] - - - [560, 256, 1, 256] - - [97, 8.889] - - - [2304, 1280, 1, 256] - - [392, 47.863] - - - [7680, 6144, 1, 256] - - [32, 72.08] - - - [15920, 2560, 1, 256] - - [35, 64.2] - - - [17456, 10240, 1, 256] - - [51, 63.79] - - - [14080, 3328, 1, 256] - - [55, 70.137] - - - [13360, 10240, 1, 256] - - [25, 64.032] - - - [8448, 5632, 1, 256] - - [47, 72.533] - - - [17408, 3584, 1, 256] - - [35, 73.531] - - - [6704, 2865, 1, 256] - - [289, 68.026] - - - [12592, 9472, 1, 256] - - [34, 64.017] - - - [18992, 10240, 1, 256] - - [42, 63.555] - - - [5376, 2865, 1, 256] - - [304, 68.978] - - - [18480, 5120, 1, 256] - - [35, 62.77] - - - [14336, 2048, 1, 256] - - [357, 77.842] - - - [7424, 3328, 1, 256] - - [308, 76.048] - - - [256, 49, 1, 256] - - [97, 0.778] - - - [12288, 1280, 1, 256] - - [360, 68.746] - - - [13568, 3329, 1, 256] - - [23, 67.986] - - - [15360, 1792, 1, 256] - - [299, 77.524] - - - [7168, 3584, 1, 256] - - [299, 76.832] - - - [10240, 3328, 1, 256] - - [390, 79.123] - - - [6400, 1281, 1, 256] - - [287, 62.123] - - - [11008, 6144, 1, 256] - - [73, 72.223] - - - [512, 513, 1, 256] - - [98, 15.777] - - - [19248, 256, 1, 256] - - [278, 51.008] - - - [2608, 256, 1, 256] - - [96, 30.8] - - - [16688, 3584, 1, 256] - - [27, 65.642] - - - [17920, 4864, 1, 256] - - [29, 74.051] - - - [18688, 1281, 1, 256] - - [289, 70.297] - - - [18224, 4864, 1, 256] - - [34, 63.328] - - - [10496, 2816, 1, 256] - - [424, 77.597] - - - [12288, 4864, 1, 256] - - [35, 73.366] - - - [9216, 2865, 1, 256] - - [299, 74.454] - - - [17664, 3329, 1, 256] - - [25, 68.974] - - - [3632, 512, 1, 256] - - [351, 36.652] - - - [11776, 3329, 1, 256] - - [55, 68.167] - - - [19456, 1792, 1, 256] - - [299, 80.004] - - - [12592, 256, 1, 256] - - [307, 48.429] - - - [10752, 3072, 1, 256] - - [299, 80.36] - - - [10800, 2816, 1, 256] - - [299, 74.904] - - - [6192, 3072, 1, 256] - - [388, 68.696] - - - [17152, 1536, 1, 256] - - [348, 76.403] - - - [2096, 2097, 1, 256] - - [419, 51.062] - - - [8192, 4608, 1, 256] - - [304, 78.747] - - - [13056, 6144, 1, 256] - - [25, 73.792] - - - [16640, 10240, 1, 256] - - [23, 74.759] - - - [12592, 9728, 1, 256] - - [56, 63.963] - - - [18176, 2816, 1, 256] - - [25, 71.857] - - - [18176, 4864, 1, 256] - - [23, 73.781] - - - [18944, 6144, 1, 256] - - [22, 74.1] - - - [12544, 256, 1, 256] - - [25, 50.485] - - - [13056, 1281, 1, 256] - - [301, 69.537] - - - [304, 49, 1, 256] - - [147, 0.914] - - - [17920, 2816, 1, 256] - - [62, 71.961] - - - [4656, 256, 1, 256] - - [96, 39.388] - - - [7728, 2865, 1, 256] - - [391, 70.935] - - - [15872, 1280, 1, 256] - - [296, 72.678] - - - [17456, 256, 1, 256] - - [312, 49.232] - - - [18176, 4608, 1, 256] - - [47, 73.101] - - - [7168, 5632, 1, 256] - - [47, 72.334] - - - [13616, 256, 1, 256] - - [392, 45.98] - - - [15104, 3329, 1, 256] - - [35, 68.888] - - - [19712, 3328, 1, 256] - - [65, 70.314] - - - [10032, 7168, 1, 256] - - [51, 63.028] - - - [11008, 3328, 1, 256] - - [390, 79.905] - - - [10496, 6144, 1, 256] - - [33, 72.906] - - - [6656, 2865, 1, 256] - - [301, 71.972] - - - [17664, 3840, 1, 256] - - [57, 72.723] - - - [6960, 4096, 1, 256] - - [299, 73.887] - - - [4608, 256, 1, 256] - - [155, 39.897] - - - [10496, 6912, 1, 256] - - [32, 73.557] - - - [16128, 2560, 1, 256] - - [57, 71.226] - - - [15872, 256, 1, 256] - - [272, 52.601] - - - [6656, 1281, 1, 256] - - [419, 60.571] - - - [3584, 512, 1, 256] - - [351, 37.358] - - - [11520, 1792, 1, 256] - - [299, 76.357] - - - [11264, 8192, 1, 256] - - [35, 74.546] - - - [10752, 2048, 1, 256] - - [299, 74.554] - - - [18688, 3329, 1, 256] - - [25, 69.295] - - - [4352, 768, 1, 256] - - [357, 52.889] - - - [18432, 512, 1, 256] - - [315, 65.232] - - - [18992, 256, 1, 256] - - [385, 50.633] - - - [13568, 2865, 1, 256] - - [390, 79.965] - - - [14640, 256, 1, 256] - - [418, 47.262] - - - [11264, 256, 1, 256] - - [348, 47.252] - - - [16896, 3329, 1, 256] - - [23, 69.544] - - - [18944, 5120, 1, 256] - - [45, 74.02] - - - [768, 513, 1, 256] - - [90, 22.833] - - - [14080, 10240, 1, 256] - - [22, 73.542] - - - [15872, 2560, 1, 256] - - [47, 71.856] - - - [6912, 5632, 1, 256] - - [296, 82.365] - - - [13360, 2865, 1, 256] - - [391, 74.499] - - - [6400, 3072, 1, 256] - - [299, 75.004] - - - [15616, 3329, 1, 256] - - [25, 69.058] - - - [9264, 2816, 1, 256] - - [296, 73.058] - - - [18176, 512, 1, 256] - - [348, 65.608] - - - [11264, 1280, 1, 256] - - [390, 71.758] - - - [1328, 1329, 1, 256] - - [345, 37.886] - - - [18736, 5376, 1, 256] - - [56, 62.844] - - - [5376, 1792, 1, 256] - - [309, 65.671] - - - [6144, 4608, 1, 256] - - [289, 78.442] - - - [6400, 3328, 1, 256] - - [296, 76.779] - - - [12032, 2865, 1, 256] - - [417, 76.515] - - - [12288, 4608, 1, 256] - - [25, 72.894] - - - [16128, 3072, 1, 256] - - [57, 71.509] - - - [2048, 256, 1, 256] - - [105, 25.3] - - - [4096, 256, 1, 256] - - [166, 36.316] - - - [5888, 2304, 1, 256] - - [287, 71.821] - - - [2816, 2561, 1, 256] - - [420, 62.391] - - - [3072, 1536, 1, 256] - - [267, 55.496] - - - [2304, 1281, 1, 256] - - [293, 47.209] - - - [15616, 2048, 1, 256] - - [299, 79.295] - - - [12800, 1024, 1, 256] - - [357, 70.997] - - - [8960, 3328, 1, 256] - - [357, 78.636] - - - [18432, 1280, 1, 256] - - [296, 74.48] - - - [8448, 2048, 1, 256] - - [299, 71.186] - - - [19712, 6400, 1, 256] - - [22, 73.172] - - - [14384, 1280, 1, 256] - - [312, 67.776] - - - [6448, 2816, 1, 256] - - [300, 70.16] - - - [18176, 2048, 1, 256] - - [299, 81.701] - - - [3072, 1792, 1, 256] - - [418, 57.493] - - - [12080, 8960, 1, 256] - - [34, 63.832] - - - [13312, 1281, 1, 256] - - [282, 66.907] - - - [16688, 2816, 1, 256] - - [25, 64.939] - - - [6400, 256, 1, 256] - - [163, 46.542] - - - [2048, 2048, 1, 256] - - [392, 51.623] - - - [14336, 256, 1, 256] - - [392, 49.932] - - - [11520, 2816, 1, 256] - - [299, 79.078] - - - [14384, 10240, 1, 256] - - [25, 63.156] - - - [7472, 256, 1, 256] - - [290, 36.306] - - - [1280, 1280, 1, 256] - - [103, 47.358] - - - [8704, 1024, 1, 256] - - [268, 63.635] - - - [9520, 2865, 1, 256] - - [320, 72.115] - - - [18480, 256, 1, 256] - - [291, 49.442] - - - [18176, 256, 1, 256] - - [418, 54.111] - - - [15872, 6144, 1, 256] - - [25, 74.514] - - - [304, 256, 1, 256] - - [98, 4.8] - - - [13568, 5888, 1, 256] - - [47, 73.35] - - - [3328, 3328, 1, 256] - - [348, 70.694] - - - [6656, 5120, 1, 256] - - [390, 79.758] - - - [9520, 2816, 1, 256] - - [314, 72.492] - - - [1536, 1537, 1, 256] - - [290, 43.89] - - - [3072, 2865, 1, 256] - - [298, 61.597] - - - [10032, 2816, 1, 256] - - [424, 75.06] - - - [12032, 9216, 1, 256] - - [52, 73.508] - - - [13872, 10240, 1, 256] - - [42, 63.139] - - - [13824, 2048, 1, 256] - - [299, 77.082] - - - [12544, 9728, 1, 256] - - [35, 74.151] - - - [17664, 4352, 1, 256] - - [55, 73.069] - - - [4096, 1281, 1, 256] - - [420, 53.157] - - - [17408, 1280, 1, 256] - - [357, 74.591] - - - [18432, 2816, 1, 256] - - [53, 72.467] - - - [5120, 256, 1, 256] - - [277, 33.476] - - - [18736, 2865, 1, 256] - - [56, 60.889] - - - [19200, 256, 1, 256] - - [267, 55.623] - - - [2048, 512, 1, 256] - - [154, 36.208] - - - [11008, 7680, 1, 256] - - [69, 73.369] - - - [5888, 3072, 1, 256] - - [421, 73.01] - - - [11776, 8192, 1, 256] - - [25, 74.076] - - - [5888, 512, 1, 256] - - [348, 48.145] - - - [7936, 2816, 1, 256] - - [296, 74.684] - - - [5120, 2865, 1, 256] - - [422, 70.841] - - - [12032, 2816, 1, 256] - - [299, 79.903] - - - [256, 257, 1, 256] - - [159, 4.192] - - - [13104, 2865, 1, 256] - - [299, 76.004] - - - [5680, 2865, 1, 256] - - [320, 66.767] - - - [15408, 10240, 1, 256] - - [27, 63.988] - - - [18432, 4864, 1, 256] - - [25, 74.451] - - - [17712, 2865, 1, 256] - - [25, 60.28] - - - [768, 256, 1, 256] - - [97, 12.668] - - - [9728, 3328, 1, 256] - - [348, 79.82] - - - [12848, 9728, 1, 256] - - [51, 64.542] - - - [2304, 2305, 1, 256] - - [392, 55.309] - - - [10240, 6144, 1, 256] - - [27, 73.496] - - - [13312, 1280, 1, 256] - - [296, 71.368] - - - [9008, 5888, 1, 256] - - [34, 63.331] - - - [7424, 3840, 1, 256] - - [289, 78.274] - - - [12032, 1280, 1, 256] - - [301, 69.975] - - - [18480, 2816, 1, 256] - - [27, 62.353] - - - [18432, 5120, 1, 256] - - [23, 74.354] - - - [7424, 4608, 1, 256] - - [390, 80.458] - - - [9776, 2865, 1, 256] - - [423, 71.794] - - - [5632, 2560, 1, 256] - - [348, 73.383] - - - [7680, 2048, 1, 256] - - [299, 73.265] - - - [6704, 2816, 1, 256] - - [391, 71.046] - - - [13872, 2816, 1, 256] - - [34, 63.15] - - - [17968, 2816, 1, 256] - - [42, 62.628] - - - [4144, 2865, 1, 256] - - [308, 64.617] - - - [14640, 1280, 1, 256] - - [422, 69.87] - - - [16432, 2816, 1, 256] - - [34, 67.138] - - - [16128, 1280, 1, 256] - - [296, 72.547] - - - [8240, 5120, 1, 256] - - [25, 65.285] - - - [4352, 2816, 1, 256] - - [419, 70.073] - - - [12544, 2865, 1, 256] - - [299, 78.74] - - - [6144, 2048, 1, 256] - - [348, 68.727] - - - [13616, 512, 1, 256] - - [291, 57.329] - - - [5632, 2048, 1, 256] - - [357, 70.024] - - - [13312, 2048, 1, 256] - - [390, 78.733] - - - [9728, 1281, 1, 256] - - [284, 64.571] - - - [7424, 1281, 1, 256] - - [284, 61.748] - - - [10800, 256, 1, 256] - - [418, 42.841] - - - [2048, 1281, 1, 256] - - [357, 42.405] - - - [5376, 1280, 1, 256] - - [421, 59.683] - - - [15664, 2816, 1, 256] - - [34, 63.151] - - - [256, 256, 1, 256] - - [147, 4.319] - - - [2048, 1280, 1, 256] - - [348, 42.655] - - - [9776, 256, 1, 256] - - [290, 43.335] - - - [4096, 3329, 1, 256] - - [360, 66.928] - - - [9728, 2304, 1, 256] - - [421, 75.049] - - - [19968, 2865, 1, 256] - - [23, 70.059] - - - [13568, 6144, 1, 256] - - [27, 73.621] - - - [15360, 2304, 1, 256] - - [390, 79.934] - - - [9264, 6400, 1, 256] - - [25, 63.431] - - - [19200, 2048, 1, 256] - - [21, 69.891] - - - [11520, 4096, 1, 256] - - [38, 70.668] - - - [18688, 5632, 1, 256] - - [23, 74.26] - - - [11776, 256, 1, 256] - - [298, 47.8] - - - [17152, 256, 1, 256] - - [266, 52.507] - - - [5120, 1280, 1, 256] - - [322, 61.555] - - - [14896, 1792, 1, 256] - - [299, 73.36] - - - [10288, 2816, 1, 256] - - [299, 73.19] - - - [7984, 2865, 1, 256] - - [390, 71.983] - - - [4864, 1281, 1, 256] - - [418, 56.21] - - - [7216, 256, 1, 256] - - [354, 35.755] - - - [5888, 3328, 1, 256] - - [296, 74.287] - - - [7424, 2816, 1, 256] - - [289, 76.795] - - - [15360, 3328, 1, 256] - - [33, 72.12] - - - [10544, 256, 1, 256] - - [296, 41.33] - - - [9776, 2816, 1, 256] - - [289, 75.811] - - - [8240, 2816, 1, 256] - - [313, 76.045] - - - [6656, 3072, 1, 256] - - [390, 73.783] - - - [18224, 10240, 1, 256] - - [51, 63.963] - - - [13824, 2865, 1, 256] - - [33, 69.183] - - - [5376, 1281, 1, 256] - - [349, 59.66] - - - [13568, 9984, 1, 256] - - [35, 74.874] - - - [18176, 4352, 1, 256] - - [35, 73.61] - - - [11776, 1281, 1, 256] - - [417, 67.318] - - - [15616, 6144, 1, 256] - - [35, 73.791] - - - [4400, 256, 1, 256] - - [105, 37.764] - - - [18992, 2816, 1, 256] - - [34, 63.728] - - - [14640, 10240, 1, 256] - - [56, 63.516] - - - [5120, 2048, 1, 256] - - [421, 65.857] - - - [19968, 10240, 1, 256] - - [25, 75.102] - - - [19200, 2865, 1, 256] - - [57, 69.04] - - - [15152, 2816, 1, 256] - - [51, 63.055] - - - [2560, 2560, 1, 256] - - [392, 61.343] - - - [8448, 2816, 1, 256] - - [296, 74.92] - - - [8704, 5632, 1, 256] - - [53, 72.624] - - - [1024, 769, 1, 256] - - [416, 22.723] - - - [17200, 4096, 1, 256] - - [42, 63.61] - - - [5376, 256, 1, 256] - - [163, 40.594] - - - [6656, 256, 1, 256] - - [347, 38.619] - - - [18688, 3328, 1, 256] - - [33, 71.918] - - - [13056, 256, 1, 256] - - [325, 51.325] - - - [13104, 2816, 1, 256] - - [424, 77.377] - - - [7424, 1792, 1, 256] - - [391, 70.925] - - - [14592, 2816, 1, 256] - - [57, 70.771] - - - [12336, 2865, 1, 256] - - [299, 75.794] - - - [17920, 256, 1, 256] - - [270, 53.394] - - - [12800, 2048, 1, 256] - - [299, 77.992] - - - [3632, 256, 1, 256] - - [103, 32.495] - - - [18688, 768, 1, 256] - - [301, 69.674] - - - [16384, 2816, 1, 256] - - [20, 62.061] - - - [14896, 10240, 1, 256] - - [23, 63.546] - - - [816, 817, 1, 256] - - [103, 30.034] - - - [9008, 2865, 1, 256] - - [422, 72.346] - - - [14848, 1024, 1, 256] - - [390, 70.459] - - - [16640, 256, 1, 256] - - [345, 49.529] - - - [7424, 256, 1, 256] - - [259, 38.044] - - - [10240, 3329, 1, 256] - - [299, 76.438] - - - [18176, 5120, 1, 256] - - [25, 73.841] - - - [6912, 2865, 1, 256] - - [287, 72.727] - - - [1024, 1025, 1, 256] - - [345, 27.794] - - - [5632, 4096, 1, 256] - - [419, 76.073] - - - [12544, 1024, 1, 256] - - [357, 69.564] - - - [2864, 2609, 1, 256] - - [308, 57.169] - - - [16896, 2048, 1, 256] - - [289, 79.321] - - - [3840, 1280, 1, 256] - - [418, 57.465] - - - [11008, 1281, 1, 256] - - [419, 68.244] - - - [15104, 1281, 1, 256] - - [315, 69.472] - - - [7168, 3329, 1, 256] - - [296, 73.138] - - - [12800, 5120, 1, 256] - - [57, 73.081] - - - [512, 257, 1, 256] - - [147, 8.338] - - - [12288, 8960, 1, 256] - - [25, 75.243] - - - [9728, 6912, 1, 256] - - [47, 74.008] - - - [9728, 6656, 1, 256] - - [32, 73.905] - - - [2560, 2304, 1, 256] - - [420, 60.437] - - - [10544, 7424, 1, 256] - - [25, 63.368] - - - [5888, 3329, 1, 256] - - [419, 70.775] - - - [3888, 2816, 1, 256] - - [348, 65.749] - - - [18944, 10240, 1, 256] - - [45, 75.238] - - - [17200, 10240, 1, 256] - - [27, 63.374] - - - [4144, 1280, 1, 256] - - [290, 54.97] - - - [9728, 1280, 1, 256] - - [272, 68.079] - - - [14896, 1536, 1, 256] - - [309, 71.392] - - - [5888, 4352, 1, 256] - - [299, 77.029] - - - [1024, 1024, 1, 256] - - [152, 36.317] - - - [4912, 2816, 1, 256] - - [289, 67.541] - - - [19456, 3329, 1, 256] - - [25, 69.935] - - - [7680, 4608, 1, 256] - - [390, 80.335] - - - [8496, 2865, 1, 256] - - [306, 70.933] - - - [3584, 2048, 1, 256] - - [392, 60.636] - - - [9984, 3329, 1, 256] - - [300, 76.151] - - - [10800, 7680, 1, 256] - - [42, 64.193] - - - [13616, 2816, 1, 256] - - [325, 77.443] - - - [15104, 10240, 1, 256] - - [25, 74.662] - - - [10240, 6656, 1, 256] - - [27, 74.233] - - - [16128, 1281, 1, 256] - - [421, 69.132] - - - [16896, 1280, 1, 256] - - [390, 77.002] - - - [12544, 9472, 1, 256] - - [23, 74.527] - - - [11008, 7424, 1, 256] - - [79, 73.158] - - - [9472, 3329, 1, 256] - - [299, 75.557] - - - [6912, 2816, 1, 256] - - [296, 74.758] - - - [2048, 1841, 1, 256] - - [293, 48.956] - - - [17152, 4096, 1, 256] - - [52, 72.113] - - - [12544, 5120, 1, 256] - - [47, 72.708] - - - [13824, 3328, 1, 256] - - [47, 71.534] - - - [6912, 2048, 1, 256] - - [348, 70.542] - - - [9472, 256, 1, 256] - - [293, 44.785] - - - [9216, 1281, 1, 256] - - [387, 64.514] - - - [7168, 1281, 1, 256] - - [300, 63.145] - - - [10752, 7424, 1, 256] - - [27, 74.342] - - - [16176, 3072, 1, 256] - - [35, 65.657] - - - [12288, 9216, 1, 256] - - [27, 74.52] - - - [14336, 512, 1, 256] - - [289, 60.783] - - - [14336, 3328, 1, 256] - - [23, 72.042] - - - [4864, 1280, 1, 256] - - [392, 60.261] - - - [19760, 2865, 1, 256] - - [34, 60.832] - - - [8240, 256, 1, 256] - - [347, 39.291] - - - [18688, 1024, 1, 256] - - [309, 73.238] - - - [16128, 10240, 1, 256] - - [25, 74.965] - - - [5632, 256, 1, 256] - - [163, 42.857] - - - [5680, 2560, 1, 256] - - [424, 70.284] - - - [7680, 1281, 1, 256] - - [419, 62.077] - - - [17408, 2048, 1, 256] - - [390, 80.891] - - - [10752, 2865, 1, 256] - - [390, 75.525] - - - [14848, 1281, 1, 256] - - [417, 70.001] - - - [560, 512, 1, 256] - - [91, 17.046] - - - [19968, 1280, 1, 256] - - [296, 76.975] - - - [16384, 10240, 1, 256] - - [20, 63.0] - - - [512, 305, 1, 256] - - [150, 9.788] - - - [19200, 6144, 1, 256] - - [27, 73.718] - - - [8448, 5120, 1, 256] - - [23, 72.211] - - - [13824, 3329, 1, 256] - - [27, 68.838] - - - [7984, 2816, 1, 256] - - [299, 73.977] - - - [17920, 3329, 1, 256] - - [49, 69.263] - - - [16688, 2865, 1, 256] - - [47, 63.796] - - - [12032, 256, 1, 256] - - [418, 49.451] - - - [7424, 2865, 1, 256] - - [299, 73.395] - - - [14336, 10240, 1, 256] - - [23, 75.425] - - - [17152, 2048, 1, 256] - - [299, 82.068] - - - [14896, 2816, 1, 256] - - [42, 62.421] - - - [16384, 2048, 1, 256] - - [302, 67.27] - - - [8192, 2816, 1, 256] - - [309, 73.744] - - - [6192, 256, 1, 256] - - [162, 42.027] - - - [2304, 768, 1, 256] - - [264, 40.189] - - - [18688, 256, 1, 256] - - [266, 54.782] - - - [8960, 1281, 1, 256] - - [300, 66.072] - - - [19968, 6400, 1, 256] - - [32, 74.803] - - - [8752, 2816, 1, 256] - - [390, 74.149] - - - [19456, 3328, 1, 256] - - [47, 72.717] - - - [2560, 2561, 1, 256] - - [293, 58.133] - - - [15920, 2865, 1, 256] - - [25, 61.767] - - - [12544, 6144, 1, 256] - - [23, 73.28] - - - [19200, 3328, 1, 256] - - [51, 71.858] - - - [3328, 2865, 1, 256] - - [420, 65.375] - - - [7936, 3329, 1, 256] - - [299, 75.851] - - - [11264, 2865, 1, 256] - - [424, 76.868] - - - [6144, 3329, 1, 256] - - [426, 72.711] - - - [16128, 3329, 1, 256] - - [23, 68.765] - - - [12800, 9728, 1, 256] - - [72, 74.84] - - - [512, 256, 1, 256] - - [147, 8.445] - - - [11264, 2816, 1, 256] - - [390, 80.451] - - - [12544, 3329, 1, 256] - - [25, 67.735] - - - [14848, 3329, 1, 256] - - [25, 69.124] - - - [1328, 256, 1, 256] - - [90, 19.803] - - - [3120, 256, 1, 256] - - [105, 35.025] - - - [1024, 768, 1, 256] - - [103, 35.569] - - - [7728, 2816, 1, 256] - - [299, 72.491] - - - [1024, 817, 1, 256] - - [416, 22.496] - - - [10288, 7424, 1, 256] - - [23, 63.45] - - - [19968, 6144, 1, 256] - - [39, 74.294] - - - [13616, 10240, 1, 256] - - [34, 63.511] - - - [1536, 1329, 1, 256] - - [301, 37.878] - - - [9984, 3328, 1, 256] - - [390, 80.232] - - - [9472, 5888, 1, 256] - - [27, 72.702] - - - [11264, 7936, 1, 256] - - [53, 74.907] - - - [8496, 256, 1, 256] - - [293, 39.408] - - - [17664, 1792, 1, 256] - - [299, 79.272] - - - [11824, 2816, 1, 256] - - [299, 75.762] - - - [16944, 2816, 1, 256] - - [27, 64.078] - - - [19968, 6912, 1, 256] - - [25, 74.983] - - - [3376, 2865, 1, 256] - - [420, 62.807] - - - [3840, 2560, 1, 256] - - [287, 67.779] - - - [11776, 8448, 1, 256] - - [25, 74.414] - - - [19248, 6144, 1, 256] - - [56, 63.548] - - - [14080, 512, 1, 256] - - [291, 60.22] - - - [16128, 3328, 1, 256] - - [47, 71.602] - - - [6656, 2048, 1, 256] - - [357, 68.893] - - - [15664, 256, 1, 256] - - [297, 50.686] - - - [17664, 1280, 1, 256] - - [296, 77.221] - - - [16384, 6144, 1, 256] - - [20, 63.174] - - - [9984, 256, 1, 256] - - [357, 42.399] - - - [14592, 1281, 1, 256] - - [421, 70.2] - - - [4608, 3329, 1, 256] - - [357, 69.702] - - - [8960, 2048, 1, 256] - - [390, 74.001] - - - [2864, 2865, 1, 256] - - [420, 61.077] - - - [2816, 2609, 1, 256] - - [294, 61.52] - - - [14080, 1281, 1, 256] - - [300, 69.075] - - - [1792, 1536, 1, 256] - - [291, 45.074] - - - [10240, 7424, 1, 256] - - [35, 74.721] - - - [5936, 2816, 1, 256] - - [419, 69.519] - - - [19712, 256, 1, 256] - - [290, 54.854] - - - [18944, 5888, 1, 256] - - [29, 74.241] - - - [9728, 3329, 1, 256] - - [390, 76.51] - - - [19248, 2816, 1, 256] - - [51, 62.64] - - - [13568, 1792, 1, 256] - - [299, 78.591] - - - [1584, 1585, 1, 256] - - [290, 39.705] - - - [8704, 2048, 1, 256] - - [357, 74.487] - - - [13056, 9728, 1, 256] - - [23, 74.687] - - - [12800, 2865, 1, 256] - - [299, 77.93] - - - [14336, 6144, 1, 256] - - [25, 74.178] - - - [5120, 1536, 1, 256] - - [293, 62.166] - - - [18432, 1281, 1, 256] - - [300, 70.815] - - - [10240, 256, 1, 256] - - [325, 42.32] - - - [12544, 9216, 1, 256] - - [35, 73.968] - - - [12800, 1281, 1, 256] - - [295, 68.393] - - - [8704, 5888, 1, 256] - - [23, 72.741] - - - [15360, 3329, 1, 256] - - [25, 69.087] - - - [11312, 8448, 1, 256] - - [23, 63.968] - - - [17152, 3328, 1, 256] - - [23, 71.81] - - - [16384, 3328, 1, 256] - - [36, 61.752] - - - [13824, 2816, 1, 256] - - [390, 81.246] - - - [560, 305, 1, 256] - - [95, 10.423] - - - [16432, 256, 1, 256] - - [291, 50.991] - - - [3632, 2865, 1, 256] - - [417, 63.355] - - - [3584, 3328, 1, 256] - - [300, 69.174] - - - [3840, 768, 1, 256] - - [293, 46.795] - - - [19504, 256, 1, 256] - - [308, 51.137] - - - [1280, 1073, 1, 256] - - [169, 40.001] - - - [17712, 10240, 1, 256] - - [51, 63.117] - - - [2816, 1536, 1, 256] - - [344, 52.347] - - - [12800, 6144, 1, 256] - - [57, 73.468] - - - [4656, 2816, 1, 256] - - [298, 69.668] - - - [17920, 10240, 1, 256] - - [35, 75.114] - - - [9984, 2816, 1, 256] - - [289, 79.952] - - - [4352, 256, 1, 256] - - [105, 38.47] - - - [11312, 2865, 1, 256] - - [390, 73.479] - - - [18432, 3328, 1, 256] - - [27, 72.629] - - - [4096, 1280, 1, 256] - - [284, 53.349] - - - [4864, 3329, 1, 256] - - [421, 71.17] - - - [14640, 2865, 1, 256] - - [35, 59.354] - - - [17152, 2816, 1, 256] - - [32, 71.888] - - - [7680, 1280, 1, 256] - - [267, 67.638] - - - [1584, 1536, 1, 256] - - [384, 42.014] - - - [14080, 1280, 1, 256] - - [425, 72.98] - - - [13824, 512, 1, 256] - - [291, 60.167] - - - [7936, 256, 1, 256] - - [259, 38.821] - - - [12592, 2865, 1, 256] - - [320, 74.867] - - - [2816, 2560, 1, 256] - - [287, 62.967] - - - [6912, 1536, 1, 256] - - [298, 66.487] - - - [12800, 9984, 1, 256] - - [23, 75.144] - - - [10496, 256, 1, 256] - - [418, 43.327] - - - [18176, 2865, 1, 256] - - [25, 69.649] - - - [4608, 1536, 1, 256] - - [421, 61.845] - - - [3328, 2816, 1, 256] - - [295, 65.823] - - - [3840, 1024, 1, 256] - - [370, 50.972] - - - [13824, 1280, 1, 256] - - [289, 73.278] - - - [3840, 1281, 1, 256] - - [418, 51.519] - - - [17152, 1281, 1, 256] - - [419, 71.17] - - - [13568, 1281, 1, 256] - - [419, 69.51] - - - [14848, 1792, 1, 256] - - [299, 76.869] - - - [13056, 9472, 1, 256] - - [23, 74.981] - - - [18176, 1281, 1, 256] - - [419, 70.525] - - - [5680, 256, 1, 256] - - [162, 39.74] - - - [13056, 2816, 1, 256] - - [298, 82.923] - - - [11824, 8704, 1, 256] - - [56, 64.185] - - - [7936, 4352, 1, 256] - - [390, 80.991] - - - [8704, 256, 1, 256] - - [418, 41.548] - - - [5424, 2304, 1, 256] - - [299, 67.773] - - - [14128, 768, 1, 256] - - [421, 64.11] - - - [10752, 1024, 1, 256] - - [299, 68.91] - - - [9264, 6144, 1, 256] - - [25, 63.829] - - - [4352, 3328, 1, 256] - - [298, 73.678] - - - [18944, 5632, 1, 256] - - [65, 74.526] - - - [12032, 8704, 1, 256] - - [47, 74.067] - - - [2048, 2049, 1, 256] - - [419, 49.291] - - - [6400, 3329, 1, 256] - - [390, 72.689] - - - [15616, 2560, 1, 256] - - [33, 71.001] - - - [7472, 2865, 1, 256] - - [390, 70.751] - - - [14848, 1536, 1, 256] - - [299, 77.981] - - - [18736, 10240, 1, 256] - - [56, 63.377] - - - [6400, 1024, 1, 256] - - [426, 60.907] - - - [7936, 5120, 1, 256] - - [35, 71.824] - - - [4656, 1536, 1, 256] - - [419, 59.879] - - - [3328, 256, 1, 256] - - [96, 37.36] - - - [3072, 1280, 1, 256] - - [293, 52.515] - - - [2864, 2816, 1, 256] - - [420, 61.084] - - - [9472, 6144, 1, 256] - - [53, 72.458] - - - [3840, 2304, 1, 256] - - [418, 64.444] - - - [17408, 2865, 1, 256] - - [23, 69.641] - - - [16384, 2560, 1, 256] - - [36, 60.625] - - - [16384, 3329, 1, 256] - - [27, 58.92] - - - [16688, 10240, 1, 256] - - [23, 65.624] - - - [18688, 2048, 1, 256] - - [390, 81.364] - - - [7936, 4608, 1, 256] - - [299, 83.245] - - - [9472, 6400, 1, 256] - - [32, 73.356] - - - [14336, 3329, 1, 256] - - [27, 69.103] - - - [4608, 1024, 1, 256] - - [289, 56.288] - - - [16896, 6144, 1, 256] - - [25, 74.464] - - - [10752, 3329, 1, 256] - - [390, 77.3] - - - [6704, 256, 1, 256] - - [293, 33.146] - - - [17408, 6144, 1, 256] - - [23, 74.582] - - - [9984, 2048, 1, 256] - - [289, 74.338] - - - [17968, 4864, 1, 256] - - [34, 63.918] - - - [5120, 3584, 1, 256] - - [422, 74.451] - - - [14336, 2865, 1, 256] - - [35, 69.431] - - - [18736, 256, 1, 256] - - [293, 50.93] - - - [13568, 2048, 1, 256] - - [299, 79.309] - - - [17456, 4352, 1, 256] - - [23, 64.048] - - - [5424, 2816, 1, 256] - - [299, 70.669] - - - [17664, 4608, 1, 256] - - [57, 72.479] - - - [7984, 256, 1, 256] - - [290, 38.082] - - - [6400, 3584, 1, 256] - - [289, 78.707] - - - [19712, 3329, 1, 256] - - [45, 67.597] - - - [10752, 2816, 1, 256] - - [390, 80.179] - - - [17152, 1280, 1, 256] - - [298, 75.302] - - - [560, 561, 1, 256] - - [96, 18.392] - - - [8192, 1281, 1, 256] - - [295, 62.594] - - - [4864, 1792, 1, 256] - - [290, 64.139] - - - [3632, 2816, 1, 256] - - [284, 63.042] - - - [11520, 8448, 1, 256] - - [27, 73.729] - - - [5168, 2865, 1, 256] - - [419, 65.249] - - - [13568, 10240, 1, 256] - - [27, 74.74] - - - [12544, 2816, 1, 256] - - [299, 80.932] - - - [19504, 6144, 1, 256] - - [27, 63.873] - - - [11776, 2048, 1, 256] - - [299, 77.588] - - - [18688, 2865, 1, 256] - - [32, 69.555] - - - [14336, 768, 1, 256] - - [300, 68.355] - - - [18688, 6144, 1, 256] - - [27, 74.125] - - - [11776, 2816, 1, 256] - - [390, 79.432] - - - [12288, 9472, 1, 256] - - [35, 75.314] - - - [5120, 1792, 1, 256] - - [299, 65.339] - - - [16128, 512, 1, 256] - - [291, 60.731] - - - [5376, 3329, 1, 256] - - [420, 70.86] - - - [9216, 3329, 1, 256] - - [390, 76.266] - - - [9008, 2816, 1, 256] - - [423, 73.205] - - - [6448, 3328, 1, 256] - - [289, 73.837] - - - [19968, 3329, 1, 256] - - [25, 69.474] - - - [11520, 8704, 1, 256] - - [65, 74.005] - - - [13824, 256, 1, 256] - - [293, 47.825] - - - [1584, 256, 1, 256] - - [98, 23.152] - - - [10496, 7168, 1, 256] - - [35, 72.807] - - - [5376, 2304, 1, 256] - - [419, 71.732] - - - [10752, 7168, 1, 256] - - [25, 73.268] - - - [18432, 2048, 1, 256] - - [299, 82.183] - - - [12080, 256, 1, 256] - - [426, 46.225] - - - [8704, 3328, 1, 256] - - [299, 80.204] - - - [4608, 1280, 1, 256] - - [294, 59.721] - - - [6192, 3328, 1, 256] - - [391, 70.655] - - - [8704, 3329, 1, 256] - - [299, 76.849] - - - [5424, 2560, 1, 256] - - [289, 68.944] - - - [11008, 2816, 1, 256] - - [299, 77.8] - - - [11776, 4352, 1, 256] - - [57, 72.393] - - - [11008, 1536, 1, 256] - - [294, 72.807] - - - [13312, 3328, 1, 256] - - [25, 71.901] - - - [7168, 4096, 1, 256] - - [299, 79.786] - - - [9216, 256, 1, 256] - - [270, 42.936] - - - [19504, 2865, 1, 256] - - [35, 61.604] - - - [5936, 2865, 1, 256] - - [298, 68.5] - - - [1840, 1792, 1, 256] - - [419, 47.486] - - - [19968, 2816, 1, 256] - - [47, 72.336] - - - [9008, 256, 1, 256] - - [293, 40.084] - - - [9728, 256, 1, 256] - - [392, 45.479] - - - [11056, 7936, 1, 256] - - [25, 64.522] - - - [7680, 3329, 1, 256] - - [298, 75.9] - - - [1792, 256, 1, 256] - - [257, 13.823] - - - [17664, 10240, 1, 256] - - [49, 74.207] - - - [11776, 2865, 1, 256] - - [299, 77.757] - - - [512, 512, 1, 256] - - [156, 15.996] - - - [16640, 768, 1, 256] - - [421, 68.212] - - - [4352, 2048, 1, 256] - - [419, 63.528] - - - [19504, 2816, 1, 256] - - [34, 63.411] - - - [12080, 2865, 1, 256] - - [390, 74.518] - - - [14080, 768, 1, 256] - - [421, 67.979] - - - [7936, 512, 1, 256] - - [38, 44.612] - - - [5376, 2560, 1, 256] - - [421, 69.629] - - - [5632, 3329, 1, 256] - - [298, 71.835] - - - [5120, 3840, 1, 256] - - [308, 75.098] - - - [6192, 2816, 1, 256] - - [425, 70.136] - - - [4608, 3072, 1, 256] - - [390, 72.452] - - - [19712, 6656, 1, 256] - - [57, 73.027] - - - [14896, 256, 1, 256] - - [293, 47.693] - - - [6400, 1280, 1, 256] - - [291, 63.849] - - - [12800, 9216, 1, 256] - - [72, 74.395] - - - [15616, 256, 1, 256] - - [290, 52.008] - - - [17920, 4608, 1, 256] - - [59, 73.278] - - - [7936, 2865, 1, 256] - - [299, 74.851] - - - [13312, 3329, 1, 256] - - [25, 68.915] - - - [5168, 2304, 1, 256] - - [419, 66.699] - - - [14128, 10240, 1, 256] - - [56, 63.75] - - - [3840, 2816, 1, 256] - - [296, 68.054] - - - [8960, 1536, 1, 256] - - [419, 71.822] - - - [3328, 3073, 1, 256] - - [300, 64.24] - - - [4096, 3328, 1, 256] - - [294, 69.246] - - - [14592, 2048, 1, 256] - - [299, 78.89] - - - [9728, 2048, 1, 256] - - [296, 74.819] - - - [13312, 5888, 1, 256] - - [53, 74.216] - - - [15616, 2304, 1, 256] - - [298, 80.723] - - - [19712, 2816, 1, 256] - - [49, 70.242] - - - [9216, 5888, 1, 256] - - [33, 73.129] - - - [7168, 4352, 1, 256] - - [289, 79.086] - - - [9520, 6400, 1, 256] - - [34, 62.601] - - - [13568, 3328, 1, 256] - - [32, 70.67] - - - [17408, 4352, 1, 256] - - [32, 74.236] - - - [11056, 2865, 1, 256] - - [299, 74.071] - - - [18480, 2865, 1, 256] - - [27, 60.402] - - - [13824, 768, 1, 256] - - [295, 66.372] - - - [17664, 6144, 1, 256] - - [65, 73.429] - - - [7216, 4352, 1, 256] - - [426, 76.164] - - - [14128, 2865, 1, 256] - - [35, 60.047] - - - [11520, 3328, 1, 256] - - [357, 81.55] - - - [18992, 5888, 1, 256] - - [42, 63.717] - - - [17408, 10240, 1, 256] - - [35, 75.574] - - - [15104, 6144, 1, 256] - - [25, 73.579] - - - [16640, 1280, 1, 256] - - [299, 75.458] - - - [13056, 2865, 1, 256] - - [390, 79.042] - - - [11776, 8960, 1, 256] - - [47, 74.797] - - - [11312, 2816, 1, 256] - - [299, 75.703] - - - [11264, 3328, 1, 256] - - [390, 80.89] - - - [8192, 512, 1, 256] - - [392, 51.017] - - - [14848, 6144, 1, 256] - - [35, 73.826] - - - [10496, 7680, 1, 256] - - [35, 73.763] - - - [2816, 2817, 1, 256] - - [349, 60.826] - - - [15104, 2865, 1, 256] - - [25, 69.417] - - - [18176, 3329, 1, 256] - - [23, 69.35] - - - [3328, 1792, 1, 256] - - [300, 58.432] - - - [6144, 3328, 1, 256] - - [299, 74.254] - - - [12288, 6144, 1, 256] - - [35, 73.86] - - - [8960, 5888, 1, 256] - - [55, 71.764] - - - [3584, 1280, 1, 256] - - [392, 54.16] - - - [7728, 4608, 1, 256] - - [325, 76.861] - - - [18176, 6144, 1, 256] - - [27, 74.114] - - - [16944, 256, 1, 256] - - [417, 49.351] - - - [3888, 768, 1, 256] - - [304, 45.093] - - - [8448, 3329, 1, 256] - - [296, 75.523] - - - [3072, 3073, 1, 256] - - [357, 64.668] - - - [4912, 256, 1, 256] - - [91, 40.173] - - - [5936, 3072, 1, 256] - - [300, 70.78] - - - [7168, 2865, 1, 256] - - [299, 73.463] - - - [19456, 10240, 1, 256] - - [23, 75.579] - - - [1840, 1585, 1, 256] - - [426, 44.816] - - - [18992, 5632, 1, 256] - - [42, 64.215] - - - [4912, 1792, 1, 256] - - [293, 61.294] - - - [8704, 6144, 1, 256] - - [32, 72.779] - - - [816, 768, 1, 256] - - [105, 29.143] - - - [18432, 2865, 1, 256] - - [23, 69.917] - - - [3120, 2865, 1, 256] - - [422, 60.361] - - - [6448, 2865, 1, 256] - - [306, 69.337] - - - [12080, 2816, 1, 256] - - [320, 76.662] - - - [10496, 3328, 1, 256] - - [390, 80.206] - - - [15920, 10240, 1, 256] - - [51, 64.072] - - - [15872, 2048, 1, 256] - - [390, 80.186] - - - [11568, 2816, 1, 256] - - [325, 75.685] - - - [19200, 10240, 1, 256] - - [35, 74.775] - - - [13312, 5632, 1, 256] - - [35, 74.331] - - - [15360, 2816, 1, 256] - - [27, 71.927] - - - [12288, 2865, 1, 256] - - [296, 76.668] - - - [19760, 6400, 1, 256] - - [42, 64.027] - - - [19968, 256, 1, 256] - - [290, 54.327] - - - [7680, 4352, 1, 256] - - [390, 80.839] - - - [11008, 3584, 1, 256] - - [73, 70.462] - - - [3072, 2817, 1, 256] - - [298, 60.678] - - - [11264, 6144, 1, 256] - - [27, 73.603] - - - [5424, 256, 1, 256] - - [92, 38.221] - - - [13568, 1280, 1, 256] - - [357, 72.353] - - - [3840, 2048, 1, 256] - - [425, 62.03] - - - [6144, 3072, 1, 256] - - [296, 74.304] - - - [19200, 1536, 1, 256] - - [357, 77.506] - - - [10240, 1280, 1, 256] - - [296, 73.092] - - - [3376, 512, 1, 256] - - [351, 34.79] - - - [12544, 1281, 1, 256] - - [421, 68.524] - - - [9776, 6656, 1, 256] - - [42, 63.826] - - - [7680, 2865, 1, 256] - - [419, 74.537] - - - [10544, 7680, 1, 256] - - [51, 63.447] - - - [15616, 3328, 1, 256] - - [23, 71.606] - - - [3328, 1280, 1, 256] - - [392, 52.615] - - - [2560, 1024, 1, 256] - - [291, 42.794] - - - [17456, 4096, 1, 256] - - [56, 64.105] - - - [6912, 3328, 1, 256] - - [296, 78.251] - - - [3584, 2816, 1, 256] - - [293, 66.423] - - - [17408, 3328, 1, 256] - - [23, 72.343] - - - [19200, 2816, 1, 256] - - [53, 71.241] - - - [15104, 1792, 1, 256] - - [299, 78.639] - - - [6144, 256, 1, 256] - - [162, 44.901] - - - [8192, 6144, 1, 256] - - [25, 72.019] - - - [12032, 4608, 1, 256] - - [33, 72.088] - - - [1840, 256, 1, 256] - - [90, 22.542] - - - [13312, 256, 1, 256] - - [423, 52.626] - - - [9216, 1792, 1, 256] - - [304, 73.48] - - - [14592, 3329, 1, 256] - - [55, 68.572] - - - [8448, 1280, 1, 256] - - [298, 66.849] - - - [11520, 8192, 1, 256] - - [55, 73.806] - - - [2608, 2560, 1, 256] - - [301, 57.668] - - - [5120, 2304, 1, 256] - - [287, 68.237] - - - [13056, 3328, 1, 256] - - [33, 70.928] - - - [11008, 8192, 1, 256] - - [73, 72.75] - - - [14896, 2865, 1, 256] - - [25, 60.701] - - - [6704, 3840, 1, 256] - - [307, 74.15] - - - [15872, 3329, 1, 256] - - [23, 69.435] - - - [7168, 1792, 1, 256] - - [299, 70.735] - - - [4656, 2865, 1, 256] - - [421, 65.486] - - - [18736, 5632, 1, 256] - - [56, 63.345] - - - [768, 512, 1, 256] - - [90, 23.02] - - - [16432, 3072, 1, 256] - - [36, 65.493] - - - [14848, 2865, 1, 256] - - [25, 68.873] - - - [4864, 1536, 1, 256] - - [295, 63.626] - - - [9472, 2865, 1, 256] - - [390, 76.432] - - - [10496, 2048, 1, 256] - - [390, 77.355] - - - [14336, 1024, 1, 256] - - [296, 71.962] - - - [18432, 256, 1, 256] - - [392, 53.918] - - - [16896, 3840, 1, 256] - - [65, 73.437] - - - [10240, 512, 1, 256] - - [291, 55.69] - - - [15664, 2304, 1, 256] - - [312, 76.692] - - - [10496, 3329, 1, 256] - - [390, 76.66] - - - [19456, 1536, 1, 256] - - [299, 80.032] - - - [17920, 1281, 1, 256] - - [300, 71.062] - - - [8960, 256, 1, 256] - - [278, 41.456] - - - [10496, 768, 1, 256] - - [290, 62.414] - - - [5120, 2816, 1, 256] - - [299, 71.386] - - - [12288, 2048, 1, 256] - - [296, 75.89] - - - [11568, 8704, 1, 256] - - [34, 63.361] - - - [10496, 1024, 1, 256] - - [357, 66.507] - - - [10288, 256, 1, 256] - - [290, 40.643] - - - [5168, 2048, 1, 256] - - [390, 63.472] - - - [11776, 3328, 1, 256] - - [55, 71.261] - - - [15152, 10240, 1, 256] - - [51, 63.573] - - - [14384, 2865, 1, 256] - - [25, 60.698] - - - [12288, 512, 1, 256] - - [350, 58.246] - - - [16688, 256, 1, 256] - - [293, 49.226] - - - [6912, 4096, 1, 256] - - [390, 79.317] - - - [4864, 2048, 1, 256] - - [313, 67.769] - - - [4096, 1024, 1, 256] - - [290, 51.441] - - - [12848, 9984, 1, 256] - - [51, 64.505] - - - [16896, 1281, 1, 256] - - [300, 70.093] - - - [768, 561, 1, 256] - - [96, 24.478] - - - [16896, 3584, 1, 256] - - [25, 73.171] - - - [14592, 6144, 1, 256] - - [65, 72.431] - - - [17664, 4096, 1, 256] - - [45, 71.584] - - - [8448, 2865, 1, 256] - - [357, 75.116] - - - [18432, 768, 1, 256] - - [289, 71.494] - - - [12032, 512, 1, 256] - - [290, 59.122] - - - [11008, 256, 1, 256] - - [309, 45.291] - - - [15360, 1536, 1, 256] - - [421, 76.97] - - - [5888, 2048, 1, 256] - - [299, 68.557] - - - [13104, 256, 1, 256] - - [315, 49.468] - - - [11264, 7680, 1, 256] - - [27, 74.867] - - - [19248, 2865, 1, 256] - - [56, 60.917] - - - [17200, 2865, 1, 256] - - [25, 61.583] - - - [8192, 2048, 1, 256] - - [350, 69.765] - - - [7472, 4608, 1, 256] - - [299, 76.009] - - - [7168, 2048, 1, 256] - - [296, 72.404] - - - [13360, 2816, 1, 256] - - [325, 76.679] - - - [17920, 4352, 1, 256] - - [55, 73.831] - - - [15408, 256, 1, 256] - - [293, 48.011] - - - [19200, 1281, 1, 256] - - [419, 70.835] - - - [15360, 256, 1, 256] - - [392, 51.704] - - - [9984, 6400, 1, 256] - - [33, 73.374] - - - [18944, 2865, 1, 256] - - [35, 69.777] - - - [3840, 2865, 1, 256] - - [417, 67.733] - - - [8192, 3328, 1, 256] - - [360, 75.704] - - - [5888, 256, 1, 256] - - [170, 43.899] - - - [15616, 2816, 1, 256] - - [25, 71.289] - - - [17664, 2865, 1, 256] - - [57, 69.269] - - - [14592, 768, 1, 256] - - [421, 67.793] - - - [18944, 1281, 1, 256] - - [349, 71.027] - - - [11264, 1536, 1, 256] - - [421, 73.608] - - - [8496, 5632, 1, 256] - - [32, 64.342] - - - [17664, 3328, 1, 256] - - [29, 71.465] - - - [14848, 2048, 1, 256] - - [390, 79.201] - - - [15408, 2865, 1, 256] - - [23, 61.565] - - - [4096, 2048, 1, 256] - - [294, 63.338] - - - [14128, 1024, 1, 256] - - [320, 68.841] - - - [1072, 817, 1, 256] - - [105, 30.424] - - - [17152, 3840, 1, 256] - - [23, 72.897] - - - [17664, 2048, 1, 256] - - [357, 80.648] - - - [16896, 256, 1, 256] - - [418, 51.343] - - - [2304, 2097, 1, 256] - - [419, 54.609] - - - [5888, 2560, 1, 256] - - [289, 75.582] - - - [9472, 1792, 1, 256] - - [289, 72.661] - - - [1328, 1280, 1, 256] - - [344, 37.404] - - - [19200, 1280, 1, 256] - - [390, 76.383] - - - [12544, 1280, 1, 256] - - [267, 72.532] - - - [16432, 3328, 1, 256] - - [26, 66.565] - - - [17920, 1280, 1, 256] - - [298, 77.724] - - - [8752, 5632, 1, 256] - - [27, 63.332] - - - [7936, 2048, 1, 256] - - [348, 73.376] - - - [9472, 1280, 1, 256] - - [298, 68.595] - - - [16896, 1024, 1, 256] - - [299, 73.105] - - - [6656, 3329, 1, 256] - - [419, 73.697] - - - [17456, 2865, 1, 256] - - [27, 62.344] - - - [5632, 2304, 1, 256] - - [348, 71.202] - - - [14080, 1024, 1, 256] - - [309, 70.189] - - - [15872, 3328, 1, 256] - - [32, 72.279] - - - [5168, 2816, 1, 256] - - [298, 69.577] - - - [13312, 9728, 1, 256] - - [23, 75.464] - - - [1584, 1329, 1, 256] - - [388, 37.767] - - - [15664, 2560, 1, 256] - - [35, 62.896] - - - [2048, 768, 1, 256] - - [99, 44.901] - - - [17712, 2816, 1, 256] - - [51, 63.326] - - - [16128, 2865, 1, 256] - - [25, 68.888] - - - [15872, 2816, 1, 256] - - [33, 72.252] - - - [18224, 2816, 1, 256] - - [42, 62.778] - - - [5632, 4352, 1, 256] - - [296, 77.958] - - - [1792, 1281, 1, 256] - - [417, 41.905] - - - [6656, 2816, 1, 256] - - [299, 73.827] - - - [16640, 1281, 1, 256] - - [421, 70.912] - - - [13056, 10240, 1, 256] - - [27, 74.986] - - - [17968, 256, 1, 256] - - [348, 49.92] - - - [5376, 4096, 1, 256] - - [357, 76.276] - - - [15152, 2048, 1, 256] - - [391, 76.255] - - - [13568, 2816, 1, 256] - - [298, 81.484] - - - [12800, 2816, 1, 256] - - [390, 82.121] - - - [6960, 2816, 1, 256] - - [390, 72.067] - - - [17968, 4608, 1, 256] - - [42, 63.285] - - - [15104, 3328, 1, 256] - - [21, 71.526] - - - [7472, 4352, 1, 256] - - [299, 76.418] - - - [15872, 2304, 1, 256] - - [299, 81.284] - - - [4400, 2816, 1, 256] - - [296, 66.694] - - - [16128, 6144, 1, 256] - - [25, 73.909] - - - [18944, 2816, 1, 256] - - [65, 72.175] - - - [5424, 2865, 1, 256] - - [421, 69.528] - - - [8192, 768, 1, 256] - - [420, 57.578] - - - [12848, 256, 1, 256] - - [38, 48.34] - - - [12288, 1281, 1, 256] - - [417, 66.326] - - - [13872, 512, 1, 256] - - [293, 58.684] - - - [5888, 1280, 1, 256] - - [392, 62.63] - - - [2816, 1281, 1, 256] - - [294, 51.498] - - - [19200, 3329, 1, 256] - - [65, 68.61] - - - [12800, 3328, 1, 256] - - [55, 71.351] - - - [15360, 2865, 1, 256] - - [25, 69.192] - - - [17152, 3584, 1, 256] - - [53, 72.518] - - - [17456, 2816, 1, 256] - - [27, 63.271] - - - [18176, 10240, 1, 256] - - [23, 74.987] - - - [6144, 2816, 1, 256] - - [289, 72.539] - - - [18176, 1280, 1, 256] - - [298, 76.618] - - - [16384, 1281, 1, 256] - - [302, 64.246] - - - [9216, 3328, 1, 256] - - [390, 78.836] - - - [14080, 2816, 1, 256] - - [65, 70.205] - - - [18688, 2816, 1, 256] - - [33, 71.742] - - - [15872, 10240, 1, 256] - - [23, 75.281] - - - [10800, 7936, 1, 256] - - [42, 64.944] - - - [9984, 1281, 1, 256] - - [300, 66.688] - - - [4144, 256, 1, 256] - - [92, 36.091] - - - [16640, 6144, 1, 256] - - [25, 73.468] - - - [11776, 4096, 1, 256] - - [72, 71.643] - - - [11056, 8192, 1, 256] - - [27, 64.782] - - - [5376, 2816, 1, 256] - - [299, 71.255] - - - [19712, 1280, 1, 256] - - [306, 74.132] - - - [4608, 2816, 1, 256] - - [289, 72.066] - - - [19456, 2865, 1, 256] - - [25, 70.001] - - - [14080, 256, 1, 256] - - [266, 47.734] - - - [7216, 4096, 1, 256] - - [391, 75.092] - - - [2816, 1280, 1, 256] - - [287, 51.754] - - - [10496, 3072, 1, 256] - - [390, 80.978] - - - [12544, 4864, 1, 256] - - [53, 72.899] - - - [9984, 6912, 1, 256] - - [32, 73.374] - - - [4912, 2048, 1, 256] - - [291, 62.34] - - - [9984, 2304, 1, 256] - - [357, 77.148] - - - [19248, 5888, 1, 256] - - [34, 63.654] - - - [19712, 2048, 1, 256] - - [78, 67.452] - - - [9728, 2816, 1, 256] - - [390, 78.226] - - - [19504, 6400, 1, 256] - - [42, 64.123] - - - [16896, 3072, 1, 256] - - [32, 72.467] - - - [15104, 1536, 1, 256] - - [299, 78.331] - - - [2608, 2609, 1, 256] - - [301, 57.532] - - - [14384, 256, 1, 256] - - [418, 46.033] - - - [17664, 2816, 1, 256] - - [29, 71.206] - - - [9776, 6912, 1, 256] - - [42, 64.424] - - - [1792, 512, 1, 256] - - [90, 33.381] - - - [13312, 1536, 1, 256] - - [299, 74.841] - - - [1072, 1073, 1, 256] - - [90, 36.546] - - - [9472, 2048, 1, 256] - - [299, 75.524] - - - [4608, 3328, 1, 256] - - [301, 71.38] - - - [9984, 2560, 1, 256] - - [299, 77.947] - - - [6912, 5376, 1, 256] - - [299, 81.699] - - - [16640, 2865, 1, 256] - - [32, 68.013] - - - [4352, 3072, 1, 256] - - [299, 73.494] - - - [5632, 3328, 1, 256] - - [299, 74.523] - - - [9216, 5632, 1, 256] - - [27, 73.616] - - - [3328, 3329, 1, 256] - - [284, 68.438] - - - [13824, 10240, 1, 256] - - [49, 74.807] - - - [12288, 3329, 1, 256] - - [25, 68.787] - - - [2864, 256, 1, 256] - - [92, 33.16] - - - [19712, 1792, 1, 256] - - [299, 77.94] - - - [6656, 1280, 1, 256] - - [290, 62.931] - - - [13056, 2048, 1, 256] - - [289, 78.993] - - - [6912, 1281, 1, 256] - - [421, 61.482] - - - [16176, 2816, 1, 256] - - [23, 65.223] - - - [14592, 3328, 1, 256] - - [55, 70.537] - - - [10496, 2865, 1, 256] - - [299, 76.432] - - - [9728, 6400, 1, 256] - - [32, 73.801] - - - [5888, 4608, 1, 256] - - [299, 78.459] - - - [16432, 10240, 1, 256] - - [52, 68.845] - - - [19456, 5888, 1, 256] - - [25, 74.612] - - - [3888, 256, 1, 256] - - [92, 34.576] - - - [12336, 2816, 1, 256] - - [299, 77.022] - - - [19456, 256, 1, 256] - - [308, 55.814] - - - [14384, 2816, 1, 256] - - [27, 62.004] - - - [14384, 1024, 1, 256] - - [390, 66.445] - - - [16640, 2816, 1, 256] - - [32, 70.493] - - - [3840, 3329, 1, 256] - - [296, 69.401] - - - [2304, 2304, 1, 256] - - [270, 56.54] - - - [10240, 2816, 1, 256] - - [299, 79.187] - - - [13104, 10240, 1, 256] - - [34, 64.67] - - - [1536, 256, 1, 256] - - [98, 22.905] - - - [11008, 2865, 1, 256] - - [421, 76.789] - - - [13104, 9984, 1, 256] - - [42, 64.266] - - - [10240, 7168, 1, 256] - - [23, 73.574] - - - [3888, 2865, 1, 256] - - [304, 64.893] - - - [8192, 4864, 1, 256] - - [23, 70.814] - - - [15920, 256, 1, 256] - - [392, 50.563] - - - [6448, 3584, 1, 256] - - [299, 73.768] - - - [16128, 2304, 1, 256] - - [296, 80.588] - - - [9728, 2865, 1, 256] - - [299, 76.416] - - - [6144, 4864, 1, 256] - - [298, 79.516] - - - [14848, 256, 1, 256] - - [290, 50.099] - - - [4352, 1024, 1, 256] - - [300, 53.213] - - - [15360, 10240, 1, 256] - - [27, 75.504] - - - [19504, 10240, 1, 256] - - [26, 63.61] - - - [3328, 3072, 1, 256] - - [421, 66.48] - - - [1536, 1281, 1, 256] - - [284, 37.654] - - - [19760, 6656, 1, 256] - - [56, 63.827] - - - [3584, 3329, 1, 256] - - [419, 68.236] - - - [14848, 2816, 1, 256] - - [32, 71.299] - - - [4400, 2865, 1, 256] - - [426, 66.835] - - - [3888, 1024, 1, 256] - - [370, 49.874] - - - [16640, 2048, 1, 256] - - [299, 79.355] - - - [4096, 2816, 1, 256] - - [306, 68.527] - - - [14640, 2816, 1, 256] - - [42, 62.79] - - - [9472, 1281, 1, 256] - - [289, 66.141] - - - [8192, 1280, 1, 256] - - [357, 63.488] - - - [8960, 2865, 1, 256] - - [296, 75.706] - - - [4144, 2816, 1, 256] - - [417, 64.872] - - - [10288, 7168, 1, 256] - - [23, 62.857] - - - [14592, 256, 1, 256] - - [270, 48.689] - - - [10240, 2048, 1, 256] - - [357, 76.107] - - - [17920, 2865, 1, 256] - - [45, 69.514] - - - [12592, 2816, 1, 256] - - [325, 76.832] - - - [14592, 1536, 1, 256] - - [390, 76.766] - - - [11568, 256, 1, 256] - - [392, 45.512] - - - [6704, 3584, 1, 256] - - [422, 74.297] - - - [5120, 3328, 1, 256] - - [300, 72.562] - - - [4400, 1536, 1, 256] - - [420, 57.112] - - - [18944, 256, 1, 256] - - [392, 54.288] - - - [19712, 5888, 1, 256] - - [39, 72.495] - - - [7984, 5120, 1, 256] - - [32, 64.029] - - - [8240, 2865, 1, 256] - - [296, 74.374] - - - [6144, 1280, 1, 256] - - [293, 61.713] - - - [8496, 2816, 1, 256] - - [320, 74.059] - - - [14592, 1024, 1, 256] - - [299, 70.575] - - - [14592, 2865, 1, 256] - - [29, 68.747] - - - [13360, 256, 1, 256] - - [392, 44.418] - - - [8448, 256, 1, 256] - - [392, 40.66] - - - [16896, 2816, 1, 256] - - [53, 72.165] - - - [15152, 2865, 1, 256] - - [35, 61.182] - - - [11056, 2816, 1, 256] - - [423, 76.256] - - - [15616, 1280, 1, 256] - - [298, 74.76] - - - [8192, 5120, 1, 256] - - [25, 70.894] - - - [17408, 256, 1, 256] - - [298, 52.673] - - - [18432, 10240, 1, 256] - - [35, 75.394] - - - [14592, 1280, 1, 256] - - [298, 73.688] - - - [3328, 512, 1, 256] - - [161, 48.048] - - - [14336, 1280, 1, 256] - - [313, 73.903] - - - [13616, 2865, 1, 256] - - [299, 75.182] - - - [8192, 256, 1, 256] - - [349, 38.34] - - - [10240, 1281, 1, 256] - - [287, 67.487] - - - [1840, 1841, 1, 256] - - [294, 47.747] - - - [12800, 9472, 1, 256] - - [27, 75.03] - - - [17664, 256, 1, 256] - - [38, 48.45] - - - [768, 769, 1, 256] - - [355, 16.767] - - - [19456, 2048, 1, 256] - - [34, 69.833] - - - [13056, 3329, 1, 256] - - [35, 67.803] - - - [11056, 256, 1, 256] - - [370, 42.769] - - - [7424, 6144, 1, 256] - - [33, 71.699] - - - [14848, 3328, 1, 256] - - [27, 71.856] - - - [6656, 3328, 1, 256] - - [301, 75.624] - - - [10752, 1281, 1, 256] - - [421, 66.27] - - - [9984, 2865, 1, 256] - - [298, 75.823] - - - [14080, 3329, 1, 256] - - [55, 68.263] - - - [17920, 3328, 1, 256] - - [74, 72.159] - - - [13312, 10240, 1, 256] - - [25, 75.296] - - - [16640, 3584, 1, 256] - - [32, 71.857] - - - [17408, 3840, 1, 256] - - [35, 73.865] - - - [12032, 8960, 1, 256] - - [32, 74.177] - - - [10800, 2865, 1, 256] - - [307, 73.698] - - - [3072, 2816, 1, 256] - - [418, 63.187] - - - [14128, 2816, 1, 256] - - [56, 61.885] - - - [11312, 8192, 1, 256] - - [27, 64.421] - - - [2560, 2305, 1, 256] - - [420, 58.296] - - - [16640, 3072, 1, 256] - - [25, 70.804] - - - [16128, 2048, 1, 256] - - [390, 79.823] - - - [6144, 512, 1, 256] - - [357, 48.833] - - - [18688, 4864, 1, 256] - - [33, 73.833] - - - [17200, 256, 1, 256] - - [392, 49.809] - - - [8752, 2865, 1, 256] - - [307, 71.967] - - - [18944, 1280, 1, 256] - - [298, 76.075] - - - [16640, 3328, 1, 256] - - [46, 70.911] - - - [304, 305, 1, 256] - - [101, 5.569] - - - [15104, 256, 1, 256] - - [270, 50.397] - - - [7680, 3328, 1, 256] - - [296, 77.446] - - - [12336, 9216, 1, 256] - - [23, 63.141] - - - [14080, 6144, 1, 256] - - [65, 72.26] - - - [7168, 5888, 1, 256] - - [25, 72.398] - - - [7424, 1280, 1, 256] - - [357, 65.898] - - - [4864, 3584, 1, 256] - - [296, 74.491] - - - [1280, 1025, 1, 256] - - [347, 31.237] - - - [10240, 2865, 1, 256] - - [299, 76.6] - - - [18480, 10240, 1, 256] - - [23, 62.84] - - - [7680, 256, 1, 256] - - [343, 38.202] - - - [9472, 6656, 1, 256] - - [32, 73.381] - - - [12032, 6144, 1, 256] - - [27, 72.815] - - - [5120, 3329, 1, 256] - - [296, 70.643] - - - [10752, 256, 1, 256] - - [345, 43.633] - - - [6960, 256, 1, 256] - - [355, 33.521] - - - [9008, 6144, 1, 256] - - [56, 63.58] - - - [7424, 2048, 1, 256] - - [299, 71.756] - - - [5632, 1280, 1, 256] - - [392, 61.682] - - - [19712, 10240, 1, 256] - - [39, 73.761] - - - [6400, 768, 1, 256] - - [293, 56.586] - - - [10752, 3328, 1, 256] - - [299, 81.099] - - - [18432, 5376, 1, 256] - - [25, 74.701] - - - [9520, 256, 1, 256] - - [291, 41.641] - - - [5680, 2816, 1, 256] - - [391, 70.447] - - - [11008, 3329, 1, 256] - - [421, 77.281] - - - [4608, 2865, 1, 256] - - [419, 69.269] - - - [6448, 256, 1, 256] - - [171, 43.664] - - - [3584, 256, 1, 256] - - [257, 23.591] - - - [12336, 9472, 1, 256] - - [35, 63.293] - - - [1280, 1281, 1, 256] - - [264, 34.532] - - - [7936, 6144, 1, 256] - - [57, 71.94] - - - [15152, 1792, 1, 256] - - [325, 74.395] - - - [4352, 1281, 1, 256] - - [418, 56.011] - - - [12848, 2865, 1, 256] - - [299, 74.676] - - - [16944, 3584, 1, 256] - - [34, 64.55] - - - [8752, 256, 1, 256] - - [418, 40.333] - - - [6912, 256, 1, 256] - - [256, 35.599] - - - [14336, 1281, 1, 256] - - [301, 68.194] - - - [2304, 2049, 1, 256] - - [287, 52.814] - - - [9216, 6144, 1, 256] - - [27, 73.135] - - - [1072, 1024, 1, 256] - - [92, 36.172] - - - [10752, 6144, 1, 256] - - [25, 73.602] - - - [1792, 1537, 1, 256] - - [312, 43.718] - - - [17968, 2865, 1, 256] - - [23, 61.071] - - - [8448, 4864, 1, 256] - - [32, 72.259] - - - [15408, 2048, 1, 256] - - [423, 74.717] - - - [15104, 1280, 1, 256] - - [422, 73.837] - - - [9264, 2865, 1, 256] - - [423, 71.703] - - - [15616, 2865, 1, 256] - - [33, 69.029] - - - [16896, 10240, 1, 256] - - [23, 75.376] - - - [15104, 2816, 1, 256] - - [58, 71.358] - - - [13872, 2865, 1, 256] - - [23, 60.894] - - - [13056, 1280, 1, 256] - - [298, 75.728] - - - [12288, 3328, 1, 256] - - [25, 71.562] - - - [3840, 3328, 1, 256] - - [299, 71.002] - - - [9216, 1280, 1, 256] - - [421, 67.322] - - - [8448, 1281, 1, 256] - - [287, 63.56] - - - [2560, 256, 1, 256] - - [105, 31.104] - - - [4608, 1281, 1, 256] - - [420, 57.494] - - - [6144, 2865, 1, 256] - - [299, 71.237] - - - [5888, 2816, 1, 256] - - [391, 73.523] - - - [3584, 1281, 1, 256] - - [419, 53.264] - - - [18688, 5120, 1, 256] - - [23, 73.865] - - - [12288, 2816, 1, 256] - - [357, 80.061] - - - [4864, 2865, 1, 256] - - [299, 68.817] - - - [9216, 2048, 1, 256] - - [390, 75.288] - - - [13872, 768, 1, 256] - - [300, 63.933] - - - [10496, 7424, 1, 256] - - [23, 73.594] - - - [16384, 512, 1, 256] - - [418, 54.998] - - - [14848, 10240, 1, 256] - - [25, 74.949] - - - [17920, 2048, 1, 256] - - [390, 81.276] - - - [11008, 7936, 1, 256] - - [69, 73.234] - - - [1792, 1792, 1, 256] - - [293, 50.637] - - - [7680, 4864, 1, 256] - - [299, 82.053] - - - [19760, 256, 1, 256] - - [56, 49.25] - - - [15616, 1792, 1, 256] - - [299, 78.686] - - - [1792, 1793, 1, 256] - - [293, 48.001] - - - [8192, 3329, 1, 256] - - [295, 72.699] - - - [2560, 1280, 1, 256] - - [293, 50.662] - - - [1328, 1073, 1, 256] - - [385, 32.955] - - - [16896, 2865, 1, 256] - - [35, 69.675] - - - [8960, 1280, 1, 256] - - [298, 70.997] - - - [6960, 2865, 1, 256] - - [289, 69.834] - - - [1280, 1024, 1, 256] - - [347, 32.389] - - - [6400, 2048, 1, 256] - - [390, 71.921] - - - [18480, 5376, 1, 256] - - [27, 62.704] - - - [18944, 2048, 1, 256] - - [299, 81.145] - - - [9520, 6656, 1, 256] - - [34, 63.438] - - - [4352, 1536, 1, 256] - - [308, 61.962] - - - [19712, 6144, 1, 256] - - [49, 72.546] - - - [6400, 2816, 1, 256] - - [422, 74.104] - - - [1792, 1585, 1, 256] - - [299, 45.139] - - - [13312, 6144, 1, 256] - - [23, 74.102] - - - [17408, 4096, 1, 256] - - [25, 72.71] - - - [16128, 256, 1, 256] - - [293, 49.983] - - - [15104, 2048, 1, 256] - - [299, 79.914] - - - [8704, 2865, 1, 256] - - [289, 74.751] - - - [6144, 768, 1, 256] - - [287, 54.535] - - - [10496, 1280, 1, 256] - - [299, 68.971] - - - [816, 561, 1, 256] - - [91, 21.462] - - - [6912, 3840, 1, 256] - - [390, 79.742] - - - [8704, 1281, 1, 256] - - [421, 64.369] - - - [13312, 1792, 1, 256] - - [390, 76.142] - - - [5120, 1281, 1, 256] - - [417, 57.06] - - - [10496, 1281, 1, 256] - - [300, 67.128] - - - [8448, 6144, 1, 256] - - [32, 72.701] - - - [2560, 2353, 1, 256] - - [287, 58.609] - - - [4352, 1280, 1, 256] - - [293, 58.143] - - - [12336, 256, 1, 256] - - [291, 46.351] - - - [21504, 10240, 1, 256] - - [27, 75.583] - - - [31744, 6144, 1, 256] - - [25, 74.524] - - - [27648, 1280, 1, 256] - - [296, 80.263] - - - [22272, 512, 1, 256] - - [299, 69.316] - - - [29184, 256, 1, 256] - - [293, 60.865] - - - [23808, 4096, 1, 256] - - [72, 72.707] - - - [30720, 7168, 1, 256] - - [35, 73.869] - - - [29440, 2865, 1, 256] - - [25, 70.006] - - - [25600, 5632, 1, 256] - - [25, 75.075] - - - [24832, 10240, 1, 256] - - [25, 74.967] - - - [22784, 2865, 1, 256] - - [27, 69.758] - - - [24368, 768, 1, 256] - - [428, 69.991] - - - [21760, 8192, 1, 256] - - [35, 74.528] - - - [29184, 10240, 1, 256] - - [22, 74.869] - - - [26368, 6144, 1, 256] - - [25, 74.073] - - - [23088, 10240, 1, 256] - - [51, 63.352] - - - [29952, 4096, 1, 256] - - [27, 72.803] - - - [24320, 6144, 1, 256] - - [28, 74.167] - - - [32256, 2048, 1, 256] - - [63, 71.243] - - - [29488, 2865, 1, 256] - - [35, 60.926] - - - [21808, 2816, 1, 256] - - [56, 63.226] - - - [32000, 7936, 1, 256] - - [25, 74.425] - - - [23040, 10240, 1, 256] - - [23, 75.226] - - - [31792, 10240, 1, 256] - - [26, 63.633] - - - [24320, 1281, 1, 256] - - [421, 72.011] - - - [27136, 3072, 1, 256] - - [49, 73.24] - - - [31488, 6144, 1, 256] - - [25, 74.077] - - - [34096, 2865, 1, 256] - - [42, 62.013] - - - [33024, 8960, 1, 256] - - [28, 74.681] - - - [28928, 1280, 1, 256] - - [296, 79.882] - - - [31488, 5632, 1, 256] - - [25, 74.077] - - - [27696, 4096, 1, 256] - - [23, 62.899] - - - [31488, 10240, 1, 256] - - [35, 74.538] - - - [28928, 10240, 1, 256] - - [27, 74.419] - - - [26160, 10240, 1, 256] - - [34, 62.983] - - - [26112, 3328, 1, 256] - - [82, 73.184] - - - [28928, 4864, 1, 256] - - [27, 73.672] - - - [27904, 3328, 1, 256] - - [77, 72.375] - - - [29184, 5376, 1, 256] - - [28, 74.353] - - - [29952, 1281, 1, 256] - - [349, 74.629] - - - [24832, 6144, 1, 256] - - [22, 74.183] - - - [28160, 4096, 1, 256] - - [40, 73.284] - - - [24320, 1280, 1, 256] - - [298, 78.752] - - - [34816, 768, 1, 256] - - [390, 77.283] - - - [34816, 1281, 1, 256] - - [25, 63.991] - - - [27136, 2816, 1, 256] - - [55, 73.176] - - - [32256, 8192, 1, 256] - - [28, 74.855] - - - [26624, 2865, 1, 256] - - [25, 70.664] - - - [23808, 3840, 1, 256] - - [25, 73.539] - - - [29440, 5376, 1, 256] - - [55, 74.171] - - - [30464, 10240, 1, 256] - - [22, 73.683] - - - [29232, 10240, 1, 256] - - [34, 62.969] - - - [27136, 1280, 1, 256] - - [296, 78.655] - - - [27904, 6144, 1, 256] - - [26, 73.746] - - - [33024, 2816, 1, 256] - - [40, 72.269] - - - [34816, 3329, 1, 256] - - [23, 70.685] - - - [34048, 1792, 1, 256] - - [29, 71.4] - - - [21248, 7424, 1, 256] - - [35, 74.819] - - - [29952, 256, 1, 256] - - [293, 59.977] - - - [34560, 256, 1, 256] - - [418, 62.58] - - - [26368, 3072, 1, 256] - - [35, 72.767] - - - [23600, 2865, 1, 256] - - [35, 61.367] - - - [30720, 512, 1, 256] - - [289, 71.902] - - - [30768, 10240, 1, 256] - - [39, 62.81] - - - [28928, 1024, 1, 256] - - [390, 77.33] - - - [26624, 256, 1, 256] - - [290, 58.907] - - - [26928, 10240, 1, 256] - - [51, 63.381] - - - [21248, 7936, 1, 256] - - [25, 74.601] - - - [34304, 2816, 1, 256] - - [57, 73.087] - - - [29696, 3840, 1, 256] - - [25, 74.502] - - - [27696, 10240, 1, 256] - - [25, 63.529] - - - [24064, 2048, 1, 256] - - [70, 70.68] - - - [33536, 6144, 1, 256] - - [28, 74.074] - - - [32512, 8704, 1, 256] - - [45, 74.767] - - - [21552, 2816, 1, 256] - - [51, 63.568] - - - [27648, 10240, 1, 256] - - [35, 75.468] - - - [22272, 2048, 1, 256] - - [31, 70.036] - - - [28976, 5632, 1, 256] - - [51, 63.367] - - - [30720, 10240, 1, 256] - - [23, 75.334] - - - [26112, 2816, 1, 256] - - [25, 73.015] - - - [20528, 10240, 1, 256] - - [26, 62.326] - - - [29696, 1536, 1, 256] - - [53, 70.945] - - - [31536, 2865, 1, 256] - - [51, 61.567] - - - [32000, 3328, 1, 256] - - [47, 72.394] - - - [20784, 2865, 1, 256] - - [23, 60.364] - - - [33280, 9984, 1, 256] - - [49, 75.53] - - - [25600, 3329, 1, 256] - - [30, 70.116] - - - [27904, 4096, 1, 256] - - [63, 72.935] - - - [29488, 256, 1, 256] - - [297, 55.118] - - - [32048, 10240, 1, 256] - - [25, 63.133] - - - [31280, 2865, 1, 256] - - [51, 61.474] - - - [32816, 2816, 1, 256] - - [26, 65.244] - - - [34096, 2816, 1, 256] - - [42, 63.755] - - - [20992, 3328, 1, 256] - - [23, 72.666] - - - [32768, 1281, 1, 256] - - [86, 51.657] - - - [24576, 4864, 1, 256] - - [36, 70.216] - - - [30464, 3328, 1, 256] - - [57, 71.563] - - - [28208, 256, 1, 256] - - [290, 55.098] - - - [23552, 1280, 1, 256] - - [296, 79.612] - - - [20528, 7168, 1, 256] - - [23, 61.991] - - - [34560, 2865, 1, 256] - - [35, 70.251] - - - [20736, 2816, 1, 256] - - [32, 72.019] - - - [26880, 3328, 1, 256] - - [23, 72.345] - - - [31536, 8192, 1, 256] - - [56, 62.985] - - - [31744, 8448, 1, 256] - - [23, 75.166] - - - [20224, 2865, 1, 256] - - [25, 69.451] - - - [22528, 2048, 1, 256] - - [59, 69.864] - - - [24320, 2048, 1, 256] - - [67, 70.443] - - - [32512, 8960, 1, 256] - - [45, 74.928] - - - [33072, 10240, 1, 256] - - [27, 64.214] - - - [24880, 10240, 1, 256] - - [27, 63.243] - - - [21040, 7680, 1, 256] - - [34, 64.019] - - - [26368, 10240, 1, 256] - - [23, 74.681] - - - [32304, 8704, 1, 256] - - [56, 63.345] - - - [33536, 1281, 1, 256] - - [72, 63.723] - - - [27136, 1024, 1, 256] - - [299, 78.445] - - - [33792, 1281, 1, 256] - - [39, 64.358] - - - [33584, 256, 1, 256] - - [25, 54.154] - - - [20528, 7424, 1, 256] - - [25, 62.655] - - - [28928, 2865, 1, 256] - - [27, 69.869] - - - [22016, 2048, 1, 256] - - [46, 70.212] - - - [29440, 3328, 1, 256] - - [59, 72.706] - - - [30208, 2048, 1, 256] - - [42, 71.22] - - - [20480, 2816, 1, 256] - - [25, 72.352] - - - [25904, 256, 1, 256] - - [291, 55.538] - - - [20736, 10240, 1, 256] - - [25, 74.9] - - - [32816, 256, 1, 256] - - [291, 56.857] - - - [33792, 3328, 1, 256] - - [23, 73.479] - - - [22272, 1281, 1, 256] - - [285, 73.045] - - - [25600, 1280, 1, 256] - - [296, 79.19] - - - [33280, 3329, 1, 256] - - [22, 70.393] - - - [22784, 1281, 1, 256] - - [349, 72.287] - - - [25392, 2816, 1, 256] - - [51, 63.703] - - - [33280, 3328, 1, 256] - - [27, 73.324] - - - [21760, 1280, 1, 256] - - [296, 78.983] - - - [33024, 768, 1, 256] - - [299, 75.201] - - - [25088, 1792, 1, 256] - - [53, 71.376] - - - [26368, 3329, 1, 256] - - [52, 69.349] - - - [34560, 3328, 1, 256] - - [59, 72.635] - - - [23040, 6144, 1, 256] - - [45, 74.371] - - - [30464, 2048, 1, 256] - - [82, 68.048] - - - [28672, 3328, 1, 256] - - [27, 72.77] - - - [30464, 6912, 1, 256] - - [28, 73.595] - - - [32048, 2816, 1, 256] - - [34, 63.364] - - - [33792, 9728, 1, 256] - - [23, 75.271] - - - [27392, 1536, 1, 256] - - [29, 66.59] - - - [24112, 512, 1, 256] - - [423, 65.164] - - - [28160, 256, 1, 256] - - [392, 60.638] - - - [34816, 2048, 1, 256] - - [26, 70.698] - - - [25648, 10240, 1, 256] - - [39, 63.438] - - - [20992, 10240, 1, 256] - - [23, 75.383] - - - [22528, 1281, 1, 256] - - [390, 72.285] - - - [25904, 2304, 1, 256] - - [51, 62.878] - - - [27952, 2865, 1, 256] - - [56, 61.685] - - - [30976, 768, 1, 256] - - [390, 76.061] - - - [20480, 3329, 1, 256] - - [30, 69.603] - - - [33072, 256, 1, 256] - - [418, 57.314] - - - [26624, 2560, 1, 256] - - [23, 73.237] - - - [28208, 2865, 1, 256] - - [56, 61.173] - - - [26672, 3328, 1, 256] - - [23, 61.969] - - - [26880, 2865, 1, 256] - - [52, 69.839] - - - [26112, 2304, 1, 256] - - [35, 72.648] - - - [29184, 5120, 1, 256] - - [35, 74.128] - - - [29744, 6144, 1, 256] - - [39, 62.89] - - - [30464, 3329, 1, 256] - - [68, 68.208] - - - [22272, 2560, 1, 256] - - [29, 72.039] - - - [25344, 2048, 1, 256] - - [37, 69.446] - - - [31792, 256, 1, 256] - - [418, 57.409] - - - [21248, 2816, 1, 256] - - [33, 71.886] - - - [32816, 10240, 1, 256] - - [72, 65.69] - - - [27136, 3840, 1, 256] - - [57, 74.215] - - - [34096, 10240, 1, 256] - - [39, 62.84] - - - [24576, 4608, 1, 256] - - [20, 69.32] - - - [32256, 1281, 1, 256] - - [37, 64.32] - - - [26928, 2865, 1, 256] - - [51, 60.727] - - - [20784, 7424, 1, 256] - - [34, 63.384] - - - [24112, 2816, 1, 256] - - [56, 64.168] - - - [22272, 256, 1, 256] - - [392, 56.705] - - - [30208, 1281, 1, 256] - - [349, 74.278] - - - [28720, 2816, 1, 256] - - [23, 61.361] - - - [20992, 1280, 1, 256] - - [299, 78.052] - - - [31488, 1536, 1, 256] - - [27, 70.299] - - - [21296, 8192, 1, 256] - - [51, 63.224] - - - [30512, 7168, 1, 256] - - [56, 62.559] - - - [27136, 2865, 1, 256] - - [23, 70.677] - - - [25088, 3329, 1, 256] - - [49, 69.78] - - - [29696, 3329, 1, 256] - - [35, 70.496] - - - [23040, 1280, 1, 256] - - [348, 76.364] - - - [30000, 256, 1, 256] - - [308, 55.615] - - - [20224, 3329, 1, 256] - - [35, 69.242] - - - [29232, 2816, 1, 256] - - [42, 63.791] - - - [31232, 7424, 1, 256] - - [45, 75.246] - - - [29488, 2816, 1, 256] - - [56, 63.082] - - - [25904, 2865, 1, 256] - - [34, 61.333] - - - [30512, 2816, 1, 256] - - [51, 63.905] - - - [20736, 768, 1, 256] - - [299, 73.166] - - - [20480, 256, 1, 256] - - [418, 53.074] - - - [28672, 6144, 1, 256] - - [25, 74.135] - - - [26624, 2816, 1, 256] - - [35, 73.242] - - - [28928, 768, 1, 256] - - [289, 75.448] - - - [27648, 256, 1, 256] - - [392, 58.948] - - - [32256, 6144, 1, 256] - - [45, 74.469] - - - [30720, 6144, 1, 256] - - [35, 74.708] - - - [32560, 2865, 1, 256] - - [51, 63.06] - - - [23088, 9728, 1, 256] - - [42, 63.913] - - - [22784, 9728, 1, 256] - - [26, 74.598] - - - [33024, 6144, 1, 256] - - [41, 73.922] - - - [27392, 2865, 1, 256] - - [39, 67.581] - - - [21504, 1280, 1, 256] - - [390, 77.666] - - - [30720, 6656, 1, 256] - - [25, 75.049] - - - [24880, 2865, 1, 256] - - [34, 62.105] - - - [25392, 1792, 1, 256] - - [42, 61.999] - - - [20224, 2816, 1, 256] - - [53, 71.778] - - - [20224, 256, 1, 256] - - [290, 54.053] - - - [25856, 3329, 1, 256] - - [23, 69.116] - - - [30976, 256, 1, 256] - - [290, 60.795] - - - [26880, 6144, 1, 256] - - [35, 74.125] - - - [26672, 2816, 1, 256] - - [35, 62.709] - - - [25600, 256, 1, 256] - - [299, 59.425] - - - [28160, 1281, 1, 256] - - [285, 73.114] - - - [20480, 10240, 1, 256] - - [25, 75.272] - - - [21504, 7936, 1, 256] - - [23, 75.534] - - - [20272, 7168, 1, 256] - - [42, 62.663] - - - [24880, 2816, 1, 256] - - [42, 64.271] - - - [23296, 9728, 1, 256] - - [27, 74.626] - - - [34816, 2865, 1, 256] - - [25, 71.003] - - - [31792, 2865, 1, 256] - - [27, 62.045] - - - [29488, 6144, 1, 256] - - [42, 62.997] - - - [23856, 2865, 1, 256] - - [34, 60.587] - - - [25088, 256, 1, 256] - - [290, 59.285] - - - [22016, 8960, 1, 256] - - [29, 75.62] - - - [23040, 3072, 1, 256] - - [25, 72.891] - - - [23856, 512, 1, 256] - - [325, 63.79] - - - [33792, 3329, 1, 256] - - [25, 70.407] - - - [22784, 9216, 1, 256] - - [26, 74.474] - - - [30720, 4864, 1, 256] - - [35, 74.739] - - - [32000, 8192, 1, 256] - - [25, 74.349] - - - [28160, 3329, 1, 256] - - [30, 69.566] - - - [28672, 256, 1, 256] - - [418, 59.775] - - - [27648, 1281, 1, 256] - - [349, 73.376] - - - [23808, 6144, 1, 256] - - [35, 74.136] - - - [23344, 10240, 1, 256] - - [34, 62.842] - - - [20736, 7680, 1, 256] - - [23, 74.828] - - - [33024, 9216, 1, 256] - - [40, 74.582] - - - [26160, 2816, 1, 256] - - [56, 63.826] - - - [24064, 10240, 1, 256] - - [45, 75.204] - - - [24320, 768, 1, 256] - - [300, 74.212] - - - [28208, 10240, 1, 256] - - [72, 62.743] - - - [34560, 1024, 1, 256] - - [299, 79.683] - - - [33792, 1792, 1, 256] - - [35, 72.448] - - - [30720, 2816, 1, 256] - - [27, 73.516] - - - [24624, 2816, 1, 256] - - [39, 62.429] - - - [20736, 3329, 1, 256] - - [27, 68.948] - - - [21760, 1792, 1, 256] - - [289, 80.565] - - - [21760, 8704, 1, 256] - - [25, 75.019] - - - [34608, 10240, 1, 256] - - [42, 63.342] - - - [22784, 9472, 1, 256] - - [65, 74.935] - - - [31536, 2816, 1, 256] - - [56, 63.566] - - - [27904, 4352, 1, 256] - - [57, 73.693] - - - [23552, 2865, 1, 256] - - [27, 70.547] - - - [24064, 256, 1, 256] - - [290, 58.123] - - - [34304, 2048, 1, 256] - - [39, 71.565] - - - [30464, 1280, 1, 256] - - [267, 78.866] - - - [29440, 5632, 1, 256] - - [29, 74.412] - - - [21808, 8704, 1, 256] - - [34, 63.111] - - - [30464, 6656, 1, 256] - - [22, 73.443] - - - [20736, 1024, 1, 256] - - [299, 76.619] - - - [24832, 1024, 1, 256] - - [390, 76.628] - - - [24576, 1024, 1, 256] - - [360, 69.679] - - - [29184, 2048, 1, 256] - - [72, 71.095] - - - [30976, 4864, 1, 256] - - [65, 72.422] - - - [25344, 1536, 1, 256] - - [299, 80.482] - - - [22016, 1280, 1, 256] - - [299, 78.985] - - - [32560, 8960, 1, 256] - - [51, 64.641] - - - [31536, 7936, 1, 256] - - [51, 64.014] - - - [26880, 3072, 1, 256] - - [35, 72.76] - - - [28464, 2865, 1, 256] - - [51, 61.458] - - - [20224, 6400, 1, 256] - - [25, 74.537] - - - [26624, 3328, 1, 256] - - [53, 73.411] - - - [24320, 512, 1, 256] - - [296, 67.76] - - - [34352, 768, 1, 256] - - [424, 74.97] - - - [30720, 768, 1, 256] - - [299, 75.853] - - - [34560, 10240, 1, 256] - - [35, 74.662] - - - [22016, 3328, 1, 256] - - [32, 72.823] - - - [20480, 1281, 1, 256] - - [417, 70.317] - - - [31232, 2816, 1, 256] - - [35, 72.918] - - - [31232, 6144, 1, 256] - - [28, 74.376] - - - [27136, 256, 1, 256] - - [293, 58.325] - - - [23344, 256, 1, 256] - - [320, 52.566] - - - [30208, 4352, 1, 256] - - [49, 74.258] - - - [32000, 6144, 1, 256] - - [27, 74.002] - - - [29184, 6144, 1, 256] - - [25, 74.248] - - - [29232, 5632, 1, 256] - - [56, 63.1] - - - [22576, 2816, 1, 256] - - [35, 62.212] - - - [31488, 1280, 1, 256] - - [32, 70.391] - - - [23856, 2816, 1, 256] - - [51, 63.075] - - - [29184, 2865, 1, 256] - - [55, 70.156] - - - [21248, 6144, 1, 256] - - [25, 74.121] - - - [30720, 4608, 1, 256] - - [35, 74.138] - - - [27952, 256, 1, 256] - - [293, 55.497] - - - [32512, 10240, 1, 256] - - [28, 74.736] - - - [31744, 3328, 1, 256] - - [23, 73.375] - - - [22528, 3328, 1, 256] - - [27, 73.174] - - - [34048, 3329, 1, 256] - - [71, 69.195] - - - [31744, 2816, 1, 256] - - [32, 73.414] - - - [27904, 256, 1, 256] - - [291, 59.236] - - - [21552, 256, 1, 256] - - [297, 51.77] - - - [29952, 6144, 1, 256] - - [22, 74.05] - - - [22784, 3328, 1, 256] - - [40, 72.078] - - - [20784, 256, 1, 256] - - [392, 50.528] - - - [30208, 2816, 1, 256] - - [84, 72.979] - - - [31232, 5376, 1, 256] - - [46, 74.579] - - - [30256, 256, 1, 256] - - [308, 56.198] - - - [21248, 1280, 1, 256] - - [296, 77.666] - - - [28160, 1280, 1, 256] - - [296, 79.922] - - - [30720, 3329, 1, 256] - - [35, 70.407] - - - [34560, 3329, 1, 256] - - [30, 69.597] - - - [31024, 2816, 1, 256] - - [56, 63.748] - - - [32000, 256, 1, 256] - - [266, 62.164] - - - [20528, 256, 1, 256] - - [293, 51.05] - - - [24624, 10240, 1, 256] - - [39, 63.948] - - - [21504, 7680, 1, 256] - - [23, 75.495] - - - [33536, 9728, 1, 256] - - [22, 74.661] - - - [33280, 6144, 1, 256] - - [49, 74.485] - - - [20480, 2865, 1, 256] - - [27, 70.005] - - - [30720, 1281, 1, 256] - - [25, 63.715] - - - [21760, 6144, 1, 256] - - [35, 74.339] - - - [30976, 6912, 1, 256] - - [39, 73.297] - - - [27648, 2816, 1, 256] - - [23, 73.351] - - - [20992, 3329, 1, 256] - - [27, 69.637] - - - [26672, 3072, 1, 256] - - [23, 62.488] - - - [24832, 2816, 1, 256] - - [55, 72.466] - - - [23552, 9728, 1, 256] - - [23, 75.537] - - - [26880, 1280, 1, 256] - - [289, 78.439] - - - [25088, 1280, 1, 256] - - [296, 78.35] - - - [33280, 9472, 1, 256] - - [22, 75.514] - - - [27136, 3328, 1, 256] - - [29, 73.302] - - - [28416, 2816, 1, 256] - - [57, 72.194] - - - [20480, 3328, 1, 256] - - [33, 72.255] - - - [31232, 256, 1, 256] - - [293, 61.17] - - - [33328, 9728, 1, 256] - - [56, 63.695] - - - [26416, 256, 1, 256] - - [418, 54.457] - - - [31744, 2865, 1, 256] - - [23, 70.877] - - - [22784, 6144, 1, 256] - - [22, 73.854] - - - [32000, 5888, 1, 256] - - [25, 73.811] - - - [28160, 4864, 1, 256] - - [23, 74.305] - - - [34352, 2865, 1, 256] - - [42, 61.36] - - - [29696, 256, 1, 256] - - [293, 59.692] - - - [26112, 2048, 1, 256] - - [54, 70.991] - - - [25088, 5376, 1, 256] - - [65, 74.491] - - - [29952, 3329, 1, 256] - - [35, 69.674] - - - [21296, 10240, 1, 256] - - [56, 63.537] - - - [31744, 1280, 1, 256] - - [27, 71.007] - - - [21760, 256, 1, 256] - - [418, 56.155] - - - [31488, 2048, 1, 256] - - [59, 70.757] - - - [30976, 1281, 1, 256] - - [82, 62.501] - - - [23040, 256, 1, 256] - - [291, 56.509] - - - [34304, 6144, 1, 256] - - [41, 74.528] - - - [31744, 3329, 1, 256] - - [52, 70.267] - - - [31744, 5888, 1, 256] - - [35, 74.707] - - - [29184, 1281, 1, 256] - - [349, 73.31] - - - [23856, 10240, 1, 256] - - [51, 63.188] - - - [23808, 1792, 1, 256] - - [47, 71.018] - - - [32000, 1792, 1, 256] - - [25, 71.451] - - - [26880, 2816, 1, 256] - - [33, 72.335] - - - [28416, 3328, 1, 256] - - [57, 72.264] - - - [27136, 6144, 1, 256] - - [25, 74.561] - - - [28416, 4608, 1, 256] - - [49, 72.828] - - - [33536, 1280, 1, 256] - - [32, 70.538] - - - [27440, 2865, 1, 256] - - [51, 61.881] - - - [25088, 2865, 1, 256] - - [22, 70.235] - - - [30976, 2816, 1, 256] - - [65, 70.984] - - - [26672, 10240, 1, 256] - - [39, 62.731] - - - [34048, 10240, 1, 256] - - [45, 74.096] - - - [34352, 2816, 1, 256] - - [56, 63.786] - - - [22064, 2865, 1, 256] - - [34, 61.128] - - - [28208, 4864, 1, 256] - - [42, 63.705] - - - [22528, 1280, 1, 256] - - [296, 78.138] - - - [26624, 3072, 1, 256] - - [35, 73.536] - - - [33072, 2865, 1, 256] - - [27, 62.888] - - - [22576, 256, 1, 256] - - [291, 52.823] - - - [34560, 2048, 1, 256] - - [76, 71.126] - - - [29440, 5888, 1, 256] - - [29, 73.964] - - - [34560, 1280, 1, 256] - - [23, 70.838] - - - [32000, 10240, 1, 256] - - [27, 74.514] - - - [32304, 2816, 1, 256] - - [56, 64.167] - - - [30976, 2865, 1, 256] - - [55, 68.745] - - - [30208, 6400, 1, 256] - - [45, 74.958] - - - [29232, 2865, 1, 256] - - [51, 61.737] - - - [33072, 2816, 1, 256] - - [51, 64.978] - - - [30512, 2865, 1, 256] - - [51, 61.791] - - - [20016, 2816, 1, 256] - - [56, 63.492] - - - [28416, 4352, 1, 256] - - [65, 73.602] - - - [25648, 2816, 1, 256] - - [56, 63.004] - - - [25344, 1280, 1, 256] - - [298, 78.848] - - - [24576, 10240, 1, 256] - - [36, 70.414] - - - [33024, 1281, 1, 256] - - [84, 63.336] - - - [33584, 10240, 1, 256] - - [23, 63.044] - - - [28416, 4864, 1, 256] - - [29, 73.651] - - - [23296, 3329, 1, 256] - - [30, 69.072] - - - [30464, 4352, 1, 256] - - [28, 72.6] - - - [29696, 5632, 1, 256] - - [35, 75.118] - - - [25136, 256, 1, 256] - - [315, 56.15] - - - [20528, 2865, 1, 256] - - [25, 60.85] - - - [27440, 2816, 1, 256] - - [42, 64.396] - - - [28160, 2048, 1, 256] - - [72, 70.872] - - - [24320, 2816, 1, 256] - - [57, 72.422] - - - [20736, 6144, 1, 256] - - [35, 73.931] - - - [28416, 5120, 1, 256] - - [29, 73.518] - - - [21552, 8448, 1, 256] - - [34, 64.011] - - - [20736, 1281, 1, 256] - - [285, 72.112] - - - [28464, 4864, 1, 256] - - [56, 63.713] - - - [30512, 10240, 1, 256] - - [39, 62.92] - - - [34304, 512, 1, 256] - - [289, 73.175] - - - [22784, 10240, 1, 256] - - [57, 74.611] - - - [25648, 2048, 1, 256] - - [42, 62.668] - - - [25856, 10240, 1, 256] - - [35, 74.853] - - - [32256, 8960, 1, 256] - - [22, 75.392] - - - [20736, 2865, 1, 256] - - [35, 69.727] - - - [20992, 7680, 1, 256] - - [35, 75.084] - - - [31024, 10240, 1, 256] - - [59, 62.465] - - - [26112, 256, 1, 256] - - [291, 59.203] - - - [30000, 2865, 1, 256] - - [51, 61.012] - - - [25904, 2560, 1, 256] - - [56, 63.309] - - - [24832, 768, 1, 256] - - [289, 72.674] - - - [25088, 6144, 1, 256] - - [49, 74.353] - - - [24624, 1280, 1, 256] - - [425, 76.355] - - - [22016, 8192, 1, 256] - - [35, 74.983] - - - [29952, 3328, 1, 256] - - [55, 72.644] - - - [31232, 2048, 1, 256] - - [41, 71.486] - - - [30256, 6656, 1, 256] - - [34, 63.547] - - - [20992, 2816, 1, 256] - - [35, 72.698] - - - [33792, 1536, 1, 256] - - [27, 71.307] - - - [20224, 1280, 1, 256] - - [298, 76.39] - - - [25600, 5888, 1, 256] - - [25, 74.704] - - - [26624, 768, 1, 256] - - [289, 74.407] - - - [32256, 2816, 1, 256] - - [65, 73.279] - - - [21760, 1281, 1, 256] - - [349, 71.886] - - - [25392, 10240, 1, 256] - - [56, 63.224] - - - [32768, 256, 1, 256] - - [318, 54.964] - - - [22528, 3329, 1, 256] - - [25, 70.282] - - - [23552, 3329, 1, 256] - - [25, 70.102] - - - [33024, 2865, 1, 256] - - [52, 69.197] - - - [29696, 2816, 1, 256] - - [23, 73.527] - - - [27392, 10240, 1, 256] - - [39, 74.262] - - - [23040, 2048, 1, 256] - - [67, 70.576] - - - [27648, 6144, 1, 256] - - [25, 74.648] - - - [22016, 2304, 1, 256] - - [32, 71.872] - - - [34560, 1281, 1, 256] - - [76, 64.272] - - - [27136, 1281, 1, 256] - - [390, 73.602] - - - [32000, 1281, 1, 256] - - [35, 63.609] - - - [27184, 3840, 1, 256] - - [56, 63.75] - - - [24880, 1536, 1, 256] - - [314, 77.817] - - - [28672, 768, 1, 256] - - [289, 72.165] - - - [34816, 2816, 1, 256] - - [20, 73.569] - - - [26160, 256, 1, 256] - - [348, 54.358] - - - [30464, 7168, 1, 256] - - [22, 72.343] - - - [30208, 3328, 1, 256] - - [41, 73.144] - - - [32304, 10240, 1, 256] - - [35, 63.119] - - - [26624, 1280, 1, 256] - - [298, 80.938] - - - [29696, 10240, 1, 256] - - [25, 75.448] - - - [32000, 8704, 1, 256] - - [35, 74.667] - - - [27392, 1281, 1, 256] - - [421, 73.432] - - - [26416, 2865, 1, 256] - - [34, 60.973] - - - [26160, 2560, 1, 256] - - [34, 63.321] - - - [28672, 3329, 1, 256] - - [25, 69.719] - - - [23808, 256, 1, 256] - - [418, 57.022] - - - [27184, 10240, 1, 256] - - [51, 63.078] - - - [33280, 2048, 1, 256] - - [41, 71.202] - - - [33280, 2816, 1, 256] - - [49, 73.287] - - - [23040, 9984, 1, 256] - - [25, 75.6] - - - [26112, 1280, 1, 256] - - [390, 77.945] - - - [33328, 9984, 1, 256] - - [34, 64.001] - - - [32560, 9216, 1, 256] - - [35, 63.849] - - - [22832, 9728, 1, 256] - - [56, 63.859] - - - [27904, 1280, 1, 256] - - [298, 80.61] - - - [33280, 1281, 1, 256] - - [45, 64.269] - - - [33280, 1280, 1, 256] - - [25, 70.983] - - - [32048, 256, 1, 256] - - [293, 58.571] - - - [27184, 2865, 1, 256] - - [34, 61.678] - - - [26880, 3329, 1, 256] - - [30, 69.209] - - - [20784, 7680, 1, 256] - - [51, 63.953] - - - [24832, 3329, 1, 256] - - [28, 69.453] - - - [25856, 1280, 1, 256] - - [299, 77.515] - - - [34560, 2816, 1, 256] - - [25, 72.636] - - - [20016, 256, 1, 256] - - [290, 50.956] - - - [23600, 256, 1, 256] - - [392, 53.403] - - - [22576, 9216, 1, 256] - - [39, 62.894] - - - [25344, 5632, 1, 256] - - [63, 73.279] - - - [28928, 5632, 1, 256] - - [25, 74.019] - - - [31024, 256, 1, 256] - - [370, 57.569] - - - [21552, 2865, 1, 256] - - [25, 61.492] - - - [29184, 3072, 1, 256] - - [23, 72.996] - - - [24320, 2865, 1, 256] - - [55, 70.104] - - - [20480, 6656, 1, 256] - - [25, 74.673] - - - [33536, 10240, 1, 256] - - [49, 74.79] - - - [20736, 1280, 1, 256] - - [313, 77.32] - - - [24832, 1280, 1, 256] - - [390, 78.598] - - - [29488, 10240, 1, 256] - - [35, 63.086] - - - [27392, 6144, 1, 256] - - [26, 73.273] - - - [29440, 3329, 1, 256] - - [45, 69.625] - - - [25856, 1281, 1, 256] - - [349, 73.265] - - - [34560, 768, 1, 256] - - [390, 77.023] - - - [31488, 7680, 1, 256] - - [23, 74.434] - - - [29184, 5632, 1, 256] - - [25, 74.567] - - - [32512, 512, 1, 256] - - [296, 68.877] - - - [26112, 2865, 1, 256] - - [35, 70.579] - - - [32512, 1280, 1, 256] - - [53, 70.012] - - - [20992, 1024, 1, 256] - - [299, 76.418] - - - [27904, 10240, 1, 256] - - [27, 74.585] - - - [29952, 6656, 1, 256] - - [49, 74.395] - - - [21248, 2048, 1, 256] - - [38, 70.141] - - - [34352, 256, 1, 256] - - [293, 58.159] - - - [24064, 512, 1, 256] - - [357, 68.067] - - - [32816, 2865, 1, 256] - - [26, 62.415] - - - [33840, 256, 1, 256] - - [392, 57.489] - - - [33792, 1280, 1, 256] - - [33, 71.087] - - - [21296, 7936, 1, 256] - - [42, 63.676] - - - [34096, 256, 1, 256] - - [297, 56.755] - - - [32256, 8704, 1, 256] - - [45, 75.201] - - - [30464, 1281, 1, 256] - - [284, 73.686] - - - [28464, 2816, 1, 256] - - [56, 63.429] - - - [25136, 2865, 1, 256] - - [34, 61.245] - - - [31792, 8448, 1, 256] - - [61, 64.158] - - - [24320, 4608, 1, 256] - - [55, 73.571] - - - [25088, 5120, 1, 256] - - [29, 74.212] - - - [31744, 2048, 1, 256] - - [59, 71.217] - - - [30720, 1280, 1, 256] - - [35, 70.942] - - - [34048, 256, 1, 256] - - [392, 62.67] - - - [28416, 512, 1, 256] - - [348, 71.914] - - - [22272, 10240, 1, 256] - - [45, 74.77] - - - [32512, 3328, 1, 256] - - [63, 72.772] - - - [29744, 10240, 1, 256] - - [39, 63.34] - - - [22784, 2048, 1, 256] - - [37, 70.375] - - - [23552, 2048, 1, 256] - - [39, 70.837] - - - [25344, 2816, 1, 256] - - [55, 71.619] - - - [27440, 3840, 1, 256] - - [56, 64.087] - - - [21552, 10240, 1, 256] - - [25, 63.385] - - - [21808, 256, 1, 256] - - [293, 52.863] - - - [24576, 6144, 1, 256] - - [30, 69.698] - - - [29744, 256, 1, 256] - - [297, 55.145] - - - [31488, 3328, 1, 256] - - [39, 72.631] - - - [33536, 3329, 1, 256] - - [23, 69.443] - - - [21040, 256, 1, 256] - - [293, 51.326] - - - [22272, 9216, 1, 256] - - [39, 74.35] - - - [27648, 4096, 1, 256] - - [39, 73.226] - - - [29440, 1280, 1, 256] - - [296, 80.782] - - - [31744, 7936, 1, 256] - - [25, 75.389] - - - [26624, 1281, 1, 256] - - [390, 72.386] - - - [28672, 2048, 1, 256] - - [35, 68.271] - - - [24064, 3328, 1, 256] - - [55, 72.942] - - - [25344, 3329, 1, 256] - - [45, 68.833] - - - [33280, 9728, 1, 256] - - [40, 75.157] - - - [22320, 8960, 1, 256] - - [51, 64.407] - - - [30464, 6144, 1, 256] - - [49, 72.887] - - - [34304, 2304, 1, 256] - - [29, 72.816] - - - [28928, 256, 1, 256] - - [418, 61.123] - - - [27392, 1280, 1, 256] - - [296, 75.861] - - - [26672, 2865, 1, 256] - - [27, 61.236] - - - [28720, 10240, 1, 256] - - [39, 62.448] - - - [25088, 2816, 1, 256] - - [32, 72.671] - - - [31280, 256, 1, 256] - - [293, 57.652] - - - [29488, 5888, 1, 256] - - [56, 63.407] - - - [30720, 2048, 1, 256] - - [72, 70.334] - - - [21808, 10240, 1, 256] - - [51, 63.114] - - - [24576, 2865, 1, 256] - - [52, 65.98] - - - [23808, 1280, 1, 256] - - [298, 77.883] - - - [33280, 1024, 1, 256] - - [299, 79.337] - - - [25856, 256, 1, 256] - - [392, 59.62] - - - [25648, 2304, 1, 256] - - [51, 63.621] - - - [29952, 2865, 1, 256] - - [23, 70.007] - - - [23040, 1024, 1, 256] - - [299, 75.649] - - - [34304, 3328, 1, 256] - - [40, 73.509] - - - [31792, 8192, 1, 256] - - [23, 63.352] - - - [24576, 2816, 1, 256] - - [36, 68.802] - - - [27648, 1536, 1, 256] - - [27, 70.642] - - - [23296, 9472, 1, 256] - - [25, 75.084] - - - [24624, 256, 1, 256] - - [423, 54.55] - - - [20736, 2048, 1, 256] - - [61, 70.538] - - - [28720, 5376, 1, 256] - - [39, 62.018] - - - [20480, 512, 1, 256] - - [296, 64.895] - - - [33840, 2865, 1, 256] - - [23, 62.278] - - - [24064, 2865, 1, 256] - - [45, 70.099] - - - [24064, 2816, 1, 256] - - [29, 72.665] - - - [20992, 256, 1, 256] - - [293, 55.935] - - - [33328, 256, 1, 256] - - [392, 57.371] - - - [28928, 5120, 1, 256] - - [23, 73.722] - - - [34304, 256, 1, 256] - - [290, 62.651] - - - [34304, 1281, 1, 256] - - [59, 64.906] - - - [31744, 1281, 1, 256] - - [26, 64.217] - - - [33584, 2816, 1, 256] - - [51, 63.774] - - - [24064, 4352, 1, 256] - - [57, 74.349] - - - [20224, 6912, 1, 256] - - [23, 74.537] - - - [21504, 1281, 1, 256] - - [421, 72.562] - - - [33536, 3328, 1, 256] - - [39, 72.67] - - - [34816, 3328, 1, 256] - - [35, 73.475] - - - [31024, 7680, 1, 256] - - [56, 63.222] - - - [22016, 3329, 1, 256] - - [27, 69.596] - - - [25344, 1281, 1, 256] - - [300, 73.116] - - - [31744, 7680, 1, 256] - - [25, 75.385] - - - [27952, 10240, 1, 256] - - [42, 63.351] - - - [23808, 2048, 1, 256] - - [58, 70.427] - - - [32768, 2816, 1, 256] - - [36, 59.065] - - - [34816, 256, 1, 256] - - [293, 61.823] - - - [27904, 2865, 1, 256] - - [29, 69.642] - - - [31232, 1280, 1, 256] - - [53, 70.796] - - - [22016, 1281, 1, 256] - - [419, 72.364] - - - [22528, 8704, 1, 256] - - [23, 75.669] - - - [22528, 9216, 1, 256] - - [25, 74.957] - - - [34816, 1280, 1, 256] - - [23, 71.284] - - - [23808, 10240, 1, 256] - - [25, 75.037] - - - [32512, 2048, 1, 256] - - [40, 70.782] - - - [34816, 1024, 1, 256] - - [390, 80.524] - - - [34048, 2048, 1, 256] - - [84, 70.237] - - - [30768, 2816, 1, 256] - - [23, 61.946] - - - [22272, 3329, 1, 256] - - [49, 68.958] - - - [25600, 3328, 1, 256] - - [32, 73.271] - - - [34048, 2816, 1, 256] - - [55, 72.312] - - - [22064, 8704, 1, 256] - - [34, 63.169] - - - [25648, 256, 1, 256] - - [423, 55.397] - - - [22784, 768, 1, 256] - - [421, 73.598] - - - [27904, 2048, 1, 256] - - [21, 70.333] - - - [22528, 9472, 1, 256] - - [25, 75.892] - - - [21504, 2865, 1, 256] - - [25, 70.242] - - - [28672, 5376, 1, 256] - - [25, 74.327] - - - [22576, 9472, 1, 256] - - [23, 62.863] - - - [24576, 256, 1, 256] - - [304, 55.488] - - - [28672, 5120, 1, 256] - - [25, 74.274] - - - [24576, 3328, 1, 256] - - [20, 68.462] - - - [32816, 9472, 1, 256] - - [26, 66.519] - - - [27440, 256, 1, 256] - - [291, 54.736] - - - [22272, 8704, 1, 256] - - [49, 74.842] - - - [30000, 2816, 1, 256] - - [51, 63.72] - - - [26928, 2816, 1, 256] - - [56, 63.237] - - - [22064, 2816, 1, 256] - - [56, 63.866] - - - [23552, 3328, 1, 256] - - [25, 73.147] - - - [28416, 256, 1, 256] - - [418, 60.72] - - - [28928, 6144, 1, 256] - - [23, 73.774] - - - [32768, 512, 1, 256] - - [263, 55.494] - - - [22272, 2865, 1, 256] - - [35, 69.352] - - - [26928, 256, 1, 256] - - [297, 53.415] - - - [21760, 10240, 1, 256] - - [25, 74.893] - - - [26368, 512, 1, 256] - - [293, 67.564] - - - [26672, 256, 1, 256] - - [392, 54.485] - - - [33328, 2865, 1, 256] - - [42, 62.95] - - - [30720, 3328, 1, 256] - - [32, 73.371] - - - [25856, 2865, 1, 256] - - [27, 69.722] - - - [25088, 3328, 1, 256] - - [65, 72.895] - - - [28416, 2560, 1, 256] - - [55, 72.27] - - - [33536, 9472, 1, 256] - - [22, 74.945] - - - [20480, 1280, 1, 256] - - [296, 76.223] - - - [30208, 6144, 1, 256] - - [28, 74.374] - - - [34864, 1024, 1, 256] - - [325, 75.206] - - - [33280, 256, 1, 256] - - [291, 60.848] - - - [23296, 3328, 1, 256] - - [35, 72.158] - - - [32560, 256, 1, 256] - - [370, 58.536] - - - [32560, 2816, 1, 256] - - [42, 64.998] - - - [33536, 256, 1, 256] - - [293, 62.467] - - - [34608, 768, 1, 256] - - [320, 72.606] - - - [24832, 5120, 1, 256] - - [23, 73.988] - - - [25856, 2048, 1, 256] - - [76, 70.029] - - - [30768, 256, 1, 256] - - [370, 56.154] - - - [30000, 6656, 1, 256] - - [42, 63.377] - - - [24320, 1024, 1, 256] - - [299, 76.286] - - - [33280, 9216, 1, 256] - - [40, 74.887] - - - [31488, 5376, 1, 256] - - [25, 73.859] - - - [28416, 1281, 1, 256] - - [349, 74.067] - - - [27392, 3584, 1, 256] - - [39, 71.904] - - - [26368, 2048, 1, 256] - - [72, 70.201] - - - [22528, 256, 1, 256] - - [293, 57.219] - - - [32768, 2048, 1, 256] - - [20, 56.064] - - - [30256, 6912, 1, 256] - - [34, 63.907] - - - [28672, 512, 1, 256] - - [360, 70.846] - - - [21760, 8448, 1, 256] - - [23, 74.631] - - - [34560, 6144, 1, 256] - - [27, 74.161] - - - [27696, 2816, 1, 256] - - [42, 63.638] - - - [29952, 2048, 1, 256] - - [21, 70.524] - - - [22576, 10240, 1, 256] - - [39, 62.743] - - - [25600, 1792, 1, 256] - - [32, 71.951] - - - [28976, 10240, 1, 256] - - [51, 62.845] - - - [29952, 1280, 1, 256] - - [299, 81.197] - - - [26368, 2816, 1, 256] - - [32, 72.464] - - - [26416, 3072, 1, 256] - - [56, 62.558] - - - [27648, 3329, 1, 256] - - [35, 70.276] - - - [34560, 2560, 1, 256] - - [27, 72.91] - - - [32048, 8448, 1, 256] - - [51, 63.167] - - - [30464, 2865, 1, 256] - - [71, 68.347] - - - [34048, 3328, 1, 256] - - [41, 72.364] - - - [23808, 2865, 1, 256] - - [23, 69.84] - - - [25600, 2816, 1, 256] - - [23, 73.099] - - - [20736, 6912, 1, 256] - - [23, 74.702] - - - [24576, 512, 1, 256] - - [289, 62.272] - - - [33792, 256, 1, 256] - - [290, 62.608] - - - [22576, 2865, 1, 256] - - [35, 60.853] - - - [30464, 256, 1, 256] - - [348, 61.154] - - - [24368, 2816, 1, 256] - - [56, 63.888] - - - [20224, 512, 1, 256] - - [291, 65.45] - - - [30512, 6912, 1, 256] - - [56, 64.4] - - - [20272, 2816, 1, 256] - - [34, 63.653] - - - [23296, 256, 1, 256] - - [293, 56.092] - - - [27904, 2816, 1, 256] - - [55, 72.231] - - - [29184, 1280, 1, 256] - - [298, 80.263] - - - [24112, 10240, 1, 256] - - [34, 63.496] - - - [31280, 7680, 1, 256] - - [39, 63.157] - - - [24064, 6144, 1, 256] - - [28, 74.387] - - - [26624, 6144, 1, 256] - - [27, 74.59] - - - [30768, 2865, 1, 256] - - [27, 61.038] - - - [20528, 2816, 1, 256] - - [23, 61.315] - - - [25392, 2865, 1, 256] - - [27, 61.179] - - - [22272, 6144, 1, 256] - - [28, 73.834] - - - [25088, 10240, 1, 256] - - [28, 75.251] - - - [25344, 2865, 1, 256] - - [29, 69.458] - - - [23552, 1792, 1, 256] - - [53, 71.439] - - - [23296, 3584, 1, 256] - - [32, 73.272] - - - [28160, 2816, 1, 256] - - [65, 72.65] - - - [20272, 2865, 1, 256] - - [27, 60.656] - - - [22832, 9472, 1, 256] - - [42, 64.024] - - - [21760, 7936, 1, 256] - - [27, 74.725] - - - [26928, 3328, 1, 256] - - [51, 63.435] - - - [33072, 9472, 1, 256] - - [42, 64.768] - - - [33024, 1280, 1, 256] - - [32, 69.175] - - - [34352, 512, 1, 256] - - [308, 69.698] - - - [26368, 2865, 1, 256] - - [25, 70.011] - - - [27952, 4352, 1, 256] - - [51, 63.569] - - - [21504, 8192, 1, 256] - - [25, 75.141] - - - [22320, 9216, 1, 256] - - [42, 63.672] - - - [31232, 2865, 1, 256] - - [25, 70.381] - - - [21248, 7680, 1, 256] - - [35, 74.765] - - - [24368, 256, 1, 256] - - [418, 55.086] - - - [25648, 2865, 1, 256] - - [25, 60.967] - - - [21248, 2865, 1, 256] - - [25, 69.655] - - - [28416, 2865, 1, 256] - - [23, 69.557] - - - [24320, 3329, 1, 256] - - [49, 69.541] - - - [27648, 2048, 1, 256] - - [80, 71.036] - - - [27648, 2865, 1, 256] - - [27, 70.739] - - - [26880, 2048, 1, 256] - - [42, 70.716] - - - [28672, 2560, 1, 256] - - [35, 72.702] - - - [24064, 1280, 1, 256] - - [299, 78.108] - - - [30256, 2865, 1, 256] - - [34, 61.639] - - - [22064, 10240, 1, 256] - - [25, 63.217] - - - [30464, 4608, 1, 256] - - [45, 72.193] - - - [22016, 6144, 1, 256] - - [27, 74.548] - - - [29440, 2816, 1, 256] - - [29, 72.647] - - - [25392, 2048, 1, 256] - - [42, 62.586] - - - [20992, 2048, 1, 256] - - [56, 70.296] - - - [33024, 3329, 1, 256] - - [30, 69.227] - - - [20224, 3328, 1, 256] - - [32, 71.953] - - - [28208, 4608, 1, 256] - - [34, 63.809] - - - [25344, 6144, 1, 256] - - [26, 72.892] - - - [30464, 512, 1, 256] - - [289, 70.952] - - - [21248, 3329, 1, 256] - - [25, 69.165] - - - [29696, 6144, 1, 256] - - [27, 74.778] - - - [20992, 7936, 1, 256] - - [23, 75.164] - - - [33024, 9472, 1, 256] - - [63, 74.823] - - - [32000, 3329, 1, 256] - - [52, 69.733] - - - [21248, 1281, 1, 256] - - [285, 71.773] - - - [24624, 1024, 1, 256] - - [423, 73.69] - - - [22272, 2816, 1, 256] - - [57, 71.99] - - - [29440, 1281, 1, 256] - - [349, 73.649] - - - [30464, 6400, 1, 256] - - [62, 73.497] - - - [25136, 10240, 1, 256] - - [51, 63.177] - - - [23040, 9472, 1, 256] - - [49, 75.599] - - - [33840, 2816, 1, 256] - - [61, 63.323] - - - [30976, 1024, 1, 256] - - [357, 78.131] - - - [34048, 6144, 1, 256] - - [49, 73.573] - - - [32000, 2048, 1, 256] - - [56, 69.97] - - - [32048, 2865, 1, 256] - - [56, 61.728] - - - [33328, 10240, 1, 256] - - [27, 63.577] - - - [25088, 1536, 1, 256] - - [299, 81.614] - - - [30512, 256, 1, 256] - - [290, 56.591] - - - [20480, 6912, 1, 256] - - [23, 74.992] - - - [34608, 2816, 1, 256] - - [42, 63.396] - - - [22064, 256, 1, 256] - - [418, 53.659] - - - [25600, 2865, 1, 256] - - [27, 70.481] - - - [26880, 1024, 1, 256] - - [299, 78.038] - - - [27392, 2048, 1, 256] - - [76, 69.256] - - - [30208, 10240, 1, 256] - - [45, 75.228] - - - [20016, 10240, 1, 256] - - [56, 63.271] - - - [26880, 10240, 1, 256] - - [35, 74.707] - - - [28160, 3328, 1, 256] - - [39, 72.965] - - - [33536, 2048, 1, 256] - - [78, 70.838] - - - [31232, 7936, 1, 256] - - [45, 75.186] - - - [31536, 10240, 1, 256] - - [42, 62.995] - - - [24832, 1536, 1, 256] - - [299, 81.775] - - - [32768, 768, 1, 256] - - [321, 59.507] - - - [29440, 6144, 1, 256] - - [49, 74.249] - - - [26112, 2560, 1, 256] - - [35, 73.095] - - - [33792, 6144, 1, 256] - - [35, 74.818] - - - [22528, 10240, 1, 256] - - [23, 75.441] - - - [20480, 768, 1, 256] - - [348, 69.426] - - - [22320, 256, 1, 256] - - [418, 54.008] - - - [23808, 3328, 1, 256] - - [25, 72.471] - - - [28464, 256, 1, 256] - - [297, 54.851] - - - [27136, 2048, 1, 256] - - [46, 70.97] - - - [29744, 6400, 1, 256] - - [51, 63.787] - - - [20480, 7168, 1, 256] - - [27, 73.728] - - - [22832, 256, 1, 256] - - [291, 52.482] - - - [21552, 8192, 1, 256] - - [25, 63.331] - - - [25856, 2560, 1, 256] - - [35, 72.187] - - - [28160, 6144, 1, 256] - - [39, 74.237] - - - [31280, 2816, 1, 256] - - [34, 63.623] - - - [23600, 10240, 1, 256] - - [26, 63.677] - - - [26368, 1281, 1, 256] - - [349, 73.105] - - - [24576, 1280, 1, 256] - - [309, 70.895] - - - [33536, 1536, 1, 256] - - [33, 70.649] - - - [23088, 2816, 1, 256] - - [51, 63.162] - - - [26624, 2048, 1, 256] - - [59, 70.108] - - - [29952, 2816, 1, 256] - - [23, 72.497] - - - [21760, 2048, 1, 256] - - [61, 70.422] - - - [30976, 6144, 1, 256] - - [26, 72.357] - - - [29696, 1280, 1, 256] - - [298, 80.632] - - - [30208, 4096, 1, 256] - - [39, 73.684] - - - [24832, 2865, 1, 256] - - [23, 69.873] - - - [31488, 1281, 1, 256] - - [38, 64.289] - - - [34304, 2865, 1, 256] - - [49, 70.546] - - - [32512, 256, 1, 256] - - [296, 61.35] - - - [25136, 1536, 1, 256] - - [390, 76.802] - - - [26112, 3329, 1, 256] - - [52, 70.013] - - - [24880, 1280, 1, 256] - - [425, 75.006] - - - [28208, 2816, 1, 256] - - [34, 63.499] - - - [29184, 5888, 1, 256] - - [45, 74.176] - - - [28160, 4352, 1, 256] - - [28, 74.11] - - - [34352, 10240, 1, 256] - - [26, 62.689] - - - [23856, 256, 1, 256] - - [315, 53.797] - - - [25344, 10240, 1, 256] - - [26, 73.908] - - - [20992, 1281, 1, 256] - - [300, 72.451] - - - [26624, 512, 1, 256] - - [348, 68.31] - - - [21040, 10240, 1, 256] - - [35, 62.806] - - - [23040, 3328, 1, 256] - - [76, 72.72] - - - [30976, 7168, 1, 256] - - [39, 72.418] - - - [25856, 2304, 1, 256] - - [23, 71.728] - - - [24368, 1024, 1, 256] - - [428, 75.458] - - - [33280, 2865, 1, 256] - - [49, 70.866] - - - [23296, 1536, 1, 256] - - [289, 80.233] - - - [21504, 6144, 1, 256] - - [25, 74.607] - - - [23552, 2816, 1, 256] - - [25, 73.152] - - - [30464, 2816, 1, 256] - - [29, 71.467] - - - [22832, 2865, 1, 256] - - [34, 61.219] - - - [24576, 2048, 1, 256] - - [25, 64.874] - - - [22272, 8448, 1, 256] - - [49, 74.457] - - - [32256, 1280, 1, 256] - - [65, 71.001] - - - [25856, 5888, 1, 256] - - [25, 73.834] - - - [30976, 5120, 1, 256] - - [39, 72.377] - - - [29184, 3329, 1, 256] - - [71, 69.841] - - - [24112, 2865, 1, 256] - - [56, 61.649] - - - [29744, 2816, 1, 256] - - [51, 63.337] - - - [21760, 2816, 1, 256] - - [33, 72.102] - - - [25600, 2048, 1, 256] - - [39, 71.033] - - - [32000, 1280, 1, 256] - - [32, 70.416] - - - [25856, 3328, 1, 256] - - [39, 72.256] - - - [20016, 6656, 1, 256] - - [51, 64.154] - - - [32256, 2865, 1, 256] - - [30, 70.537] - - - [22272, 3328, 1, 256] - - [65, 72.025] - - - [21504, 3328, 1, 256] - - [23, 72.919] - - - [31232, 5120, 1, 256] - - [22, 74.185] - - - [24112, 256, 1, 256] - - [392, 54.183] - - - [30208, 1280, 1, 256] - - [296, 80.703] - - - [22064, 8960, 1, 256] - - [56, 64.15] - - - [28160, 10240, 1, 256] - - [27, 74.999] - - - [21504, 1536, 1, 256] - - [391, 78.625] - - - [31744, 5632, 1, 256] - - [20, 74.894] - - - [20272, 6912, 1, 256] - - [56, 63.868] - - - [29952, 1792, 1, 256] - - [23, 71.606] - - - [25904, 10240, 1, 256] - - [26, 62.643] - - - [25344, 1792, 1, 256] - - [65, 70.475] - - - [32512, 8448, 1, 256] - - [63, 74.783] - - - [25088, 2048, 1, 256] - - [56, 70.772] - - - [23808, 9984, 1, 256] - - [27, 75.123] - - - [32768, 3329, 1, 256] - - [89, 56.152] - - - [34816, 6144, 1, 256] - - [35, 74.652] - - - [32256, 256, 1, 256] - - [322, 61.95] - - - [26368, 3328, 1, 256] - - [25, 72.308] - - - [23296, 1280, 1, 256] - - [296, 78.778] - - - [34608, 1024, 1, 256] - - [299, 77.285] - - - [30976, 1280, 1, 256] - - [57, 69.531] - - - [22528, 6144, 1, 256] - - [25, 74.61] - - - [21248, 10240, 1, 256] - - [25, 74.773] - - - [22528, 2865, 1, 256] - - [35, 70.368] - - - [22528, 768, 1, 256] - - [300, 73.231] - - - [22016, 8704, 1, 256] - - [45, 75.464] - - - [30720, 6912, 1, 256] - - [25, 75.396] - - - [33024, 2048, 1, 256] - - [40, 70.653] - - - [31232, 3329, 1, 256] - - [35, 69.855] - - - [33024, 3328, 1, 256] - - [41, 72.644] - - - [30976, 7424, 1, 256] - - [39, 73.418] - - - [27136, 3584, 1, 256] - - [23, 73.896] - - - [34048, 1280, 1, 256] - - [57, 70.127] - - - [34864, 1280, 1, 256] - - [23, 62.169] - - - [25600, 2304, 1, 256] - - [27, 72.842] - - - [21760, 3329, 1, 256] - - [27, 69.145] - - - [26928, 3584, 1, 256] - - [51, 63.394] - - - [28976, 2816, 1, 256] - - [51, 63.623] - - - [24832, 4864, 1, 256] - - [55, 74.103] - - - [21248, 1536, 1, 256] - - [390, 79.48] - - - [23808, 2816, 1, 256] - - [32, 72.377] - - - [32768, 9472, 1, 256] - - [20, 59.284] - - - [27392, 3328, 1, 256] - - [41, 71.508] - - - [26880, 3584, 1, 256] - - [23, 73.397] - - - [23552, 1281, 1, 256] - - [299, 72.529] - - - [27648, 3840, 1, 256] - - [35, 74.357] - - - [22016, 10240, 1, 256] - - [45, 75.337] - - - [34816, 2560, 1, 256] - - [23, 73.621] - - - [31536, 256, 1, 256] - - [392, 57.005] - - - [34816, 10240, 1, 256] - - [25, 75.289] - - - [27904, 1792, 1, 256] - - [55, 71.214] - - - [33792, 10240, 1, 256] - - [25, 75.354] - - - [23296, 2816, 1, 256] - - [25, 72.116] - - - [31024, 7424, 1, 256] - - [34, 63.587] - - - [22784, 1280, 1, 256] - - [390, 76.661] - - - [30976, 2048, 1, 256] - - [85, 68.691] - - - [27392, 4096, 1, 256] - - [41, 72.243] - - - [33792, 2816, 1, 256] - - [35, 73.562] - - - [32560, 10240, 1, 256] - - [35, 64.059] - - - [20736, 7424, 1, 256] - - [27, 74.802] - - - [28672, 2865, 1, 256] - - [23, 70.166] - - - [31488, 256, 1, 256] - - [418, 62.551] - - - [20992, 7424, 1, 256] - - [25, 75.243] - - - [21504, 1792, 1, 256] - - [296, 81.536] - - - [27696, 2865, 1, 256] - - [35, 61.694] - - - [33024, 1024, 1, 256] - - [299, 78.437] - - - [22016, 256, 1, 256] - - [272, 56.337] - - - [23088, 256, 1, 256] - - [425, 52.181] - - - [28976, 256, 1, 256] - - [418, 55.743] - - - [27392, 256, 1, 256] - - [290, 56.39] - - - [34304, 3329, 1, 256] - - [45, 69.933] - - - [32512, 9216, 1, 256] - - [41, 74.559] - - - [31488, 3329, 1, 256] - - [52, 69.544] - - - [20016, 2865, 1, 256] - - [34, 61.493] - - - [22016, 8448, 1, 256] - - [49, 75.105] - - - [31024, 2865, 1, 256] - - [51, 61.593] - - - [29440, 256, 1, 256] - - [291, 60.551] - - - [34608, 2865, 1, 256] - - [34, 61.135] - - - [20480, 2048, 1, 256] - - [25, 67.869] - - - [28160, 2865, 1, 256] - - [52, 69.887] - - - [28416, 2304, 1, 256] - - [53, 71.87] - - - [23552, 6144, 1, 256] - - [23, 74.767] - - - [21296, 256, 1, 256] - - [291, 50.983] - - - [28672, 4864, 1, 256] - - [23, 74.407] - - - [27648, 1792, 1, 256] - - [35, 72.137] - - - [31488, 7424, 1, 256] - - [35, 74.459] - - - [23040, 2865, 1, 256] - - [28, 69.926] - - - [30976, 3328, 1, 256] - - [29, 71.027] - - - [25856, 1792, 1, 256] - - [32, 71.058] - - - [33536, 9984, 1, 256] - - [49, 74.907] - - - [24832, 1281, 1, 256] - - [300, 72.273] - - - [29184, 3328, 1, 256] - - [59, 72.913] - - - [32000, 2816, 1, 256] - - [25, 72.454] - - - [34304, 768, 1, 256] - - [299, 78.025] - - - [24576, 1281, 1, 256] - - [302, 66.079] - - - [25088, 1281, 1, 256] - - [285, 72.769] - - - [29744, 2865, 1, 256] - - [51, 61.455] - - - [25136, 2816, 1, 256] - - [56, 63.662] - - - [29696, 1281, 1, 256] - - [285, 74.081] - - - [27392, 3329, 1, 256] - - [78, 67.992] - - - [31488, 2816, 1, 256] - - [35, 72.59] - - - [30976, 10240, 1, 256] - - [26, 73.744] - - - [26624, 3329, 1, 256] - - [23, 70.492] - - - [34304, 1280, 1, 256] - - [57, 71.143] - - - [25392, 256, 1, 256] - - [348, 56.437] - - - [26624, 10240, 1, 256] - - [27, 75.441] - - - [26112, 6144, 1, 256] - - [23, 74.593] - - - [29696, 3328, 1, 256] - - [27, 73.404] - - - [32304, 2865, 1, 256] - - [56, 62.87] - - - [24368, 2865, 1, 256] - - [42, 62.435] - - - [31488, 8192, 1, 256] - - [35, 74.293] - - - [20224, 6656, 1, 256] - - [23, 74.229] - - - [31232, 1281, 1, 256] - - [31, 64.494] - - - [21296, 2865, 1, 256] - - [51, 61.548] - - - [24112, 768, 1, 256] - - [325, 70.354] - - - [32000, 8448, 1, 256] - - [25, 74.291] - - - [23552, 1536, 1, 256] - - [299, 79.772] - - - [30976, 7680, 1, 256] - - [73, 72.959] - - - [31280, 10240, 1, 256] - - [39, 63.142] - - - [23344, 9984, 1, 256] - - [42, 63.465] - - - [21248, 8192, 1, 256] - - [27, 74.558] - - - [29696, 6400, 1, 256] - - [25, 75.374] - - - [32304, 8960, 1, 256] - - [56, 64.13] - - - [27184, 256, 1, 256] - - [418, 53.634] - - - [28464, 10240, 1, 256] - - [39, 62.755] - - - [20736, 256, 1, 256] - - [293, 54.71] - - - [31232, 10240, 1, 256] - - [49, 75.057] - - - [25856, 6144, 1, 256] - - [35, 74.009] - - - [27440, 10240, 1, 256] - - [35, 62.863] - - - [23088, 2865, 1, 256] - - [34, 61.571] - - - [29696, 3584, 1, 256] - - [23, 74.288] - - - [23040, 9728, 1, 256] - - [63, 75.267] - - - [31744, 10240, 1, 256] - - [25, 75.225] - - - [31744, 1792, 1, 256] - - [33, 72.252] - - - [24320, 256, 1, 256] - - [418, 58.223] - - - [27696, 256, 1, 256] - - [418, 54.948] - - - [29696, 2865, 1, 256] - - [25, 70.8] - - - [22784, 3072, 1, 256] - - [25, 72.397] - - - [29952, 5888, 1, 256] - - [57, 73.889] - - - [28928, 2816, 1, 256] - - [25, 72.431] - - - [30768, 7424, 1, 256] - - [35, 62.446] - - - [27440, 4096, 1, 256] - - [51, 63.269] - - - [24064, 4096, 1, 256] - - [84, 73.037] - - - [32256, 3329, 1, 256] - - [28, 70.183] - - - [30976, 3329, 1, 256] - - [57, 68.061] - - - [25600, 10240, 1, 256] - - [23, 75.488] - - - [20224, 6144, 1, 256] - - [23, 73.983] - - - [21040, 7936, 1, 256] - - [38, 64.061] - - - [26368, 2560, 1, 256] - - [47, 72.493] - - - [32512, 1281, 1, 256] - - [84, 63.536] - - - [28928, 3072, 1, 256] - - [25, 72.55] - - - [34864, 2865, 1, 256] - - [25, 61.273] - - - [23552, 9984, 1, 256] - - [25, 75.93] - - - [21040, 2865, 1, 256] - - [51, 61.177] - - - [34048, 1281, 1, 256] - - [82, 63.624] - - - [23296, 10240, 1, 256] - - [27, 74.732] - - - [32768, 6144, 1, 256] - - [20, 58.987] - - - [25904, 2816, 1, 256] - - [42, 63.446] - - - [31232, 1024, 1, 256] - - [390, 79.258] - - - [27648, 3328, 1, 256] - - [23, 73.377] - - - [34864, 256, 1, 256] - - [290, 58.395] - - - [21248, 256, 1, 256] - - [293, 56.769] - - - [26416, 10240, 1, 256] - - [51, 62.855] - - - [27184, 3584, 1, 256] - - [56, 63.863] - - - [23296, 2048, 1, 256] - - [51, 69.765] - - - [34048, 512, 1, 256] - - [390, 72.497] - - - [21760, 2865, 1, 256] - - [23, 69.561] - - - [28672, 2816, 1, 256] - - [23, 72.765] - - - [28672, 4608, 1, 256] - - [23, 73.631] - - - [34560, 512, 1, 256] - - [320, 72.014] - - - [32768, 2865, 1, 256] - - [89, 56.916] - - - [30208, 6912, 1, 256] - - [45, 75.246] - - - [32512, 6144, 1, 256] - - [28, 74.065] - - - [24832, 3328, 1, 256] - - [55, 72.518] - - - [27392, 2816, 1, 256] - - [74, 70.787] - - - [32768, 8704, 1, 256] - - [36, 59.402] - - - [23552, 10240, 1, 256] - - [25, 75.547] - - - [32816, 9216, 1, 256] - - [26, 66.618] - - - [33024, 10240, 1, 256] - - [41, 74.697] - - - [34608, 256, 1, 256] - - [392, 57.926] - - - [20736, 3328, 1, 256] - - [27, 72.065] - - - [31232, 7680, 1, 256] - - [45, 75.166] - - - [22528, 512, 1, 256] - - [289, 69.609] - - - [30208, 2865, 1, 256] - - [35, 70.349] - - - [22272, 2304, 1, 256] - - [32, 71.492] - - - [32512, 2816, 1, 256] - - [22, 72.51] - - - [31488, 7936, 1, 256] - - [35, 74.414] - - - [28416, 2048, 1, 256] - - [46, 69.489] - - - [22784, 3329, 1, 256] - - [27, 69.258] - - - [23040, 2816, 1, 256] - - [33, 72.336] - - - [24320, 3328, 1, 256] - - [57, 72.697] - - - [24064, 1281, 1, 256] - - [349, 73.057] - - - [33072, 9728, 1, 256] - - [51, 64.079] - - - [29440, 10240, 1, 256] - - [28, 74.925] - - - [30208, 6656, 1, 256] - - [28, 74.77] - - - [32768, 3328, 1, 256] - - [20, 58.272] - - - [28416, 6144, 1, 256] - - [25, 73.683] - - - [27904, 4608, 1, 256] - - [26, 73.253] - - - [27184, 2816, 1, 256] - - [56, 63.994] - - - [29184, 1024, 1, 256] - - [299, 78.212] - - - [31744, 1536, 1, 256] - - [53, 71.1] - - - [28416, 10240, 1, 256] - - [45, 74.509] - - - [24368, 10240, 1, 256] - - [27, 63.847] - - - [27904, 3329, 1, 256] - - [28, 69.138] - - - [25344, 3328, 1, 256] - - [57, 71.605] - - - [29952, 6400, 1, 256] - - [22, 74.528] - - - [29440, 2048, 1, 256] - - [72, 70.786] - - - [28928, 1281, 1, 256] - - [390, 73.577] - - - [30208, 3329, 1, 256] - - [35, 69.876] - - - [23088, 9984, 1, 256] - - [56, 64.288] - - - [29184, 2816, 1, 256] - - [22, 72.862] - - - [22528, 2560, 1, 256] - - [25, 72.954] - - - [33328, 2816, 1, 256] - - [56, 64.013] - - - [26368, 256, 1, 256] - - [293, 57.838] - - - [22832, 10240, 1, 256] - - [56, 63.212] - - - [31792, 2816, 1, 256] - - [38, 63.597] - - - [24832, 2048, 1, 256] - - [70, 70.487] - - - [24880, 256, 1, 256] - - [418, 55.831] - - - [33840, 10240, 1, 256] - - [39, 63.711] - - - [33584, 9984, 1, 256] - - [42, 63.613] - - - [28672, 10240, 1, 256] - - [25, 74.999] - - - [24832, 256, 1, 256] - - [293, 58.455] - - - [31488, 2865, 1, 256] - - [35, 70.05] - - - [30720, 7424, 1, 256] - - [27, 75.416] - - - [33536, 2816, 1, 256] - - [55, 72.7] - - - [30000, 6400, 1, 256] - - [56, 63.97] - - - [20224, 1281, 1, 256] - - [300, 72.065] - - - [22832, 2816, 1, 256] - - [51, 63.64] - - - [25600, 6144, 1, 256] - - [25, 74.742] - - - [24320, 4352, 1, 256] - - [55, 73.984] - - - [32768, 10240, 1, 256] - - [64, 58.859] - - - [26880, 768, 1, 256] - - [299, 74.198] - - - [24576, 3329, 1, 256] - - [30, 65.155] - - - [27904, 3840, 1, 256] - - [55, 73.388] - - - [30256, 2816, 1, 256] - - [51, 63.427] - - - [23296, 1281, 1, 256] - - [419, 72.298] - - - [26880, 256, 1, 256] - - [291, 57.934] - - - [23344, 2816, 1, 256] - - [34, 63.552] - - - [33792, 2048, 1, 256] - - [80, 71.385] - - - [21504, 3329, 1, 256] - - [27, 70.177] - - - [20272, 256, 1, 256] - - [290, 50.631] - - - [32768, 1280, 1, 256] - - [36, 56.265] - - - [32256, 10240, 1, 256] - - [49, 75.088] - - - [27952, 2816, 1, 256] - - [56, 63.766] - - - [28928, 5376, 1, 256] - - [45, 73.7] - - - [20992, 6144, 1, 256] - - [27, 74.373] - - - [20224, 2048, 1, 256] - - [21, 69.842] - - - [33280, 10240, 1, 256] - - [25, 75.116] - - - [24064, 3329, 1, 256] - - [49, 69.727] - - - [32768, 9216, 1, 256] - - [64, 58.928] - - - [20016, 6912, 1, 256] - - [51, 64.096] - - - [22320, 10240, 1, 256] - - [42, 63.577] - - - [22784, 256, 1, 256] - - [418, 56.605] - - - [34816, 512, 1, 256] - - [423, 72.913] - - - [32048, 8704, 1, 256] - - [51, 63.329] - - - [29232, 5888, 1, 256] - - [34, 63.527] - - - [24064, 768, 1, 256] - - [300, 74.325] - - - [33792, 9984, 1, 256] - - [27, 75.596] - - - [32512, 3329, 1, 256] - - [71, 69.597] - - - [21504, 2048, 1, 256] - - [39, 70.357] - - - [28160, 2304, 1, 256] - - [25, 72.19] - - - [20784, 10240, 1, 256] - - [35, 63.029] - - - [20224, 7168, 1, 256] - - [27, 73.51] - - - [28976, 2865, 1, 256] - - [42, 61.422] - - - [21296, 2816, 1, 256] - - [42, 63.29] - - - [23552, 256, 1, 256] - - [293, 57.581] - - - [26160, 2865, 1, 256] - - [34, 61.461] - - - [23600, 2816, 1, 256] - - [34, 63.383] - - - [20480, 7424, 1, 256] - - [23, 75.186] - - - [28928, 3329, 1, 256] - - [52, 69.267] - - - [20784, 2816, 1, 256] - - [51, 63.356] - - - [25344, 256, 1, 256] - - [418, 58.528] - - - [20224, 10240, 1, 256] - - [35, 74.867] - - - [28672, 1280, 1, 256] - - [306, 78.47] - - - [29232, 256, 1, 256] - - [370, 55.656] - - - [28720, 2865, 1, 256] - - [23, 59.965] - - - [22016, 2816, 1, 256] - - [25, 72.506] - - - [25600, 1536, 1, 256] - - [32, 70.369] - - - [26112, 10240, 1, 256] - - [25, 75.339] - - - [27136, 10240, 1, 256] - - [49, 75.178] - - - [31744, 8192, 1, 256] - - [25, 74.947] - - - [24320, 10240, 1, 256] - - [49, 74.908] - - - [29952, 10240, 1, 256] - - [28, 74.674] - - - [23296, 9984, 1, 256] - - [23, 75.04] - - - [34560, 2304, 1, 256] - - [33, 72.547] - - - [32000, 2865, 1, 256] - - [52, 70.214] - - - [25088, 1024, 1, 256] - - [299, 76.962] - - - [20272, 10240, 1, 256] - - [42, 63.816] - - - [25344, 5376, 1, 256] - - [40, 73.025] - - - [21760, 3328, 1, 256] - - [23, 72.312] - - - [32768, 8960, 1, 256] - - [36, 59.008] - - - [29952, 3840, 1, 256] - - [29, 73.676] - - - [32512, 2865, 1, 256] - - [71, 69.874] - - - [23344, 2865, 1, 256] - - [34, 60.816] - - - [24576, 768, 1, 256] - - [302, 66.763] - - - [27648, 3584, 1, 256] - - [23, 74.08] - - - [27952, 4608, 1, 256] - - [51, 63.343] - - - [29440, 3584, 1, 256] - - [25, 73.461] - - - [34096, 512, 1, 256] - - [312, 68.74] - - - [32304, 256, 1, 256] - - [418, 58.959] - - - [21040, 2816, 1, 256] - - [34, 64.179] - - - [22784, 1024, 1, 256] - - [296, 76.909] - - - [22784, 2816, 1, 256] - - [33, 71.985] - - - [25856, 2816, 1, 256] - - [47, 72.336] - - - [23296, 6144, 1, 256] - - [25, 74.154] - - - [28160, 4608, 1, 256] - - [84, 73.703] - - - [25136, 1792, 1, 256] - - [27, 62.163] - - - [30208, 256, 1, 256] - - [392, 59.67] - - - [23808, 1281, 1, 256] - - [419, 72.422] - - - [26368, 2304, 1, 256] - - [33, 71.856] - - - [27648, 4352, 1, 256] - - [25, 74.703] - - - [31280, 7936, 1, 256] - - [42, 63.951] - - - [22320, 2865, 1, 256] - - [23, 61.292] - - - [22320, 2816, 1, 256] - - [56, 63.845] - - - [28720, 5120, 1, 256] - - [26, 61.738] - - - [22272, 1280, 1, 256] - - [298, 77.805] - - - [31232, 3328, 1, 256] - - [39, 73.148] - - - [29696, 2048, 1, 256] - - [39, 71.313] - - - [34048, 9984, 1, 256] - - [49, 74.355] - - - [28416, 1280, 1, 256] - - [289, 80.301] - - - [21504, 2816, 1, 256] - - [47, 72.798] - - - [33536, 2865, 1, 256] - - [35, 69.879] - - - [23552, 3840, 1, 256] - - [35, 74.249] - - - [31744, 256, 1, 256] - - [290, 62.148] - - - [25600, 1281, 1, 256] - - [299, 72.845] - - - [30768, 7168, 1, 256] - - [26, 62.311] - - - [23808, 3329, 1, 256] - - [25, 69.505] - - - [32256, 3328, 1, 256] - - [22, 73.232] - - - [23040, 9216, 1, 256] - - [39, 74.942] - - - [33024, 256, 1, 256] - - [392, 56.756] - - - [33584, 2865, 1, 256] - - [56, 62.267] - - - [21504, 8448, 1, 256] - - [25, 75.477] - - - [27904, 1281, 1, 256] - - [349, 73.498] - - - [34304, 10240, 1, 256] - - [41, 75.096] - - - [20992, 2865, 1, 256] - - [35, 70.053] - - - [22528, 8960, 1, 256] - - [27, 75.783] - - - [28928, 3328, 1, 256] - - [23, 72.199] - - - [21808, 2865, 1, 256] - - [51, 61.674] - - - [26416, 2816, 1, 256] - - [51, 63.431] - - - [27392, 3840, 1, 256] - - [39, 72.145] - - - [26112, 1281, 1, 256] - - [349, 72.431] - - - [34864, 10240, 1, 256] - - [26, 62.895] - - - [29440, 1536, 1, 256] - - [33, 70.113] - - - [30256, 10240, 1, 256] - - [42, 62.69] - - - [22528, 2816, 1, 256] - - [35, 73.081] - - - [28928, 2048, 1, 256] - - [37, 69.845] - - - [28976, 5376, 1, 256] - - [34, 63.383] - - - [20736, 7168, 1, 256] - - [23, 73.681] - - - [22016, 2865, 1, 256] - - [25, 70.061] - - - [26368, 1280, 1, 256] - - [296, 78.657] - - - [24624, 2865, 1, 256] - - [53, 59.672] - - - [23040, 3329, 1, 256] - - [71, 69.455] - - - [23296, 2865, 1, 256] - - [25, 69.699] - - - [28416, 3329, 1, 256] - - [52, 69.052] - - - [23040, 1281, 1, 256] - - [419, 71.733] - - - [21808, 8448, 1, 256] - - [34, 64.072] - - - [30720, 2865, 1, 256] - - [25, 70.964] - - - [22272, 8960, 1, 256] - - [29, 74.925] - - - [34864, 2816, 1, 256] - - [27, 62.188] - - - [31232, 7168, 1, 256] - - [49, 73.728] - - - [27696, 4352, 1, 256] - - [23, 63.284] - - - [21504, 256, 1, 256] - - [418, 56.28] - - - [28672, 1281, 1, 256] - - [348, 71.961] - - - [29696, 1792, 1, 256] - - [25, 72.277] - - - [28464, 5120, 1, 256] - - [42, 63.253] - - - [27136, 3329, 1, 256] - - [45, 70.101] - - - [21248, 3328, 1, 256] - - [53, 72.083] - - - [26880, 1281, 1, 256] - - [299, 73.516] - - - [32256, 8448, 1, 256] - - [45, 75.036] - - - [20480, 6144, 1, 256] - - [25, 74.369] - - - [34048, 2865, 1, 256] - - [71, 69.561] - - - [29696, 5888, 1, 256] - - [35, 74.749] - - - [28720, 256, 1, 256] - - [293, 55.263] - - - [33792, 2865, 1, 256] - - [35, 71.066] - - - [22784, 8960, 1, 256] - - [25, 74.796] - - - [30720, 256, 1, 256] - - [418, 61.227] - - - [23808, 512, 1, 256] - - [293, 67.41] - - - [33024, 9728, 1, 256] - - [41, 74.736] - - - [42624, 13824, 1, 384] - - [26, 88.278] - - - [33024, 3840, 1, 384] - - [34, 89.14] - - - [33408, 15360, 1, 384] - - [27, 90.765] - - - [44160, 8832, 1, 384] - - [51, 90.868] - - - [31488, 2688, 1, 384] - - [34, 89.239] - - - [39168, 3072, 1, 384] - - [23, 89.312] - - - [31872, 5760, 1, 384] - - [35, 90.116] - - - [36096, 13440, 1, 384] - - [26, 89.824] - - - [41856, 1152, 1, 384] - - [51, 87.469] - - - [32256, 1153, 1, 384] - - [51, 78.124] - - - [44160, 1153, 1, 384] - - [56, 78.2] - - - [31488, 7296, 1, 384] - - [27, 90.342] - - - [43008, 9216, 1, 384] - - [26, 88.337] - - - [31872, 6144, 1, 384] - - [35, 89.989] - - - [32640, 7297, 1, 384] - - [52, 85.715] - - - [33792, 1152, 1, 384] - - [25, 85.923] - - - [43776, 13441, 1, 384] - - [52, 87.285] - - - [36480, 1153, 1, 384] - - [34, 78.52] - - - [37632, 1152, 1, 384] - - [34, 85.178] - - - [37248, 8448, 1, 384] - - [25, 90.158] - - - [31872, 7297, 1, 384] - - [27, 88.03] - - - [41856, 7296, 1, 384] - - [56, 90.634] - - - [39936, 7297, 1, 384] - - [25, 87.892] - - - [35712, 1153, 1, 384] - - [34, 77.05] - - - [35712, 3072, 1, 384] - - [35, 89.547] - - - [31488, 1153, 1, 384] - - [23, 76.43] - - - [36480, 1152, 1, 384] - - [51, 85.411] - - - [36864, 9216, 1, 384] - - [52, 86.953] - - - [42624, 15360, 1, 384] - - [26, 87.319] - - - [37632, 8832, 1, 384] - - [56, 90.755] - - - [32640, 1153, 1, 384] - - [76, 73.418] - - - [36864, 3072, 1, 384] - - [27, 87.687] - - - [32640, 6912, 1, 384] - - [35, 88.737] - - - [31872, 13440, 1, 384] - - [25, 90.709] - - - [39168, 3840, 1, 384] - - [27, 89.866] - - - [39168, 10368, 1, 384] - - [51, 90.644] - - - [33792, 3072, 1, 384] - - [35, 88.025] - - - [39552, 1536, 1, 384] - - [25, 87.432] - - - [38784, 7296, 1, 384] - - [34, 90.329] - - - [40320, 1153, 1, 384] - - [34, 78.245] - - - [42240, 1152, 1, 384] - - [42, 87.93] - - - [43776, 14976, 1, 384] - - [39, 89.905] - - - [38784, 9216, 1, 384] - - [56, 90.782] - - - [33024, 4224, 1, 384] - - [51, 89.263] - - - [43776, 7297, 1, 384] - - [52, 86.204] - - - [34560, 9216, 1, 384] - - [54, 90.365] - - - [43392, 8064, 1, 384] - - [25, 90.729] - - - [34944, 7296, 1, 384] - - [25, 90.377] - - - [38400, 7296, 1, 384] - - [35, 90.343] - - - [41856, 6912, 1, 384] - - [42, 90.762] - - - [40704, 3072, 1, 384] - - [27, 89.188] - - - [41472, 12672, 1, 384] - - [23, 91.04] - - - [36864, 1920, 1, 384] - - [35, 87.498] - - - [43008, 1920, 1, 384] - - [25, 88.654] - - - [43008, 13824, 1, 384] - - [27, 90.147] - - - [31104, 13441, 1, 384] - - [27, 89.23] - - - [41472, 12288, 1, 384] - - [59, 89.607] - - - [31488, 7297, 1, 384] - - [23, 87.888] - - - [35712, 6912, 1, 384] - - [23, 90.541] - - - [40704, 5376, 1, 384] - - [56, 90.373] - - - [36480, 9216, 1, 384] - - [42, 90.768] - - - [38784, 13440, 1, 384] - - [35, 90.948] - - - [36096, 15360, 1, 384] - - [28, 89.555] - - - [41856, 15360, 1, 384] - - [34, 91.147] - - - [37632, 2688, 1, 384] - - [42, 88.713] - - - [33792, 4608, 1, 384] - - [25, 88.828] - - - [38400, 13440, 1, 384] - - [35, 90.905] - - - [31104, 3072, 1, 384] - - [23, 88.396] - - - [33792, 13440, 1, 384] - - [23, 90.724] - - - [34176, 5376, 1, 384] - - [34, 89.995] - - - [31872, 3072, 1, 384] - - [23, 88.853] - - - [33792, 1920, 1, 384] - - [27, 88.614] - - - [34560, 1153, 1, 384] - - [34, 77.466] - - - [43392, 15360, 1, 384] - - [25, 90.778] - - - [39168, 4224, 1, 384] - - [23, 90.179] - - - [43776, 1153, 1, 384] - - [49, 76.234] - - - [41472, 6528, 1, 384] - - [56, 90.214] - - - [42240, 1153, 1, 384] - - [23, 78.784] - - - [36480, 13441, 1, 384] - - [25, 88.921] - - - [31488, 5760, 1, 384] - - [27, 90.215] - - - [34560, 13440, 1, 384] - - [23, 90.784] - - - [32256, 3072, 1, 384] - - [23, 88.739] - - - [37632, 15360, 1, 384] - - [23, 90.795] - - - [43776, 8448, 1, 384] - - [39, 89.663] - - - [37248, 13440, 1, 384] - - [35, 90.849] - - - [34944, 13440, 1, 384] - - [27, 90.894] - - - [41088, 3072, 1, 384] - - [30, 86.816] - - - [43008, 14208, 1, 384] - - [23, 90.513] - - - [33792, 7296, 1, 384] - - [23, 90.356] - - - [43392, 8448, 1, 384] - - [51, 90.947] - - - [31104, 7297, 1, 384] - - [35, 88.048] - - - [31104, 2304, 1, 384] - - [51, 88.538] - - - [35712, 1152, 1, 384] - - [34, 86.709] - - - [39552, 13440, 1, 384] - - [23, 90.967] - - - [37632, 2304, 1, 384] - - [51, 88.3] - - - [31872, 1153, 1, 384] - - [42, 77.391] - - - [39552, 3072, 1, 384] - - [23, 89.481] - - - [36864, 15360, 1, 384] - - [30, 88.677] - - - [33408, 4608, 1, 384] - - [56, 89.822] - - - [43392, 7297, 1, 384] - - [25, 88.191] - - - [32256, 7296, 1, 384] - - [25, 90.393] - - - [41472, 7296, 1, 384] - - [23, 90.726] - - - [38016, 9216, 1, 384] - - [34, 90.741] - - - [38784, 1153, 1, 384] - - [23, 77.728] - - - [34944, 2688, 1, 384] - - [42, 88.484] - - - [36864, 1152, 1, 384] - - [23, 85.952] - - - [39168, 7297, 1, 384] - - [35, 88.429] - - - [33024, 768, 1, 384] - - [393, 81.112] - - - [34560, 13441, 1, 384] - - [25, 89.15] - - - [33792, 7680, 1, 384] - - [27, 90.317] - - - [36864, 1153, 1, 384] - - [23, 76.423] - - - [40320, 4992, 1, 384] - - [51, 90.42] - - - [31488, 13440, 1, 384] - - [23, 90.896] - - - [39552, 10752, 1, 384] - - [34, 91.09] - - - [36096, 1152, 1, 384] - - [52, 83.95] - - - [44160, 1152, 1, 384] - - [35, 86.378] - - - [37632, 9216, 1, 384] - - [42, 90.767] - - - [37248, 15360, 1, 384] - - [35, 90.911] - - - [34944, 5760, 1, 384] - - [25, 90.259] - - - [41088, 15360, 1, 384] - - [59, 89.764] - - - [41088, 11904, 1, 384] - - [72, 89.957] - - - [35328, 6528, 1, 384] - - [27, 90.045] - - - [32640, 15360, 1, 384] - - [52, 89.211] - - - [33024, 7297, 1, 384] - - [30, 86.942] - - - [31104, 1153, 1, 384] - - [23, 75.891] - - - [40704, 1153, 1, 384] - - [34, 78.6] - - - [42240, 13440, 1, 384] - - [56, 91.201] - - - [41472, 7297, 1, 384] - - [35, 88.199] - - - [33408, 3072, 1, 384] - - [23, 89.593] - - - [40704, 13440, 1, 384] - - [23, 91.016] - - - [39168, 7296, 1, 384] - - [23, 90.527] - - - [34176, 9216, 1, 384] - - [54, 90.031] - - - [35328, 15360, 1, 384] - - [25, 90.572] - - - [38400, 1152, 1, 384] - - [51, 86.421] - - - [37248, 3072, 1, 384] - - [25, 89.373] - - - [31488, 2304, 1, 384] - - [51, 87.961] - - - [40704, 1152, 1, 384] - - [27, 85.429] - - - [39168, 768, 1, 384] - - [34, 85.104] - - - [34944, 1153, 1, 384] - - [51, 78.011] - - - [39936, 13440, 1, 384] - - [23, 90.764] - - - [43008, 7297, 1, 384] - - [23, 87.793] - - - [33024, 15360, 1, 384] - - [25, 90.317] - - - [34176, 1920, 1, 384] - - [23, 87.757] - - - [40320, 15360, 1, 384] - - [27, 90.659] - - - [37632, 3072, 1, 384] - - [23, 88.886] - - - [40320, 11136, 1, 384] - - [42, 90.855] - - - [34944, 1152, 1, 384] - - [56, 85.364] - - - [44160, 14976, 1, 384] - - [42, 90.918] - - - [33792, 1536, 1, 384] - - [27, 86.629] - - - [38016, 13441, 1, 384] - - [27, 88.959] - - - [37632, 7296, 1, 384] - - [35, 90.485] - - - [41856, 6528, 1, 384] - - [34, 90.77] - - - [36096, 6912, 1, 384] - - [49, 89.39] - - - [39936, 15360, 1, 384] - - [30, 89.281] - - - [43776, 9216, 1, 384] - - [72, 89.701] - - - [38400, 9600, 1, 384] - - [27, 90.744] - - - [39552, 15360, 1, 384] - - [35, 90.804] - - - [37248, 2304, 1, 384] - - [27, 88.9] - - - [33792, 1153, 1, 384] - - [25, 78.249] - - - [42624, 1152, 1, 384] - - [27, 85.746] - - - [35328, 3072, 1, 384] - - [23, 88.672] - - - [37632, 13440, 1, 384] - - [25, 90.935] - - - [38400, 3072, 1, 384] - - [23, 88.81] - - - [32640, 1152, 1, 384] - - [35, 82.416] - - - [31872, 1152, 1, 384] - - [42, 84.902] - - - [40320, 3072, 1, 384] - - [23, 89.67] - - - [38016, 15360, 1, 384] - - [27, 91.029] - - - [35712, 9216, 1, 384] - - [56, 90.628] - - - [33024, 13441, 1, 384] - - [30, 88.236] - - - [36096, 3072, 1, 384] - - [28, 87.825] - - - [36864, 13440, 1, 384] - - [23, 90.3] - - - [33408, 13441, 1, 384] - - [25, 89.408] - - - [37248, 9216, 1, 384] - - [27, 90.012] - - - [31488, 1152, 1, 384] - - [27, 84.3] - - - [31488, 3072, 1, 384] - - [25, 88.897] - - - [35328, 1152, 1, 384] - - [35, 86.19] - - - [37248, 7297, 1, 384] - - [22, 87.954] - - - [34944, 6144, 1, 384] - - [25, 89.899] - - - [36480, 1536, 1, 384] - - [27, 85.898] - - - [39168, 15360, 1, 384] - - [35, 90.611] - - - [43392, 13441, 1, 384] - - [23, 89.041] - - - [42624, 1536, 1, 384] - - [30, 86.032] - - - [36480, 7296, 1, 384] - - [56, 90.429] - - - [33792, 9216, 1, 384] - - [59, 88.885] - - - [36096, 768, 1, 384] - - [60, 81.715] - - - [33408, 1536, 1, 384] - - [25, 86.603] - - - [31872, 13441, 1, 384] - - [23, 88.734] - - - [43008, 13440, 1, 384] - - [35, 90.679] - - - [33024, 1152, 1, 384] - - [51, 83.603] - - - [34560, 5376, 1, 384] - - [34, 89.95] - - - [32640, 3840, 1, 384] - - [27, 87.392] - - - [33408, 1153, 1, 384] - - [25, 77.608] - - - [32256, 1152, 1, 384] - - [25, 85.845] - - - [41856, 13440, 1, 384] - - [42, 91.068] - - - [43776, 2688, 1, 384] - - [72, 87.926] - - - [34560, 8832, 1, 384] - - [27, 90.698] - - - [32256, 6528, 1, 384] - - [23, 89.983] - - - [33408, 13440, 1, 384] - - [23, 90.855] - - - [36096, 7296, 1, 384] - - [28, 89.202] - - - [43776, 3072, 1, 384] - - [28, 86.725] - - - [38784, 7297, 1, 384] - - [23, 87.841] - - - [39936, 7296, 1, 384] - - [25, 90.369] - - - [37632, 8448, 1, 384] - - [51, 90.624] - - - [43392, 9216, 1, 384] - - [56, 90.379] - - - [41856, 13056, 1, 384] - - [51, 91.0] - - - [30720, 13441, 1, 384] - - [27, 88.74] - - - [36864, 7680, 1, 384] - - [23, 89.447] - - - [41472, 1152, 1, 384] - - [42, 86.836] - - - [39168, 13440, 1, 384] - - [35, 90.989] - - - [43776, 2304, 1, 384] - - [72, 87.775] - - - [34176, 15360, 1, 384] - - [23, 90.734] - - - [36096, 7297, 1, 384] - - [28, 86.378] - - - [33792, 4992, 1, 384] - - [25, 90.089] - - - [35712, 15360, 1, 384] - - [27, 90.899] - - - [39168, 9984, 1, 384] - - [23, 90.689] - - - [36096, 9216, 1, 384] - - [72, 89.158] - - - [43008, 1536, 1, 384] - - [27, 86.668] - - - [33408, 9216, 1, 384] - - [60, 90.217] - - - [40704, 7296, 1, 384] - - [56, 90.655] - - - [38016, 2688, 1, 384] - - [34, 89.371] - - - [39168, 13441, 1, 384] - - [27, 89.187] - - - [39168, 9216, 1, 384] - - [35, 89.958] - - - [38400, 15360, 1, 384] - - [35, 90.558] - - - [43392, 2304, 1, 384] - - [42, 89.144] - - - [38400, 13441, 1, 384] - - [23, 88.937] - - - [43008, 1152, 1, 384] - - [23, 86.658] - - - [39936, 4608, 1, 384] - - [27, 88.688] - - - [43392, 14592, 1, 384] - - [25, 90.978] - - - [34176, 13441, 1, 384] - - [28, 88.89] - - - [38784, 9984, 1, 384] - - [27, 90.687] - - - [44160, 13441, 1, 384] - - [35, 88.866] - - - [31488, 5376, 1, 384] - - [42, 90.263] - - - [39936, 13441, 1, 384] - - [35, 88.989] - - - [34176, 1152, 1, 384] - - [35, 86.48] - - - [32640, 3072, 1, 384] - - [23, 86.224] - - - [34560, 15360, 1, 384] - - [27, 90.733] - - - [34944, 15360, 1, 384] - - [35, 90.744] - - - [37632, 13441, 1, 384] - - [25, 89.069] - - - [40320, 5376, 1, 384] - - [34, 90.308] - - - [41856, 12672, 1, 384] - - [42, 91.102] - - - [34176, 4992, 1, 384] - - [27, 89.892] - - - [42624, 7297, 1, 384] - - [49, 84.477] - - - [41856, 1153, 1, 384] - - [23, 78.384] - - - [41472, 9216, 1, 384] - - [72, 89.586] - - - [40704, 2304, 1, 384] - - [51, 88.393] - - - [36864, 8064, 1, 384] - - [23, 90.015] - - - [40704, 5760, 1, 384] - - [35, 90.322] - - - [41088, 7297, 1, 384] - - [52, 86.365] - - - [38784, 1152, 1, 384] - - [23, 87.044] - - - [38784, 3072, 1, 384] - - [27, 89.002] - - - [34560, 2304, 1, 384] - - [35, 88.169] - - - [36096, 1153, 1, 384] - - [60, 76.971] - - - [35712, 13440, 1, 384] - - [27, 90.934] - - - [39936, 1152, 1, 384] - - [25, 86.321] - - - [43392, 14208, 1, 384] - - [23, 90.644] - - - [39552, 1153, 1, 384] - - [34, 77.084] - - - [35712, 6528, 1, 384] - - [42, 90.358] - - - [31104, 5376, 1, 384] - - [56, 90.199] - - - [31104, 9216, 1, 384] - - [42, 90.311] - - - [33024, 9216, 1, 384] - - [59, 89.636] - - - [39936, 11136, 1, 384] - - [23, 90.513] - - - [43008, 3072, 1, 384] - - [23, 87.535] - - - [41856, 768, 1, 384] - - [51, 85.598] - - - [43776, 1152, 1, 384] - - [72, 85.56] - - - [34176, 7297, 1, 384] - - [22, 87.794] - - - [38016, 7297, 1, 384] - - [51, 88.208] - - - [36480, 7680, 1, 384] - - [56, 90.618] - - - [38400, 7297, 1, 384] - - [35, 88.136] - - - [44160, 2688, 1, 384] - - [42, 89.787] - - - [33792, 15360, 1, 384] - - [52, 89.277] - - - [40704, 2688, 1, 384] - - [34, 89.517] - - - [38784, 3840, 1, 384] - - [51, 90.014] - - - [44160, 7296, 1, 384] - - [56, 90.668] - - - [41088, 2688, 1, 384] - - [34, 88.083] - - - [38016, 3072, 1, 384] - - [27, 89.602] - - - [42240, 7296, 1, 384] - - [34, 90.815] - - - [41856, 9216, 1, 384] - - [56, 90.762] - - - [32640, 13440, 1, 384] - - [25, 89.719] - - - [40320, 13441, 1, 384] - - [25, 88.933] - - - [36480, 13440, 1, 384] - - [25, 90.863] - - - [41856, 7297, 1, 384] - - [56, 88.115] - - - [41088, 7296, 1, 384] - - [36, 89.193] - - - [33408, 1152, 1, 384] - - [34, 85.116] - - - [43392, 1920, 1, 384] - - [25, 88.388] - - - [31104, 1920, 1, 384] - - [35, 86.98] - - - [31488, 15360, 1, 384] - - [319, 91.493] - - - [31872, 7296, 1, 384] - - [23, 90.056] - - - [43008, 7680, 1, 384] - - [27, 90.127] - - - [35328, 13440, 1, 384] - - [23, 90.941] - - - [43776, 15360, 1, 384] - - [59, 89.613] - - - [34944, 3072, 1, 384] - - [23, 89.225] - - - [37248, 1153, 1, 384] - - [34, 77.474] - - - [31104, 1152, 1, 384] - - [56, 86.317] - - - [34560, 7297, 1, 384] - - [35, 88.375] - - - [43776, 14592, 1, 384] - - [39, 89.832] - - - [33408, 7296, 1, 384] - - [23, 90.328] - - - [33024, 7296, 1, 384] - - [27, 89.596] - - - [33024, 13440, 1, 384] - - [25, 90.417] - - - [31104, 7296, 1, 384] - - [23, 90.339] - - - [42240, 9216, 1, 384] - - [51, 91.084] - - - [34944, 13441, 1, 384] - - [35, 89.08] - - - [33792, 7297, 1, 384] - - [35, 88.085] - - - [35328, 13441, 1, 384] - - [23, 89.221] - - - [34176, 7296, 1, 384] - - [35, 90.146] - - - [40320, 1920, 1, 384] - - [51, 89.136] - - - [31872, 15360, 1, 384] - - [23, 90.709] - - - [39168, 1153, 1, 384] - - [51, 78.445] - - - [31104, 4992, 1, 384] - - [35, 89.792] - - - [41088, 1152, 1, 384] - - [72, 84.162] - - - [39552, 10368, 1, 384] - - [42, 90.976] - - - [40704, 11520, 1, 384] - - [23, 90.92] - - - [36864, 7297, 1, 384] - - [25, 87.1] - - - [42240, 15360, 1, 384] - - [34, 91.13] - - - [34560, 1152, 1, 384] - - [25, 84.888] - - - [31104, 13440, 1, 384] - - [27, 90.806] - - - [31488, 9216, 1, 384] - - [27, 89.809] - - - [34176, 3072, 1, 384] - - [27, 88.904] - - - [41088, 1153, 1, 384] - - [56, 76.435] - - - [43392, 1153, 1, 384] - - [51, 78.69] - - - [42240, 6912, 1, 384] - - [34, 90.914] - - - [43008, 15360, 1, 384] - - [52, 89.299] - - - [42240, 7297, 1, 384] - - [25, 88.344] - - - [43776, 7296, 1, 384] - - [39, 89.302] - - - [35712, 7296, 1, 384] - - [35, 90.409] - - - [38400, 9216, 1, 384] - - [54, 89.62] - - - [39936, 9216, 1, 384] - - [59, 89.072] - - - [32256, 6144, 1, 384] - - [25, 89.891] - - - [42624, 7680, 1, 384] - - [23, 88.093] - - - [33408, 4224, 1, 384] - - [34, 89.662] - - - [38784, 768, 1, 384] - - [42, 84.311] - - - [38016, 7296, 1, 384] - - [27, 90.545] - - - [34560, 5760, 1, 384] - - [27, 90.265] - - - [34944, 7297, 1, 384] - - [25, 87.974] - - - [38016, 8832, 1, 384] - - [35, 90.763] - - - [39936, 1920, 1, 384] - - [25, 88.364] - - - [40320, 11520, 1, 384] - - [25, 91.004] - - - [32256, 7297, 1, 384] - - [25, 88.391] - - - [33792, 13441, 1, 384] - - [23, 88.944] - - - [41472, 3072, 1, 384] - - [35, 89.299] - - - [33024, 1153, 1, 384] - - [56, 76.254] - - - [36864, 7296, 1, 384] - - [35, 89.855] - - - [38016, 1153, 1, 384] - - [25, 78.612] - - - [40320, 7297, 1, 384] - - [27, 88.106] - - - [42624, 13441, 1, 384] - - [28, 84.193] - - - [43008, 13441, 1, 384] - - [23, 88.6] - - - [39552, 9216, 1, 384] - - [42, 90.808] - - - [35328, 9216, 1, 384] - - [72, 89.367] - - - [42624, 3072, 1, 384] - - [23, 88.928] - - - [40320, 13440, 1, 384] - - [25, 91.065] - - - [42240, 13441, 1, 384] - - [27, 89.158] - - - [39936, 10752, 1, 384] - - [35, 90.046] - - - [41472, 6144, 1, 384] - - [27, 89.786] - - - [36864, 1536, 1, 384] - - [23, 85.867] - - - [33408, 7297, 1, 384] - - [25, 88.0] - - - [31872, 2688, 1, 384] - - [51, 88.748] - - - [41472, 1153, 1, 384] - - [25, 78.0] - - - [38400, 1153, 1, 384] - - [34, 77.123] - - - [38400, 3456, 1, 384] - - [34, 89.537] - - - [41856, 13441, 1, 384] - - [27, 88.889] - - - [43392, 1152, 1, 384] - - [34, 87.397] - - - [39552, 4608, 1, 384] - - [42, 89.957] - - - [40704, 15360, 1, 384] - - [27, 90.792] - - - [42240, 3072, 1, 384] - - [27, 89.467] - - - [32640, 3456, 1, 384] - - [25, 86.766] - - - [35712, 768, 1, 384] - - [42, 82.475] - - - [31104, 15360, 1, 384] - - [331, 91.195] - - - [40704, 13441, 1, 384] - - [25, 89.252] - - - [32640, 7296, 1, 384] - - [27, 88.652] - - - [34176, 8448, 1, 384] - - [85, 90.045] - - - [32640, 13441, 1, 384] - - [52, 87.325] - - - [36864, 13441, 1, 384] - - [23, 87.818] - - - [34176, 13440, 1, 384] - - [35, 90.629] - - - [37248, 1152, 1, 384] - - [51, 87.015] - - - [44160, 7297, 1, 384] - - [42, 88.025] - - - [41088, 6144, 1, 384] - - [59, 88.988] - - - [39936, 1536, 1, 384] - - [23, 86.0] - - - [44160, 15360, 1, 384] - - [42, 90.988] - - - [35712, 7297, 1, 384] - - [23, 88.087] - - - [35328, 6144, 1, 384] - - [27, 89.727] - - - [42624, 7296, 1, 384] - - [48, 88.397] - - - [33408, 7680, 1, 384] - - [27, 90.515] - - - [41472, 13441, 1, 384] - - [35, 89.163] - - - [43776, 8832, 1, 384] - - [72, 89.563] - - - [32256, 15360, 1, 384] - - [23, 90.612] - - - [32256, 9216, 1, 384] - - [25, 89.77] - - - [31872, 9216, 1, 384] - - [24, 90.157] - - - [37248, 7296, 1, 384] - - [35, 90.367] - - - [40320, 1152, 1, 384] - - [42, 87.104] - - - [34560, 8448, 1, 384] - - [60, 90.284] - - - [38784, 3456, 1, 384] - - [51, 89.687] - - - [41472, 15360, 1, 384] - - [23, 90.467] - - - [41856, 3072, 1, 384] - - [27, 89.414] - - - [41088, 13441, 1, 384] - - [52, 87.66] - - - [39936, 1153, 1, 384] - - [25, 77.349] - - - [37248, 1920, 1, 384] - - [25, 88.403] - - - [39552, 7296, 1, 384] - - [34, 90.452] - - - [40320, 2304, 1, 384] - - [42, 89.112] - - - [34560, 2688, 1, 384] - - [51, 89.033] - - - [42240, 13056, 1, 384] - - [35, 90.791] - - - [40320, 9216, 1, 384] - - [34, 90.595] - - - [40704, 7297, 1, 384] - - [51, 88.281] - - - [43776, 13440, 1, 384] - - [26, 90.065] - - - [39936, 4992, 1, 384] - - [25, 90.119] - - - [42624, 13440, 1, 384] - - [20, 87.56] - - - [37632, 1153, 1, 384] - - [42, 78.129] - - - [33024, 3072, 1, 384] - - [27, 88.041] - - - [40704, 9216, 1, 384] - - [51, 90.812] - - - [42624, 1153, 1, 384] - - [23, 77.163] - - - [43392, 13440, 1, 384] - - [23, 90.96] - - - [36480, 3072, 1, 384] - - [35, 88.881] - - - [41088, 12288, 1, 384] - - [72, 89.769] - - - [39168, 1152, 1, 384] - - [23, 85.477] - - - [39936, 3072, 1, 384] - - [52, 87.229] - - - [35712, 13441, 1, 384] - - [23, 89.034] - - - [41088, 13440, 1, 384] - - [35, 90.307] - - - [43392, 3072, 1, 384] - - [25, 89.924] - - - [33792, 8064, 1, 384] - - [35, 90.4] - - - [32256, 13440, 1, 384] - - [25, 90.922] - - - [35328, 7297, 1, 384] - - [23, 88.144] - - - [40704, 11904, 1, 384] - - [23, 90.964] - - - [33024, 6912, 1, 384] - - [35, 89.944] - - - [38784, 15360, 1, 384] - - [23, 90.615] - - - [42240, 768, 1, 384] - - [51, 86.132] - - - [44160, 13440, 1, 384] - - [34, 91.059] - - - [39552, 7297, 1, 384] - - [23, 88.337] - - - [32640, 768, 1, 384] - - [299, 83.864] - - - [44160, 9216, 1, 384] - - [56, 90.934] - - - [32640, 6528, 1, 384] - - [27, 88.489] - - - [39552, 13441, 1, 384] - - [23, 88.971] - - - [31488, 13441, 1, 384] - - [27, 89.238] - - - [43008, 7296, 1, 384] - - [27, 90.267] - - - [41088, 5760, 1, 384] - - [52, 88.944] - - - [41472, 13440, 1, 384] - - [23, 90.964] - - - [43392, 7296, 1, 384] - - [23, 90.595] - - - [34944, 9216, 1, 384] - - [34, 90.258] - - - [43008, 1153, 1, 384] - - [23, 78.043] - - - [32640, 9216, 1, 384] - - [26, 88.653] - - - [36096, 13441, 1, 384] - - [30, 87.233] - - - [39552, 1152, 1, 384] - - [34, 86.13] - - - [37632, 7297, 1, 384] - - [25, 88.148] - - - [42624, 9216, 1, 384] - - [45, 86.819] - - - [43008, 8064, 1, 384] - - [25, 90.492] - - - [38784, 9600, 1, 384] - - [42, 90.879] - - - [37248, 8064, 1, 384] - - [35, 90.624] - - - [30720, 15360, 1, 384] - - [319, 91.732] - - - [38016, 13440, 1, 384] - - [27, 90.967] - - - [34944, 8832, 1, 384] - - [23, 90.609] - - - [37248, 13441, 1, 384] - - [27, 88.869] - - - [34560, 7296, 1, 384] - - [35, 90.444] - - - [44160, 3072, 1, 384] - - [27, 89.316] - - - [40320, 7296, 1, 384] - - [35, 90.462] - - - [34176, 2304, 1, 384] - - [27, 88.599] - - - [41088, 9216, 1, 384] - - [72, 89.557] - - - [34176, 1153, 1, 384] - - [23, 76.738] - - - [39552, 4224, 1, 384] - - [42, 90.185] - - - [38784, 13441, 1, 384] - - [23, 88.737] - - - [36480, 7297, 1, 384] - - [23, 87.916] - - - [32256, 3456, 1, 384] - - [25, 89.567] - - - [34176, 8064, 1, 384] - - [23, 90.295] - - - [36480, 15360, 1, 384] - - [56, 91.186] - - - [34560, 3072, 1, 384] - - [25, 89.269] - - - [35328, 7296, 1, 384] - - [27, 90.268] - - - [32256, 13441, 1, 384] - - [23, 89.398] - - - [38016, 1152, 1, 384] - - [56, 85.896] - - - [35328, 1153, 1, 384] - - [23, 76.639] - - - [23040, 7296, 1, 384] - - [23, 89.767] - - - [12672, 7296, 1, 384] - - [23, 88.855] - - - [4224, 4225, 1, 384] - - [360, 81.052] - - - [19968, 13440, 1, 384] - - [27, 90.89] - - - [16128, 3072, 1, 384] - - [23, 86.309] - - - [19968, 9216, 1, 384] - - [51, 89.746] - - - [24576, 13440, 1, 384] - - [36, 87.615] - - - [17280, 3072, 1, 384] - - [25, 87.102] - - - [16512, 9216, 1, 384] - - [59, 87.679] - - - [21120, 1536, 1, 384] - - [35, 85.882] - - - [18432, 13441, 1, 384] - - [27, 88.684] - - - [21120, 9216, 1, 384] - - [54, 90.106] - - - [27264, 3072, 1, 384] - - [23, 88.364] - - - [12288, 4608, 1, 384] - - [25, 86.933] - - - [22272, 5376, 1, 384] - - [42, 89.631] - - - [7296, 6912, 1, 384] - - [25, 88.378] - - - [26880, 9216, 1, 384] - - [42, 90.633] - - - [3072, 2688, 1, 384] - - [304, 72.889] - - - [16512, 2688, 1, 384] - - [78, 82.82] - - - [8064, 7680, 1, 384] - - [25, 87.628] - - - [22656, 1153, 1, 384] - - [42, 73.989] - - - [24960, 8064, 1, 384] - - [27, 90.562] - - - [23808, 9216, 1, 384] - - [51, 90.726] - - - [29568, 15360, 1, 384] - - [49, 90.385] - - - [1920, 1152, 1, 384] - - [387, 56.041] - - - [11136, 10752, 1, 384] - - [25, 89.659] - - - [25728, 1152, 1, 384] - - [25, 84.086] - - - [19584, 3072, 1, 384] - - [35, 87.287] - - - [3840, 1153, 1, 384] - - [305, 61.557] - - - [15360, 7296, 1, 384] - - [319, 91.611] - - - [13056, 12673, 1, 384] - - [27, 89.234] - - - [5376, 5377, 1, 384] - - [27, 82.103] - - - [28416, 13440, 1, 384] - - [34, 91.161] - - - [11904, 4224, 1, 384] - - [42, 88.137] - - - [24576, 10752, 1, 384] - - [52, 86.561] - - - [20352, 7297, 1, 384] - - [23, 88.385] - - - [16512, 7296, 1, 384] - - [39, 87.442] - - - [17280, 13441, 1, 384] - - [27, 89.218] - - - [24192, 10368, 1, 384] - - [42, 90.498] - - - [20352, 6528, 1, 384] - - [42, 90.017] - - - [1920, 1536, 1, 384] - - [392, 55.824] - - - [15744, 8064, 1, 384] - - [324, 92.375] - - - [13056, 3072, 1, 384] - - [25, 85.097] - - - [20352, 7296, 1, 384] - - [27, 89.816] - - - [10368, 1152, 1, 384] - - [389, 77.29] - - - [16128, 1152, 1, 384] - - [302, 81.981] - - - [13440, 7297, 1, 384] - - [25, 87.271] - - - [19200, 13441, 1, 384] - - [25, 89.246] - - - [13440, 13441, 1, 384] - - [56, 88.895] - - - [7680, 7297, 1, 384] - - [23, 86.162] - - - [27648, 14208, 1, 384] - - [23, 90.692] - - - [23424, 9216, 1, 384] - - [25, 89.747] - - - [24960, 1153, 1, 384] - - [34, 76.623] - - - [28032, 2304, 1, 384] - - [34, 88.371] - - - [30720, 3072, 1, 384] - - [27, 88.071] - - - [11904, 1152, 1, 384] - - [303, 79.842] - - - [24576, 3072, 1, 384] - - [27, 84.927] - - - [26112, 1153, 1, 384] - - [56, 75.83] - - - [10368, 10369, 1, 384] - - [23, 87.889] - - - [14976, 1536, 1, 384] - - [23, 82.728] - - - [11520, 7296, 1, 384] - - [34, 89.153] - - - [5376, 5376, 1, 384] - - [23, 82.183] - - - [28800, 7296, 1, 384] - - [42, 90.21] - - - [22656, 3072, 1, 384] - - [25, 87.879] - - - [11904, 7296, 1, 384] - - [25, 88.549] - - - [13824, 3072, 1, 384] - - [25, 85.697] - - - [21504, 13440, 1, 384] - - [35, 90.62] - - - [28800, 13440, 1, 384] - - [42, 90.974] - - - [13824, 7296, 1, 384] - - [23, 89.863] - - - [28416, 13441, 1, 384] - - [25, 89.206] - - - [20736, 7296, 1, 384] - - [35, 90.329] - - - [4992, 4608, 1, 384] - - [23, 84.18] - - - [21888, 1153, 1, 384] - - [419, 76.047] - - - [6912, 3072, 1, 384] - - [324, 84.104] - - - [7680, 7680, 1, 384] - - [23, 88.391] - - - [11904, 11905, 1, 384] - - [23, 88.706] - - - [9600, 1920, 1, 384] - - [25, 79.204] - - - [25728, 2688, 1, 384] - - [34, 87.846] - - - [29568, 3840, 1, 384] - - [22, 89.362] - - - [9984, 7297, 1, 384] - - [23, 86.231] - - - [13056, 2688, 1, 384] - - [51, 84.981] - - - [3456, 1920, 1, 384] - - [300, 75.644] - - - [19200, 1152, 1, 384] - - [35, 80.905] - - - [15744, 2304, 1, 384] - - [23, 84.376] - - - [17664, 7296, 1, 384] - - [51, 89.781] - - - [3072, 3072, 1, 384] - - [309, 74.933] - - - [21888, 7296, 1, 384] - - [62, 87.508] - - - [16128, 13440, 1, 384] - - [25, 90.31] - - - [23040, 1153, 1, 384] - - [56, 74.924] - - - [21504, 9216, 1, 384] - - [59, 88.698] - - - [21120, 4608, 1, 384] - - [54, 89.236] - - - [10368, 1153, 1, 384] - - [350, 72.909] - - - [29184, 13441, 1, 384] - - [45, 89.21] - - - [8832, 1536, 1, 384] - - [393, 79.335] - - - [30336, 3072, 1, 384] - - [25, 89.232] - - - [24192, 1153, 1, 384] - - [34, 74.563] - - - [16128, 2304, 1, 384] - - [23, 85.472] - - - [20736, 13440, 1, 384] - - [25, 90.841] - - - [24960, 7297, 1, 384] - - [23, 88.598] - - - [18048, 1536, 1, 384] - - [23, 82.772] - - - [19200, 5760, 1, 384] - - [23, 89.282] - - - [13440, 13056, 1, 384] - - [23, 90.019] - - - [6144, 1152, 1, 384] - - [303, 70.094] - - - [1920, 1920, 1, 384] - - [301, 61.041] - - - [18816, 5376, 1, 384] - - [34, 88.84] - - - [28800, 2688, 1, 384] - - [51, 88.798] - - - [20352, 3840, 1, 384] - - [25, 88.528] - - - [3840, 3841, 1, 384] - - [27, 76.356] - - - [17280, 768, 1, 384] - - [309, 80.559] - - - [21888, 2304, 1, 384] - - [52, 85.589] - - - [28416, 14592, 1, 384] - - [56, 90.909] - - - [18816, 3072, 1, 384] - - [23, 86.239] - - - [25344, 13440, 1, 384] - - [27, 91.119] - - - [20736, 6912, 1, 384] - - [51, 89.894] - - - [26880, 1152, 1, 384] - - [35, 83.551] - - - [29952, 3072, 1, 384] - - [35, 88.052] - - - [24960, 8448, 1, 384] - - [51, 90.766] - - - [15360, 8064, 1, 384] - - [27, 89.763] - - - [27648, 1920, 1, 384] - - [23, 87.035] - - - [3456, 2304, 1, 384] - - [350, 73.711] - - - [23040, 6528, 1, 384] - - [34, 89.69] - - - [14208, 1153, 1, 384] - - [320, 75.209] - - - [27648, 1153, 1, 384] - - [35, 76.656] - - - [1920, 1921, 1, 384] - - [392, 57.007] - - - [19584, 13441, 1, 384] - - [25, 89.274] - - - [8448, 3072, 1, 384] - - [27, 82.728] - - - [16512, 13441, 1, 384] - - [30, 86.662] - - - [4992, 768, 1, 384] - - [293, 59.832] - - - [28416, 14976, 1, 384] - - [56, 90.971] - - - [8448, 1152, 1, 384] - - [267, 75.399] - - - [20352, 9216, 1, 384] - - [56, 90.003] - - - [19584, 1153, 1, 384] - - [315, 77.134] - - - [20736, 768, 1, 384] - - [51, 81.601] - - - [28416, 2688, 1, 384] - - [51, 88.293] - - - [27264, 13440, 1, 384] - - [319, 91.225] - - - [16128, 7296, 1, 384] - - [27, 89.025] - - - [27648, 13440, 1, 384] - - [319, 91.636] - - - [26880, 13056, 1, 384] - - [318, 90.855] - - - [6528, 1920, 1, 384] - - [310, 78.574] - - - [20352, 13441, 1, 384] - - [23, 89.041] - - - [12288, 7297, 1, 384] - - [23, 87.098] - - - [21120, 7680, 1, 384] - - [23, 90.502] - - - [13824, 13441, 1, 384] - - [25, 88.762] - - - [26112, 13440, 1, 384] - - [35, 90.947] - - - [16512, 7297, 1, 384] - - [30, 84.44] - - - [6144, 5761, 1, 384] - - [23, 82.18] - - - [24960, 1152, 1, 384] - - [23, 85.839] - - - [9600, 9216, 1, 384] - - [34, 88.337] - - - [22272, 1153, 1, 384] - - [389, 75.206] - - - [24960, 2304, 1, 384] - - [35, 88.378] - - - [11136, 7296, 1, 384] - - [25, 88.208] - - - [28800, 3072, 1, 384] - - [25, 87.63] - - - [6912, 2688, 1, 384] - - [303, 82.762] - - - [25728, 3072, 1, 384] - - [27, 88.902] - - - [15744, 13441, 1, 384] - - [25, 89.503] - - - [18816, 7296, 1, 384] - - [34, 89.387] - - - [18816, 7297, 1, 384] - - [23, 87.989] - - - [13440, 13440, 1, 384] - - [23, 90.071] - - - [29184, 3456, 1, 384] - - [23, 89.49] - - - [8064, 768, 1, 384] - - [301, 69.346] - - - [4992, 4609, 1, 384] - - [23, 78.626] - - - [26496, 13056, 1, 384] - - [319, 91.047] - - - [21504, 4608, 1, 384] - - [27, 88.464] - - - [18048, 9216, 1, 384] - - [331, 91.513] - - - [14592, 13441, 1, 384] - - [35, 89.353] - - - [22656, 1152, 1, 384] - - [42, 83.362] - - - [14976, 3072, 1, 384] - - [27, 86.849] - - - [24960, 13441, 1, 384] - - [34, 89.752] - - - [768, 768, 1, 384] - - [90, 32.393] - - - [12672, 4992, 1, 384] - - [23, 87.364] - - - [11136, 3072, 1, 384] - - [25, 86.907] - - - [19584, 1152, 1, 384] - - [25, 82.418] - - - [16896, 3456, 1, 384] - - [27, 87.275] - - - [23040, 1152, 1, 384] - - [25, 84.389] - - - [6528, 6528, 1, 384] - - [25, 86.8] - - - [25344, 3072, 1, 384] - - [35, 87.695] - - - [2688, 1536, 1, 384] - - [392, 66.899] - - - [5760, 1536, 1, 384] - - [293, 72.299] - - - [6144, 5760, 1, 384] - - [25, 85.977] - - - [21504, 8064, 1, 384] - - [35, 89.897] - - - [12288, 12288, 1, 384] - - [27, 88.86] - - - [16128, 13441, 1, 384] - - [25, 89.135] - - - [25344, 8448, 1, 384] - - [27, 90.218] - - - [23808, 7297, 1, 384] - - [56, 88.195] - - - [15744, 7296, 1, 384] - - [23, 89.763] - - - [16896, 13441, 1, 384] - - [23, 89.152] - - - [15360, 1920, 1, 384] - - [23, 83.634] - - - [21504, 1152, 1, 384] - - [422, 85.161] - - - [6912, 1152, 1, 384] - - [304, 72.576] - - - [16512, 3072, 1, 384] - - [30, 82.448] - - - [28800, 1153, 1, 384] - - [51, 76.181] - - - [21888, 8064, 1, 384] - - [48, 87.855] - - - [20736, 7297, 1, 384] - - [25, 88.128] - - - [10752, 10753, 1, 384] - - [27, 88.712] - - - [8832, 7297, 1, 384] - - [34, 86.031] - - - [28032, 7297, 1, 384] - - [27, 88.513] - - - [23424, 9600, 1, 384] - - [23, 90.657] - - - [23040, 13440, 1, 384] - - [35, 90.83] - - - [26880, 13441, 1, 384] - - [365, 89.574] - - - [4224, 4224, 1, 384] - - [425, 80.914] - - - [9600, 9600, 1, 384] - - [23, 88.649] - - - [26112, 1152, 1, 384] - - [56, 85.034] - - - [29568, 3456, 1, 384] - - [51, 88.517] - - - [28032, 9216, 1, 384] - - [51, 90.625] - - - [27648, 9216, 1, 384] - - [59, 88.785] - - - [17664, 1153, 1, 384] - - [311, 75.097] - - - [12672, 12289, 1, 384] - - [51, 87.114] - - - [21888, 1152, 1, 384] - - [27, 82.958] - - - [21888, 9216, 1, 384] - - [26, 87.296] - - - [10752, 10369, 1, 384] - - [35, 88.454] - - - [22656, 7296, 1, 384] - - [27, 90.219] - - - [13440, 13057, 1, 384] - - [42, 88.973] - - - [10752, 1153, 1, 384] - - [389, 72.243] - - - [12672, 3072, 1, 384] - - [23, 86.363] - - - [23424, 13440, 1, 384] - - [35, 90.874] - - - [29952, 3840, 1, 384] - - [51, 89.78] - - - [18432, 1920, 1, 384] - - [27, 85.746] - - - [26112, 7297, 1, 384] - - [27, 88.499] - - - [18816, 1153, 1, 384] - - [303, 75.563] - - - [17664, 4224, 1, 384] - - [25, 88.407] - - - [11520, 11521, 1, 384] - - [27, 88.63] - - - [30720, 1920, 1, 384] - - [23, 87.964] - - - [15360, 13441, 1, 384] - - [23, 89.017] - - - [17664, 13441, 1, 384] - - [51, 89.74] - - - [26496, 3072, 1, 384] - - [25, 87.645] - - - [20736, 4224, 1, 384] - - [56, 89.19] - - - [18816, 13441, 1, 384] - - [42, 89.24] - - - [18048, 13441, 1, 384] - - [34, 89.388] - - - [20352, 3072, 1, 384] - - [25, 88.042] - - - [1152, 768, 1, 384] - - [256, 43.744] - - - [16896, 7296, 1, 384] - - [23, 89.528] - - - [28800, 9216, 1, 384] - - [56, 90.373] - - - [9600, 1152, 1, 384] - - [311, 76.962] - - - [29952, 1153, 1, 384] - - [35, 76.04] - - - [20736, 1153, 1, 384] - - [423, 77.32] - - - [19584, 5760, 1, 384] - - [35, 89.61] - - - [29568, 7296, 1, 384] - - [51, 89.833] - - - [7296, 3072, 1, 384] - - [25, 82.095] - - - [27264, 1152, 1, 384] - - [23, 83.817] - - - [12288, 4992, 1, 384] - - [35, 86.87] - - - [5760, 5376, 1, 384] - - [35, 83.708] - - - [30720, 1152, 1, 384] - - [27, 85.402] - - - [14208, 13441, 1, 384] - - [42, 89.455] - - - [21504, 7296, 1, 384] - - [25, 89.468] - - - [7296, 6913, 1, 384] - - [25, 85.463] - - - [23808, 6912, 1, 384] - - [51, 90.007] - - - [20352, 768, 1, 384] - - [51, 79.924] - - - [2688, 2688, 1, 384] - - [311, 71.101] - - - [13056, 12672, 1, 384] - - [27, 90.482] - - - [29568, 13440, 1, 384] - - [34, 90.69] - - - [11904, 1153, 1, 384] - - [311, 73.665] - - - [2688, 2689, 1, 384] - - [287, 69.819] - - - [9984, 9985, 1, 384] - - [25, 87.492] - - - [22272, 13440, 1, 384] - - [25, 90.599] - - - [30336, 15360, 1, 384] - - [319, 91.719] - - - [21504, 7680, 1, 384] - - [27, 90.021] - - - [24192, 13441, 1, 384] - - [35, 89.274] - - - [15360, 1536, 1, 384] - - [25, 80.28] - - - [24576, 7297, 1, 384] - - [35, 83.448] - - - [11136, 3456, 1, 384] - - [25, 85.406] - - - [9600, 1153, 1, 384] - - [418, 70.269] - - - [18048, 7297, 1, 384] - - [27, 87.754] - - - [6144, 1153, 1, 384] - - [428, 65.47] - - - [23040, 9600, 1, 384] - - [23, 90.44] - - - [26880, 1153, 1, 384] - - [56, 75.08] - - - [10752, 7297, 1, 384] - - [23, 86.643] - - - [6912, 6529, 1, 384] - - [23, 85.275] - - - [29184, 9216, 1, 384] - - [54, 89.646] - - - [20736, 9216, 1, 384] - - [56, 90.051] - - - [23808, 1152, 1, 384] - - [42, 82.731] - - - [11136, 1153, 1, 384] - - [297, 71.038] - - - [25344, 1152, 1, 384] - - [23, 83.112] - - - [25344, 13441, 1, 384] - - [22, 89.36] - - - [14976, 7296, 1, 384] - - [319, 91.797] - - - [14592, 13440, 1, 384] - - [23, 90.678] - - - [7680, 7681, 1, 384] - - [23, 85.928] - - - [29568, 768, 1, 384] - - [70, 81.546] - - - [5760, 1152, 1, 384] - - [286, 69.229] - - - [21888, 13441, 1, 384] - - [26, 84.654] - - - [17664, 768, 1, 384] - - [393, 78.99] - - - [25728, 11904, 1, 384] - - [42, 90.922] - - - [9984, 2688, 1, 384] - - [34, 85.359] - - - [28416, 1153, 1, 384] - - [23, 75.623] - - - [17664, 3072, 1, 384] - - [23, 86.289] - - - [23040, 7297, 1, 384] - - [25, 88.179] - - - [8448, 8448, 1, 384] - - [51, 88.552] - - - [4608, 4225, 1, 384] - - [360, 82.482] - - - [4224, 2688, 1, 384] - - [350, 77.346] - - - [3072, 1152, 1, 384] - - [392, 68.888] - - - [29184, 1152, 1, 384] - - [23, 85.452] - - - [13440, 3072, 1, 384] - - [35, 87.428] - - - [6912, 6913, 1, 384] - - [35, 86.541] - - - [18432, 13440, 1, 384] - - [35, 90.518] - - - [14208, 7296, 1, 384] - - [319, 91.627] - - - [5376, 768, 1, 384] - - [418, 61.935] - - - [29184, 7296, 1, 384] - - [23, 90.039] - - - [20352, 1152, 1, 384] - - [25, 80.551] - - - [2304, 1153, 1, 384] - - [418, 54.02] - - - [23808, 9984, 1, 384] - - [23, 90.845] - - - [8448, 8065, 1, 384] - - [25, 86.48] - - - [24576, 1152, 1, 384] - - [23, 82.609] - - - [1536, 1537, 1, 384] - - [300, 51.851] - - - [4224, 3072, 1, 384] - - [25, 74.892] - - - [19968, 7296, 1, 384] - - [23, 90.127] - - - [19200, 5376, 1, 384] - - [34, 88.93] - - - [4608, 1152, 1, 384] - - [297, 63.834] - - - [18432, 4992, 1, 384] - - [23, 88.376] - - - [26880, 7297, 1, 384] - - [56, 88.499] - - - [15744, 3072, 1, 384] - - [25, 87.77] - - - [22272, 7296, 1, 384] - - [23, 89.771] - - - [20352, 6912, 1, 384] - - [25, 89.537] - - - [26880, 13440, 1, 384] - - [319, 90.962] - - - [4224, 3840, 1, 384] - - [303, 81.647] - - - [23424, 13441, 1, 384] - - [45, 88.888] - - - [16512, 13440, 1, 384] - - [30, 89.118] - - - [21120, 1152, 1, 384] - - [25, 82.711] - - - [10368, 3072, 1, 384] - - [25, 85.551] - - - [28032, 13440, 1, 384] - - [56, 90.785] - - - [14208, 6528, 1, 384] - - [34, 89.104] - - - [768, 769, 1, 384] - - [351, 22.402] - - - [3456, 1152, 1, 384] - - [14, 60.521] - - - [12672, 1152, 1, 384] - - [296, 80.337] - - - [7680, 3072, 1, 384] - - [393, 84.375] - - - [19200, 2304, 1, 384] - - [51, 86.548] - - - [13056, 1153, 1, 384] - - [307, 75.2] - - - [27264, 1153, 1, 384] - - [25, 75.153] - - - [29568, 1153, 1, 384] - - [56, 77.447] - - - [11520, 11136, 1, 384] - - [51, 89.532] - - - [9216, 9216, 1, 384] - - [23, 87.436] - - - [18048, 1153, 1, 384] - - [350, 76.294] - - - [8064, 1152, 1, 384] - - [419, 72.833] - - - [22272, 7297, 1, 384] - - [42, 88.518] - - - [22272, 13441, 1, 384] - - [56, 89.718] - - - [22656, 2688, 1, 384] - - [51, 88.542] - - - [19584, 6144, 1, 384] - - [56, 89.72] - - - [8064, 7297, 1, 384] - - [25, 85.223] - - - [8064, 7681, 1, 384] - - [23, 86.659] - - - [23808, 7296, 1, 384] - - [27, 90.104] - - - [24960, 7296, 1, 384] - - [25, 89.816] - - - [14208, 6912, 1, 384] - - [323, 91.718] - - - [19968, 6528, 1, 384] - - [42, 89.457] - - - [28416, 7296, 1, 384] - - [42, 90.064] - - - [29952, 13440, 1, 384] - - [35, 90.917] - - - [17280, 7297, 1, 384] - - [27, 87.711] - - - [1536, 1152, 1, 384] - - [271, 46.914] - - - [8832, 1153, 1, 384] - - [360, 70.585] - - - [28032, 1153, 1, 384] - - [25, 77.624] - - - [2688, 2305, 1, 384] - - [293, 64.063] - - - [8064, 3072, 1, 384] - - [25, 84.009] - - - [28032, 3072, 1, 384] - - [35, 89.145] - - - [3840, 3456, 1, 384] - - [312, 80.816] - - - [21888, 1920, 1, 384] - - [36, 83.085] - - - [11904, 11520, 1, 384] - - [23, 89.564] - - - [9600, 9601, 1, 384] - - [56, 88.003] - - - [21120, 13440, 1, 384] - - [27, 90.747] - - - [19584, 2688, 1, 384] - - [34, 86.563] - - - [6912, 6528, 1, 384] - - [34, 85.626] - - - [29568, 1152, 1, 384] - - [52, 85.277] - - - [23808, 3072, 1, 384] - - [35, 88.181] - - - [18816, 4992, 1, 384] - - [51, 88.521] - - - [29952, 9216, 1, 384] - - [56, 90.581] - - - [22656, 13440, 1, 384] - - [25, 90.994] - - - [20352, 3456, 1, 384] - - [51, 87.447] - - - [3456, 1153, 1, 384] - - [418, 59.68] - - - [3840, 3457, 1, 384] - - [310, 77.16] - - - [15744, 8448, 1, 384] - - [51, 90.086] - - - [26112, 3072, 1, 384] - - [27, 88.29] - - - [28032, 14208, 1, 384] - - [365, 91.378] - - - [21504, 1536, 1, 384] - - [35, 83.608] - - - [11520, 768, 1, 384] - - [418, 72.512] - - - [6528, 6144, 1, 384] - - [25, 85.453] - - - [18432, 1153, 1, 384] - - [393, 74.41] - - - [3072, 1920, 1, 384] - - [417, 66.871] - - - [25344, 9216, 1, 384] - - [56, 90.665] - - - [30336, 7297, 1, 384] - - [35, 87.864] - - - [8832, 1152, 1, 384] - - [290, 72.966] - - - [26112, 9216, 1, 384] - - [56, 90.379] - - - [29952, 7296, 1, 384] - - [25, 90.178] - - - [11520, 11137, 1, 384] - - [27, 89.105] - - - [16896, 13440, 1, 384] - - [25, 90.773] - - - [29568, 13441, 1, 384] - - [49, 88.47] - - - [30336, 9216, 1, 384] - - [51, 90.28] - - - [2688, 1152, 1, 384] - - [392, 62.165] - - - [10368, 10368, 1, 384] - - [56, 89.575] - - - [25344, 11520, 1, 384] - - [25, 91.069] - - - [24576, 1920, 1, 384] - - [27, 83.311] - - - [11904, 4608, 1, 384] - - [51, 87.354] - - - [12672, 5376, 1, 384] - - [51, 88.794] - - - [11520, 3072, 1, 384] - - [23, 85.887] - - - [3072, 3073, 1, 384] - - [368, 72.327] - - - [24960, 11136, 1, 384] - - [51, 90.638] - - - [9984, 9600, 1, 384] - - [23, 88.848] - - - [19200, 2688, 1, 384] - - [51, 87.33] - - - [26496, 7296, 1, 384] - - [23, 90.121] - - - [23040, 3072, 1, 384] - - [35, 87.388] - - - [5760, 5761, 1, 384] - - [25, 84.571] - - - [5760, 5377, 1, 384] - - [27, 83.087] - - - [26880, 768, 1, 384] - - [51, 81.248] - - - [13824, 7297, 1, 384] - - [27, 88.168] - - - [13440, 7296, 1, 384] - - [23, 89.037] - - - [16128, 8448, 1, 384] - - [34, 89.625] - - - [24960, 3072, 1, 384] - - [25, 88.551] - - - [6144, 6144, 1, 384] - - [25, 86.096] - - - [27648, 13441, 1, 384] - - [319, 90.086] - - - [10368, 7297, 1, 384] - - [34, 87.132] - - - [22272, 2304, 1, 384] - - [23, 87.15] - - - [30720, 1153, 1, 384] - - [35, 77.37] - - - [24192, 13440, 1, 384] - - [25, 90.816] - - - [9984, 9984, 1, 384] - - [25, 89.105] - - - [29952, 1152, 1, 384] - - [23, 84.039] - - - [26112, 12672, 1, 384] - - [319, 91.15] - - - [8448, 7296, 1, 384] - - [25, 87.405] - - - [19584, 13440, 1, 384] - - [35, 90.462] - - - [21120, 1153, 1, 384] - - [317, 77.159] - - - [8832, 8449, 1, 384] - - [51, 86.262] - - - [28032, 13441, 1, 384] - - [25, 89.389] - - - [7680, 1153, 1, 384] - - [298, 69.481] - - - [19584, 9216, 1, 384] - - [51, 89.89] - - - [28800, 1152, 1, 384] - - [56, 84.104] - - - [29952, 768, 1, 384] - - [51, 83.522] - - - [12288, 1152, 1, 384] - - [25, 80.098] - - - [9600, 9217, 1, 384] - - [60, 86.678] - - - [14976, 13441, 1, 384] - - [27, 89.511] - - - [25344, 8832, 1, 384] - - [25, 90.401] - - - [18432, 4608, 1, 384] - - [23, 87.45] - - - [2304, 1920, 1, 384] - - [392, 61.393] - - - [11520, 4224, 1, 384] - - [25, 88.351] - - - [26496, 1153, 1, 384] - - [51, 77.269] - - - [28416, 2304, 1, 384] - - [51, 87.819] - - - [19200, 3072, 1, 384] - - [35, 87.757] - - - [26112, 7296, 1, 384] - - [27, 90.453] - - - [21504, 7297, 1, 384] - - [23, 87.803] - - - [4224, 1152, 1, 384] - - [305, 65.853] - - - [17664, 3840, 1, 384] - - [56, 88.524] - - - [6144, 1536, 1, 384] - - [360, 75.101] - - - [28032, 14592, 1, 384] - - [34, 91.167] - - - [8064, 8064, 1, 384] - - [25, 87.259] - - - [11136, 1152, 1, 384] - - [25, 74.068] - - - [13056, 7297, 1, 384] - - [23, 87.919] - - - [19968, 3456, 1, 384] - - [51, 87.779] - - - [25344, 7297, 1, 384] - - [27, 88.445] - - - [17280, 3840, 1, 384] - - [25, 88.77] - - - [28416, 1152, 1, 384] - - [25, 83.942] - - - [21120, 3072, 1, 384] - - [23, 88.775] - - - [28416, 7297, 1, 384] - - [35, 88.08] - - - [6528, 6529, 1, 384] - - [25, 83.475] - - - [26496, 9216, 1, 384] - - [34, 90.68] - - - [14592, 7296, 1, 384] - - [318, 91.305] - - - [14208, 1152, 1, 384] - - [25, 76.797] - - - [24576, 1536, 1, 384] - - [35, 83.575] - - - [18048, 7296, 1, 384] - - [35, 89.317] - - - [4608, 3072, 1, 384] - - [350, 78.649] - - - [28800, 14976, 1, 384] - - [42, 90.941] - - - [17664, 1152, 1, 384] - - [34, 80.529] - - - [24576, 7680, 1, 384] - - [35, 86.674] - - - [16896, 9216, 1, 384] - - [42, 89.757] - - - [20736, 3840, 1, 384] - - [42, 88.309] - - - [27264, 9216, 1, 384] - - [23, 89.624] - - - [21888, 3072, 1, 384] - - [28, 84.822] - - - [24576, 11136, 1, 384] - - [20, 87.405] - - - [14592, 1153, 1, 384] - - [393, 74.217] - - - [23424, 7296, 1, 384] - - [23, 90.291] - - - [22272, 3072, 1, 384] - - [25, 88.543] - - - [8832, 8832, 1, 384] - - [35, 88.346] - - - [8064, 7296, 1, 384] - - [25, 88.017] - - - [22656, 8832, 1, 384] - - [56, 90.025] - - - [22272, 2688, 1, 384] - - [42, 87.261] - - - [6528, 1152, 1, 384] - - [303, 73.597] - - - [8832, 8833, 1, 384] - - [34, 87.997] - - - [28800, 15360, 1, 384] - - [34, 90.849] - - - [23424, 1153, 1, 384] - - [25, 75.606] - - - [13440, 1152, 1, 384] - - [51, 79.284] - - - [10752, 10368, 1, 384] - - [34, 89.934] - - - [3456, 3456, 1, 384] - - [302, 76.565] - - - [4608, 4608, 1, 384] - - [23, 83.853] - - - [4224, 1153, 1, 384] - - [291, 58.505] - - - [12672, 2304, 1, 384] - - [23, 83.227] - - - [25728, 7297, 1, 384] - - [56, 88.005] - - - [5376, 1153, 1, 384] - - [291, 63.931] - - - [30720, 4992, 1, 384] - - [23, 89.461] - - - [27264, 7297, 1, 384] - - [27, 87.552] - - - [21504, 1920, 1, 384] - - [25, 87.192] - - - [11136, 11136, 1, 384] - - [56, 90.048] - - - [22656, 6144, 1, 384] - - [56, 89.491] - - - [26496, 13440, 1, 384] - - [319, 91.13] - - - [9216, 7296, 1, 384] - - [27, 87.749] - - - [17280, 7296, 1, 384] - - [35, 90.133] - - - [23040, 13441, 1, 384] - - [27, 89.079] - - - [23808, 13441, 1, 384] - - [34, 89.697] - - - [30336, 4224, 1, 384] - - [27, 89.287] - - - [6144, 1920, 1, 384] - - [303, 75.345] - - - [11904, 11904, 1, 384] - - [56, 90.052] - - - [30336, 13441, 1, 384] - - [23, 89.106] - - - [11904, 1536, 1, 384] - - [35, 78.25] - - - [24576, 9216, 1, 384] - - [30, 84.679] - - - [9984, 2304, 1, 384] - - [56, 83.947] - - - [18048, 4608, 1, 384] - - [51, 88.225] - - - [18432, 7297, 1, 384] - - [27, 87.867] - - - [11136, 3840, 1, 384] - - [25, 87.16] - - - [12288, 11904, 1, 384] - - [35, 89.561] - - - [19584, 7296, 1, 384] - - [25, 89.672] - - - [3072, 2689, 1, 384] - - [303, 69.934] - - - [2304, 2305, 1, 384] - - [392, 63.073] - - - [26496, 7297, 1, 384] - - [27, 88.044] - - - [15744, 1152, 1, 384] - - [23, 77.976] - - - [6912, 6912, 1, 384] - - [27, 87.066] - - - [4992, 3072, 1, 384] - - [25, 79.07] - - - [15744, 13440, 1, 384] - - [27, 90.395] - - - [2688, 2304, 1, 384] - - [417, 69.744] - - - [8448, 7297, 1, 384] - - [25, 86.977] - - - [25344, 11904, 1, 384] - - [25, 90.852] - - - [18432, 7296, 1, 384] - - [25, 89.726] - - - [8448, 8449, 1, 384] - - [27, 86.577] - - - [30720, 1536, 1, 384] - - [25, 85.263] - - - [9216, 1153, 1, 384] - - [293, 69.747] - - - [24192, 9216, 1, 384] - - [56, 90.538] - - - [25344, 2688, 1, 384] - - [56, 88.692] - - - [24576, 1153, 1, 384] - - [27, 73.377] - - - [14208, 7297, 1, 384] - - [51, 87.622] - - - [12672, 1920, 1, 384] - - [35, 82.978] - - - [4608, 4224, 1, 384] - - [422, 82.615] - - - [27264, 1536, 1, 384] - - [25, 84.468] - - - [24576, 13441, 1, 384] - - [23, 84.434] - - - [21504, 4992, 1, 384] - - [25, 89.361] - - - [21888, 4992, 1, 384] - - [48, 87.042] - - - [18432, 3072, 1, 384] - - [35, 86.579] - - - [19968, 6144, 1, 384] - - [25, 88.894] - - - [24192, 1536, 1, 384] - - [25, 85.442] - - - [9600, 7297, 1, 384] - - [25, 86.663] - - - [13824, 6528, 1, 384] - - [324, 90.979] - - - [2304, 2304, 1, 384] - - [270, 63.907] - - - [23424, 9984, 1, 384] - - [27, 90.684] - - - [18816, 1152, 1, 384] - - [296, 84.86] - - - [1152, 769, 1, 384] - - [264, 30.993] - - - [23424, 768, 1, 384] - - [423, 82.165] - - - [17280, 1153, 1, 384] - - [393, 74.342] - - - [9600, 2304, 1, 384] - - [25, 81.095] - - - [29184, 7297, 1, 384] - - [27, 88.119] - - - [26880, 3072, 1, 384] - - [27, 88.803] - - - [11520, 11520, 1, 384] - - [23, 90.049] - - - [23040, 6144, 1, 384] - - [56, 89.401] - - - [18048, 13440, 1, 384] - - [56, 90.675] - - - [30336, 1536, 1, 384] - - [25, 86.95] - - - [14976, 7680, 1, 384] - - [323, 92.364] - - - [14976, 1152, 1, 384] - - [23, 80.267] - - - [15360, 7680, 1, 384] - - [319, 92.178] - - - [28800, 13441, 1, 384] - - [23, 88.768] - - - [28032, 1920, 1, 384] - - [23, 88.185] - - - [16128, 2688, 1, 384] - - [56, 84.65] - - - [6144, 6145, 1, 384] - - [23, 81.33] - - - [10368, 7296, 1, 384] - - [25, 89.3] - - - [5760, 3072, 1, 384] - - [424, 82.938] - - - [24960, 9216, 1, 384] - - [34, 90.634] - - - [14592, 768, 1, 384] - - [302, 77.187] - - - [14208, 768, 1, 384] - - [317, 76.023] - - - [6912, 1153, 1, 384] - - [310, 67.65] - - - [21888, 13440, 1, 384] - - [26, 88.568] - - - [13056, 5760, 1, 384] - - [25, 89.082] - - - [12288, 1920, 1, 384] - - [25, 80.703] - - - [13056, 13056, 1, 384] - - [25, 90.324] - - - [6528, 1153, 1, 384] - - [312, 65.677] - - - [22272, 8448, 1, 384] - - [56, 90.311] - - - [7296, 1153, 1, 384] - - [291, 67.393] - - - [17280, 3456, 1, 384] - - [34, 87.076] - - - [27264, 13441, 1, 384] - - [319, 89.707] - - - [9216, 7297, 1, 384] - - [25, 87.23] - - - [4992, 4992, 1, 384] - - [35, 85.007] - - - [16128, 7297, 1, 384] - - [25, 87.628] - - - [20352, 13440, 1, 384] - - [23, 90.851] - - - [30336, 1153, 1, 384] - - [34, 76.888] - - - [13056, 7296, 1, 384] - - [27, 89.592] - - - [27648, 1152, 1, 384] - - [27, 85.05] - - - [13824, 6144, 1, 384] - - [23, 88.162] - - - [9216, 1920, 1, 384] - - [25, 81.957] - - - [17280, 13440, 1, 384] - - [25, 90.76] - - - [21888, 5376, 1, 384] - - [28, 85.822] - - - [3456, 3072, 1, 384] - - [389, 75.18] - - - [13440, 1153, 1, 384] - - [389, 72.913] - - - [24192, 7680, 1, 384] - - [23, 90.064] - - - [29952, 4224, 1, 384] - - [25, 89.448] - - - [8832, 3072, 1, 384] - - [27, 81.695] - - - [5760, 5760, 1, 384] - - [23, 85.038] - - - [23424, 6912, 1, 384] - - [25, 90.398] - - - [24192, 3072, 1, 384] - - [23, 87.871] - - - [18048, 3072, 1, 384] - - [23, 87.997] - - - [27264, 7296, 1, 384] - - [27, 90.021] - - - [11520, 3840, 1, 384] - - [27, 86.779] - - - [18432, 1536, 1, 384] - - [23, 84.098] - - - [11136, 10753, 1, 384] - - [25, 87.819] - - - [9600, 7296, 1, 384] - - [25, 89.09] - - - [26496, 13441, 1, 384] - - [25, 89.289] - - - [29568, 9216, 1, 384] - - [34, 89.74] - - - [25728, 7296, 1, 384] - - [25, 89.917] - - - [6528, 3072, 1, 384] - - [389, 82.255] - - - [18816, 9216, 1, 384] - - [331, 91.774] - - - [1920, 1153, 1, 384] - - [265, 49.333] - - - [1152, 1153, 1, 384] - - [161, 46.328] - - - [16896, 1153, 1, 384] - - [296, 76.03] - - - [4992, 1153, 1, 384] - - [287, 64.674] - - - [22656, 13441, 1, 384] - - [56, 89.502] - - - [9984, 1152, 1, 384] - - [315, 78.936] - - - [26496, 768, 1, 384] - - [34, 80.366] - - - [25344, 2304, 1, 384] - - [25, 87.575] - - - [14592, 6912, 1, 384] - - [324, 91.959] - - - [9216, 8833, 1, 384] - - [27, 87.479] - - - [19584, 7297, 1, 384] - - [23, 88.16] - - - [8448, 1153, 1, 384] - - [293, 68.915] - - - [21120, 7297, 1, 384] - - [45, 88.024] - - - [11520, 7297, 1, 384] - - [27, 87.053] - - - [12288, 7296, 1, 384] - - [27, 89.211] - - - [4224, 3841, 1, 384] - - [47, 76.561] - - - [9984, 9601, 1, 384] - - [25, 88.172] - - - [2304, 1152, 1, 384] - - [418, 51.103] - - - [21120, 7296, 1, 384] - - [25, 89.903] - - - [15360, 1153, 1, 384] - - [325, 75.147] - - - [27648, 3072, 1, 384] - - [27, 86.696] - - - [19200, 1153, 1, 384] - - [25, 74.708] - - - [28032, 1152, 1, 384] - - [56, 86.259] - - - [12672, 12288, 1, 384] - - [51, 90.041] - - - [22272, 5760, 1, 384] - - [23, 89.478] - - - [26496, 1152, 1, 384] - - [34, 86.187] - - - [26880, 7296, 1, 384] - - [25, 90.428] - - - [6528, 2304, 1, 384] - - [298, 83.007] - - - [9984, 7296, 1, 384] - - [25, 88.426] - - - [19968, 1152, 1, 384] - - [56, 83.787] - - - [10368, 9984, 1, 384] - - [27, 89.268] - - - [3840, 3840, 1, 384] - - [298, 81.645] - - - [5376, 1152, 1, 384] - - [300, 70.174] - - - [24192, 7296, 1, 384] - - [23, 90.513] - - - [14592, 3072, 1, 384] - - [27, 87.402] - - - [27648, 7297, 1, 384] - - [25, 87.86] - - - [23424, 1152, 1, 384] - - [25, 85.461] - - - [3456, 3457, 1, 384] - - [389, 76.389] - - - [13056, 2304, 1, 384] - - [27, 85.22] - - - [23808, 768, 1, 384] - - [299, 82.877] - - - [18048, 1152, 1, 384] - - [25, 81.926] - - - [28416, 9216, 1, 384] - - [51, 90.634] - - - [21888, 7297, 1, 384] - - [65, 83.447] - - - [25728, 12288, 1, 384] - - [56, 90.662] - - - [21120, 4224, 1, 384] - - [25, 88.889] - - - [20736, 3072, 1, 384] - - [27, 87.451] - - - [3840, 2688, 1, 384] - - [297, 73.959] - - - [29568, 7297, 1, 384] - - [45, 87.324] - - - [13824, 1153, 1, 384] - - [317, 73.681] - - - [15744, 1153, 1, 384] - - [389, 74.737] - - - [11136, 768, 1, 384] - - [303, 74.149] - - - [17664, 7297, 1, 384] - - [27, 88.339] - - - [24192, 7297, 1, 384] - - [27, 88.31] - - - [25344, 1153, 1, 384] - - [34, 77.647] - - - [30720, 4608, 1, 384] - - [23, 88.764] - - - [25728, 9216, 1, 384] - - [51, 90.516] - - - [29184, 1153, 1, 384] - - [42, 77.332] - - - [30336, 1152, 1, 384] - - [42, 84.968] - - - [24960, 13440, 1, 384] - - [42, 90.738] - - - [18432, 9216, 1, 384] - - [331, 91.744] - - - [15360, 13440, 1, 384] - - [25, 90.359] - - - [12288, 1536, 1, 384] - - [23, 80.308] - - - [8832, 8448, 1, 384] - - [25, 88.439] - - - [19968, 7297, 1, 384] - - [35, 88.003] - - - [19968, 3072, 1, 384] - - [35, 86.678] - - - [24960, 1920, 1, 384] - - [25, 87.186] - - - [15360, 1152, 1, 384] - - [25, 81.733] - - - [30720, 7296, 1, 384] - - [25, 90.195] - - - [14976, 1153, 1, 384] - - [393, 75.442] - - - [25344, 7296, 1, 384] - - [25, 90.358] - - - [16512, 8832, 1, 384] - - [39, 87.877] - - - [26112, 13441, 1, 384] - - [27, 89.566] - - - [22272, 1152, 1, 384] - - [27, 82.343] - - - [27648, 1536, 1, 384] - - [25, 85.661] - - - [15744, 1920, 1, 384] - - [23, 85.498] - - - [5760, 1153, 1, 384] - - [300, 65.752] - - - [29952, 13441, 1, 384] - - [25, 89.176] - - - [12672, 1153, 1, 384] - - [423, 72.883] - - - [13440, 2688, 1, 384] - - [56, 83.727] - - - [18816, 13440, 1, 384] - - [42, 90.572] - - - [22656, 9216, 1, 384] - - [34, 90.68] - - - [9216, 1152, 1, 384] - - [389, 74.841] - - - [20736, 1152, 1, 384] - - [23, 81.729] - - - [8832, 7296, 1, 384] - - [25, 88.738] - - - [15744, 7297, 1, 384] - - [23, 88.247] - - - [16512, 1153, 1, 384] - - [302, 74.455] - - - [29952, 7297, 1, 384] - - [35, 88.269] - - - [11136, 7297, 1, 384] - - [25, 87.578] - - - [9600, 3072, 1, 384] - - [23, 83.769] - - - [28800, 7297, 1, 384] - - [23, 87.828] - - - [27648, 13824, 1, 384] - - [319, 91.62] - - - [23808, 10368, 1, 384] - - [51, 90.899] - - - [13824, 13440, 1, 384] - - [25, 90.055] - - - [9216, 1536, 1, 384] - - [23, 80.124] - - - [23808, 1153, 1, 384] - - [42, 76.937] - - - [15360, 3072, 1, 384] - - [25, 85.325] - - - [12288, 3072, 1, 384] - - [25, 86.147] - - - [28416, 3072, 1, 384] - - [25, 88.406] - - - [30336, 13440, 1, 384] - - [23, 90.835] - - - [1152, 1152, 1, 384] - - [416, 39.991] - - - [21504, 3072, 1, 384] - - [35, 86.952] - - - [23040, 9216, 1, 384] - - [56, 90.364] - - - [22656, 7297, 1, 384] - - [35, 88.163] - - - [22656, 5760, 1, 384] - - [23, 89.68] - - - [12288, 11905, 1, 384] - - [25, 88.072] - - - [28032, 7296, 1, 384] - - [27, 90.255] - - - [29184, 3072, 1, 384] - - [27, 88.836] - - - [7680, 1152, 1, 384] - - [297, 72.503] - - - [16896, 7297, 1, 384] - - [35, 88.076] - - - [13056, 5376, 1, 384] - - [25, 89.113] - - - [5376, 4993, 1, 384] - - [25, 80.619] - - - [17280, 9216, 1, 384] - - [56, 89.968] - - - [8448, 8064, 1, 384] - - [23, 88.899] - - - [4608, 1153, 1, 384] - - [297, 61.829] - - - [19200, 9216, 1, 384] - - [331, 91.877] - - - [30720, 7297, 1, 384] - - [23, 88.048] - - - [13440, 5760, 1, 384] - - [23, 89.431] - - - [9984, 3072, 1, 384] - - [35, 82.616] - - - [29952, 15360, 1, 384] - - [23, 90.909] - - - [3840, 1152, 1, 384] - - [313, 61.326] - - - [10368, 9985, 1, 384] - - [25, 87.627] - - - [14592, 7297, 1, 384] - - [326, 89.442] - - - [3456, 3073, 1, 384] - - [25, 68.565] - - - [22272, 9216, 1, 384] - - [51, 90.666] - - - [8064, 8065, 1, 384] - - [35, 86.659] - - - [1536, 1536, 1, 384] - - [285, 52.7] - - - [30336, 4608, 1, 384] - - [51, 89.816] - - - [26112, 12288, 1, 384] - - [51, 90.415] - - - [11904, 11521, 1, 384] - - [51, 89.158] - - - [13440, 6144, 1, 384] - - [34, 89.227] - - - [19200, 13440, 1, 384] - - [35, 90.448] - - - [17280, 1152, 1, 384] - - [296, 84.38] - - - [23424, 3072, 1, 384] - - [25, 88.984] - - - [2304, 1921, 1, 384] - - [306, 61.184] - - - [12672, 7297, 1, 384] - - [23, 86.974] - - - [16896, 1152, 1, 384] - - [23, 82.505] - - - [18432, 1152, 1, 384] - - [25, 83.095] - - - [27264, 13824, 1, 384] - - [318, 91.359] - - - [10752, 1152, 1, 384] - - [310, 77.368] - - - [30336, 7296, 1, 384] - - [27, 90.3] - - - [11904, 3072, 1, 384] - - [35, 84.928] - - - [2304, 768, 1, 384] - - [384, 47.119] - - - [14592, 1152, 1, 384] - - [25, 78.634] - - - [20736, 13441, 1, 384] - - [51, 89.398] - - - [10752, 10752, 1, 384] - - [25, 89.21] - - - [23808, 13440, 1, 384] - - [23, 90.934] - - - [5376, 4992, 1, 384] - - [35, 85.527] - - - [10752, 3072, 1, 384] - - [23, 84.298] - - - [24576, 7296, 1, 384] - - [25, 87.228] - - - [7296, 7296, 1, 384] - - [35, 87.616] - - - [19200, 7296, 1, 384] - - [27, 89.94] - - - [25728, 8832, 1, 384] - - [27, 90.353] - - - [18048, 4224, 1, 384] - - [23, 88.169] - - - [4992, 1152, 1, 384] - - [392, 68.045] - - - [22272, 8832, 1, 384] - - [23, 90.167] - - - [21504, 1153, 1, 384] - - [298, 75.749] - - - [14208, 13440, 1, 384] - - [35, 90.637] - - - [10752, 7296, 1, 384] - - [35, 88.717] - - - [24192, 1152, 1, 384] - - [23, 83.698] - - - [7296, 1152, 1, 384] - - [300, 72.82] - - - [16128, 1153, 1, 384] - - [317, 73.252] - - - [19200, 7297, 1, 384] - - [25, 87.529] - - - [4992, 4993, 1, 384] - - [25, 79.64] - - - [12672, 12673, 1, 384] - - [25, 89.458] - - - [14208, 3072, 1, 384] - - [25, 85.44] - - - [23424, 6528, 1, 384] - - [23, 89.85] - - - [24576, 8064, 1, 384] - - [20, 87.204] - - - [6528, 6145, 1, 384] - - [35, 82.886] - - - [1920, 1537, 1, 384] - - [348, 55.48] - - - [21888, 8448, 1, 384] - - [26, 87.485] - - - [3072, 1536, 1, 384] - - [419, 65.798] - - - [7680, 7296, 1, 384] - - [25, 86.882] - - - [16896, 3072, 1, 384] - - [23, 87.3] - - - [24960, 11520, 1, 384] - - [25, 90.87] - - - [13824, 1152, 1, 384] - - [27, 81.264] - - - [25728, 1153, 1, 384] - - [25, 75.264] - - - [19968, 13441, 1, 384] - - [25, 89.301] - - - [13056, 13057, 1, 384] - - [23, 89.144] - - - [29184, 13440, 1, 384] - - [27, 90.843] - - - [23424, 7297, 1, 384] - - [28, 87.659] - - - [9216, 8832, 1, 384] - - [35, 88.217] - - - [11520, 1153, 1, 384] - - [304, 73.108] - - - [19968, 1153, 1, 384] - - [303, 76.236] - - - [14976, 13440, 1, 384] - - [23, 90.603] - - - [9216, 3072, 1, 384] - - [25, 84.6] - - - [24192, 10752, 1, 384] - - [25, 90.661] - - - [16128, 8832, 1, 384] - - [35, 89.401] - - - [9984, 1153, 1, 384] - - [303, 71.976] - - - [8064, 1153, 1, 384] - - [349, 70.44] - - - [12672, 12672, 1, 384] - - [25, 89.862] - - - [25728, 13441, 1, 384] - - [34, 89.082] - - - [11520, 1152, 1, 384] - - [422, 80.734] - - - [26496, 12672, 1, 384] - - [34, 90.847] - - - [1920, 768, 1, 384] - - [385, 51.55] - - - [20352, 1153, 1, 384] - - [299, 75.852] - - - [10368, 2688, 1, 384] - - [34, 83.77] - - - [6912, 2304, 1, 384] - - [42, 81.625] - - - [17664, 13440, 1, 384] - - [51, 90.735] - - - [17664, 9216, 1, 384] - - [51, 90.066] - - - [25728, 13440, 1, 384] - - [34, 90.911] - - - [10752, 3456, 1, 384] - - [23, 86.158] - - - [6144, 3072, 1, 384] - - [424, 82.19] - - - [9216, 9217, 1, 384] - - [26, 84.232] - - - [3840, 2304, 1, 384] - - [291, 71.59] - - - [12288, 12289, 1, 384] - - [52, 83.752] - - - [11136, 11137, 1, 384] - - [23, 88.387] - - - [11904, 7297, 1, 384] - - [25, 86.522] - - - [29568, 3072, 1, 384] - - [49, 87.919] - - - [12288, 1153, 1, 384] - - [360, 71.488] - - - [18816, 1920, 1, 384] - - [25, 83.941] - - - [13056, 1152, 1, 384] - - [425, 82.408] - - - [8448, 768, 1, 384] - - [336, 69.026] - - - [18816, 2304, 1, 384] - - [42, 85.066] - - - [5376, 3072, 1, 384] - - [25, 77.57] - - - [16512, 1152, 1, 384] - - [389, 79.581] - - - [27648, 7296, 1, 384] - - [35, 90.303] - - - [7296, 2688, 1, 384] - - [25, 83.159] - - - [29184, 15360, 1, 384] - - [25, 90.741] - - - [4608, 4609, 1, 384] - - [25, 77.982] - - - [7296, 7297, 1, 384] - - [23, 84.469] - - - [30720, 9216, 1, 384] - - [72, 88.322] - - - [16384, 3072, 1, 256] - - [20, 62.582] - - - [42496, 10240, 1, 256] - - [27, 74.992] - - - [20992, 7168, 1, 256] - - [25, 73.905] - - - [8960, 5632, 1, 256] - - [57, 72.47] - - - [4864, 256, 1, 256] - - [172, 41.505] - - - [23552, 3584, 1, 256] - - [23, 73.921] - - - [2560, 1281, 1, 256] - - [417, 47.467] - - - [7168, 1280, 1, 256] - - [289, 64.229] - - - [1536, 1153, 1, 384] - - [354, 42.235] - - - [18224, 256, 1, 256] - - [293, 50.589] - - - [13441, 128, 1, 384] - - [173, 47.394] - - - [10753, 128, 1, 384] - - [174, 40.289] - - - [12289, 128, 1, 384] - - [175, 42.921] - - - [385, 128, 1, 384] - - [183, 2.268] - - - [11136, 128, 1, 384] - - [162, 48.351] - - - [13440, 128, 1, 384] - - [351, 42.818] - - - [1153, 128, 1, 384] - - [113, 10.092] - - - [6145, 128, 1, 384] - - [129, 37.382] - - - [4225, 128, 1, 384] - - [132, 26.504] - - - [1537, 128, 1, 384] - - [113, 13.779] - - - [8064, 128, 1, 384] - - [90, 42.016] - - - [3072, 128, 1, 384] - - [103, 28.226] - - - [3457, 128, 1, 384] - - [100, 26.58] - - - [5760, 128, 1, 384] - - [90, 39.611] - - - [8449, 128, 1, 384] - - [121, 38.339] - - - [2305, 128, 1, 384] - - [131, 18.904] - - - [11520, 128, 1, 384] - - [163, 49.823] - - - [11521, 128, 1, 384] - - [176, 42.238] - - - [6528, 128, 1, 384] - - [347, 32.274] - - - [14208, 128, 1, 384] - - [354, 42.589] - - - [768, 128, 1, 384] - - [92, 7.268] - - - [12672, 128, 1, 384] - - [93, 53.145] - - - [9216, 128, 1, 384] - - [106, 45.249] - - - [8448, 128, 1, 384] - - [103, 44.434] - - - [6144, 128, 1, 384] - - [96, 41.479] - - - [2689, 128, 1, 384] - - [100, 21.972] - - - [4224, 128, 1, 384] - - [92, 29.979] - - - [9601, 128, 1, 384] - - [121, 41.933] - - - [13056, 128, 1, 384] - - [177, 54.343] - - - [8065, 128, 1, 384] - - [108, 36.521] - - - [2304, 128, 1, 384] - - [90, 21.347] - - - [8833, 128, 1, 384] - - [132, 39.514] - - - [13824, 128, 1, 384] - - [354, 42.075] - - - [7680, 128, 1, 384] - - [258, 32.081] - - - [3840, 128, 1, 384] - - [257, 19.7] - - - [1920, 128, 1, 384] - - [96, 18.247] - - - [5761, 128, 1, 384] - - [118, 35.338] - - - [7681, 128, 1, 384] - - [178, 34.854] - - - [4608, 128, 1, 384] - - [105, 32.496] - - - [10369, 128, 1, 384] - - [140, 44.325] - - - [3841, 128, 1, 384] - - [132, 24.094] - - - [7296, 128, 1, 384] - - [92, 38.194] - - - [7297, 128, 1, 384] - - [129, 33.523] - - - [10752, 128, 1, 384] - - [125, 47.522] - - - [1536, 128, 1, 384] - - [258, 8.411] - - - [11137, 128, 1, 384] - - [179, 40.762] - - - [2688, 128, 1, 384] - - [96, 24.495] - - - [4609, 128, 1, 384] - - [144, 28.912] - - - [6529, 128, 1, 384] - - [129, 39.286] - - - [11905, 128, 1, 384] - - [179, 43.004] - - - [6912, 128, 1, 384] - - [96, 46.101] - - - [769, 128, 1, 384] - - [183, 4.234] - - - [12288, 128, 1, 384] - - [170, 52.227] - - - [15360, 128, 1, 384] - - [258, 45.333] - - - [9600, 128, 1, 384] - - [96, 48.868] - - - [13057, 128, 1, 384] - - [180, 46.04] - - - [10368, 128, 1, 384] - - [347, 40.008] - - - [12673, 128, 1, 384] - - [116, 45.262] - - - [9217, 128, 1, 384] - - [178, 38.95] - - - [4993, 128, 1, 384] - - [122, 31.057] - - - [9984, 128, 1, 384] - - [118, 49.458] - - - [6913, 128, 1, 384] - - [144, 40.93] - - - [8832, 128, 1, 384] - - [152, 45.694] - - - [3073, 128, 1, 384] - - [181, 23.96] - - - [14976, 128, 1, 384] - - [259, 44.207] - - - [384, 128, 1, 384] - - [91, 3.649] - - - [5377, 128, 1, 384] - - [129, 33.352] - - - [1152, 128, 1, 384] - - [257, 6.478] - - - [9985, 128, 1, 384] - - [140, 42.849] - - - [14592, 128, 1, 384] - - [355, 45.218] - - - [4992, 128, 1, 384] - - [90, 34.871] - - - [3456, 128, 1, 384] - - [152, 30.612] - - - [1921, 128, 1, 384] - - [146, 15.697] - - - [5376, 128, 1, 384] - - [91, 37.318] - - - [11904, 128, 1, 384] - - [170, 50.594] - - - [44544, 2048, 1, 384] - - [51, 88.554] - - - [39552, 512, 1, 384] - - [42, 79.797] - - - [38016, 22145, 1, 384] - - [35, 89.555] - - - [39552, 23297, 1, 384] - - [35, 89.357] - - - [39552, 23681, 1, 384] - - [27, 89.45] - - - [36864, 2048, 1, 384] - - [35, 82.377] - - - [44544, 28673, 1, 384] - - [30, 87.03] - - - [43776, 512, 1, 384] - - [24, 79.402] - - - [43392, 1024, 1, 384] - - [34, 86.883] - - - [42240, 4096, 1, 384] - - [34, 90.427] - - - [42624, 26369, 1, 384] - - [49, 83.329] - - - [35328, 1024, 1, 384] - - [56, 83.966] - - - [36096, 384, 1, 384] - - [322, 79.689] - - - [38784, 4096, 1, 384] - - [34, 89.674] - - - [39552, 384, 1, 384] - - [323, 78.339] - - - [42240, 8192, 1, 384] - - [51, 91.007] - - - [42240, 25985, 1, 384] - - [23, 89.72] - - - [38016, 4096, 1, 384] - - [56, 89.941] - - - [39168, 4096, 1, 384] - - [51, 89.71] - - - [35328, 19457, 1, 384] - - [30, 87.338] - - - [43392, 2048, 1, 384] - - [34, 88.584] - - - [38400, 4096, 1, 384] - - [42, 88.905] - - - [35712, 1024, 1, 384] - - [51, 84.814] - - - [36480, 2048, 1, 384] - - [42, 87.979] - - - [40704, 512, 1, 384] - - [25, 81.464] - - - [36864, 20609, 1, 384] - - [35, 88.091] - - - [37632, 21761, 1, 384] - - [35, 89.191] - - - [38016, 2048, 1, 384] - - [42, 87.947] - - - [44160, 2048, 1, 384] - - [51, 88.431] - - - [35328, 384, 1, 384] - - [320, 78.358] - - - [43392, 384, 1, 384] - - [424, 81.819] - - - [39168, 512, 1, 384] - - [313, 83.751] - - - [38784, 1024, 1, 384] - - [42, 84.387] - - - [35328, 2048, 1, 384] - - [56, 87.335] - - - [44544, 8192, 1, 384] - - [23, 89.941] - - - [40704, 384, 1, 384] - - [323, 80.143] - - - [39936, 512, 1, 384] - - [25, 80.271] - - - [41472, 25217, 1, 384] - - [27, 89.363] - - - [42240, 2048, 1, 384] - - [56, 89.372] - - - [37632, 512, 1, 384] - - [325, 81.32] - - - [37248, 1024, 1, 384] - - [34, 84.781] - - - [42240, 26369, 1, 384] - - [25, 89.63] - - - [43776, 384, 1, 384] - - [327, 78.6] - - - [44160, 8192, 1, 384] - - [56, 90.909] - - - [39936, 1024, 1, 384] - - [56, 86.121] - - - [43392, 27137, 1, 384] - - [23, 88.823] - - - [39936, 384, 1, 384] - - [325, 78.888] - - - [41472, 25601, 1, 384] - - [30, 87.287] - - - [36864, 4096, 1, 384] - - [30, 84.148] - - - [43392, 8192, 1, 384] - - [56, 90.761] - - - [36096, 512, 1, 384] - - [25, 78.538] - - - [36480, 4096, 1, 384] - - [34, 90.149] - - - [40320, 512, 1, 384] - - [35, 80.915] - - - [41088, 4096, 1, 384] - - [72, 88.437] - - - [43776, 27521, 1, 384] - - [52, 88.034] - - - [35328, 19073, 1, 384] - - [27, 89.484] - - - [44160, 384, 1, 384] - - [319, 79.612] - - - [36864, 8192, 1, 384] - - [30, 87.044] - - - [41088, 2048, 1, 384] - - [59, 86.645] - - - [38016, 21761, 1, 384] - - [23, 89.316] - - - [41856, 1024, 1, 384] - - [34, 87.124] - - - [39552, 8192, 1, 384] - - [56, 90.847] - - - [37632, 4096, 1, 384] - - [42, 90.013] - - - [41856, 384, 1, 384] - - [320, 79.821] - - - [44160, 28289, 1, 384] - - [25, 89.054] - - - [43008, 26753, 1, 384] - - [52, 88.861] - - - [38400, 512, 1, 384] - - [269, 81.761] - - - [39168, 384, 1, 384] - - [322, 81.447] - - - [37632, 1024, 1, 384] - - [42, 85.403] - - - [44544, 4096, 1, 384] - - [72, 88.698] - - - [42240, 512, 1, 384] - - [425, 84.238] - - - [43008, 2048, 1, 384] - - [72, 86.187] - - - [36480, 20609, 1, 384] - - [51, 89.24] - - - [36864, 512, 1, 384] - - [51, 80.085] - - - [43008, 384, 1, 384] - - [424, 81.874] - - - [43392, 4096, 1, 384] - - [34, 90.249] - - - [38400, 22145, 1, 384] - - [35, 89.347] - - - [39936, 23681, 1, 384] - - [23, 89.105] - - - [36096, 19841, 1, 384] - - [52, 87.621] - - - [44544, 512, 1, 384] - - [25, 82.624] - - - [38400, 2048, 1, 384] - - [34, 88.327] - - - [41856, 25985, 1, 384] - - [25, 89.011] - - - [42624, 2048, 1, 384] - - [35, 83.262] - - - [38400, 1024, 1, 384] - - [51, 86.389] - - - [36480, 512, 1, 384] - - [25, 79.706] - - - [42624, 26753, 1, 384] - - [49, 83.126] - - - [43776, 27905, 1, 384] - - [30, 88.04] - - - [37248, 2048, 1, 384] - - [56, 87.981] - - - [35712, 19841, 1, 384] - - [27, 89.178] - - - [43392, 27521, 1, 384] - - [23, 89.004] - - - [43008, 1024, 1, 384] - - [34, 86.036] - - - [42624, 512, 1, 384] - - [25, 79.899] - - - [41472, 384, 1, 384] - - [313, 81.27] - - - [40704, 2048, 1, 384] - - [51, 88.243] - - - [36096, 2048, 1, 384] - - [60, 85.766] - - - [39936, 4096, 1, 384] - - [59, 87.649] - - - [40320, 2048, 1, 384] - - [56, 89.135] - - - [41088, 8192, 1, 384] - - [72, 89.419] - - - [35328, 8192, 1, 384] - - [27, 90.273] - - - [40320, 4096, 1, 384] - - [56, 90.213] - - - [41856, 512, 1, 384] - - [325, 83.613] - - - [39552, 4096, 1, 384] - - [56, 90.197] - - - [35712, 2048, 1, 384] - - [34, 88.452] - - - [39936, 24065, 1, 384] - - [35, 88.775] - - - [36480, 20225, 1, 384] - - [27, 89.099] - - - [38016, 1024, 1, 384] - - [34, 85.94] - - - [43008, 512, 1, 384] - - [23, 80.274] - - - [40704, 24833, 1, 384] - - [35, 88.984] - - - [37248, 4096, 1, 384] - - [51, 89.751] - - - [41856, 4096, 1, 384] - - [34, 89.878] - - - [41472, 512, 1, 384] - - [315, 83.016] - - - [39552, 2048, 1, 384] - - [56, 89.065] - - - [41088, 384, 1, 384] - - [326, 80.352] - - - [36480, 8192, 1, 384] - - [56, 90.778] - - - [37632, 2048, 1, 384] - - [51, 88.738] - - - [40704, 8192, 1, 384] - - [51, 90.642] - - - [36864, 20993, 1, 384] - - [35, 87.812] - - - [35328, 512, 1, 384] - - [312, 82.449] - - - [40320, 384, 1, 384] - - [321, 79.411] - - - [36096, 1024, 1, 384] - - [85, 83.824] - - - [42624, 8192, 1, 384] - - [23, 89.493] - - - [38784, 22529, 1, 384] - - [30, 87.014] - - - [44160, 4096, 1, 384] - - [34, 90.077] - - - [41472, 4096, 1, 384] - - [85, 88.798] - - - [36480, 1024, 1, 384] - - [34, 85.97] - - - [38784, 2048, 1, 384] - - [51, 87.715] - - - [44544, 1024, 1, 384] - - [34, 85.806] - - - [41088, 24833, 1, 384] - - [52, 88.119] - - - [36864, 384, 1, 384] - - [306, 80.185] - - - [43392, 512, 1, 384] - - [25, 81.03] - - - [39168, 8192, 1, 384] - - [25, 90.446] - - - [42624, 4096, 1, 384] - - [23, 87.761] - - - [40320, 24065, 1, 384] - - [27, 89.226] - - - [44160, 512, 1, 384] - - [424, 82.788] - - - [38016, 384, 1, 384] - - [319, 76.045] - - - [38016, 512, 1, 384] - - [424, 82.252] - - - [37248, 512, 1, 384] - - [23, 80.906] - - - [43776, 2048, 1, 384] - - [59, 86.412] - - - [35712, 8192, 1, 384] - - [23, 90.561] - - - [38400, 384, 1, 384] - - [309, 76.836] - - - [42240, 1024, 1, 384] - - [51, 87.472] - - - [35712, 19457, 1, 384] - - [30, 86.829] - - - [41856, 2048, 1, 384] - - [56, 88.873] - - - [41472, 1024, 1, 384] - - [42, 86.378] - - - [37632, 384, 1, 384] - - [324, 75.413] - - - [40704, 1024, 1, 384] - - [42, 84.888] - - - [43008, 27137, 1, 384] - - [30, 88.817] - - - [40704, 4096, 1, 384] - - [56, 90.082] - - - [36096, 20225, 1, 384] - - [30, 87.899] - - - [39936, 8192, 1, 384] - - [72, 88.219] - - - [38784, 384, 1, 384] - - [319, 77.186] - - - [38784, 8192, 1, 384] - - [34, 90.697] - - - [42624, 384, 1, 384] - - [309, 80.845] - - - [35712, 4096, 1, 384] - - [56, 89.523] - - - [37632, 8192, 1, 384] - - [51, 90.488] - - - [38784, 22913, 1, 384] - - [25, 89.099] - - - [36864, 1024, 1, 384] - - [42, 84.734] - - - [37248, 384, 1, 384] - - [322, 80.941] - - - [39168, 23297, 1, 384] - - [25, 89.091] - - - [40704, 24449, 1, 384] - - [25, 89.212] - - - [41472, 2048, 1, 384] - - [56, 88.036] - - - [44160, 27905, 1, 384] - - [23, 88.95] - - - [44160, 1024, 1, 384] - - [51, 85.41] - - - [36480, 384, 1, 384] - - [323, 79.846] - - - [42240, 384, 1, 384] - - [324, 82.169] - - - [44544, 28289, 1, 384] - - [30, 89.099] - - - [37248, 21377, 1, 384] - - [23, 88.605] - - - [36096, 4096, 1, 384] - - [72, 87.498] - - - [38784, 512, 1, 384] - - [309, 83.883] - - - [35712, 384, 1, 384] - - [321, 78.476] - - - [43776, 1024, 1, 384] - - [72, 84.107] - - - [41088, 25217, 1, 384] - - [30, 88.117] - - - [40320, 8192, 1, 384] - - [23, 90.614] - - - [39168, 22913, 1, 384] - - [27, 89.058] - - - [38400, 8192, 1, 384] - - [35, 90.376] - - - [41088, 512, 1, 384] - - [22, 79.633] - - - [42624, 1024, 1, 384] - - [23, 83.626] - - - [39168, 2048, 1, 384] - - [34, 88.606] - - - [43008, 4096, 1, 384] - - [59, 87.372] - - - [35712, 512, 1, 384] - - [423, 82.384] - - - [41856, 8192, 1, 384] - - [34, 90.813] - - - [43008, 8192, 1, 384] - - [52, 87.95] - - - [41472, 8192, 1, 384] - - [27, 90.544] - - - [41088, 1024, 1, 384] - - [37, 83.173] - - - [37248, 20993, 1, 384] - - [25, 88.42] - - - [44544, 384, 1, 384] - - [327, 79.871] - - - [36096, 8192, 1, 384] - - [49, 88.991] - - - [43776, 8192, 1, 384] - - [59, 89.499] - - - [41856, 25601, 1, 384] - - [30, 87.17] - - - [37632, 21377, 1, 384] - - [35, 89.33] - - - [40320, 24449, 1, 384] - - [23, 89.215] - - - [43776, 4096, 1, 384] - - [72, 88.614] - - - [35328, 4096, 1, 384] - - [54, 88.908] - - - [39552, 1024, 1, 384] - - [56, 85.957] - - - [38016, 8192, 1, 384] - - [42, 90.868] - - - [38400, 22529, 1, 384] - - [52, 87.23] - - - [39936, 2048, 1, 384] - - [72, 86.374] - - - [39168, 1024, 1, 384] - - [51, 85.143] - - - [37248, 8192, 1, 384] - - [27, 90.594] - - - [40320, 1024, 1, 384] - - [56, 87.081] - - - [26112, 1024, 1, 384] - - [56, 85.092] - - - [24192, 2048, 1, 384] - - [56, 86.935] - - - [13440, 5761, 1, 384] - - [23, 87.024] - - - [3456, 384, 1, 384] - - [271, 51.608] - - - [21888, 4096, 1, 384] - - [22, 84.551] - - - [384, 384, 1, 384] - - [256, 10.767] - - - [21120, 1024, 1, 384] - - [51, 83.912] - - - [30336, 4096, 1, 384] - - [56, 89.998] - - - [31488, 512, 1, 384] - - [424, 80.456] - - - [2304, 1793, 1, 384] - - [313, 57.018] - - - [16896, 9217, 1, 384] - - [51, 86.173] - - - [9216, 1024, 1, 384] - - [425, 74.472] - - - [29568, 1024, 1, 384] - - [54, 83.999] - - - [27264, 11393, 1, 384] - - [23, 88.608] - - - [33408, 17537, 1, 384] - - [23, 89.304] - - - [18816, 1024, 1, 384] - - [34, 81.935] - - - [5760, 1024, 1, 384] - - [300, 65.991] - - - [31104, 14849, 1, 384] - - [35, 89.596] - - - [18816, 4096, 1, 384] - - [42, 88.971] - - - [11136, 1024, 1, 384] - - [424, 77.75] - - - [17664, 9985, 1, 384] - - [34, 88.697] - - - [9216, 512, 1, 384] - - [390, 65.198] - - - [17664, 1024, 1, 384] - - [34, 78.042] - - - [17664, 512, 1, 384] - - [293, 73.017] - - - [31488, 384, 1, 384] - - [311, 78.426] - - - [15744, 8065, 1, 384] - - [35, 88.222] - - - [5760, 3841, 1, 384] - - [23, 80.878] - - - [24192, 1024, 1, 384] - - [424, 85.38] - - - [20352, 384, 1, 384] - - [302, 74.267] - - - [21888, 2048, 1, 384] - - [23, 80.794] - - - [7680, 2048, 1, 384] - - [389, 80.221] - - - [2688, 512, 1, 384] - - [163, 48.391] - - - [13056, 1024, 1, 384] - - [424, 81.317] - - - [22656, 14977, 1, 384] - - [51, 89.781] - - - [10752, 6785, 1, 384] - - [23, 86.162] - - - [6912, 2048, 1, 384] - - [42, 80.824] - - - [15360, 512, 1, 384] - - [425, 72.804] - - - [31104, 384, 1, 384] - - [317, 77.79] - - - [30720, 14465, 1, 384] - - [23, 89.054] - - - [17280, 2048, 1, 384] - - [56, 85.776] - - - [34176, 1024, 1, 384] - - [34, 84.779] - - - [16896, 2048, 1, 384] - - [51, 84.012] - - - [17664, 384, 1, 384] - - [419, 67.11] - - - [21504, 512, 1, 384] - - [390, 76.165] - - - [18048, 10369, 1, 384] - - [35, 88.443] - - - [15744, 1024, 1, 384] - - [391, 81.026] - - - [33408, 4096, 1, 384] - - [51, 89.946] - - - [11904, 4096, 1, 384] - - [34, 85.832] - - - [18816, 512, 1, 384] - - [314, 75.047] - - - [34944, 4096, 1, 384] - - [56, 89.626] - - - [13824, 2048, 1, 384] - - [42, 84.822] - - - [3840, 512, 1, 384] - - [392, 44.633] - - - [4992, 1024, 1, 384] - - [293, 60.966] - - - [11136, 7553, 1, 384] - - [25, 87.11] - - - [16512, 1024, 1, 384] - - [302, 78.69] - - - [17280, 9217, 1, 384] - - [85, 85.933] - - - [29184, 1024, 1, 384] - - [34, 84.597] - - - [18048, 512, 1, 384] - - [357, 72.919] - - - [6528, 384, 1, 384] - - [287, 60.501] - - - [28416, 1024, 1, 384] - - [51, 82.714] - - - [2688, 1153, 1, 384] - - [392, 61.295] - - - [34560, 18305, 1, 384] - - [35, 89.206] - - - [20736, 384, 1, 384] - - [300, 75.215] - - - [11520, 512, 1, 384] - - [428, 65.78] - - - [26112, 8192, 1, 384] - - [42, 90.445] - - - [31872, 384, 1, 384] - - [314, 78.822] - - - [24192, 512, 1, 384] - - [51, 78.806] - - - [19968, 2048, 1, 384] - - [34, 86.677] - - - [32256, 8192, 1, 384] - - [35, 90.325] - - - [11520, 384, 1, 384] - - [293, 69.974] - - - [1920, 1409, 1, 384] - - [298, 50.525] - - - [25728, 9857, 1, 384] - - [27, 88.779] - - - [9216, 5633, 1, 384] - - [35, 84.683] - - - [28032, 12161, 1, 384] - - [27, 89.361] - - - [28800, 8192, 1, 384] - - [56, 90.586] - - - [28416, 12161, 1, 384] - - [51, 89.491] - - - [23040, 15361, 1, 384] - - [52, 87.18] - - - [31488, 15617, 1, 384] - - [35, 89.332] - - - [22272, 14209, 1, 384] - - [56, 89.567] - - - [1536, 512, 1, 384] - - [91, 41.102] - - - [1152, 257, 1, 384] - - [151, 21.611] - - - [21120, 2048, 1, 384] - - [24, 86.878] - - - [32256, 16001, 1, 384] - - [25, 89.634] - - - [9600, 6017, 1, 384] - - [27, 86.235] - - - [32640, 384, 1, 384] - - [306, 79.683] - - - [34176, 512, 1, 384] - - [56, 80.817] - - - [10368, 512, 1, 384] - - [293, 63.549] - - - [21120, 384, 1, 384] - - [303, 76.068] - - - [29568, 4096, 1, 384] - - [24, 88.237] - - - [31872, 2048, 1, 384] - - [51, 87.371] - - - [8832, 384, 1, 384] - - [293, 66.166] - - - [4224, 384, 1, 384] - - [278, 55.472] - - - [33408, 8192, 1, 384] - - [25, 90.609] - - - [768, 257, 1, 384] - - [151, 14.592] - - - [10368, 6401, 1, 384] - - [25, 86.647] - - - [13824, 384, 1, 384] - - [293, 64.262] - - - [29568, 512, 1, 384] - - [42, 77.505] - - - [28032, 1024, 1, 384] - - [34, 85.796] - - - [19200, 384, 1, 384] - - [293, 71.557] - - - [23040, 2048, 1, 384] - - [34, 85.758] - - - [8448, 4481, 1, 384] - - [25, 83.925] - - - [22272, 14593, 1, 384] - - [56, 89.97] - - - [26496, 10241, 1, 384] - - [56, 85.965] - - - [19584, 384, 1, 384] - - [300, 72.581] - - - [4992, 3457, 1, 384] - - [25, 80.037] - - - [22656, 384, 1, 384] - - [307, 75.998] - - - [15360, 1024, 1, 384] - - [34, 80.098] - - - [7296, 2048, 1, 384] - - [424, 82.125] - - - [30720, 384, 1, 384] - - [310, 77.384] - - - [6144, 2177, 1, 384] - - [310, 76.277] - - - [30720, 14849, 1, 384] - - [35, 88.592] - - - [23424, 2048, 1, 384] - - [42, 85.898] - - - [5760, 384, 1, 384] - - [284, 55.312] - - - [6144, 2561, 1, 384] - - [27, 73.855] - - - [12672, 384, 1, 384] - - [293, 67.419] - - - [16128, 8065, 1, 384] - - [319, 90.412] - - - [10752, 7169, 1, 384] - - [85, 83.986] - - - [2304, 384, 1, 384] - - [256, 43.869] - - - [18816, 2048, 1, 384] - - [42, 85.312] - - - [22272, 4096, 1, 384] - - [56, 89.256] - - - [12672, 4993, 1, 384] - - [23, 86.778] - - - [12288, 512, 1, 384] - - [296, 69.133] - - - [13056, 4993, 1, 384] - - [35, 85.015] - - - [19584, 512, 1, 384] - - [305, 77.907] - - - [30336, 14465, 1, 384] - - [25, 89.258] - - - [5376, 3841, 1, 384] - - [25, 81.167] - - - [17664, 9601, 1, 384] - - [42, 88.671] - - - [29952, 2048, 1, 384] - - [56, 86.831] - - - [8832, 512, 1, 384] - - [390, 62.517] - - - [9984, 512, 1, 384] - - [290, 62.819] - - - [19200, 1024, 1, 384] - - [299, 83.114] - - - [24192, 8321, 1, 384] - - [56, 88.409] - - - [26112, 10241, 1, 384] - - [30, 86.108] - - - [17280, 9601, 1, 384] - - [56, 88.482] - - - [7296, 384, 1, 384] - - [289, 56.179] - - - [16512, 8449, 1, 384] - - [30, 85.255] - - - [11904, 4225, 1, 384] - - [51, 85.165] - - - [24576, 4096, 1, 384] - - [23, 82.017] - - - [6912, 2945, 1, 384] - - [25, 79.917] - - - [33024, 16769, 1, 384] - - [30, 88.675] - - - [24576, 8705, 1, 384] - - [27, 83.163] - - - [16128, 2048, 1, 384] - - [51, 83.245] - - - [13824, 6145, 1, 384] - - [27, 84.121] - - - [28800, 512, 1, 384] - - [25, 75.38] - - - [33792, 8192, 1, 384] - - [59, 88.394] - - - [27648, 11393, 1, 384] - - [25, 88.687] - - - [21888, 384, 1, 384] - - [305, 73.907] - - - [12672, 4096, 1, 384] - - [34, 87.969] - - - [23040, 14977, 1, 384] - - [23, 89.321] - - - [11904, 384, 1, 384] - - [293, 64.364] - - - [7680, 3713, 1, 384] - - [27, 80.768] - - - [24576, 8192, 1, 384] - - [30, 84.517] - - - [34176, 384, 1, 384] - - [318, 75.55] - - - [17664, 2048, 1, 384] - - [42, 84.228] - - - [29952, 4096, 1, 384] - - [56, 89.17] - - - [9984, 6017, 1, 384] - - [35, 84.842] - - - [33408, 2048, 1, 384] - - [56, 88.753] - - - [21120, 4096, 1, 384] - - [54, 88.935] - - - [34560, 4096, 1, 384] - - [34, 89.877] - - - [19200, 11521, 1, 384] - - [35, 88.647] - - - [21120, 13057, 1, 384] - - [25, 89.01] - - - [25728, 384, 1, 384] - - [293, 77.151] - - - [28800, 12929, 1, 384] - - [27, 88.677] - - - [20736, 1024, 1, 384] - - [34, 83.369] - - - [18816, 10753, 1, 384] - - [23, 89.03] - - - [34560, 8192, 1, 384] - - [23, 90.331] - - - [23040, 512, 1, 384] - - [291, 74.125] - - - [30336, 2048, 1, 384] - - [56, 87.643] - - - [17280, 512, 1, 384] - - [291, 71.561] - - - [19200, 2048, 1, 384] - - [51, 86.532] - - - [12288, 4225, 1, 384] - - [25, 84.633] - - - [15744, 7681, 1, 384] - - [356, 90.151] - - - [30720, 4096, 1, 384] - - [72, 86.995] - - - [10752, 384, 1, 384] - - [293, 67.277] - - - [15744, 512, 1, 384] - - [348, 73.656] - - - [24960, 384, 1, 384] - - [297, 76.014] - - - [768, 384, 1, 384] - - [259, 21.264] - - - [6912, 3329, 1, 384] - - [393, 81.545] - - - [8064, 512, 1, 384] - - [293, 61.443] - - - [26496, 384, 1, 384] - - [312, 75.859] - - - [24960, 4096, 1, 384] - - [51, 89.572] - - - [19584, 11905, 1, 384] - - [56, 89.153] - - - [16512, 8833, 1, 384] - - [52, 85.805] - - - [18816, 384, 1, 384] - - [290, 70.244] - - - [23808, 1024, 1, 384] - - [42, 83.004] - - - [16512, 384, 1, 384] - - [293, 66.449] - - - [8448, 4865, 1, 384] - - [35, 83.744] - - - [34944, 1024, 1, 384] - - [56, 86.21] - - - [29184, 4096, 1, 384] - - [42, 88.774] - - - [8832, 2048, 1, 384] - - [56, 77.616] - - - [9984, 1024, 1, 384] - - [268, 72.525] - - - [22272, 1024, 1, 384] - - [56, 83.249] - - - [14592, 6913, 1, 384] - - [324, 89.529] - - - [9216, 2048, 1, 384] - - [21, 78.405] - - - [7296, 1024, 1, 384] - - [358, 69.573] - - - [26880, 8192, 1, 384] - - [51, 90.358] - - - [26880, 10625, 1, 384] - - [42, 89.218] - - - [28800, 12545, 1, 384] - - [27, 88.809] - - - [18048, 1024, 1, 384] - - [34, 79.175] - - - [27264, 11009, 1, 384] - - [25, 88.429] - - - [12288, 2048, 1, 384] - - [25, 79.298] - - - [19200, 4096, 1, 384] - - [51, 88.577] - - - [32256, 384, 1, 384] - - [312, 79.89] - - - [9216, 5249, 1, 384] - - [35, 84.69] - - - [29952, 14081, 1, 384] - - [28, 89.172] - - - [7680, 384, 1, 384] - - [290, 58.932] - - - [19200, 11137, 1, 384] - - [45, 88.991] - - - [14976, 1024, 1, 384] - - [42, 78.752] - - - [25728, 1024, 1, 384] - - [56, 83.763] - - - [3456, 1921, 1, 384] - - [420, 65.583] - - - [21120, 13441, 1, 384] - - [45, 89.157] - - - [15360, 2048, 1, 384] - - [59, 82.111] - - - [34560, 512, 1, 384] - - [313, 81.375] - - - [31872, 8192, 1, 384] - - [27, 90.406] - - - [32640, 16769, 1, 384] - - [30, 87.83] - - - [26496, 1024, 1, 384] - - [51, 81.816] - - - [12672, 1024, 1, 384] - - [309, 79.745] - - - [3072, 384, 1, 384] - - [271, 46.928] - - - [31104, 4096, 1, 384] - - [51, 89.842] - - - [25344, 4096, 1, 384] - - [34, 89.495] - - - [4224, 2689, 1, 384] - - [310, 72.701] - - - [24576, 1024, 1, 384] - - [30, 78.371] - - - [8448, 512, 1, 384] - - [32, 58.027] - - - [1536, 1025, 1, 384] - - [182, 50.72] - - - [14208, 6145, 1, 384] - - [54, 85.307] - - - [27264, 384, 1, 384] - - [314, 77.856] - - - [34560, 1024, 1, 384] - - [51, 85.683] - - - [14976, 6913, 1, 384] - - [25, 87.539] - - - [21504, 2048, 1, 384] - - [59, 84.642] - - - [14208, 4096, 1, 384] - - [51, 87.288] - - - [14592, 4096, 1, 384] - - [56, 86.938] - - - [6528, 2561, 1, 384] - - [25, 77.922] - - - [34176, 18305, 1, 384] - - [49, 89.147] - - - [19968, 384, 1, 384] - - [301, 73.431] - - - [30720, 8192, 1, 384] - - [72, 87.502] - - - [14592, 512, 1, 384] - - [370, 70.688] - - - [25728, 2048, 1, 384] - - [51, 86.383] - - - [23424, 4096, 1, 384] - - [25, 87.192] - - - [27264, 2048, 1, 384] - - [51, 84.397] - - - [21504, 1024, 1, 384] - - [51, 80.566] - - - [30336, 384, 1, 384] - - [297, 78.286] - - - [2688, 1024, 1, 384] - - [289, 51.292] - - - [22656, 4096, 1, 384] - - [34, 89.102] - - - [20352, 2048, 1, 384] - - [51, 84.991] - - - [33408, 384, 1, 384] - - [314, 74.322] - - - [15360, 4096, 1, 384] - - [35, 86.111] - - - [22272, 512, 1, 384] - - [350, 78.34] - - - [14208, 384, 1, 384] - - [290, 66.866] - - - [32640, 512, 1, 384] - - [350, 79.249] - - - [23808, 512, 1, 384] - - [297, 75.838] - - - [24960, 1024, 1, 384] - - [56, 81.728] - - - [4608, 512, 1, 384] - - [392, 51.755] - - - [25344, 2048, 1, 384] - - [34, 87.652] - - - [11904, 1024, 1, 384] - - [51, 78.276] - - - [28416, 12545, 1, 384] - - [27, 89.264] - - - [14208, 6529, 1, 384] - - [23, 87.106] - - - [13824, 5761, 1, 384] - - [27, 85.86] - - - [26112, 9857, 1, 384] - - [27, 89.098] - - - [9600, 2048, 1, 384] - - [56, 83.301] - - - [33024, 1024, 1, 384] - - [59, 83.484] - - - [34944, 18689, 1, 384] - - [23, 89.205] - - - [13824, 512, 1, 384] - - [370, 67.951] - - - [26880, 384, 1, 384] - - [313, 76.793] - - - [15744, 384, 1, 384] - - [297, 71.954] - - - [29568, 8192, 1, 384] - - [51, 89.975] - - - [24960, 9089, 1, 384] - - [56, 89.011] - - - [28032, 2048, 1, 384] - - [51, 88.256] - - - [19968, 11905, 1, 384] - - [35, 89.202] - - - [6528, 2945, 1, 384] - - [350, 81.201] - - - [20352, 12289, 1, 384] - - [85, 87.032] - - - [5376, 512, 1, 384] - - [296, 51.256] - - - [5376, 3457, 1, 384] - - [309, 79.161] - - - [21504, 384, 1, 384] - - [304, 73.055] - - - [11520, 1024, 1, 384] - - [393, 74.578] - - - [3840, 1921, 1, 384] - - [421, 70.918] - - - [18432, 4096, 1, 384] - - [59, 85.855] - - - [28416, 2048, 1, 384] - - [56, 87.239] - - - [3456, 512, 1, 384] - - [355, 43.93] - - - [2688, 384, 1, 384] - - [271, 42.327] - - - [28032, 4096, 1, 384] - - [56, 89.737] - - - [16128, 384, 1, 384] - - [297, 72.906] - - - [33792, 17537, 1, 384] - - [23, 89.18] - - - [2688, 1793, 1, 384] - - [287, 64.712] - - - [27648, 1024, 1, 384] - - [51, 84.344] - - - [13440, 1024, 1, 384] - - [34, 78.549] - - - [28032, 8192, 1, 384] - - [56, 90.546] - - - [34560, 18689, 1, 384] - - [35, 89.298] - - - [16896, 512, 1, 384] - - [297, 70.67] - - - [13056, 2048, 1, 384] - - [51, 84.795] - - - [3072, 1537, 1, 384] - - [420, 64.951] - - - [3072, 512, 1, 384] - - [161, 53.669] - - - [25344, 9089, 1, 384] - - [49, 88.616] - - - [9600, 384, 1, 384] - - [294, 64.963] - - - [26880, 512, 1, 384] - - [56, 78.239] - - - [33024, 512, 1, 384] - - [291, 76.933] - - - [21888, 1024, 1, 384] - - [23, 78.92] - - - [18048, 384, 1, 384] - - [293, 73.352] - - - [16896, 4096, 1, 384] - - [34, 87.888] - - - [23808, 384, 1, 384] - - [310, 75.552] - - - [26496, 4096, 1, 384] - - [51, 89.126] - - - [20736, 13057, 1, 384] - - [35, 89.631] - - - [24576, 512, 1, 384] - - [35, 78.16] - - - [14592, 6529, 1, 384] - - [35, 87.597] - - - [6528, 512, 1, 384] - - [51, 61.445] - - - [22656, 14593, 1, 384] - - [56, 89.381] - - - [26112, 2048, 1, 384] - - [51, 87.556] - - - [25728, 9473, 1, 384] - - [23, 88.602] - - - [15744, 2048, 1, 384] - - [56, 86.199] - - - [31488, 1024, 1, 384] - - [34, 86.085] - - - [11136, 2048, 1, 384] - - [56, 83.043] - - - [4608, 2689, 1, 384] - - [425, 76.172] - - - [30720, 1024, 1, 384] - - [51, 84.187] - - - [1920, 512, 1, 384] - - [105, 40.881] - - - [25728, 8192, 1, 384] - - [56, 90.683] - - - [31104, 2048, 1, 384] - - [34, 87.584] - - - [3456, 1024, 1, 384] - - [300, 58.23] - - - [25344, 384, 1, 384] - - [297, 76.621] - - - [27264, 8192, 1, 384] - - [23, 90.393] - - - [16128, 4096, 1, 384] - - [56, 87.503] - - - [20736, 12673, 1, 384] - - [35, 89.313] - - - [4224, 2305, 1, 384] - - [418, 70.347] - - - [27648, 11777, 1, 384] - - [25, 88.558] - - - [6144, 512, 1, 384] - - [33, 58.515] - - - [24576, 2048, 1, 384] - - [27, 79.488] - - - [15360, 384, 1, 384] - - [293, 70.587] - - - [34944, 19073, 1, 384] - - [23, 89.106] - - - [33792, 384, 1, 384] - - [304, 75.202] - - - [15360, 7681, 1, 384] - - [319, 90.032] - - - [34176, 17921, 1, 384] - - [22, 88.898] - - - [10368, 1024, 1, 384] - - [51, 77.886] - - - [34176, 8192, 1, 384] - - [25, 90.199] - - - [34176, 2048, 1, 384] - - [85, 88.035] - - - [7680, 4097, 1, 384] - - [42, 82.31] - - - [10752, 1024, 1, 384] - - [299, 76.477] - - - [9984, 2048, 1, 384] - - [42, 80.888] - - - [5760, 2048, 1, 384] - - [317, 75.895] - - - [30336, 1024, 1, 384] - - [51, 83.496] - - - [23424, 384, 1, 384] - - [309, 77.634] - - - [13440, 5377, 1, 384] - - [27, 85.376] - - - [14592, 2048, 1, 384] - - [42, 84.533] - - - [31872, 4096, 1, 384] - - [51, 89.55] - - - [6528, 2048, 1, 384] - - [391, 80.928] - - - [8064, 384, 1, 384] - - [291, 61.594] - - - [31872, 16001, 1, 384] - - [27, 89.19] - - - [16896, 1024, 1, 384] - - [42, 80.371] - - - [15360, 7297, 1, 384] - - [319, 89.724] - - - [33792, 4096, 1, 384] - - [72, 87.283] - - - [16896, 384, 1, 384] - - [295, 73.144] - - - [29952, 1024, 1, 384] - - [34, 82.722] - - - [768, 512, 1, 384] - - [91, 28.582] - - - [24576, 384, 1, 384] - - [311, 77.179] - - - [9984, 384, 1, 384] - - [293, 59.035] - - - [28416, 4096, 1, 384] - - [51, 89.71] - - - [11904, 7937, 1, 384] - - [25, 87.153] - - - [22656, 512, 1, 384] - - [306, 79.459] - - - [32640, 16385, 1, 384] - - [30, 85.468] - - - [14592, 1024, 1, 384] - - [34, 77.259] - - - [29952, 13697, 1, 384] - - [35, 89.409] - - - [32640, 1024, 1, 384] - - [72, 79.399] - - - [24960, 512, 1, 384] - - [357, 77.872] - - - [24192, 384, 1, 384] - - [294, 76.515] - - - [10752, 512, 1, 384] - - [293, 65.53] - - - [25344, 8192, 1, 384] - - [34, 90.336] - - - [32256, 16385, 1, 384] - - [52, 85.964] - - - [18432, 10753, 1, 384] - - [35, 88.005] - - - [27648, 512, 1, 384] - - [27, 79.871] - - - [28800, 4096, 1, 384] - - [42, 89.288] - - - [13440, 512, 1, 384] - - [291, 67.16] - - - [22272, 2048, 1, 384] - - [51, 86.157] - - - [29184, 2048, 1, 384] - - [51, 86.734] - - - [29952, 8192, 1, 384] - - [56, 90.427] - - - [384, 385, 1, 384] - - [257, 10.485] - - - [33408, 17153, 1, 384] - - [25, 89.285] - - - [27264, 512, 1, 384] - - [51, 78.612] - - - [33792, 1024, 1, 384] - - [42, 83.815] - - - [12288, 384, 1, 384] - - [293, 66.011] - - - [4224, 1024, 1, 384] - - [267, 59.417] - - - [13056, 5377, 1, 384] - - [27, 86.897] - - - [9600, 5633, 1, 384] - - [56, 85.785] - - - [30336, 512, 1, 384] - - [25, 79.244] - - - [7680, 1024, 1, 384] - - [299, 72.521] - - - [14976, 384, 1, 384] - - [293, 69.592] - - - [11904, 512, 1, 384] - - [298, 67.693] - - - [16128, 512, 1, 384] - - [370, 70.57] - - - [16128, 8449, 1, 384] - - [27, 88.103] - - - [18432, 2048, 1, 384] - - [59, 83.507] - - - [32256, 1024, 1, 384] - - [51, 84.016] - - - [16896, 8833, 1, 384] - - [35, 89.018] - - - [11136, 7169, 1, 384] - - [60, 84.227] - - - [8832, 4865, 1, 384] - - [25, 84.203] - - - [13440, 4096, 1, 384] - - [51, 87.546] - - - [10752, 2048, 1, 384] - - [56, 80.532] - - - [27264, 1024, 1, 384] - - [25, 82.207] - - - [1536, 384, 1, 384] - - [264, 33.465] - - - [20352, 1024, 1, 384] - - [34, 81.987] - - - [30720, 512, 1, 384] - - [42, 80.004] - - - [16512, 512, 1, 384] - - [268, 68.539] - - - [20736, 4096, 1, 384] - - [51, 88.305] - - - [23424, 15745, 1, 384] - - [35, 89.763] - - - [24960, 2048, 1, 384] - - [42, 86.756] - - - [32256, 2048, 1, 384] - - [34, 87.968] - - - [10368, 384, 1, 384] - - [293, 60.869] - - - [14976, 7297, 1, 384] - - [319, 89.902] - - - [23040, 4096, 1, 384] - - [56, 88.538] - - - [16512, 4096, 1, 384] - - [74, 83.999] - - - [20736, 512, 1, 384] - - [268, 73.442] - - - [34560, 384, 1, 384] - - [312, 76.501] - - - [23040, 1024, 1, 384] - - [34, 80.485] - - - [5376, 384, 1, 384] - - [282, 52.152] - - - [11136, 512, 1, 384] - - [291, 66.86] - - - [19200, 512, 1, 384] - - [424, 76.228] - - - [19584, 11521, 1, 384] - - [25, 89.051] - - - [21504, 4096, 1, 384] - - [72, 86.954] - - - [25728, 4096, 1, 384] - - [51, 89.093] - - - [4992, 512, 1, 384] - - [291, 49.917] - - - [26880, 4096, 1, 384] - - [51, 88.944] - - - [31488, 15233, 1, 384] - - [22, 89.348] - - - [2304, 1409, 1, 384] - - [419, 55.733] - - - [28800, 1024, 1, 384] - - [34, 83.054] - - - [25344, 9473, 1, 384] - - [23, 89.006] - - - [13824, 4096, 1, 384] - - [56, 87.375] - - - [18048, 2048, 1, 384] - - [51, 85.495] - - - [13056, 512, 1, 384] - - [391, 73.457] - - - [31104, 8192, 1, 384] - - [25, 90.454] - - - [1152, 641, 1, 384] - - [96, 39.307] - - - [8064, 1024, 1, 384] - - [15, 71.094] - - - [7296, 512, 1, 384] - - [290, 59.112] - - - [12672, 4609, 1, 384] - - [23, 84.709] - - - [27264, 4096, 1, 384] - - [25, 87.94] - - - [11520, 2048, 1, 384] - - [51, 80.803] - - - [15744, 4096, 1, 384] - - [42, 88.574] - - - [19968, 512, 1, 384] - - [290, 72.939] - - - [5760, 2177, 1, 384] - - [422, 76.181] - - - [3840, 384, 1, 384] - - [277, 51.55] - - - [30336, 8192, 1, 384] - - [56, 90.763] - - - [28416, 8192, 1, 384] - - [51, 90.413] - - - [25344, 512, 1, 384] - - [423, 79.161] - - - [7296, 3713, 1, 384] - - [25, 80.855] - - - [28416, 384, 1, 384] - - [297, 75.08] - - - [19584, 2048, 1, 384] - - [51, 85.165] - - - [10368, 2048, 1, 384] - - [42, 83.477] - - - [33024, 4096, 1, 384] - - [51, 88.724] - - - [4224, 512, 1, 384] - - [291, 49.035] - - - [26496, 8192, 1, 384] - - [34, 90.525] - - - [768, 385, 1, 384] - - [256, 20.97] - - - [23040, 384, 1, 384] - - [308, 76.898] - - - [11520, 7937, 1, 384] - - [23, 87.554] - - - [28800, 384, 1, 384] - - [297, 75.824] - - - [8064, 4481, 1, 384] - - [25, 82.664] - - - [28032, 384, 1, 384] - - [313, 78.83] - - - [31104, 512, 1, 384] - - [350, 79.659] - - - [23808, 16129, 1, 384] - - [34, 90.108] - - - [29184, 384, 1, 384] - - [297, 76.231] - - - [9600, 512, 1, 384] - - [423, 66.46] - - - [26112, 512, 1, 384] - - [23, 76.532] - - - [31488, 8192, 1, 384] - - [27, 90.292] - - - [8448, 384, 1, 384] - - [293, 63.938] - - - [34944, 8192, 1, 384] - - [35, 90.6] - - - [4608, 3073, 1, 384] - - [350, 78.024] - - - [30720, 2048, 1, 384] - - [72, 85.262] - - - [34944, 512, 1, 384] - - [313, 82.104] - - - [27648, 8192, 1, 384] - - [72, 88.499] - - - [33024, 2048, 1, 384] - - [51, 87.021] - - - [26112, 4096, 1, 384] - - [34, 88.996] - - - [17280, 384, 1, 384] - - [294, 74.322] - - - [33024, 17153, 1, 384] - - [30, 88.628] - - - [14208, 2048, 1, 384] - - [51, 82.778] - - - [13440, 2048, 1, 384] - - [51, 82.8] - - - [1536, 641, 1, 384] - - [96, 39.611] - - - [8064, 4097, 1, 384] - - [54, 81.905] - - - [26496, 10625, 1, 384] - - [23, 88.98] - - - [33024, 384, 1, 384] - - [312, 73.665] - - - [26112, 384, 1, 384] - - [306, 75.246] - - - [23424, 15361, 1, 384] - - [54, 85.936] - - - [34944, 2048, 1, 384] - - [56, 88.407] - - - [32256, 512, 1, 384] - - [425, 81.851] - - - [23808, 15745, 1, 384] - - [51, 90.117] - - - [5760, 512, 1, 384] - - [290, 55.295] - - - [16128, 1024, 1, 384] - - [299, 80.882] - - - [31488, 4096, 1, 384] - - [56, 89.699] - - - [29568, 13313, 1, 384] - - [52, 86.436] - - - [18816, 11137, 1, 384] - - [23, 89.01] - - - [26496, 2048, 1, 384] - - [51, 86.456] - - - [1920, 384, 1, 384] - - [256, 37.415] - - - [31872, 1024, 1, 384] - - [34, 83.409] - - - [12672, 512, 1, 384] - - [391, 70.337] - - - [13056, 4096, 1, 384] - - [42, 87.556] - - - [17280, 1024, 1, 384] - - [42, 82.053] - - - [12288, 1024, 1, 384] - - [315, 76.728] - - - [1152, 512, 1, 384] - - [90, 32.6] - - - [31104, 15233, 1, 384] - - [23, 89.393] - - - [4608, 384, 1, 384] - - [345, 47.214] - - - [21888, 512, 1, 384] - - [425, 74.447] - - - [33408, 1024, 1, 384] - - [34, 86.352] - - - [8448, 2048, 1, 384] - - [56, 80.242] - - - [7296, 3329, 1, 384] - - [302, 81.951] - - - [10368, 6785, 1, 384] - - [23, 86.949] - - - [8832, 1024, 1, 384] - - [25, 67.434] - - - [31104, 1024, 1, 384] - - [51, 85.253] - - - [11520, 7553, 1, 384] - - [27, 86.704] - - - [34176, 4096, 1, 384] - - [56, 89.32] - - - [20352, 512, 1, 384] - - [293, 73.927] - - - [18432, 512, 1, 384] - - [350, 74.298] - - - [31488, 2048, 1, 384] - - [56, 88.404] - - - [9984, 6401, 1, 384] - - [25, 85.589] - - - [6144, 2048, 1, 384] - - [42, 78.784] - - - [22656, 2048, 1, 384] - - [51, 87.419] - - - [2304, 512, 1, 384] - - [90, 47.349] - - - [21504, 13441, 1, 384] - - [23, 88.86] - - - [1920, 1025, 1, 384] - - [294, 44.483] - - - [24960, 8705, 1, 384] - - [51, 88.641] - - - [16512, 2048, 1, 384] - - [59, 80.148] - - - [26880, 11009, 1, 384] - - [23, 89.218] - - - [32256, 4096, 1, 384] - - [24, 88.284] - - - [14976, 2048, 1, 384] - - [51, 82.793] - - - [21120, 512, 1, 384] - - [290, 75.598] - - - [31872, 512, 1, 384] - - [424, 80.882] - - - [8064, 2048, 1, 384] - - [27, 76.349] - - - [3072, 1024, 1, 384] - - [25, 58.286] - - - [23808, 2048, 1, 384] - - [42, 85.843] - - - [12672, 2048, 1, 384] - - [56, 83.079] - - - [19968, 4096, 1, 384] - - [56, 88.078] - - - [14976, 512, 1, 384] - - [422, 70.739] - - - [25344, 1024, 1, 384] - - [34, 82.986] - - - [31872, 15617, 1, 384] - - [35, 89.141] - - - [20352, 12673, 1, 384] - - [23, 89.358] - - - [11136, 384, 1, 384] - - [293, 68.837] - - - [32640, 8192, 1, 384] - - [30, 88.408] - - - [28800, 2048, 1, 384] - - [42, 87.385] - - - [22656, 1024, 1, 384] - - [424, 85.161] - - - [17280, 4096, 1, 384] - - [51, 87.951] - - - [17664, 4096, 1, 384] - - [56, 87.84] - - - [32640, 2048, 1, 384] - - [26, 83.394] - - - [28032, 11777, 1, 384] - - [56, 89.285] - - - [20352, 4096, 1, 384] - - [56, 88.403] - - - [33792, 512, 1, 384] - - [309, 79.419] - - - [24192, 4096, 1, 384] - - [42, 89.825] - - - [9216, 384, 1, 384] - - [293, 67.67] - - - [6912, 512, 1, 384] - - [358, 57.396] - - - [14208, 1024, 1, 384] - - [350, 80.04] - - - [26496, 512, 1, 384] - - [51, 77.278] - - - [4992, 384, 1, 384] - - [264, 57.491] - - - [33408, 512, 1, 384] - - [35, 79.233] - - - [3456, 1537, 1, 384] - - [297, 63.22] - - - [21888, 14209, 1, 384] - - [39, 84.605] - - - [24576, 8321, 1, 384] - - [23, 83.293] - - - [33792, 17921, 1, 384] - - [23, 88.783] - - - [13440, 384, 1, 384] - - [295, 70.825] - - - [18432, 384, 1, 384] - - [291, 69.384] - - - [6912, 1024, 1, 384] - - [393, 69.095] - - - [22272, 384, 1, 384] - - [306, 75.356] - - - [3840, 2305, 1, 384] - - [422, 69.872] - - - [6144, 1024, 1, 384] - - [306, 69.589] - - - [7680, 512, 1, 384] - - [418, 60.421] - - - [19584, 4096, 1, 384] - - [42, 88.664] - - - [23808, 4096, 1, 384] - - [42, 88.714] - - - [29568, 384, 1, 384] - - [297, 76.901] - - - [29184, 512, 1, 384] - - [424, 81.503] - - - [13056, 384, 1, 384] - - [296, 69.463] - - - [28032, 512, 1, 384] - - [313, 79.254] - - - [26880, 2048, 1, 384] - - [34, 87.507] - - - [18048, 9985, 1, 384] - - [25, 88.583] - - - [29952, 512, 1, 384] - - [35, 78.472] - - - [27648, 2048, 1, 384] - - [72, 85.281] - - - [29568, 13697, 1, 384] - - [28, 88.429] - - - [19584, 1024, 1, 384] - - [314, 84.384] - - - [27648, 384, 1, 384] - - [315, 78.481] - - - [6912, 384, 1, 384] - - [386, 50.992] - - - [26880, 1024, 1, 384] - - [51, 82.929] - - - [24960, 8192, 1, 384] - - [51, 90.373] - - - [13824, 1024, 1, 384] - - [51, 80.611] - - - [11904, 2048, 1, 384] - - [56, 83.021] - - - [34560, 2048, 1, 384] - - [51, 87.843] - - - [12288, 4609, 1, 384] - - [25, 83.753] - - - [21504, 13825, 1, 384] - - [27, 88.505] - - - [29184, 8192, 1, 384] - - [27, 90.304] - - - [12288, 4096, 1, 384] - - [23, 85.445] - - - [23424, 1024, 1, 384] - - [51, 81.746] - - - [14208, 512, 1, 384] - - [291, 69.863] - - - [25728, 512, 1, 384] - - [425, 79.754] - - - [29568, 2048, 1, 384] - - [85, 86.694] - - - [9600, 1024, 1, 384] - - [322, 77.018] - - - [29952, 384, 1, 384] - - [297, 77.657] - - - [18048, 4096, 1, 384] - - [51, 87.568] - - - [30336, 14081, 1, 384] - - [27, 89.07] - - - [24192, 8192, 1, 384] - - [51, 90.674] - - - [33792, 2048, 1, 384] - - [59, 85.928] - - - [6144, 384, 1, 384] - - [285, 57.994] - - - [8448, 1024, 1, 384] - - [358, 70.301] - - - [6528, 1024, 1, 384] - - [296, 71.853] - - - [18432, 10369, 1, 384] - - [27, 88.481] - - - [19968, 1024, 1, 384] - - [51, 80.777] - - - [23424, 512, 1, 384] - - [291, 74.893] - - - [20736, 2048, 1, 384] - - [56, 86.572] - - - [29184, 12929, 1, 384] - - [49, 89.169] - - - [3072, 1153, 1, 384] - - [418, 55.652] - - - [28416, 512, 1, 384] - - [325, 79.772] - - - [14592, 384, 1, 384] - - [290, 67.666] - - - [18432, 1024, 1, 384] - - [34, 80.413] - - - [29184, 13313, 1, 384] - - [52, 86.922] - - - [32640, 4096, 1, 384] - - [84, 86.09] - - - [21888, 13825, 1, 384] - - [57, 85.031] - - - [5376, 1024, 1, 384] - - [370, 63.787] - - - [4608, 1024, 1, 384] - - [305, 65.405] - - - [8832, 5249, 1, 384] - - [25, 84.238] - - - [14976, 4096, 1, 384] - - [56, 86.86] - - - [3840, 1024, 1, 384] - - [290, 60.196] - - - [24192, 16129, 1, 384] - - [56, 89.733] - - - [19968, 12289, 1, 384] - - [42, 86.208] - - - [1152, 384, 1, 384] - - [256, 30.499] - - - [27648, 4096, 1, 384] - - [59, 87.529] - - - [4992, 3073, 1, 384] - - [35, 77.085] - - - [33024, 8192, 1, 384] - - [60, 89.435] - - - [34944, 384, 1, 384] - - [319, 76.763] - - - [32, 28672, 1, 32] - - [6, 23.634] - - - [32, 24576, 1, 32] - - [5, 23.263] - - - [32, 16384, 1, 32] - - [2, 22.424] - - - [32, 20480, 1, 32] - - [4, 22.915] - - - [32, 12288, 1, 32] - - [2, 20.484] - - - [32, 8192, 1, 32] - - [3, 16.315] - - - [32, 4096, 1, 32] - - [1, 8.373] - - - [32, 32768, 1, 32] - - [0, 23.586] - - - [4224, 3840, 1, 4096] - - [13, 98.172] - - - [5376, 4096, 1, 4096] - - [14, 94.215] - - - [7040, 4096, 1, 384] - - [15, 86.756] - - - [7040, 4096, 1, 768] - - [13, 93.117] - - - [7040, 4096, 1, 1536] - - [16, 96.418] - - - [3840, 4224, 1, 4096] - - [7, 79.968] - - - [3840, 4224, 1, 4224] - - [8, 79.975] - - - [3840, 4224, 1, 4320] - - [9, 79.983] - - - [7680, 8448, 1, 8192] - - [10, 81.886] - - - [7680, 8448, 1, 8448] - - [10, 81.886] - - - [7680, 8448, 1, 8640] - - [10, 81.886] - - - [4096, 7169, 1, 512] - - [459, 88.167] - - - [4096, 7681, 1, 512] - - [326, 88.841] - - - [4096, 8193, 1, 512] - - [319, 87.86] - - - [4608, 512, 1, 512] - - [277, 61.028] - - - [4608, 8193, 1, 512] - - [453, 89.573] - - - [4608, 8705, 1, 512] - - [356, 89.242] - - - [4608, 9217, 1, 512] - - [454, 89.741] - - - [5120, 512, 1, 512] - - [278, 66.032] - - - [5120, 9217, 1, 512] - - [454, 90.059] - - - [5120, 9729, 1, 512] - - [319, 90.936] - - - [5120, 10241, 1, 512] - - [455, 90.535] - - - [5632, 512, 1, 512] - - [357, 60.928] - - - [5632, 10241, 1, 512] - - [454, 91.2] - - - [5632, 10753, 1, 512] - - [341, 91.858] - - - [5632, 11265, 1, 512] - - [458, 90.934] - - - [6144, 512, 1, 512] - - [290, 64.921] - - - [6144, 11265, 1, 512] - - [454, 91.689] - - - [6144, 11777, 1, 512] - - [326, 91.522] - - - [6144, 12289, 1, 512] - - [454, 91.486] - - - [6656, 512, 1, 512] - - [312, 68.733] - - - [6656, 12289, 1, 512] - - [454, 91.715] - - - [6656, 12801, 1, 512] - - [326, 92.349] - - - [6656, 13313, 1, 512] - - [454, 92.002] - - - [7168, 512, 1, 512] - - [358, 62.865] - - - [7168, 13313, 1, 512] - - [452, 92.196] - - - [7168, 13825, 1, 512] - - [457, 92.402] - - - [7168, 14337, 1, 512] - - [454, 92.768] - - - [7680, 512, 1, 512] - - [290, 65.861] - - - [7680, 14337, 1, 512] - - [454, 92.373] - - - [7680, 14849, 1, 512] - - [356, 92.877] - - - [7680, 15361, 1, 512] - - [455, 92.218] - - - [8192, 512, 1, 512] - - [293, 69.275] - - - [8192, 15361, 1, 512] - - [454, 92.959] - - - [8192, 15873, 1, 512] - - [332, 93.06] - - - [8192, 16385, 1, 512] - - [454, 93.033] - - - [8704, 512, 1, 512] - - [293, 71.705] - - - [8704, 16385, 1, 512] - - [454, 93.107] - - - [8704, 16897, 1, 512] - - [326, 93.129] - - - [8704, 17409, 1, 512] - - [454, 92.785] - - - [9216, 512, 1, 512] - - [458, 68.831] - - - [9216, 17409, 1, 512] - - [454, 93.08] - - - [9216, 17921, 1, 512] - - [326, 93.416] - - - [9216, 18433, 1, 512] - - [454, 92.976] - - - [9728, 512, 1, 512] - - [451, 72.309] - - - [9728, 18433, 1, 512] - - [453, 93.313] - - - [9728, 18945, 1, 512] - - [356, 93.384] - - - [9728, 19457, 1, 512] - - [451, 93.225] - - - [10240, 512, 1, 512] - - [297, 68.291] - - - [10240, 19457, 1, 512] - - [454, 93.374] - - - [10240, 19969, 1, 512] - - [361, 93.552] - - - [10240, 20481, 1, 512] - - [455, 93.181] - - - [10752, 512, 1, 512] - - [291, 69.447] - - - [10752, 20481, 1, 512] - - [455, 93.381] - - - [10752, 20993, 1, 512] - - [361, 93.522] - - - [10752, 21505, 1, 512] - - [451, 93.239] - - - [11264, 512, 1, 512] - - [293, 71.35] - - - [11264, 21505, 1, 512] - - [454, 93.509] - - - [11264, 22017, 1, 512] - - [356, 93.65] - - - [11264, 22529, 1, 512] - - [454, 93.548] - - - [11776, 512, 1, 512] - - [293, 73.606] - - - [11776, 22529, 1, 512] - - [458, 93.551] - - - [11776, 23041, 1, 512] - - [326, 93.761] - - - [11776, 23553, 1, 512] - - [454, 93.625] - - - [12288, 512, 1, 512] - - [297, 75.45] - - - [12288, 23553, 1, 512] - - [453, 93.541] - - - [12288, 24065, 1, 512] - - [361, 93.93] - - - [12288, 24577, 1, 512] - - [450, 93.635] - - - [12800, 512, 1, 512] - - [325, 75.684] - - - [12800, 24577, 1, 512] - - [454, 93.576] - - - [12800, 25089, 1, 512] - - [332, 93.881] - - - [12800, 25601, 1, 512] - - [455, 93.708] - - - [13312, 512, 1, 512] - - [363, 75.137] - - - [13312, 25601, 1, 512] - - [454, 93.705] - - - [13312, 26113, 1, 512] - - [361, 93.955] - - - [13312, 26625, 1, 512] - - [454, 93.722] - - - [13824, 512, 1, 512] - - [362, 73.227] - - - [13824, 26625, 1, 512] - - [454, 93.719] - - - [13824, 27137, 1, 512] - - [361, 94.152] - - - [13824, 27649, 1, 512] - - [453, 93.845] - - - [14336, 512, 1, 512] - - [297, 72.843] - - - [14336, 27649, 1, 512] - - [454, 93.819] - - - [14336, 28161, 1, 512] - - [319, 94.127] - - - [14336, 28673, 1, 512] - - [451, 93.934] - - - [14848, 512, 1, 512] - - [297, 74.864] - - - [14848, 28673, 1, 512] - - [451, 93.904] - - - [14848, 29185, 1, 512] - - [365, 94.19] - - - [14848, 29697, 1, 512] - - [451, 93.917] - - - [15360, 512, 1, 512] - - [297, 76.564] - - - [15360, 29697, 1, 512] - - [454, 93.898] - - - [15360, 30209, 1, 512] - - [365, 94.311] - - - [15360, 30721, 1, 512] - - [454, 93.948] - - - [15872, 512, 1, 512] - - [456, 78.265] - - - [15872, 30721, 1, 512] - - [455, 93.935] - - - [15872, 31233, 1, 512] - - [326, 94.373] - - - [15872, 31745, 1, 512] - - [454, 94.005] - - - [16384, 512, 1, 512] - - [454, 80.377] - - - [16384, 31745, 1, 512] - - [454, 93.967] - - - [16384, 32257, 1, 512] - - [361, 94.453] - - - [16384, 32769, 1, 512] - - [453, 93.942] - - - [16896, 512, 1, 512] - - [359, 78.688] - - - [16896, 32769, 1, 512] - - [451, 93.994] - - - [16896, 33281, 1, 512] - - [332, 94.381] - - - [16896, 33793, 1, 512] - - [450, 94.043] - - - [17408, 512, 1, 512] - - [318, 80.426] - - - [17408, 33793, 1, 512] - - [454, 94.061] - - - [17408, 34305, 1, 512] - - [361, 94.564] - - - [17408, 34817, 1, 512] - - [453, 94.06] - - - [17920, 512, 1, 512] - - [367, 77.178] - - - [17920, 34817, 1, 512] - - [454, 94.089] - - - [17920, 35329, 1, 512] - - [365, 94.567] - - - [17920, 35841, 1, 512] - - [451, 94.147] - - - [18432, 512, 1, 512] - - [457, 78.257] - - - [18432, 35841, 1, 512] - - [451, 94.155] - - - [18432, 36353, 1, 512] - - [361, 94.591] - - - [18432, 36865, 1, 512] - - [453, 94.111] - - - [18944, 512, 1, 512] - - [457, 80.276] - - - [18944, 36865, 1, 512] - - [453, 94.165] - - - [18944, 37377, 1, 512] - - [326, 94.622] - - - [18944, 37889, 1, 512] - - [454, 94.122] - - - [19456, 512, 1, 512] - - [455, 81.688] - - - [19456, 37889, 1, 512] - - [453, 94.133] - - - [19456, 38401, 1, 512] - - [326, 94.675] - - - [19456, 38913, 1, 512] - - [454, 94.15] - - - [19968, 512, 1, 512] - - [319, 79.452] - - - [19968, 38913, 1, 512] - - [454, 94.163] - - - [19968, 39425, 1, 512] - - [319, 94.652] - - - [19968, 39937, 1, 512] - - [454, 94.171] - - - [20480, 512, 1, 512] - - [368, 81.161] - - - [20480, 39937, 1, 512] - - [319, 94.182] - - - [20480, 40449, 1, 512] - - [319, 94.748] - - - [20480, 40961, 1, 512] - - [453, 94.182] - - - [20992, 512, 1, 512] - - [369, 82.746] - - - [20992, 40961, 1, 512] - - [453, 94.179] - - - [20992, 41473, 1, 512] - - [326, 94.748] - - - [20992, 41985, 1, 512] - - [451, 94.211] - - - [21504, 512, 1, 512] - - [450, 79.684] - - - [21504, 41985, 1, 512] - - [451, 94.229] - - - [21504, 42497, 1, 512] - - [319, 94.802] - - - [21504, 43009, 1, 512] - - [451, 94.283] - - - [22016, 512, 1, 512] - - [451, 81.34] - - - [22016, 43009, 1, 512] - - [451, 94.27] - - - [22016, 43521, 1, 512] - - [319, 94.793] - - - [22016, 44033, 1, 512] - - [451, 94.257] - - - [22528, 512, 1, 512] - - [450, 82.888] - - - [22528, 44033, 1, 512] - - [451, 94.263] - - - [22528, 44545, 1, 512] - - [361, 94.762] - - - [22528, 45057, 1, 512] - - [453, 94.261] - - - [23040, 512, 1, 512] - - [367, 79.196] - - - [23040, 45057, 1, 512] - - [458, 94.257] - - - [23040, 45569, 1, 512] - - [319, 94.882] - - - [23040, 46081, 1, 512] - - [453, 94.308] - - - [23552, 512, 1, 512] - - [332, 81.569] - - - [23552, 46081, 1, 512] - - [453, 94.276] - - - [23552, 46593, 1, 512] - - [319, 94.852] - - - [23552, 47105, 1, 512] - - [454, 94.291] - - - [24064, 512, 1, 512] - - [323, 83.075] - - - [24064, 47105, 1, 512] - - [454, 94.28] - - - [24064, 47617, 1, 512] - - [319, 94.911] - - - [24064, 48129, 1, 512] - - [454, 94.29] - - - [24576, 512, 1, 512] - - [323, 84.031] - - - [24576, 48129, 1, 512] - - [454, 94.283] - - - [24576, 48641, 1, 512] - - [319, 94.939] - - - [24576, 49153, 1, 512] - - [451, 94.311] - - - [25088, 512, 1, 512] - - [453, 82.17] - - - [25088, 49153, 1, 512] - - [453, 94.313] - - - [25088, 49665, 1, 512] - - [365, 94.917] - - - [25088, 50177, 1, 512] - - [451, 94.353] - - - [25600, 512, 1, 512] - - [455, 83.475] - - - [25600, 50177, 1, 512] - - [451, 94.346] - - - [25600, 50689, 1, 512] - - [365, 94.933] - - - [25600, 51201, 1, 512] - - [453, 94.34] - - - [26112, 512, 1, 512] - - [457, 84.967] - - - [26112, 51201, 1, 512] - - [454, 94.34] - - - [26112, 51713, 1, 512] - - [319, 94.958] - - - [26112, 52225, 1, 512] - - [451, 94.341] - - - [26624, 512, 1, 512] - - [367, 82.882] - - - [26624, 52225, 1, 512] - - [451, 94.342] - - - [26624, 52737, 1, 512] - - [319, 94.955] - - - [26624, 53249, 1, 512] - - [453, 94.363] - - - [27136, 512, 1, 512] - - [366, 83.179] - - - [27136, 53249, 1, 512] - - [453, 94.349] - - - [27136, 53761, 1, 512] - - [319, 94.96] - - - [27136, 54273, 1, 512] - - [453, 94.377] - - - [27648, 512, 1, 512] - - [323, 84.333] - - - [27648, 54273, 1, 512] - - [451, 94.371] - - - [27648, 54785, 1, 512] - - [319, 94.942] - - - [27648, 55297, 1, 512] - - [453, 94.393] - - - [28160, 512, 1, 512] - - [366, 85.343] - - - [28160, 55297, 1, 512] - - [453, 94.381] - - - [28160, 55809, 1, 512] - - [356, 94.674] - - - [28160, 56321, 1, 512] - - [451, 94.39] - - - [28672, 512, 1, 512] - - [455, 84.296] - - - [28672, 56321, 1, 512] - - [453, 94.375] - - - [28672, 56833, 1, 512] - - [319, 94.976] - - - [28672, 57345, 1, 512] - - [451, 94.387] - - - [29184, 512, 1, 512] - - [458, 85.42] - - - [29184, 57345, 1, 512] - - [451, 94.393] - - - [29184, 57857, 1, 512] - - [319, 94.988] - - - [29184, 58369, 1, 512] - - [453, 94.388] - - - [29696, 512, 1, 512] - - [326, 82.196] - - - [29696, 58369, 1, 512] - - [453, 94.395] - - - [29696, 58881, 1, 512] - - [319, 94.976] - - - [29696, 59393, 1, 512] - - [451, 94.406] - - - [30208, 512, 1, 512] - - [332, 83.301] - - - [30208, 59393, 1, 512] - - [451, 94.405] - - - [30208, 59905, 1, 512] - - [365, 94.955] - - - [30208, 60417, 1, 512] - - [451, 94.426] - - - [30720, 512, 1, 512] - - [326, 84.261] - - - [30720, 60417, 1, 512] - - [454, 94.409] - - - [30720, 60929, 1, 512] - - [319, 94.983] - - - [30720, 61441, 1, 512] - - [451, 94.417] - - - [31232, 512, 1, 512] - - [319, 85.605] - - - [31232, 61441, 1, 512] - - [453, 94.414] - - - [31232, 61953, 1, 512] - - [365, 94.923] - - - [31232, 62465, 1, 512] - - [451, 94.426] - - - [31744, 512, 1, 512] - - [457, 84.935] - - - [31744, 62465, 1, 512] - - [451, 94.415] - - - [31744, 62977, 1, 512] - - [319, 95.029] - - - [31744, 63489, 1, 512] - - [453, 94.416] - - - [32256, 512, 1, 512] - - [457, 85.881] - - - [32256, 63489, 1, 512] - - [451, 94.433] - - - [32256, 64001, 1, 512] - - [319, 95.023] - - - [32256, 64513, 1, 512] - - [453, 94.441] - - - [32768, 512, 1, 512] - - [458, 86.874] - - - [32768, 64513, 1, 512] - - [451, 94.447] - - - [32768, 65025, 1, 512] - - [365, 95.037] - - - [32768, 65537, 1, 512] - - [451, 94.427] - - - [33280, 512, 1, 512] - - [363, 83.067] - - - [33280, 65537, 1, 512] - - [451, 94.432] - - - [33280, 66049, 1, 512] - - [319, 94.996] - - - [33280, 66561, 1, 512] - - [451, 94.445] - - - [33792, 512, 1, 512] - - [331, 84.541] - - - [33792, 66561, 1, 512] - - [451, 94.465] - - - [33792, 67073, 1, 512] - - [332, 94.912] - - - [33792, 67585, 1, 512] - - [453, 94.45] - - - [34304, 512, 1, 512] - - [341, 85.298] - - - [34304, 67585, 1, 512] - - [453, 94.462] - - - [34304, 68097, 1, 512] - - [319, 94.973] - - - [34304, 68609, 1, 512] - - [451, 94.459] - - - [34816, 512, 1, 512] - - [341, 86.299] - - - [34816, 68609, 1, 512] - - [451, 94.47] - - - [34816, 69121, 1, 512] - - [326, 94.949] - - - [34816, 69633, 1, 512] - - [453, 94.467] - - - [35328, 512, 1, 512] - - [459, 86.267] - - - [35328, 69633, 1, 512] - - [453, 94.476] - - - [35328, 70145, 1, 512] - - [326, 94.951] - - - [35328, 70657, 1, 512] - - [453, 94.461] - - - [35840, 512, 1, 512] - - [458, 87.091] - - - [35840, 70657, 1, 512] - - [451, 94.486] - - - [35840, 71169, 1, 512] - - [326, 94.948] - - - [35840, 71681, 1, 512] - - [451, 94.501] - - - [36352, 512, 1, 512] - - [331, 83.258] - - - [36352, 71681, 1, 512] - - [451, 94.487] - - - [36352, 72193, 1, 512] - - [326, 94.927] - - - [36352, 72705, 1, 512] - - [451, 94.494] - - - [36864, 512, 1, 512] - - [331, 84.381] - - - [36864, 72705, 1, 512] - - [451, 94.484] - - - [36864, 73217, 1, 512] - - [332, 94.905] - - - [36864, 73729, 1, 512] - - [453, 94.489] - - - [37376, 512, 1, 512] - - [341, 85.141] - - - [37376, 73729, 1, 512] - - [453, 94.499] - - - [37376, 74241, 1, 512] - - [332, 94.923] - - - [37376, 74753, 1, 512] - - [453, 94.476] - - - [37888, 512, 1, 512] - - [361, 85.959] - - - [37888, 74753, 1, 512] - - [453, 94.492] - - - [37888, 75265, 1, 512] - - [326, 94.91] - - - [37888, 75777, 1, 512] - - [451, 94.498] - - - [38400, 512, 1, 512] - - [341, 87.188] - - - [38400, 75777, 1, 512] - - [451, 94.496] - - - [38400, 76289, 1, 512] - - [326, 94.918] - - - [38400, 76801, 1, 512] - - [373, 93.561] - - - [38912, 512, 1, 512] - - [456, 87.416] - - - [38912, 76801, 1, 512] - - [319, 93.586] - - - [38912, 77313, 1, 512] - - [326, 94.872] - - - [38912, 77825, 1, 512] - - [373, 93.387] - - - [39424, 512, 1, 512] - - [454, 83.464] - - - [39424, 77825, 1, 512] - - [373, 93.325] - - - [39424, 78337, 1, 512] - - [332, 94.835] - - - [39424, 78849, 1, 512] - - [373, 93.112] - - - [39936, 512, 1, 512] - - [458, 84.52] - - - [39936, 78849, 1, 512] - - [373, 93.438] - - - [39936, 79361, 1, 512] - - [332, 94.821] - - - [39936, 79873, 1, 512] - - [319, 93.36] - - - [40448, 512, 1, 512] - - [452, 85.041] - - - [40448, 79873, 1, 512] - - [319, 93.357] - - - [40448, 80385, 1, 512] - - [326, 94.832] - - - [40448, 80897, 1, 512] - - [373, 93.21] - - - [40960, 512, 1, 512] - - [451, 85.901] - - - [40960, 80897, 1, 512] - - [373, 93.29] - - - [40960, 81409, 1, 512] - - [332, 94.786] - - - [40960, 81921, 1, 512] - - [364, 91.125] - - - [41472, 512, 1, 512] - - [456, 86.786] - - - [41472, 81921, 1, 512] - - [364, 91.266] - - - [41472, 82433, 1, 512] - - [332, 94.779] - - - [41472, 82945, 1, 512] - - [373, 93.081] - - - [41984, 512, 1, 512] - - [458, 87.652] - - - [41984, 82945, 1, 512] - - [319, 92.87] - - - [41984, 83457, 1, 512] - - [375, 94.639] - - - [41984, 83969, 1, 512] - - [373, 92.824] - - - [42496, 512, 1, 512] - - [452, 88.572] - - - [42496, 83969, 1, 512] - - [319, 92.525] - - - [42496, 84481, 1, 512] - - [375, 94.621] - - - [42496, 84993, 1, 512] - - [373, 92.987] - - - [43008, 512, 1, 512] - - [454, 84.869] - - - [43008, 84993, 1, 512] - - [319, 92.72] - - - [43008, 85505, 1, 512] - - [375, 94.488] - - - [43008, 86017, 1, 512] - - [373, 92.774] - - - [43520, 512, 1, 512] - - [454, 85.38] - - - [43520, 86017, 1, 512] - - [373, 92.82] - - - [43520, 86529, 1, 512] - - [375, 94.487] - - - [43520, 87041, 1, 512] - - [319, 92.132] - - - [44032, 512, 1, 512] - - [455, 86.087] - - - [44032, 87041, 1, 512] - - [319, 92.595] - - - [44032, 87553, 1, 512] - - [375, 94.391] - - - [44032, 88065, 1, 512] - - [373, 92.292] - - - [44544, 512, 1, 512] - - [455, 86.966] - - - [44544, 88065, 1, 512] - - [373, 92.202] - - - [44544, 88577, 1, 512] - - [356, 94.274] - - - [44544, 89089, 1, 512] - - [356, 91.928] - - - [45056, 512, 1, 512] - - [456, 87.869] - - - [45056, 89089, 1, 512] - - [319, 92.193] - - - [45056, 89601, 1, 512] - - [356, 94.136] - - - [45056, 90113, 1, 512] - - [364, 91.652] - - - [45568, 512, 1, 512] - - [458, 88.702] - - - [45568, 90113, 1, 512] - - [364, 91.682] - - - [45568, 90625, 1, 512] - - [356, 94.145] - - - [45568, 91137, 1, 512] - - [356, 91.84] - - - [46080, 512, 1, 512] - - [451, 85.174] - - - [46080, 91137, 1, 512] - - [356, 91.935] - - - [46080, 91649, 1, 512] - - [356, 94.076] - - - [46080, 92161, 1, 512] - - [373, 91.813] - - - [46592, 512, 1, 512] - - [454, 85.815] - - - [46592, 92161, 1, 512] - - [319, 91.898] - - - [46592, 92673, 1, 512] - - [356, 93.999] - - - [46592, 93185, 1, 512] - - [364, 91.814] - - - [47104, 512, 1, 512] - - [450, 86.516] - - - [47104, 93185, 1, 512] - - [356, 91.757] - - - [47104, 93697, 1, 512] - - [356, 93.977] - - - [47104, 94209, 1, 512] - - [373, 91.808] - - - [47616, 512, 1, 512] - - [459, 87.277] - - - [47616, 94209, 1, 512] - - [364, 91.774] - - - [47616, 94721, 1, 512] - - [369, 93.91] - - - [47616, 95233, 1, 512] - - [356, 91.74] - - - [48128, 512, 1, 512] - - [454, 87.931] - - - [48128, 95233, 1, 512] - - [364, 91.733] - - - [48128, 95745, 1, 512] - - [356, 93.866] - - - [48128, 96257, 1, 512] - - [356, 91.803] - - - [48640, 512, 1, 512] - - [450, 88.711] - - - [48640, 96257, 1, 512] - - [356, 91.726] - - - [48640, 96769, 1, 512] - - [356, 93.801] - - - [48640, 97281, 1, 512] - - [339, 91.757] - - - [49152, 512, 1, 512] - - [341, 87.527] - - - [49152, 97281, 1, 512] - - [339, 91.678] - - - [49152, 97793, 1, 512] - - [369, 93.737] - - - [49152, 98305, 1, 512] - - [376, 90.818] - - - [49664, 512, 1, 512] - - [457, 86.176] - - - [49664, 98305, 1, 512] - - [376, 90.781] - - - [49664, 98817, 1, 512] - - [369, 93.678] - - - [49664, 99329, 1, 512] - - [356, 91.658] - - - [50176, 512, 1, 512] - - [457, 86.704] - - - [50176, 99329, 1, 512] - - [356, 91.653] - - - [50176, 99841, 1, 512] - - [369, 93.647] - - - [50176, 100353, 1, 512] - - [356, 91.585] - - - [50688, 512, 1, 512] - - [451, 87.298] - - - [50688, 100353, 1, 512] - - [356, 91.56] - - - [50688, 100865, 1, 512] - - [356, 93.567] - - - [50688, 101377, 1, 512] - - [356, 91.57] - - - [51200, 512, 1, 512] - - [451, 88.108] - - - [51200, 101377, 1, 512] - - [356, 91.563] - - - [51200, 101889, 1, 512] - - [369, 93.502] - - - [51200, 102401, 1, 512] - - [356, 91.443] - - - [51712, 512, 1, 512] - - [451, 88.879] - - - [51712, 102401, 1, 512] - - [356, 91.549] - - - [51712, 102913, 1, 512] - - [369, 93.47] - - - [51712, 103425, 1, 512] - - [356, 91.579] - - - [52224, 512, 1, 512] - - [458, 89.512] - - - [52224, 103425, 1, 512] - - [356, 91.494] - - - [52224, 103937, 1, 512] - - [369, 93.436] - - - [52224, 104449, 1, 512] - - [356, 91.41] - - - [52736, 512, 1, 512] - - [356, 88.847] - - - [52736, 104449, 1, 512] - - [356, 91.456] - - - [52736, 104961, 1, 512] - - [369, 93.35] - - - [52736, 105473, 1, 512] - - [356, 91.43] - - - [53248, 512, 1, 512] - - [456, 86.995] - - - [53248, 105473, 1, 512] - - [356, 91.37] - - - [53248, 105985, 1, 512] - - [369, 93.253] - - - [53248, 106497, 1, 512] - - [356, 91.059] - - - [53760, 512, 1, 512] - - [458, 87.557] - - - [53760, 106497, 1, 512] - - [356, 91.065] - - - [53760, 107009, 1, 512] - - [369, 93.249] - - - [53760, 107521, 1, 512] - - [356, 91.359] - - - [54272, 512, 1, 512] - - [452, 88.104] - - - [54272, 107521, 1, 512] - - [356, 91.276] - - - [54272, 108033, 1, 512] - - [369, 93.158] - - - [54272, 108545, 1, 512] - - [356, 91.189] - - - [54784, 512, 1, 512] - - [458, 88.877] - - - [54784, 108545, 1, 512] - - [356, 91.243] - - - [54784, 109057, 1, 512] - - [369, 93.098] - - - [54784, 109569, 1, 512] - - [356, 91.207] - - - [55296, 512, 1, 512] - - [451, 89.406] - - - [55296, 109569, 1, 512] - - [356, 91.139] - - - [55296, 110081, 1, 512] - - [369, 93.028] - - - [55296, 110593, 1, 512] - - [356, 91.032] - - - [55808, 512, 1, 512] - - [341, 88.484] - - - [55808, 110593, 1, 512] - - [356, 91.246] - - - [55808, 111105, 1, 512] - - [369, 92.994] - - - [55808, 111617, 1, 512] - - [356, 91.085] - - - [56320, 512, 1, 512] - - [356, 89.044] - - - [56320, 111617, 1, 512] - - [377, 90.97] - - - [56320, 112129, 1, 512] - - [369, 92.878] - - - [56320, 112641, 1, 512] - - [356, 90.893] - - - [56832, 512, 1, 512] - - [457, 87.658] - - - [56832, 112641, 1, 512] - - [377, 91.051] - - - [56832, 113153, 1, 512] - - [369, 92.866] - - - [56832, 113665, 1, 512] - - [377, 91.145] - - - [57344, 512, 1, 512] - - [457, 88.176] - - - [57344, 113665, 1, 512] - - [377, 91.103] - - - [57344, 114177, 1, 512] - - [369, 92.756] - - - [57344, 114689, 1, 512] - - [378, 89.228] - - - [57856, 512, 1, 512] - - [451, 88.821] - - - [57856, 114689, 1, 512] - - [379, 89.444] - - - [57856, 115201, 1, 512] - - [369, 92.668] - - - [57856, 115713, 1, 512] - - [377, 91.071] - - - [58368, 512, 1, 512] - - [457, 89.475] - - - [58368, 115713, 1, 512] - - [377, 90.936] - - - [58368, 116225, 1, 512] - - [369, 92.562] - - - [58368, 116737, 1, 512] - - [377, 90.921] - - - [58880, 512, 1, 512] - - [454, 89.984] - - - [58880, 116737, 1, 512] - - [377, 91.059] - - - [58880, 117249, 1, 512] - - [369, 92.463] - - - [58880, 117761, 1, 512] - - [377, 91.032] - - - [59392, 512, 1, 512] - - [356, 88.949] - - - [59392, 117761, 1, 512] - - [377, 90.935] - - - [59392, 118273, 1, 512] - - [369, 92.359] - - - [59392, 118785, 1, 512] - - [380, 90.655] - - - [59904, 512, 1, 512] - - [454, 87.843] - - - [59904, 118785, 1, 512] - - [380, 90.677] - - - [59904, 119297, 1, 512] - - [369, 92.282] - - - [59904, 119809, 1, 512] - - [377, 91.074] - - - [60416, 512, 1, 512] - - [457, 88.386] - - - [60416, 119809, 1, 512] - - [377, 91.021] - - - [60416, 120321, 1, 512] - - [369, 92.17] - - - [60416, 120833, 1, 512] - - [377, 90.865] - - - [60928, 512, 1, 512] - - [458, 88.947] - - - [60928, 120833, 1, 512] - - [377, 91.052] - - - [60928, 121345, 1, 512] - - [369, 92.033] - - - [60928, 121857, 1, 512] - - [377, 90.997] - - - [61440, 512, 1, 512] - - [450, 89.58] - - - [61440, 121857, 1, 512] - - [377, 91.095] - - - [61440, 122369, 1, 512] - - [369, 91.857] - - - [61440, 122881, 1, 512] - - [381, 90.354] - - - [61952, 512, 1, 512] - - [454, 90.063] - - - [61952, 122881, 1, 512] - - [381, 90.251] - - - [61952, 123393, 1, 512] - - [369, 91.691] - - - [61952, 123905, 1, 512] - - [377, 90.928] - - - [62464, 512, 1, 512] - - [341, 88.778] - - - [62464, 123905, 1, 512] - - [377, 91.029] - - - [62464, 124417, 1, 512] - - [369, 91.578] - - - [62464, 124929, 1, 512] - - [380, 90.749] - - - [62976, 512, 1, 512] - - [341, 89.013] - - - [62976, 124929, 1, 512] - - [377, 91.044] - - - [62976, 125441, 1, 512] - - [369, 91.493] - - - [62976, 125953, 1, 512] - - [377, 90.97] - - - [63488, 512, 1, 512] - - [456, 88.507] - - - [63488, 125953, 1, 512] - - [377, 91.013] - - - [63488, 126465, 1, 512] - - [382, 91.46] - - - [63488, 126977, 1, 512] - - [380, 90.619] - - - [64000, 512, 1, 512] - - [455, 88.992] - - - [64000, 126977, 1, 512] - - [380, 90.662] - - - [64000, 127489, 1, 512] - - [12, 84.172] - - - [64000, 128001, 1, 512] - - [11, 84.304] - - - [64512, 512, 1, 512] - - [456, 89.515] - - - [64512, 128001, 1, 512] - - [12, 84.344] - - - [4096, 4096, 1, 4128] - - [237, 93.126] - - - [25600, 25600, 1, 512] - - [451, 94.545] - - - [512, 512, 1, 512] - - [259, 21.149] - - - [1024, 512, 1, 512] - - [345, 33.417] - - - [1536, 512, 1, 512] - - [347, 43.936] - - - [1536, 1024, 1, 512] - - [385, 57.881] - - - [2048, 512, 1, 512] - - [351, 43.751] - - - [2048, 1024, 1, 512] - - [387, 52.725] - - - [2560, 512, 1, 512] - - [278, 43.564] - - - [2560, 1024, 1, 512] - - [289, 54.805] - - - [2560, 1536, 1, 512] - - [418, 66.513] - - - [3072, 512, 1, 512] - - [277, 58.252] - - - [3072, 1024, 1, 512] - - [390, 63.893] - - - [3072, 1536, 1, 512] - - [456, 68.831] - - - [3584, 512, 1, 512] - - [355, 55.564] - - - [3584, 1536, 1, 512] - - [291, 69.447] - - - [3584, 2048, 1, 512] - - [393, 73.443] - - - [4096, 512, 1, 512] - - [293, 55.619] - - - [4096, 1536, 1, 512] - - [420, 74.34] - - - [4096, 2048, 1, 512] - - [363, 76.685] - - - [4608, 2048, 1, 512] - - [458, 78.151] - - - [4608, 2560, 1, 512] - - [389, 81.688] - - - [5120, 2048, 1, 512] - - [331, 80.566] - - - [5120, 2560, 1, 512] - - [453, 83.759] - - - [5632, 2560, 1, 512] - - [373, 85.718] - - - [5632, 3072, 1, 512] - - [324, 84.22] - - - [6144, 2560, 1, 512] - - [318, 84.556] - - - [6144, 3072, 1, 512] - - [393, 84.809] - - - [6656, 3072, 1, 512] - - [454, 84.291] - - - [6656, 3584, 1, 512] - - [398, 86.435] - - - [7168, 3072, 1, 512] - - [454, 84.829] - - - [7168, 3584, 1, 512] - - [454, 86.932] - - - [7680, 3584, 1, 512] - - [459, 87.832] - - - [7680, 4096, 1, 512] - - [454, 89.747] - - - [8192, 3584, 1, 512] - - [458, 88.285] - - - [8192, 4096, 1, 512] - - [324, 88.705] - - - [8704, 4096, 1, 512] - - [331, 89.258] - - - [8704, 4608, 1, 512] - - [452, 91.333] - - - [9216, 4096, 1, 512] - - [363, 89.904] - - - [9216, 4608, 1, 512] - - [332, 89.962] - - - [9728, 4608, 1, 512] - - [451, 91.195] - - - [9728, 5120, 1, 512] - - [454, 91.586] - - - [10240, 4608, 1, 512] - - [453, 90.0] - - - [10240, 5120, 1, 512] - - [454, 90.725] - - - [10752, 5120, 1, 512] - - [454, 92.107] - - - [10752, 5632, 1, 512] - - [456, 91.088] - - - [11264, 5120, 1, 512] - - [331, 91.594] - - - [11264, 5632, 1, 512] - - [454, 92.54] - - - [11776, 5632, 1, 512] - - [455, 92.147] - - - [11776, 6144, 1, 512] - - [457, 91.717] - - - [12288, 5632, 1, 512] - - [455, 91.84] - - - [12288, 6144, 1, 512] - - [399, 91.959] - - - [12800, 6144, 1, 512] - - [453, 93.082] - - - [12800, 6656, 1, 512] - - [451, 93.145] - - - [13312, 6144, 1, 512] - - [454, 92.998] - - - [13312, 6656, 1, 512] - - [454, 93.233] - - - [13824, 6656, 1, 512] - - [451, 93.357] - - - [13824, 7168, 1, 512] - - [331, 92.689] - - - [14336, 6656, 1, 512] - - [451, 93.435] - - - [14336, 7168, 1, 512] - - [454, 92.858] - - - [14848, 7168, 1, 512] - - [454, 93.13] - - - [14848, 7680, 1, 512] - - [451, 92.822] - - - [15360, 7168, 1, 512] - - [454, 93.404] - - - [15360, 7680, 1, 512] - - [454, 93.201] - - - [15872, 7680, 1, 512] - - [458, 93.575] - - - [15872, 8192, 1, 512] - - [454, 93.537] - - - [16384, 7680, 1, 512] - - [332, 93.552] - - - [16384, 8192, 1, 512] - - [450, 93.168] - - - [16896, 8192, 1, 512] - - [453, 93.64] - - - [16896, 8704, 1, 512] - - [451, 93.837] - - - [17408, 8192, 1, 512] - - [456, 93.259] - - - [17408, 8704, 1, 512] - - [453, 93.583] - - - [17920, 8704, 1, 512] - - [326, 93.555] - - - [17920, 9216, 1, 512] - - [454, 93.722] - - - [18432, 8704, 1, 512] - - [454, 93.9] - - - [18432, 9216, 1, 512] - - [453, 93.63] - - - [18944, 9216, 1, 512] - - [453, 93.558] - - - [18944, 9728, 1, 512] - - [454, 93.48] - - - [19456, 9216, 1, 512] - - [460, 93.504] - - - [19456, 9728, 1, 512] - - [451, 93.502] - - - [19968, 9728, 1, 512] - - [453, 93.528] - - - [19968, 10240, 1, 512] - - [451, 93.622] - - - [20480, 9728, 1, 512] - - [453, 93.597] - - - [20480, 10240, 1, 512] - - [451, 93.733] - - - [20992, 10240, 1, 512] - - [460, 93.785] - - - [20992, 10752, 1, 512] - - [453, 93.981] - - - [21504, 10240, 1, 512] - - [453, 93.883] - - - [21504, 10752, 1, 512] - - [451, 93.762] - - - [22016, 10752, 1, 512] - - [460, 93.917] - - - [22016, 11264, 1, 512] - - [456, 93.852] - - - [22528, 10752, 1, 512] - - [454, 94.0] - - - [22528, 11264, 1, 512] - - [453, 93.977] - - - [23040, 11264, 1, 512] - - [450, 93.814] - - - [23040, 11776, 1, 512] - - [458, 93.827] - - - [23552, 11264, 1, 512] - - [453, 93.992] - - - [23552, 11776, 1, 512] - - [450, 94.036] - - - [24064, 11776, 1, 512] - - [460, 93.97] - - - [24064, 12288, 1, 512] - - [455, 94.06] - - - [24576, 11776, 1, 512] - - [454, 94.176] - - - [24576, 12288, 1, 512] - - [451, 94.082] - - - [25088, 12288, 1, 512] - - [454, 93.986] - - - [25088, 12800, 1, 512] - - [451, 94.136] - - - [25600, 12288, 1, 512] - - [451, 94.169] - - - [25600, 12800, 1, 512] - - [454, 94.153] - - - [26112, 12800, 1, 512] - - [450, 94.229] - - - [26112, 13312, 1, 512] - - [450, 94.275] - - - [26624, 12800, 1, 512] - - [454, 94.139] - - - [26624, 13312, 1, 512] - - [453, 94.186] - - - [27136, 13312, 1, 512] - - [451, 94.213] - - - [27136, 13824, 1, 512] - - [453, 94.152] - - - [27648, 13312, 1, 512] - - [451, 94.25] - - - [27648, 13824, 1, 512] - - [460, 94.193] - - - [28160, 13824, 1, 512] - - [455, 94.246] - - - [28160, 14336, 1, 512] - - [451, 94.2] - - - [28672, 13824, 1, 512] - - [453, 94.25] - - - [28672, 14336, 1, 512] - - [451, 94.274] - - - [29184, 14336, 1, 512] - - [451, 94.206] - - - [29184, 14848, 1, 512] - - [451, 94.304] - - - [29696, 14336, 1, 512] - - [451, 94.304] - - - [29696, 14848, 1, 512] - - [451, 94.304] - - - [30208, 14848, 1, 512] - - [453, 94.254] - - - [30208, 15360, 1, 512] - - [453, 94.306] - - - [30720, 14848, 1, 512] - - [454, 94.315] - - - [30720, 15360, 1, 512] - - [451, 94.352] - - - [31232, 15360, 1, 512] - - [451, 94.348] - - - [31232, 15872, 1, 512] - - [450, 94.338] - - - [31744, 15360, 1, 512] - - [455, 94.328] - - - [31744, 15872, 1, 512] - - [453, 94.351] - - - [32256, 15872, 1, 512] - - [451, 94.384] - - - [32256, 16384, 1, 512] - - [455, 94.384] - - - [32768, 15872, 1, 512] - - [451, 94.318] - - - [32768, 16384, 1, 512] - - [454, 94.258] - - - [33280, 16384, 1, 512] - - [455, 94.375] - - - [33280, 16896, 1, 512] - - [451, 94.383] - - - [33792, 16384, 1, 512] - - [458, 94.387] - - - [33792, 16896, 1, 512] - - [451, 94.409] - - - [34304, 16896, 1, 512] - - [451, 94.439] - - - [34304, 17408, 1, 512] - - [454, 94.467] - - - [34816, 16896, 1, 512] - - [453, 94.455] - - - [34816, 17408, 1, 512] - - [450, 94.424] - - - [35328, 17408, 1, 512] - - [451, 94.462] - - - [35328, 17920, 1, 512] - - [451, 94.458] - - - [35840, 17408, 1, 512] - - [451, 94.454] - - - [35840, 17920, 1, 512] - - [451, 94.441] - - - [36352, 17920, 1, 512] - - [451, 94.464] - - - [36352, 18432, 1, 512] - - [451, 94.478] - - - [36864, 17920, 1, 512] - - [451, 94.48] - - - [36864, 18432, 1, 512] - - [451, 94.445] - - - [37376, 18432, 1, 512] - - [451, 94.467] - - - [37376, 18944, 1, 512] - - [453, 94.465] - - - [37888, 18432, 1, 512] - - [451, 94.525] - - - [37888, 18944, 1, 512] - - [451, 94.461] - - - [38400, 18944, 1, 512] - - [455, 94.474] - - - [38400, 19456, 1, 512] - - [454, 94.517] - - - [38912, 18944, 1, 512] - - [454, 94.495] - - - [38912, 19456, 1, 512] - - [451, 94.527] - - - [39424, 19456, 1, 512] - - [453, 94.516] - - - [39424, 19968, 1, 512] - - [451, 94.511] - - - [39936, 19456, 1, 512] - - [453, 94.518] - - - [39936, 19968, 1, 512] - - [455, 94.504] - - - [40448, 19968, 1, 512] - - [451, 94.509] - - - [40448, 20480, 1, 512] - - [451, 94.523] - - - [40960, 19968, 1, 512] - - [451, 94.47] - - - [40960, 20480, 1, 512] - - [451, 94.513] - - - [41472, 20480, 1, 512] - - [453, 94.524] - - - [41472, 20992, 1, 512] - - [451, 94.533] - - - [41984, 20480, 1, 512] - - [454, 94.523] - - - [41984, 20992, 1, 512] - - [454, 94.521] - - - [42496, 20992, 1, 512] - - [455, 94.51] - - - [42496, 21504, 1, 512] - - [451, 94.55] - - - [43008, 20992, 1, 512] - - [454, 94.539] - - - [43008, 21504, 1, 512] - - [453, 94.537] - - - [43520, 21504, 1, 512] - - [451, 94.587] - - - [43520, 22016, 1, 512] - - [451, 94.549] - - - [44032, 21504, 1, 512] - - [451, 94.552] - - - [44032, 22016, 1, 512] - - [451, 94.554] - - - [44544, 22016, 1, 512] - - [451, 94.561] - - - [44544, 22528, 1, 512] - - [454, 94.558] - - - [45056, 22016, 1, 512] - - [454, 94.546] - - - [45056, 22528, 1, 512] - - [455, 94.557] - - - [45568, 22528, 1, 512] - - [454, 94.555] - - - [45568, 23040, 1, 512] - - [451, 94.562] - - - [46080, 22528, 1, 512] - - [451, 94.552] - - - [46080, 23040, 1, 512] - - [451, 94.577] - - - [46592, 23040, 1, 512] - - [451, 94.58] - - - [46592, 23552, 1, 512] - - [451, 94.595] - - - [47104, 23040, 1, 512] - - [453, 94.569] - - - [47104, 23552, 1, 512] - - [454, 94.582] - - - [47616, 23552, 1, 512] - - [453, 94.581] - - - [47616, 24064, 1, 512] - - [453, 94.57] - - - [48128, 23552, 1, 512] - - [454, 94.568] - - - [48128, 24064, 1, 512] - - [451, 94.578] - - - [48640, 24064, 1, 512] - - [453, 94.576] - - - [48640, 24576, 1, 512] - - [454, 94.566] - - - [49152, 24064, 1, 512] - - [453, 94.554] - - - [49152, 24576, 1, 512] - - [454, 94.536] - - - [49664, 24576, 1, 512] - - [451, 94.569] - - - [49664, 25088, 1, 512] - - [451, 94.591] - - - [50176, 24576, 1, 512] - - [454, 94.581] - - - [50176, 25088, 1, 512] - - [451, 94.602] - - - [50688, 25088, 1, 512] - - [453, 94.586] - - - [50688, 25600, 1, 512] - - [451, 94.588] - - - [51200, 25088, 1, 512] - - [451, 94.586] - - - [51200, 25600, 1, 512] - - [451, 94.582] - - - [51712, 25600, 1, 512] - - [455, 94.564] - - - [51712, 26112, 1, 512] - - [453, 94.588] - - - [52224, 25600, 1, 512] - - [451, 94.582] - - - [52224, 26112, 1, 512] - - [451, 94.586] - - - [52736, 26112, 1, 512] - - [453, 94.576] - - - [52736, 26624, 1, 512] - - [451, 94.576] - - - [53248, 26112, 1, 512] - - [453, 94.565] - - - [53248, 26624, 1, 512] - - [451, 94.583] - - - [53760, 26624, 1, 512] - - [451, 94.58] - - - [53760, 27136, 1, 512] - - [453, 94.578] - - - [54272, 26624, 1, 512] - - [454, 94.597] - - - [54272, 27136, 1, 512] - - [451, 94.586] - - - [54784, 27136, 1, 512] - - [451, 94.574] - - - [54784, 27648, 1, 512] - - [451, 94.582] - - - [55296, 27136, 1, 512] - - [361, 94.572] - - - [55296, 27648, 1, 512] - - [453, 94.598] - - - [55808, 27648, 1, 512] - - [451, 94.605] - - - [55808, 28160, 1, 512] - - [451, 94.581] - - - [56320, 27648, 1, 512] - - [453, 94.589] - - - [56320, 28160, 1, 512] - - [451, 94.58] - - - [56832, 28160, 1, 512] - - [451, 94.59] - - - [56832, 28672, 1, 512] - - [451, 94.595] - - - [57344, 28160, 1, 512] - - [458, 94.574] - - - [57344, 28672, 1, 512] - - [451, 94.595] - - - [57856, 28672, 1, 512] - - [451, 94.602] - - - [57856, 29184, 1, 512] - - [451, 94.592] - - - [58368, 28672, 1, 512] - - [451, 94.614] - - - [58368, 29184, 1, 512] - - [451, 94.609] - - - [58880, 29184, 1, 512] - - [451, 94.593] - - - [58880, 29696, 1, 512] - - [453, 94.589] - - - [59392, 29184, 1, 512] - - [451, 94.611] - - - [59392, 29696, 1, 512] - - [453, 94.602] - - - [59904, 29696, 1, 512] - - [451, 94.588] - - - [59904, 30208, 1, 512] - - [451, 94.602] - - - [60416, 29696, 1, 512] - - [451, 94.584] - - - [60416, 30208, 1, 512] - - [451, 94.606] - - - [60928, 30208, 1, 512] - - [451, 94.602] - - - [60928, 30720, 1, 512] - - [453, 94.613] - - - [61440, 30208, 1, 512] - - [451, 94.59] - - - [61440, 30720, 1, 512] - - [451, 94.602] - - - [61952, 30720, 1, 512] - - [453, 94.614] - - - [61952, 31232, 1, 512] - - [451, 94.617] - - - [62464, 30720, 1, 512] - - [451, 94.594] - - - [62464, 31232, 1, 512] - - [451, 94.615] - - - [62976, 31232, 1, 512] - - [451, 94.611] - - - [62976, 31744, 1, 512] - - [451, 94.618] - - - [63488, 31232, 1, 512] - - [451, 94.602] - - - [63488, 31744, 1, 512] - - [453, 94.595] - - - [64000, 31744, 1, 512] - - [453, 94.601] - - - [64000, 32256, 1, 512] - - [451, 94.621] - - - [64512, 31744, 1, 512] - - [453, 94.603] - - - [64512, 32256, 1, 512] - - [453, 94.621] - - - [65024, 512, 1, 512] - - [456, 90.116] - - - [65024, 32256, 1, 512] - - [451, 94.622] - - - [65024, 32768, 1, 512] - - [451, 94.609] - - - [65536, 512, 1, 512] - - [450, 87.59] - - - [65536, 32256, 1, 512] - - [453, 94.521] - - - [65536, 32768, 1, 512] - - [453, 94.491] - - - [66048, 512, 1, 512] - - [326, 88.777] - - - [66048, 32768, 1, 512] - - [451, 94.601] - - - [66048, 33280, 1, 512] - - [453, 94.607] - - - [66560, 512, 1, 512] - - [326, 89.14] - - - [66560, 32768, 1, 512] - - [451, 94.6] - - - [66560, 33280, 1, 512] - - [453, 94.607] - - - [67072, 512, 1, 512] - - [455, 88.977] - - - [67072, 33280, 1, 512] - - [453, 94.63] - - - [67072, 33792, 1, 512] - - [451, 94.621] - - - [67584, 512, 1, 512] - - [452, 89.599] - - - [67584, 33280, 1, 512] - - [453, 94.612] - - - [67584, 33792, 1, 512] - - [451, 94.614] - - - [68096, 512, 1, 512] - - [458, 90.192] - - - [68096, 33792, 1, 512] - - [451, 94.605] - - - [68096, 34304, 1, 512] - - [453, 94.617] - - - [68608, 512, 1, 512] - - [460, 90.362] - - - [68608, 33792, 1, 512] - - [451, 94.62] - - - [68608, 34304, 1, 512] - - [453, 94.624] - - - [69120, 512, 1, 512] - - [326, 88.486] - - - [69120, 34304, 1, 512] - - [451, 94.625] - - - [69120, 34816, 1, 512] - - [458, 94.598] - - - [69632, 512, 1, 512] - - [341, 89.267] - - - [69632, 34304, 1, 512] - - [453, 94.603] - - - [69632, 34816, 1, 512] - - [451, 94.62] - - - [70144, 512, 1, 512] - - [341, 89.313] - - - [70144, 34816, 1, 512] - - [451, 94.621] - - - [70144, 35328, 1, 512] - - [451, 94.606] - - - [70656, 512, 1, 512] - - [451, 89.566] - - - [70656, 34816, 1, 512] - - [451, 94.626] - - - [70656, 35328, 1, 512] - - [451, 94.628] - - - [71168, 512, 1, 512] - - [454, 90.033] - - - [71168, 35328, 1, 512] - - [453, 94.633] - - - [71168, 35840, 1, 512] - - [451, 94.618] - - - [71680, 512, 1, 512] - - [456, 90.415] - - - [71680, 35328, 1, 512] - - [453, 94.624] - - - [71680, 35840, 1, 512] - - [451, 94.638] - - - [72192, 512, 1, 512] - - [326, 88.487] - - - [72192, 35840, 1, 512] - - [451, 94.629] - - - [72192, 36352, 1, 512] - - [451, 94.616] - - - [72704, 512, 1, 512] - - [341, 89.06] - - - [72704, 35840, 1, 512] - - [451, 94.642] - - - [72704, 36352, 1, 512] - - [451, 94.629] - - - [73216, 512, 1, 512] - - [459, 89.236] - - - [73216, 36352, 1, 512] - - [453, 94.615] - - - [73216, 36864, 1, 512] - - [453, 94.626] - - - [73728, 512, 1, 512] - - [457, 89.589] - - - [73728, 36352, 1, 512] - - [451, 94.616] - - - [73728, 36864, 1, 512] - - [451, 94.605] - - - [74240, 512, 1, 512] - - [454, 90.036] - - - [74240, 36864, 1, 512] - - [451, 94.627] - - - [74240, 37376, 1, 512] - - [451, 94.62] - - - [74752, 512, 1, 512] - - [455, 90.525] - - - [74752, 36864, 1, 512] - - [453, 94.623] - - - [74752, 37376, 1, 512] - - [451, 94.621] - - - [75264, 512, 1, 512] - - [450, 90.916] - - - [75264, 37376, 1, 512] - - [451, 94.642] - - - [75264, 37888, 1, 512] - - [453, 94.621] - - - [75776, 512, 1, 512] - - [459, 88.879] - - - [75776, 37376, 1, 512] - - [451, 94.634] - - - [75776, 37888, 1, 512] - - [451, 94.61] - - - [76288, 512, 1, 512] - - [451, 89.203] - - - [76288, 37888, 1, 512] - - [451, 94.621] - - - [76288, 38400, 1, 512] - - [451, 94.637] - - - [76800, 512, 1, 512] - - [356, 89.705] - - - [76800, 37888, 1, 512] - - [451, 94.625] - - - [76800, 38400, 1, 512] - - [451, 94.633] - - - [77312, 512, 1, 512] - - [451, 90.047] - - - [77312, 38400, 1, 512] - - [451, 94.636] - - - [77312, 38912, 1, 512] - - [326, 94.093] - - - [77824, 512, 1, 512] - - [458, 90.491] - - - [77824, 38400, 1, 512] - - [451, 94.636] - - - [77824, 38912, 1, 512] - - [326, 94.041] - - - [78336, 512, 1, 512] - - [458, 90.977] - - - [78336, 38912, 1, 512] - - [326, 93.989] - - - [78336, 39424, 1, 512] - - [326, 94.056] - - - [78848, 512, 1, 512] - - [458, 88.894] - - - [78848, 38912, 1, 512] - - [326, 94.017] - - - [78848, 39424, 1, 512] - - [332, 94.354] - - - [79360, 512, 1, 512] - - [455, 89.302] - - - [79360, 39424, 1, 512] - - [361, 94.236] - - - [79360, 39936, 1, 512] - - [326, 93.803] - - - [79872, 512, 1, 512] - - [453, 89.676] - - - [79872, 39424, 1, 512] - - [332, 94.402] - - - [79872, 39936, 1, 512] - - [326, 93.715] - - - [80384, 512, 1, 512] - - [457, 90.162] - - - [80384, 39936, 1, 512] - - [326, 93.674] - - - [80384, 40448, 1, 512] - - [361, 94.177] - - - [80896, 512, 1, 512] - - [453, 90.566] - - - [80896, 39936, 1, 512] - - [326, 93.532] - - - [80896, 40448, 1, 512] - - [332, 94.323] - - - [81408, 512, 1, 512] - - [455, 90.895] - - - [81408, 40448, 1, 512] - - [361, 94.148] - - - [81408, 40960, 1, 512] - - [326, 93.922] - - - [81920, 512, 1, 512] - - [458, 88.877] - - - [81920, 40448, 1, 512] - - [326, 92.848] - - - [81920, 40960, 1, 512] - - [326, 92.602] - - - [82432, 512, 1, 512] - - [456, 89.332] - - - [82432, 40960, 1, 512] - - [326, 93.839] - - - [82432, 41472, 1, 512] - - [326, 94.069] - - - [82944, 512, 1, 512] - - [454, 89.77] - - - [82944, 40960, 1, 512] - - [326, 93.864] - - - [82944, 41472, 1, 512] - - [332, 94.266] - - - [83456, 512, 1, 512] - - [450, 90.137] - - - [83456, 41472, 1, 512] - - [332, 94.113] - - - [83456, 41984, 1, 512] - - [326, 93.797] - - - [83968, 512, 1, 512] - - [457, 90.465] - - - [83968, 41472, 1, 512] - - [332, 94.3] - - - [83968, 41984, 1, 512] - - [326, 93.696] - - - [84480, 512, 1, 512] - - [455, 90.826] - - - [84480, 41984, 1, 512] - - [326, 93.796] - - - [84480, 42496, 1, 512] - - [332, 94.105] - - - [84992, 512, 1, 512] - - [456, 90.923] - - - [84992, 41984, 1, 512] - - [395, 93.504] - - - [84992, 42496, 1, 512] - - [332, 94.229] - - - [85504, 512, 1, 512] - - [457, 89.464] - - - [85504, 42496, 1, 512] - - [326, 94.033] - - - [85504, 43008, 1, 512] - - [326, 93.547] - - - [86016, 512, 1, 512] - - [453, 89.7] - - - [86016, 42496, 1, 512] - - [332, 94.189] - - - [86016, 43008, 1, 512] - - [331, 93.236] - - - [86528, 512, 1, 512] - - [450, 90.068] - - - [86528, 43008, 1, 512] - - [326, 93.514] - - - [86528, 43520, 1, 512] - - [332, 94.028] - - - [87040, 512, 1, 512] - - [453, 90.435] - - - [87040, 43008, 1, 512] - - [395, 93.326] - - - [87040, 43520, 1, 512] - - [332, 94.139] - - - [87552, 512, 1, 512] - - [451, 90.906] - - - [87552, 43520, 1, 512] - - [332, 94.017] - - - [87552, 44032, 1, 512] - - [326, 93.791] - - - [88064, 512, 1, 512] - - [451, 91.173] - - - [88064, 43520, 1, 512] - - [332, 94.215] - - - [88064, 44032, 1, 512] - - [395, 93.594] - - - [88576, 512, 1, 512] - - [451, 89.551] - - - [88576, 44032, 1, 512] - - [326, 93.707] - - - [88576, 44544, 1, 512] - - [332, 94.007] - - - [89088, 512, 1, 512] - - [457, 89.779] - - - [89088, 44032, 1, 512] - - [331, 93.361] - - - [89088, 44544, 1, 512] - - [332, 94.087] - - - [89600, 512, 1, 512] - - [453, 90.113] - - - [89600, 44544, 1, 512] - - [326, 93.985] - - - [89600, 45056, 1, 512] - - [395, 93.377] - - - [90112, 512, 1, 512] - - [456, 90.445] - - - [90112, 44544, 1, 512] - - [326, 93.721] - - - [90112, 45056, 1, 512] - - [353, 92.682] - - - [90624, 512, 1, 512] - - [450, 90.923] - - - [90624, 45056, 1, 512] - - [395, 93.247] - - - [90624, 45568, 1, 512] - - [332, 93.884] - - - [91136, 512, 1, 512] - - [451, 91.049] - - - [91136, 45056, 1, 512] - - [395, 93.106] - - - [91136, 45568, 1, 512] - - [332, 94.047] - - - [91648, 512, 1, 512] - - [458, 89.5] - - - [91648, 45568, 1, 512] - - [332, 93.914] - - - [91648, 46080, 1, 512] - - [326, 93.506] - - - [92160, 512, 1, 512] - - [453, 89.909] - - - [92160, 45568, 1, 512] - - [332, 94.064] - - - [92160, 46080, 1, 512] - - [353, 93.318] - - - [92672, 512, 1, 512] - - [454, 90.183] - - - [92672, 46080, 1, 512] - - [326, 93.595] - - - [92672, 46592, 1, 512] - - [326, 93.903] - - - [93184, 512, 1, 512] - - [451, 90.455] - - - [93184, 46080, 1, 512] - - [353, 93.272] - - - [93184, 46592, 1, 512] - - [332, 93.942] - - - [93696, 512, 1, 512] - - [455, 90.861] - - - [93696, 46592, 1, 512] - - [326, 93.845] - - - [93696, 47104, 1, 512] - - [353, 93.108] - - - [94208, 512, 1, 512] - - [451, 91.188] - - - [94208, 46592, 1, 512] - - [332, 93.99] - - - [94208, 47104, 1, 512] - - [399, 92.897] - - - [94720, 512, 1, 512] - - [457, 91.555] - - - [94720, 47104, 1, 512] - - [395, 93.353] - - - [94720, 47616, 1, 512] - - [332, 93.811] - - - [95232, 512, 1, 512] - - [454, 89.922] - - - [95232, 47104, 1, 512] - - [399, 93.035] - - - [95232, 47616, 1, 512] - - [332, 93.915] - - - [95744, 512, 1, 512] - - [452, 90.208] - - - [95744, 47616, 1, 512] - - [332, 93.776] - - - [95744, 48128, 1, 512] - - [395, 93.291] - - - [96256, 512, 1, 512] - - [459, 90.547] - - - [96256, 47616, 1, 512] - - [332, 93.921] - - - [96256, 48128, 1, 512] - - [399, 92.916] - - - [96768, 512, 1, 512] - - [457, 90.775] - - - [96768, 48128, 1, 512] - - [395, 93.238] - - - [96768, 48640, 1, 512] - - [332, 93.718] - - - [97280, 512, 1, 512] - - [450, 91.098] - - - [97280, 48128, 1, 512] - - [399, 92.915] - - - [97280, 48640, 1, 512] - - [332, 93.74] - - - [97792, 512, 1, 512] - - [459, 91.468] - - - [97792, 48640, 1, 512] - - [326, 93.704] - - - [97792, 49152, 1, 512] - - [399, 92.662] - - - [98304, 512, 1, 512] - - [459, 89.848] - - - [98304, 48640, 1, 512] - - [326, 92.133] - - - [98304, 49152, 1, 512] - - [326, 91.717] - - - [98816, 512, 1, 512] - - [453, 90.204] - - - [98816, 49152, 1, 512] - - [372, 92.902] - - - [98816, 49664, 1, 512] - - [395, 93.685] - - - [99328, 512, 1, 512] - - [453, 90.467] - - - [99328, 49152, 1, 512] - - [399, 92.597] - - - [99328, 49664, 1, 512] - - [395, 93.637] - - - [99840, 512, 1, 512] - - [458, 90.893] - - - [99840, 49664, 1, 512] - - [375, 93.611] - - - [99840, 50176, 1, 512] - - [399, 92.745] - - - [100352, 512, 1, 512] - - [456, 91.081] - - - [100352, 49664, 1, 512] - - [375, 93.663] - - - [100352, 50176, 1, 512] - - [399, 92.79] - - - [100864, 512, 1, 512] - - [453, 91.446] - - - [100864, 50176, 1, 512] - - [399, 92.852] - - - [100864, 50688, 1, 512] - - [395, 93.521] - - - [101376, 512, 1, 512] - - [454, 91.704] - - - [101376, 50176, 1, 512] - - [399, 92.619] - - - [101376, 50688, 1, 512] - - [375, 93.594] - - - [101888, 512, 1, 512] - - [452, 90.163] - - - [101888, 50688, 1, 512] - - [375, 93.509] - - - [101888, 51200, 1, 512] - - [372, 92.706] - - - [102400, 512, 1, 512] - - [458, 90.571] - - - [102400, 50688, 1, 512] - - [375, 93.572] - - - [102400, 51200, 1, 512] - - [372, 92.687] - - - [102912, 512, 1, 512] - - [459, 90.85] - - - [102912, 51200, 1, 512] - - [372, 92.752] - - - [102912, 51712, 1, 512] - - [375, 93.489] - - - [103424, 512, 1, 512] - - [457, 91.276] - - - [103424, 51200, 1, 512] - - [399, 92.553] - - - [103424, 51712, 1, 512] - - [375, 93.526] - - - [103936, 512, 1, 512] - - [456, 91.361] - - - [103936, 51712, 1, 512] - - [395, 93.49] - - - [103936, 52224, 1, 512] - - [399, 92.549] - - - [104448, 512, 1, 512] - - [457, 91.502] - - - [104448, 51712, 1, 512] - - [375, 93.515] - - - [104448, 52224, 1, 512] - - [399, 92.539] - - - [104960, 512, 1, 512] - - [455, 90.242] - - - [104960, 52224, 1, 512] - - [399, 92.49] - - - [104960, 52736, 1, 512] - - [395, 93.387] - - - [105472, 512, 1, 512] - - [459, 90.532] - - - [105472, 52224, 1, 512] - - [399, 92.457] - - - [105472, 52736, 1, 512] - - [375, 93.388] - - - [105984, 512, 1, 512] - - [459, 90.791] - - - [105984, 52736, 1, 512] - - [375, 93.347] - - - [105984, 53248, 1, 512] - - [399, 92.514] - - - [106496, 512, 1, 512] - - [459, 91.036] - - - [106496, 52736, 1, 512] - - [395, 92.302] - - - [106496, 53248, 1, 512] - - [372, 91.45] - - - [107008, 512, 1, 512] - - [453, 91.33] - - - [107008, 53248, 1, 512] - - [399, 92.245] - - - [107008, 53760, 1, 512] - - [395, 93.317] - - - [107520, 512, 1, 512] - - [459, 91.636] - - - [107520, 53248, 1, 512] - - [399, 92.282] - - - [107520, 53760, 1, 512] - - [405, 92.371] - - - [108032, 512, 1, 512] - - [459, 90.244] - - - [108032, 53760, 1, 512] - - [375, 93.24] - - - [108032, 54272, 1, 512] - - [399, 92.339] - - - [108544, 512, 1, 512] - - [459, 90.563] - - - [108544, 53760, 1, 512] - - [405, 92.74] - - - [108544, 54272, 1, 512] - - [399, 91.992] - - - [109056, 512, 1, 512] - - [458, 90.85] - - - [109056, 54272, 1, 512] - - [399, 92.303] - - - [109056, 54784, 1, 512] - - [375, 93.11] - - - [109568, 512, 1, 512] - - [457, 91.107] - - - [109568, 54272, 1, 512] - - [399, 91.699] - - - [109568, 54784, 1, 512] - - [399, 91.884] - - - [110080, 512, 1, 512] - - [454, 91.417] - - - [110080, 54784, 1, 512] - - [375, 93.117] - - - [110080, 55296, 1, 512] - - [399, 92.089] - - - [110592, 512, 1, 512] - - [452, 91.605] - - - [110592, 54784, 1, 512] - - [375, 92.612] - - - [110592, 55296, 1, 512] - - [399, 91.688] - - - [111104, 512, 1, 512] - - [450, 90.677] - - - [111104, 55296, 1, 512] - - [399, 91.987] - - - [111104, 55808, 1, 512] - - [375, 92.947] - - - [111616, 512, 1, 512] - - [457, 90.532] - - - [111616, 55296, 1, 512] - - [399, 91.507] - - - [111616, 55808, 1, 512] - - [406, 91.232] - - - [112128, 512, 1, 512] - - [458, 90.865] - - - [112128, 55808, 1, 512] - - [375, 92.944] - - - [112128, 56320, 1, 512] - - [399, 91.869] - - - [112640, 512, 1, 512] - - [452, 91.156] - - - [112640, 55808, 1, 512] - - [399, 91.385] - - - [112640, 56320, 1, 512] - - [399, 91.134] - - - [113152, 512, 1, 512] - - [454, 91.381] - - - [113152, 56320, 1, 512] - - [399, 91.787] - - - [113152, 56832, 1, 512] - - [399, 91.745] - - - [113664, 512, 1, 512] - - [454, 91.528] - - - [113664, 56320, 1, 512] - - [399, 91.155] - - - [113664, 56832, 1, 512] - - [399, 91.51] - - - [114176, 512, 1, 512] - - [454, 91.745] - - - [114176, 56832, 1, 512] - - [399, 91.822] - - - [114176, 57344, 1, 512] - - [399, 91.335] - - - [114688, 512, 1, 512] - - [453, 90.487] - - - [114688, 56832, 1, 512] - - [326, 90.139] - - - [114688, 57344, 1, 512] - - [353, 88.626] - - - [115200, 512, 1, 512] - - [456, 90.839] - - - [115200, 57344, 1, 512] - - [399, 91.233] - - - [115200, 57856, 1, 512] - - [405, 91.741] - - - [115712, 512, 1, 512] - - [455, 91.088] - - - [115712, 57344, 1, 512] - - [399, 90.773] - - - [115712, 57856, 1, 512] - - [407, 91.291] - - - [116224, 512, 1, 512] - - [450, 91.341] - - - [116224, 57856, 1, 512] - - [407, 91.302] - - - [116224, 58368, 1, 512] - - [399, 90.986] - - - [116736, 512, 1, 512] - - [453, 91.64] - - - [116736, 57856, 1, 512] - - [407, 91.219] - - - [116736, 58368, 1, 512] - - [401, 90.79] - - - [117248, 512, 1, 512] - - [456, 91.914] - - - [117248, 58368, 1, 512] - - [410, 90.917] - - - [117248, 58880, 1, 512] - - [405, 91.904] - - - [117760, 512, 1, 512] - - [456, 90.627] - - - [117760, 58368, 1, 512] - - [411, 90.759] - - - [117760, 58880, 1, 512] - - [407, 91.234] - - - [118272, 512, 1, 512] - - [451, 90.786] - - - [118272, 58880, 1, 512] - - [407, 91.26] - - - [118272, 59392, 1, 512] - - [410, 91.152] - - - [118784, 512, 1, 512] - - [458, 91.096] - - - [118784, 58880, 1, 512] - - [407, 91.219] - - - [118784, 59392, 1, 512] - - [410, 90.634] - - - [119296, 512, 1, 512] - - [451, 91.393] - - - [119296, 59392, 1, 512] - - [410, 91.147] - - - [119296, 59904, 1, 512] - - [407, 91.26] - - - [119808, 512, 1, 512] - - [459, 91.6] - - - [119808, 59392, 1, 512] - - [410, 90.798] - - - [119808, 59904, 1, 512] - - [407, 91.232] - - - [120320, 512, 1, 512] - - [452, 91.94] - - - [120320, 59904, 1, 512] - - [407, 91.214] - - - [120320, 60416, 1, 512] - - [410, 90.921] - - - [120832, 512, 1, 512] - - [459, 91.955] - - - [120832, 59904, 1, 512] - - [407, 91.222] - - - [120832, 60416, 1, 512] - - [411, 90.758] - - - [121344, 512, 1, 512] - - [452, 90.894] - - - [121344, 60416, 1, 512] - - [401, 90.933] - - - [121344, 60928, 1, 512] - - [407, 91.238] - - - [121856, 512, 1, 512] - - [459, 91.114] - - - [121856, 60416, 1, 512] - - [411, 90.759] - - - [121856, 60928, 1, 512] - - [407, 91.213] - - - [122368, 512, 1, 512] - - [454, 91.387] - - - [122368, 60928, 1, 512] - - [407, 91.239] - - - [122368, 61440, 1, 512] - - [410, 91.101] - - - [122880, 512, 1, 512] - - [457, 91.633] - - - [122880, 60928, 1, 512] - - [409, 89.826] - - - [122880, 61440, 1, 512] - - [399, 89.109] - - - [123392, 512, 1, 512] - - [453, 91.927] - - - [123392, 61440, 1, 512] - - [410, 91.075] - - - [123392, 61952, 1, 512] - - [407, 91.252] - - - [123904, 512, 1, 512] - - [456, 92.017] - - - [123904, 61440, 1, 512] - - [410, 90.717] - - - [123904, 61952, 1, 512] - - [407, 91.176] - - - [124416, 512, 1, 512] - - [455, 90.885] - - - [124416, 61952, 1, 512] - - [407, 91.237] - - - [124416, 62464, 1, 512] - - [401, 90.967] - - - [124928, 512, 1, 512] - - [453, 91.195] - - - [124928, 61952, 1, 512] - - [406, 91.213] - - - [124928, 62464, 1, 512] - - [411, 90.761] - - - [125440, 512, 1, 512] - - [450, 91.377] - - - [125440, 62464, 1, 512] - - [410, 90.989] - - - [125440, 62976, 1, 512] - - [407, 91.247] - - - [125952, 512, 1, 512] - - [453, 91.563] - - - [125952, 62464, 1, 512] - - [401, 90.743] - - - [125952, 62976, 1, 512] - - [406, 91.21] - - - [126464, 512, 1, 512] - - [450, 91.839] - - - [126464, 62976, 1, 512] - - [407, 91.283] - - - [126464, 63488, 1, 512] - - [410, 90.926] - - - [126976, 512, 1, 512] - - [453, 92.029] - - - [126976, 62976, 1, 512] - - [407, 91.185] - - - [126976, 63488, 1, 512] - - [414, 90.545] - - - [127488, 512, 1, 512] - - [451, 92.208] - - - [127488, 63488, 1, 512] - - [401, 90.905] - - - [127488, 64000, 1, 512] - - [407, 91.224] - - - [128000, 512, 1, 512] - - [452, 91.075] - - - [128000, 63488, 1, 512] - - [414, 90.743] - - - [128000, 64000, 1, 512] - - [407, 91.229] - - - [4096, 1537, 1, 512] - - [393, 72.921] - - - [4096, 2049, 1, 512] - - [282, 74.109] - - - [4608, 2049, 1, 512] - - [454, 77.951] - - - [5120, 2049, 1, 512] - - [460, 76.654] - - - [5120, 2561, 1, 512] - - [397, 80.786] - - - [5632, 2561, 1, 512] - - [396, 81.849] - - - [6144, 2561, 1, 512] - - [460, 82.634] - - - [6144, 3073, 1, 512] - - [454, 83.913] - - - [6656, 3073, 1, 512] - - [454, 84.734] - - - [7168, 3073, 1, 512] - - [460, 84.829] - - - [7168, 3585, 1, 512] - - [459, 86.884] - - - [7680, 3585, 1, 512] - - [460, 87.65] - - - [8192, 3585, 1, 512] - - [341, 85.559] - - - [8192, 4097, 1, 512] - - [460, 87.668] - - - [8704, 4097, 1, 512] - - [454, 88.912] - - - [9216, 4097, 1, 512] - - [460, 87.103] - - - [9216, 4609, 1, 512] - - [460, 87.501] - - - [9728, 4609, 1, 512] - - [453, 88.982] - - - [10240, 4609, 1, 512] - - [460, 89.872] - - - [10240, 5121, 1, 512] - - [454, 88.816] - - - [10752, 5121, 1, 512] - - [454, 90.323] - - - [11264, 5121, 1, 512] - - [460, 89.538] - - - [11264, 5633, 1, 512] - - [458, 90.841] - - - [11776, 5633, 1, 512] - - [458, 90.555] - - - [12288, 5633, 1, 512] - - [460, 90.165] - - - [12288, 6145, 1, 512] - - [454, 90.195] - - - [12800, 6145, 1, 512] - - [450, 91.687] - - - [13312, 6145, 1, 512] - - [454, 91.638] - - - [13312, 6657, 1, 512] - - [460, 91.896] - - - [13824, 6657, 1, 512] - - [460, 90.698] - - - [14336, 6657, 1, 512] - - [460, 90.868] - - - [14336, 7169, 1, 512] - - [454, 91.642] - - - [14848, 7169, 1, 512] - - [454, 91.988] - - - [15360, 7169, 1, 512] - - [454, 92.319] - - - [15360, 7681, 1, 512] - - [456, 92.153] - - - [15872, 7681, 1, 512] - - [460, 91.613] - - - [16384, 7681, 1, 512] - - [455, 91.971] - - - [16384, 8193, 1, 512] - - [454, 92.219] - - - [16896, 8193, 1, 512] - - [454, 92.594] - - - [17408, 8193, 1, 512] - - [454, 92.311] - - - [17408, 8705, 1, 512] - - [460, 92.649] - - - [17920, 8705, 1, 512] - - [460, 92.388] - - - [18432, 8705, 1, 512] - - [460, 92.306] - - - [18432, 9217, 1, 512] - - [454, 92.837] - - - [18944, 9217, 1, 512] - - [453, 92.762] - - - [19456, 9217, 1, 512] - - [453, 92.708] - - - [19456, 9729, 1, 512] - - [460, 92.632] - - - [19968, 9729, 1, 512] - - [460, 92.705] - - - [20480, 9729, 1, 512] - - [460, 92.705] - - - [20480, 10241, 1, 512] - - [454, 92.93] - - - [20992, 10241, 1, 512] - - [456, 92.653] - - - [21504, 10241, 1, 512] - - [460, 92.763] - - - [21504, 10753, 1, 512] - - [451, 93.012] - - - [22016, 10753, 1, 512] - - [451, 92.808] - - - [22528, 10753, 1, 512] - - [451, 92.926] - - - [22528, 11265, 1, 512] - - [460, 92.935] - - - [23040, 11265, 1, 512] - - [454, 93.121] - - - [23552, 11265, 1, 512] - - [460, 92.979] - - - [23552, 11777, 1, 512] - - [451, 93.013] - - - [24064, 11777, 1, 512] - - [460, 92.934] - - - [24576, 11777, 1, 512] - - [460, 93.215] - - - [24576, 12289, 1, 512] - - [450, 93.108] - - - [25088, 12289, 1, 512] - - [460, 93.072] - - - [25600, 12289, 1, 512] - - [454, 93.355] - - - [25600, 12801, 1, 512] - - [455, 93.26] - - - [26112, 12801, 1, 512] - - [455, 93.34] - - - [26624, 12801, 1, 512] - - [455, 93.3] - - - [26624, 13313, 1, 512] - - [454, 93.398] - - - [27136, 13313, 1, 512] - - [454, 93.424] - - - [27648, 13313, 1, 512] - - [454, 93.361] - - - [27648, 13825, 1, 512] - - [450, 93.361] - - - [28160, 13825, 1, 512] - - [450, 93.439] - - - [28672, 13825, 1, 512] - - [453, 93.334] - - - [28672, 14337, 1, 512] - - [451, 93.547] - - - [29184, 14337, 1, 512] - - [451, 93.465] - - - [29696, 14337, 1, 512] - - [451, 93.549] - - - [29696, 14849, 1, 512] - - [458, 93.517] - - - [30208, 14849, 1, 512] - - [453, 93.496] - - - [30720, 14849, 1, 512] - - [451, 93.501] - - - [30720, 15361, 1, 512] - - [450, 93.593] - - - [31232, 15361, 1, 512] - - [454, 93.573] - - - [31744, 15361, 1, 512] - - [451, 93.639] - - - [31744, 15873, 1, 512] - - [451, 93.585] - - - [32256, 15873, 1, 512] - - [460, 93.581] - - - [32768, 15873, 1, 512] - - [451, 93.518] - - - [32768, 16385, 1, 512] - - [454, 93.646] - - - [33280, 16385, 1, 512] - - [454, 93.707] - - - [33792, 16385, 1, 512] - - [454, 93.698] - - - [33792, 16897, 1, 512] - - [458, 93.756] - - - [34304, 16897, 1, 512] - - [451, 93.703] - - - [34816, 16897, 1, 512] - - [450, 93.724] - - - [34816, 17409, 1, 512] - - [454, 93.781] - - - [35328, 17409, 1, 512] - - [454, 93.771] - - - [35840, 17409, 1, 512] - - [454, 93.816] - - - [35840, 17921, 1, 512] - - [455, 93.827] - - - [36352, 17921, 1, 512] - - [451, 93.828] - - - [36864, 17921, 1, 512] - - [451, 93.841] - - - [36864, 18433, 1, 512] - - [454, 93.836] - - - [37376, 18433, 1, 512] - - [454, 93.875] - - - [37888, 18433, 1, 512] - - [454, 93.857] - - - [37888, 18945, 1, 512] - - [454, 93.824] - - - [38400, 18945, 1, 512] - - [458, 93.843] - - - [38912, 18945, 1, 512] - - [451, 93.884] - - - [38912, 19457, 1, 512] - - [454, 93.931] - - - [39424, 19457, 1, 512] - - [454, 93.914] - - - [39936, 19457, 1, 512] - - [454, 93.935] - - - [39936, 19969, 1, 512] - - [451, 93.88] - - - [40448, 19969, 1, 512] - - [451, 93.91] - - - [40960, 19969, 1, 512] - - [451, 93.873] - - - [40960, 20481, 1, 512] - - [455, 93.961] - - - [41472, 20481, 1, 512] - - [454, 93.952] - - - [41984, 20481, 1, 512] - - [454, 93.964] - - - [41984, 20993, 1, 512] - - [454, 93.942] - - - [42496, 20993, 1, 512] - - [453, 93.941] - - - [43008, 20993, 1, 512] - - [454, 93.932] - - - [43008, 21505, 1, 512] - - [451, 94.033] - - - [43520, 21505, 1, 512] - - [451, 94.02] - - - [44032, 21505, 1, 512] - - [451, 94.021] - - - [44032, 22017, 1, 512] - - [453, 93.998] - - - [44544, 22017, 1, 512] - - [451, 94.002] - - - [45056, 22017, 1, 512] - - [453, 94.014] - - - [45056, 22529, 1, 512] - - [454, 94.046] - - - [45568, 22529, 1, 512] - - [454, 94.041] - - - [46080, 22529, 1, 512] - - [454, 94.05] - - - [46080, 23041, 1, 512] - - [453, 94.066] - - - [46592, 23041, 1, 512] - - [453, 94.086] - - - [47104, 23041, 1, 512] - - [453, 94.077] - - - [47104, 23553, 1, 512] - - [454, 94.063] - - - [47616, 23553, 1, 512] - - [454, 94.072] - - - [48128, 23553, 1, 512] - - [454, 94.066] - - - [48128, 24065, 1, 512] - - [451, 94.079] - - - [48640, 24065, 1, 512] - - [455, 94.034] - - - [49152, 24065, 1, 512] - - [454, 94.043] - - - [49152, 24577, 1, 512] - - [454, 94.063] - - - [49664, 24577, 1, 512] - - [451, 94.066] - - - [50176, 24577, 1, 512] - - [454, 94.095] - - - [50176, 25089, 1, 512] - - [451, 94.099] - - - [50688, 25089, 1, 512] - - [451, 94.124] - - - [51200, 25089, 1, 512] - - [451, 94.144] - - - [51200, 25601, 1, 512] - - [455, 94.128] - - - [51712, 25601, 1, 512] - - [454, 94.139] - - - [52224, 25601, 1, 512] - - [451, 94.136] - - - [52224, 26113, 1, 512] - - [451, 94.152] - - - [52736, 26113, 1, 512] - - [451, 94.102] - - - [53248, 26113, 1, 512] - - [458, 94.08] - - - [53248, 26625, 1, 512] - - [454, 94.144] - - - [53760, 26625, 1, 512] - - [454, 94.122] - - - [54272, 26625, 1, 512] - - [454, 94.133] - - - [54272, 27137, 1, 512] - - [451, 94.131] - - - [54784, 27137, 1, 512] - - [453, 94.109] - - - [55296, 27137, 1, 512] - - [454, 94.135] - - - [55296, 27649, 1, 512] - - [451, 94.162] - - - [55808, 27649, 1, 512] - - [453, 94.173] - - - [56320, 27649, 1, 512] - - [453, 94.159] - - - [56320, 28161, 1, 512] - - [451, 94.146] - - - [56832, 28161, 1, 512] - - [453, 94.141] - - - [57344, 28161, 1, 512] - - [455, 94.146] - - - [57344, 28673, 1, 512] - - [451, 94.186] - - - [57856, 28673, 1, 512] - - [451, 94.194] - - - [58368, 28673, 1, 512] - - [451, 94.195] - - - [58368, 29185, 1, 512] - - [451, 94.181] - - - [58880, 29185, 1, 512] - - [451, 94.176] - - - [59392, 29185, 1, 512] - - [451, 94.193] - - - [59392, 29697, 1, 512] - - [454, 94.176] - - - [59904, 29697, 1, 512] - - [453, 94.158] - - - [60416, 29697, 1, 512] - - [454, 94.188] - - - [60416, 30209, 1, 512] - - [453, 94.2] - - - [60928, 30209, 1, 512] - - [453, 94.18] - - - [61440, 30209, 1, 512] - - [454, 94.18] - - - [61440, 30721, 1, 512] - - [454, 94.201] - - - [61952, 30721, 1, 512] - - [454, 94.198] - - - [62464, 30721, 1, 512] - - [453, 94.223] - - - [62464, 31233, 1, 512] - - [458, 94.189] - - - [62976, 31233, 1, 512] - - [451, 94.222] - - - [63488, 31233, 1, 512] - - [451, 94.226] - - - [63488, 31745, 1, 512] - - [451, 94.214] - - - [64000, 31745, 1, 512] - - [454, 94.226] - - - [64512, 31745, 1, 512] - - [453, 94.219] - - - [64512, 32257, 1, 512] - - [451, 94.262] - - - [65024, 32257, 1, 512] - - [453, 94.258] - - - [65536, 32257, 1, 512] - - [453, 94.138] - - - [65536, 32769, 1, 512] - - [453, 94.156] - - - [66048, 32769, 1, 512] - - [451, 94.24] - - - [66560, 32769, 1, 512] - - [451, 94.235] - - - [66560, 33281, 1, 512] - - [453, 94.254] - - - [67072, 33281, 1, 512] - - [453, 94.221] - - - [67584, 33281, 1, 512] - - [453, 94.24] - - - [67584, 33793, 1, 512] - - [454, 94.253] - - - [68096, 33793, 1, 512] - - [451, 94.224] - - - [68608, 33793, 1, 512] - - [454, 94.248] - - - [68608, 34305, 1, 512] - - [451, 94.251] - - - [69120, 34305, 1, 512] - - [451, 94.246] - - - [69632, 34305, 1, 512] - - [451, 94.255] - - - [69632, 34817, 1, 512] - - [453, 94.251] - - - [70144, 34817, 1, 512] - - [451, 94.254] - - - [70656, 34817, 1, 512] - - [451, 94.258] - - - [70656, 35329, 1, 512] - - [453, 94.267] - - - [71168, 35329, 1, 512] - - [453, 94.25] - - - [71680, 35329, 1, 512] - - [451, 94.251] - - - [71680, 35841, 1, 512] - - [451, 94.294] - - - [72192, 35841, 1, 512] - - [451, 94.293] - - - [72704, 35841, 1, 512] - - [451, 94.303] - - - [72704, 36353, 1, 512] - - [451, 94.281] - - - [73216, 36353, 1, 512] - - [451, 94.279] - - - [73728, 36353, 1, 512] - - [451, 94.279] - - - [73728, 36865, 1, 512] - - [453, 94.283] - - - [74240, 36865, 1, 512] - - [451, 94.288] - - - [74752, 36865, 1, 512] - - [453, 94.288] - - - [74752, 37377, 1, 512] - - [453, 94.284] - - - [75264, 37377, 1, 512] - - [453, 94.298] - - - [75776, 37377, 1, 512] - - [458, 94.283] - - - [75776, 37889, 1, 512] - - [453, 94.308] - - - [76288, 37889, 1, 512] - - [453, 94.303] - - - [76800, 37889, 1, 512] - - [451, 94.299] - - - [76800, 38401, 1, 512] - - [332, 93.644] - - - [77312, 38401, 1, 512] - - [361, 93.507] - - - [77824, 38401, 1, 512] - - [332, 93.693] - - - [77824, 38913, 1, 512] - - [353, 91.448] - - - [78336, 38913, 1, 512] - - [326, 91.909] - - - [78848, 38913, 1, 512] - - [395, 91.037] - - - [78848, 39425, 1, 512] - - [332, 93.621] - - - [79360, 39425, 1, 512] - - [326, 93.44] - - - [79872, 39425, 1, 512] - - [332, 93.696] - - - [79872, 39937, 1, 512] - - [373, 90.877] - - - [80384, 39937, 1, 512] - - [326, 91.695] - - - [80896, 39937, 1, 512] - - [319, 90.885] - - - [80896, 40449, 1, 512] - - [332, 93.617] - - - [81408, 40449, 1, 512] - - [326, 93.427] - - - [81920, 40449, 1, 512] - - [332, 91.712] - - - [81920, 40961, 1, 512] - - [326, 90.511] - - - [82432, 40961, 1, 512] - - [353, 91.602] - - - [82944, 40961, 1, 512] - - [353, 91.358] - - - [82944, 41473, 1, 512] - - [332, 93.573] - - - [83456, 41473, 1, 512] - - [332, 93.308] - - - [83968, 41473, 1, 512] - - [332, 93.623] - - - [83968, 41985, 1, 512] - - [318, 90.156] - - - [84480, 41985, 1, 512] - - [353, 91.627] - - - [84992, 41985, 1, 512] - - [319, 90.837] - - - [84992, 42497, 1, 512] - - [332, 93.555] - - - [85504, 42497, 1, 512] - - [326, 93.266] - - - [86016, 42497, 1, 512] - - [332, 93.468] - - - [86016, 43009, 1, 512] - - [319, 90.573] - - - [86528, 43009, 1, 512] - - [353, 91.509] - - - [87040, 43009, 1, 512] - - [395, 90.146] - - - [87040, 43521, 1, 512] - - [332, 93.44] - - - [87552, 43521, 1, 512] - - [326, 93.241] - - - [88064, 43521, 1, 512] - - [332, 93.529] - - - [88064, 44033, 1, 512] - - [319, 90.845] - - - [88576, 44033, 1, 512] - - [353, 91.526] - - - [89088, 44033, 1, 512] - - [319, 90.752] - - - [89088, 44545, 1, 512] - - [332, 93.431] - - - [89600, 44545, 1, 512] - - [326, 93.252] - - - [90112, 44545, 1, 512] - - [326, 92.881] - - - [90112, 45057, 1, 512] - - [395, 90.15] - - - [90624, 45057, 1, 512] - - [372, 90.922] - - - [91136, 45057, 1, 512] - - [319, 89.698] - - - [91136, 45569, 1, 512] - - [332, 93.37] - - - [91648, 45569, 1, 512] - - [332, 93.132] - - - [92160, 45569, 1, 512] - - [332, 93.414] - - - [92160, 46081, 1, 512] - - [319, 90.737] - - - [92672, 46081, 1, 512] - - [353, 91.146] - - - [93184, 46081, 1, 512] - - [319, 89.816] - - - [93184, 46593, 1, 512] - - [332, 93.287] - - - [93696, 46593, 1, 512] - - [326, 93.15] - - - [94208, 46593, 1, 512] - - [332, 93.33] - - - [94208, 47105, 1, 512] - - [372, 90.024] - - - [94720, 47105, 1, 512] - - [353, 91.146] - - - [95232, 47105, 1, 512] - - [319, 89.821] - - - [95232, 47617, 1, 512] - - [332, 93.268] - - - [95744, 47617, 1, 512] - - [326, 93.09] - - - [96256, 47617, 1, 512] - - [332, 93.273] - - - [96256, 48129, 1, 512] - - [319, 89.51] - - - [96768, 48129, 1, 512] - - [353, 91.09] - - - [97280, 48129, 1, 512] - - [400, 89.477] - - - [97280, 48641, 1, 512] - - [332, 93.124] - - - [97792, 48641, 1, 512] - - [326, 93.002] - - - [98304, 48641, 1, 512] - - [326, 90.881] - - - [98304, 49153, 1, 512] - - [326, 89.65] - - - [98816, 49153, 1, 512] - - [372, 90.759] - - - [99328, 49153, 1, 512] - - [364, 88.096] - - - [99328, 49665, 1, 512] - - [332, 93.094] - - - [99840, 49665, 1, 512] - - [332, 92.832] - - - [100352, 49665, 1, 512] - - [332, 92.984] - - - [100352, 50177, 1, 512] - - [401, 89.171] - - - [100864, 50177, 1, 512] - - [395, 91.014] - - - [101376, 50177, 1, 512] - - [402, 88.937] - - - [101376, 50689, 1, 512] - - [375, 92.814] - - - [101888, 50689, 1, 512] - - [395, 92.728] - - - [102400, 50689, 1, 512] - - [375, 92.708] - - - [102400, 51201, 1, 512] - - [376, 89.1] - - - [102912, 51201, 1, 512] - - [372, 90.032] - - - [103424, 51201, 1, 512] - - [401, 89.017] - - - [103424, 51713, 1, 512] - - [375, 92.701] - - - [103936, 51713, 1, 512] - - [395, 92.663] - - - [104448, 51713, 1, 512] - - [375, 92.685] - - - [104448, 52225, 1, 512] - - [403, 89.025] - - - [104960, 52225, 1, 512] - - [404, 89.973] - - - [105472, 52225, 1, 512] - - [403, 89.085] - - - [105472, 52737, 1, 512] - - [405, 92.166] - - - [105984, 52737, 1, 512] - - [395, 92.58] - - - [106496, 52737, 1, 512] - - [327, 91.727] - - - [106496, 53249, 1, 512] - - [372, 89.65] - - - [107008, 53249, 1, 512] - - [404, 89.721] - - - [107520, 53249, 1, 512] - - [403, 88.422] - - - [107520, 53761, 1, 512] - - [375, 92.495] - - - [108032, 53761, 1, 512] - - [395, 92.46] - - - [108544, 53761, 1, 512] - - [375, 92.463] - - - [108544, 54273, 1, 512] - - [403, 89.015] - - - [109056, 54273, 1, 512] - - [404, 89.897] - - - [109568, 54273, 1, 512] - - [403, 88.893] - - - [109568, 54785, 1, 512] - - [405, 91.957] - - - [110080, 54785, 1, 512] - - [395, 92.337] - - - [110592, 54785, 1, 512] - - [405, 91.851] - - - [110592, 55297, 1, 512] - - [403, 88.458] - - - [111104, 55297, 1, 512] - - [404, 89.939] - - - [111616, 55297, 1, 512] - - [403, 88.902] - - - [111616, 55809, 1, 512] - - [405, 91.636] - - - [112128, 55809, 1, 512] - - [395, 92.221] - - - [112640, 55809, 1, 512] - - [407, 90.918] - - - [112640, 56321, 1, 512] - - [401, 88.692] - - - [113152, 56321, 1, 512] - - [404, 89.927] - - - [113664, 56321, 1, 512] - - [401, 88.997] - - - [113664, 56833, 1, 512] - - [407, 90.952] - - - [114176, 56833, 1, 512] - - [405, 91.61] - - - [114688, 56833, 1, 512] - - [326, 89.574] - - - [114688, 57345, 1, 512] - - [353, 87.847] - - - [115200, 57345, 1, 512] - - [408, 89.34] - - - [115712, 57345, 1, 512] - - [378, 87.387] - - - [115712, 57857, 1, 512] - - [409, 90.838] - - - [116224, 57857, 1, 512] - - [405, 91.423] - - - [116736, 57857, 1, 512] - - [407, 90.883] - - - [116736, 58369, 1, 512] - - [401, 88.849] - - - [117248, 58369, 1, 512] - - [404, 89.793] - - - [117760, 58369, 1, 512] - - [401, 89.007] - - - [117760, 58881, 1, 512] - - [407, 90.924] - - - [118272, 58881, 1, 512] - - [405, 91.199] - - - [118784, 58881, 1, 512] - - [407, 90.844] - - - [118784, 59393, 1, 512] - - [403, 88.376] - - - [119296, 59393, 1, 512] - - [404, 89.884] - - - [119808, 59393, 1, 512] - - [401, 88.935] - - - [119808, 59905, 1, 512] - - [407, 90.889] - - - [120320, 59905, 1, 512] - - [407, 90.946] - - - [120832, 59905, 1, 512] - - [407, 90.899] - - - [120832, 60417, 1, 512] - - [401, 88.84] - - - [121344, 60417, 1, 512] - - [404, 89.913] - - - [121856, 60417, 1, 512] - - [403, 88.925] - - - [121856, 60929, 1, 512] - - [407, 90.869] - - - [122368, 60929, 1, 512] - - [407, 90.994] - - - [122880, 60929, 1, 512] - - [405, 88.807] - - - [122880, 61441, 1, 512] - - [412, 86.857] - - - [123392, 61441, 1, 512] - - [413, 89.659] - - - [123904, 61441, 1, 512] - - [403, 88.534] - - - [123904, 61953, 1, 512] - - [407, 90.867] - - - [124416, 61953, 1, 512] - - [407, 90.997] - - - [124928, 61953, 1, 512] - - [407, 90.907] - - - [124928, 62465, 1, 512] - - [401, 88.927] - - - [125440, 62465, 1, 512] - - [404, 89.782] - - - [125952, 62465, 1, 512] - - [401, 88.881] - - - [125952, 62977, 1, 512] - - [407, 90.919] - - - [126464, 62977, 1, 512] - - [407, 91.033] - - - [126976, 62977, 1, 512] - - [407, 90.82] - - - [126976, 63489, 1, 512] - - [403, 88.3] - - - [127488, 63489, 1, 512] - - [404, 89.893] - - - [128000, 63489, 1, 512] - - [403, 88.924] - - - [3584, 6657, 1, 512] - - [341, 85.708] - - - [3584, 6145, 1, 512] - - [353, 83.855] - - - [3072, 5633, 1, 512] - - [319, 85.17] - - - [3072, 5121, 1, 512] - - [318, 84.434] - - - [2560, 4609, 1, 512] - - [303, 78.729] - - - [2560, 4097, 1, 512] - - [326, 81.236] - - - [2048, 3585, 1, 512] - - [310, 74.674] - - - [2048, 3073, 1, 512] - - [352, 72.101] - - - [1536, 2561, 1, 512] - - [297, 65.323] - - - [1536, 2049, 1, 512] - - [348, 65.718] - - - [1024, 1537, 1, 512] - - [346, 58.197] - - - [1024, 1025, 1, 512] - - [344, 43.401] - - - [512, 513, 1, 512] - - [343, 21.264] - - - [512, 1, 1, 512] - - [383, 0.034] - - - [6656, 4096, 1, 512] - - [19, 87.137] - - - [6144, 4992, 1, 512] - - [19, 87.63] - - - [8192, 3328, 1, 512] - - [19, 87.21] - - - [8320, 4096, 1, 512] - - [19, 87.851] - - - [7040, 4096, 1, 512] - - [18, 91.956] - - - [7040, 4096, 1, 512] - - [13, 90.075] - - - [8448, 3840, 1, 512] - - [18, 92.567] - - - [8448, 3840, 1, 512] - - [17, 90.593] - - - [7680, 4224, 1, 512] - - [18, 92.491] - - - [7680, 4224, 1, 512] - - [17, 90.538] - - - [1024, 513, 1, 512] - - [384, 33.669] - - - [1536, 513, 1, 512] - - [384, 31.35] - - - [2048, 513, 1, 512] - - [386, 44.808] - - - [2048, 1025, 1, 512] - - [295, 55.673] - - - [2560, 1025, 1, 512] - - [289, 54.447] - - - [3072, 1025, 1, 512] - - [291, 59.64] - - - [3072, 1537, 1, 512] - - [303, 68.351] - - - [3584, 1537, 1, 512] - - [392, 66.406] - - - [1024, 1, 1, 512] - - [415, 0.082] - - - [1152, 385, 1, 384] - - [264, 25.247] - - - [1536, 385, 1, 384] - - [256, 30.548] - - - [1536, 769, 1, 384] - - [416, 36.571] - - - [1920, 769, 1, 384] - - [416, 51.306] - - - [2304, 769, 1, 384] - - [344, 50.762] - - - [3456, 1536, 1, 384] - - [293, 63.928] - - - [3840, 1536, 1, 384] - - [392, 70.93] - - - [3840, 1537, 1, 384] - - [392, 67.505] - - - [3840, 1920, 1, 384] - - [303, 71.961] - - - [4224, 1920, 1, 384] - - [389, 77.453] - - - [4224, 1921, 1, 384] - - [303, 70.21] - - - [4224, 2304, 1, 384] - - [293, 76.587] - - - [4608, 1920, 1, 384] - - [314, 77.994] - - - [4608, 1921, 1, 384] - - [389, 73.504] - - - [4608, 2304, 1, 384] - - [389, 79.328] - - - [4608, 2305, 1, 384] - - [393, 72.307] - - - [4992, 2304, 1, 384] - - [296, 79.632] - - - [4992, 2305, 1, 384] - - [303, 76.379] - - - [5376, 2304, 1, 384] - - [422, 80.46] - - - [5376, 2305, 1, 384] - - [303, 75.242] - - - [5376, 2689, 1, 384] - - [303, 78.166] - - - [5760, 2689, 1, 384] - - [423, 79.724] - - - [6144, 2689, 1, 384] - - [424, 78.35] - - - [6144, 3073, 1, 384] - - [425, 79.283] - - - [6528, 3073, 1, 384] - - [425, 78.275] - - - [6528, 3456, 1, 384] - - [389, 84.946] - - - [6912, 3073, 1, 384] - - [425, 77.891] - - - [6912, 3456, 1, 384] - - [389, 84.848] - - - [6912, 3457, 1, 384] - - [324, 82.06] - - - [7296, 3456, 1, 384] - - [323, 86.422] - - - [7296, 3457, 1, 384] - - [353, 81.31] - - - [7296, 3840, 1, 384] - - [389, 86.038] - - - [7680, 3456, 1, 384] - - [324, 85.949] - - - [7680, 3457, 1, 384] - - [365, 80.973] - - - [7680, 3840, 1, 384] - - [389, 86.09] - - - [7680, 3841, 1, 384] - - [319, 84.741] - - - [8064, 3840, 1, 384] - - [389, 86.513] - - - [8064, 3841, 1, 384] - - [319, 84.29] - - - [8064, 4224, 1, 384] - - [318, 88.004] - - - [8448, 3840, 1, 384] - - [323, 88.267] - - - [8448, 3841, 1, 384] - - [318, 84.415] - - - [8448, 4224, 1, 384] - - [318, 88.229] - - - [8448, 4225, 1, 384] - - [319, 84.403] - - - [8832, 4224, 1, 384] - - [319, 88.014] - - - [8832, 4225, 1, 384] - - [319, 84.344] - - - [8832, 4608, 1, 384] - - [363, 88.443] - - - [9216, 4224, 1, 384] - - [318, 87.945] - - - [9216, 4225, 1, 384] - - [341, 84.185] - - - [9216, 4608, 1, 384] - - [324, 88.531] - - - [9216, 4609, 1, 384] - - [356, 84.645] - - - [9600, 4608, 1, 384] - - [324, 88.745] - - - [9600, 4609, 1, 384] - - [331, 84.869] - - - [9600, 4992, 1, 384] - - [323, 89.231] - - - [9984, 4608, 1, 384] - - [323, 88.984] - - - [9984, 4609, 1, 384] - - [331, 85.085] - - - [9984, 4992, 1, 384] - - [323, 89.618] - - - [9984, 4993, 1, 384] - - [365, 86.111] - - - [10368, 4992, 1, 384] - - [323, 89.817] - - - [10368, 4993, 1, 384] - - [323, 86.747] - - - [10368, 5376, 1, 384] - - [324, 90.255] - - - [10752, 4992, 1, 384] - - [323, 90.178] - - - [10752, 4993, 1, 384] - - [324, 87.092] - - - [10752, 5376, 1, 384] - - [323, 88.993] - - - [10752, 5377, 1, 384] - - [319, 87.906] - - - [11136, 5376, 1, 384] - - [324, 89.472] - - - [11136, 5377, 1, 384] - - [399, 87.185] - - - [11136, 5760, 1, 384] - - [318, 90.428] - - - [11520, 5376, 1, 384] - - [324, 89.902] - - - [11520, 5377, 1, 384] - - [331, 86.901] - - - [11520, 5760, 1, 384] - - [324, 90.917] - - - [11520, 5761, 1, 384] - - [319, 88.229] - - - [11904, 5760, 1, 384] - - [323, 89.89] - - - [11904, 5761, 1, 384] - - [319, 88.307] - - - [11904, 6144, 1, 384] - - [356, 90.669] - - - [12288, 5760, 1, 384] - - [324, 90.44] - - - [12288, 5761, 1, 384] - - [319, 87.842] - - - [12288, 6144, 1, 384] - - [324, 91.197] - - - [12288, 6145, 1, 384] - - [318, 85.506] - - - [12672, 6144, 1, 384] - - [331, 90.509] - - - [12672, 6145, 1, 384] - - [426, 85.983] - - - [12672, 6528, 1, 384] - - [324, 91.553] - - - [13056, 6144, 1, 384] - - [331, 90.829] - - - [13056, 6145, 1, 384] - - [313, 85.712] - - - [13056, 6528, 1, 384] - - [324, 91.077] - - - [13056, 6529, 1, 384] - - [319, 88.884] - - - [13440, 6528, 1, 384] - - [324, 91.506] - - - [13440, 6529, 1, 384] - - [324, 89.408] - - - [13440, 6912, 1, 384] - - [323, 91.7] - - - [13824, 6529, 1, 384] - - [331, 88.936] - - - [13824, 6912, 1, 384] - - [331, 91.274] - - - [13824, 6913, 1, 384] - - [318, 89.693] - - - [14208, 6913, 1, 384] - - [319, 89.505] - - - [15744, 7680, 1, 384] - - [323, 92.015] - - - [16128, 7680, 1, 384] - - [323, 92.486] - - - [16128, 7681, 1, 384] - - [324, 90.194] - - - [16128, 8064, 1, 384] - - [323, 92.246] - - - [16512, 8064, 1, 384] - - [324, 92.676] - - - [16512, 8065, 1, 384] - - [356, 90.185] - - - [16512, 8448, 1, 384] - - [323, 92.297] - - - [16896, 8064, 1, 384] - - [319, 92.582] - - - [16896, 8065, 1, 384] - - [319, 90.64] - - - [16896, 8448, 1, 384] - - [323, 92.308] - - - [16896, 8449, 1, 384] - - [319, 90.752] - - - [17280, 8448, 1, 384] - - [324, 92.701] - - - [17280, 8449, 1, 384] - - [331, 90.566] - - - [17280, 8832, 1, 384] - - [319, 92.525] - - - [17664, 8448, 1, 384] - - [323, 92.61] - - - [17664, 8449, 1, 384] - - [356, 90.366] - - - [17664, 8832, 1, 384] - - [319, 92.311] - - - [17664, 8833, 1, 384] - - [319, 90.272] - - - [18048, 8832, 1, 384] - - [318, 92.224] - - - [18048, 8833, 1, 384] - - [319, 90.447] - - - [18432, 8832, 1, 384] - - [319, 92.388] - - - [18432, 8833, 1, 384] - - [319, 90.586] - - - [18432, 9217, 1, 384] - - [318, 87.476] - - - [18816, 9217, 1, 384] - - [315, 87.941] - - - [18816, 9600, 1, 384] - - [319, 91.77] - - - [19200, 9217, 1, 384] - - [306, 87.68] - - - [19200, 9600, 1, 384] - - [319, 91.9] - - - [19200, 9601, 1, 384] - - [365, 89.907] - - - [19584, 9600, 1, 384] - - [319, 92.176] - - - [19584, 9601, 1, 384] - - [319, 89.913] - - - [19584, 9984, 1, 384] - - [365, 91.614] - - - [19968, 9600, 1, 384] - - [319, 91.901] - - - [19968, 9601, 1, 384] - - [365, 90.193] - - - [19968, 9984, 1, 384] - - [365, 91.658] - - - [19968, 9985, 1, 384] - - [319, 89.945] - - - [20352, 9984, 1, 384] - - [365, 91.605] - - - [20352, 9985, 1, 384] - - [319, 90.077] - - - [20352, 10368, 1, 384] - - [365, 91.185] - - - [20736, 9984, 1, 384] - - [319, 91.566] - - - [20736, 9985, 1, 384] - - [319, 90.122] - - - [20736, 10368, 1, 384] - - [365, 91.18] - - - [20736, 10369, 1, 384] - - [319, 89.815] - - - [21120, 10368, 1, 384] - - [319, 91.153] - - - [21120, 10369, 1, 384] - - [319, 89.796] - - - [21120, 10752, 1, 384] - - [318, 91.279] - - - [21504, 10368, 1, 384] - - [319, 91.366] - - - [21504, 10369, 1, 384] - - [319, 90.032] - - - [21504, 10752, 1, 384] - - [318, 91.51] - - - [21504, 10753, 1, 384] - - [319, 89.923] - - - [21888, 10752, 1, 384] - - [341, 91.499] - - - [21888, 10753, 1, 384] - - [319, 89.938] - - - [21888, 11136, 1, 384] - - [319, 91.058] - - - [22272, 10752, 1, 384] - - [318, 91.447] - - - [22272, 10753, 1, 384] - - [319, 90.046] - - - [22272, 11136, 1, 384] - - [319, 91.191] - - - [22272, 11137, 1, 384] - - [319, 89.78] - - - [22656, 11136, 1, 384] - - [365, 91.097] - - - [22656, 11137, 1, 384] - - [324, 89.585] - - - [22656, 11520, 1, 384] - - [365, 91.124] - - - [23040, 11136, 1, 384] - - [319, 91.296] - - - [23040, 11137, 1, 384] - - [319, 89.887] - - - [23040, 11520, 1, 384] - - [319, 91.189] - - - [23040, 11521, 1, 384] - - [319, 89.813] - - - [23424, 11520, 1, 384] - - [365, 91.143] - - - [23424, 11521, 1, 384] - - [319, 89.614] - - - [23424, 11904, 1, 384] - - [318, 90.942] - - - [23808, 11520, 1, 384] - - [319, 91.298] - - - [23808, 11521, 1, 384] - - [319, 89.55] - - - [23808, 11904, 1, 384] - - [319, 91.072] - - - [23808, 11905, 1, 384] - - [319, 89.627] - - - [24192, 11904, 1, 384] - - [319, 91.168] - - - [24192, 11905, 1, 384] - - [319, 89.7] - - - [24192, 12288, 1, 384] - - [331, 90.654] - - - [24576, 11904, 1, 384] - - [319, 91.348] - - - [24576, 11905, 1, 384] - - [319, 89.899] - - - [24576, 12288, 1, 384] - - [331, 90.83] - - - [24576, 12289, 1, 384] - - [319, 86.891] - - - [24960, 12288, 1, 384] - - [331, 90.658] - - - [24960, 12289, 1, 384] - - [341, 87.054] - - - [24960, 12672, 1, 384] - - [319, 91.106] - - - [25344, 12288, 1, 384] - - [356, 90.638] - - - [25344, 12289, 1, 384] - - [341, 86.951] - - - [25344, 12672, 1, 384] - - [361, 90.39] - - - [25344, 12673, 1, 384] - - [365, 89.049] - - - [25728, 12672, 1, 384] - - [321, 90.787] - - - [25728, 12673, 1, 384] - - [365, 89.312] - - - [25728, 13056, 1, 384] - - [321, 90.906] - - - [26112, 12673, 1, 384] - - [319, 89.719] - - - [26112, 13056, 1, 384] - - [319, 91.096] - - - [26112, 13057, 1, 384] - - [319, 89.816] - - - [26496, 13057, 1, 384] - - [319, 89.639] - - - [26880, 13057, 1, 384] - - [319, 89.517] - - - [27648, 13825, 1, 384] - - [319, 90.253] - - - [28032, 13824, 1, 384] - - [318, 91.635] - - - [28032, 13825, 1, 384] - - [341, 90.037] - - - [28416, 13824, 1, 384] - - [319, 90.778] - - - [28416, 13825, 1, 384] - - [319, 89.362] - - - [28416, 14208, 1, 384] - - [323, 90.7] - - - [28416, 14209, 1, 384] - - [319, 89.409] - - - [28800, 14208, 1, 384] - - [319, 91.067] - - - [28800, 14209, 1, 384] - - [319, 89.958] - - - [28800, 14592, 1, 384] - - [319, 91.328] - - - [29184, 14208, 1, 384] - - [321, 91.283] - - - [29184, 14209, 1, 384] - - [319, 90.182] - - - [29184, 14592, 1, 384] - - [319, 91.462] - - - [29184, 14593, 1, 384] - - [319, 90.145] - - - [29568, 14592, 1, 384] - - [361, 91.223] - - - [29568, 14593, 1, 384] - - [319, 89.933] - - - [29568, 14976, 1, 384] - - [319, 91.216] - - - [29952, 14592, 1, 384] - - [319, 91.391] - - - [29952, 14593, 1, 384] - - [319, 90.049] - - - [29952, 14976, 1, 384] - - [319, 91.417] - - - [29952, 14977, 1, 384] - - [319, 90.116] - - - [30336, 14976, 1, 384] - - [319, 91.59] - - - [30336, 14977, 1, 384] - - [319, 90.319] - - - [30720, 14976, 1, 384] - - [319, 91.861] - - - [30720, 14977, 1, 384] - - [319, 90.472] - - - [30720, 15361, 1, 384] - - [341, 87.78] - - - [31104, 15361, 1, 384] - - [319, 87.786] - - - [31104, 15744, 1, 384] - - [321, 91.504] - - - [31488, 15361, 1, 384] - - [319, 87.911] - - - [31488, 15744, 1, 384] - - [365, 91.78] - - - [31488, 15745, 1, 384] - - [319, 90.298] - - - [31872, 15744, 1, 384] - - [318, 91.559] - - - [31872, 15745, 1, 384] - - [319, 90.275] - - - [31872, 16128, 1, 384] - - [319, 91.411] - - - [32256, 15744, 1, 384] - - [319, 91.873] - - - [32256, 15745, 1, 384] - - [319, 90.552] - - - [32256, 16128, 1, 384] - - [365, 91.599] - - - [32256, 16129, 1, 384] - - [319, 90.58] - - - [32640, 16128, 1, 384] - - [365, 91.339] - - - [32640, 16129, 1, 384] - - [319, 90.118] - - - [32640, 16512, 1, 384] - - [319, 91.364] - - - [33024, 16128, 1, 384] - - [319, 91.623] - - - [33024, 16129, 1, 384] - - [319, 90.74] - - - [33024, 16512, 1, 384] - - [319, 91.948] - - - [33024, 16513, 1, 384] - - [319, 90.728] - - - [33408, 16512, 1, 384] - - [319, 91.999] - - - [33408, 16513, 1, 384] - - [319, 90.727] - - - [33408, 16896, 1, 384] - - [318, 92.053] - - - [33792, 16512, 1, 384] - - [321, 91.688] - - - [33792, 16513, 1, 384] - - [365, 90.563] - - - [33792, 16896, 1, 384] - - [365, 91.969] - - - [33792, 16897, 1, 384] - - [365, 90.67] - - - [34176, 16896, 1, 384] - - [365, 91.996] - - - [34176, 16897, 1, 384] - - [365, 90.551] - - - [34176, 17280, 1, 384] - - [321, 91.649] - - - [34560, 16896, 1, 384] - - [321, 91.942] - - - [34560, 16897, 1, 384] - - [319, 90.658] - - - [34560, 17280, 1, 384] - - [319, 91.866] - - - [34560, 17281, 1, 384] - - [365, 90.688] - - - [34944, 17280, 1, 384] - - [365, 91.95] - - - [34944, 17281, 1, 384] - - [319, 90.649] - - - [34944, 17664, 1, 384] - - [365, 92.0] - - - [35328, 17280, 1, 384] - - [319, 92.039] - - - [35328, 17281, 1, 384] - - [319, 90.896] - - - [35328, 17664, 1, 384] - - [319, 92.015] - - - [35328, 17665, 1, 384] - - [319, 90.848] - - - [35712, 17664, 1, 384] - - [319, 92.006] - - - [35712, 17665, 1, 384] - - [319, 90.82] - - - [35712, 18048, 1, 384] - - [319, 91.873] - - - [36096, 17664, 1, 384] - - [319, 91.969] - - - [36096, 17665, 1, 384] - - [319, 90.832] - - - [36096, 18048, 1, 384] - - [319, 91.818] - - - [36096, 18049, 1, 384] - - [319, 90.912] - - - [36480, 18048, 1, 384] - - [319, 91.776] - - - [36480, 18049, 1, 384] - - [319, 90.668] - - - [36480, 18432, 1, 384] - - [341, 91.937] - - - [36864, 18048, 1, 384] - - [321, 91.789] - - - [36864, 18049, 1, 384] - - [365, 90.841] - - - [36864, 18432, 1, 384] - - [324, 91.846] - - - [36864, 18433, 1, 384] - - [319, 88.35] - - - [37248, 18432, 1, 384] - - [341, 91.988] - - - [37248, 18433, 1, 384] - - [318, 88.591] - - - [37248, 18816, 1, 384] - - [319, 91.964] - - - [37632, 18432, 1, 384] - - [319, 91.966] - - - [37632, 18433, 1, 384] - - [318, 88.577] - - - [37632, 18816, 1, 384] - - [319, 92.008] - - - [37632, 18817, 1, 384] - - [319, 90.943] - - - [38016, 18816, 1, 384] - - [365, 92.136] - - - [38016, 18817, 1, 384] - - [365, 90.899] - - - [38016, 19200, 1, 384] - - [365, 92.093] - - - [38400, 18816, 1, 384] - - [365, 92.317] - - - [38400, 18817, 1, 384] - - [319, 91.151] - - - [38400, 19200, 1, 384] - - [365, 92.183] - - - [38400, 19201, 1, 384] - - [319, 91.082] - - - [38784, 19200, 1, 384] - - [319, 92.27] - - - [38784, 19201, 1, 384] - - [319, 91.11] - - - [38784, 19584, 1, 384] - - [319, 92.215] - - - [39168, 19200, 1, 384] - - [319, 92.209] - - - [39168, 19201, 1, 384] - - [319, 90.963] - - - [39168, 19584, 1, 384] - - [319, 92.207] - - - [39168, 19585, 1, 384] - - [319, 90.933] - - - [39552, 19584, 1, 384] - - [365, 91.883] - - - [39552, 19585, 1, 384] - - [365, 90.819] - - - [39552, 19968, 1, 384] - - [321, 92.207] - - - [39936, 19584, 1, 384] - - [319, 92.285] - - - [39936, 19585, 1, 384] - - [365, 91.031] - - - [39936, 19968, 1, 384] - - [365, 92.277] - - - [39936, 19969, 1, 384] - - [319, 91.13] - - - [40320, 19968, 1, 384] - - [365, 92.306] - - - [40320, 19969, 1, 384] - - [319, 91.26] - - - [40320, 20352, 1, 384] - - [319, 92.286] - - - [40704, 19968, 1, 384] - - [365, 92.216] - - - [40704, 19969, 1, 384] - - [319, 91.19] - - - [40704, 20352, 1, 384] - - [319, 92.247] - - - [40704, 20353, 1, 384] - - [319, 91.179] - - - [41088, 20352, 1, 384] - - [319, 92.331] - - - [41088, 20353, 1, 384] - - [319, 91.267] - - - [41088, 20736, 1, 384] - - [319, 92.306] - - - [41472, 20352, 1, 384] - - [319, 92.498] - - - [41472, 20353, 1, 384] - - [319, 91.431] - - - [41472, 20736, 1, 384] - - [319, 92.428] - - - [41472, 20737, 1, 384] - - [319, 91.473] - - - [41856, 20736, 1, 384] - - [319, 92.475] - - - [41856, 20737, 1, 384] - - [319, 91.304] - - - [41856, 21120, 1, 384] - - [319, 92.535] - - - [42240, 20736, 1, 384] - - [319, 91.857] - - - [42240, 20737, 1, 384] - - [319, 91.012] - - - [42240, 21120, 1, 384] - - [318, 92.015] - - - [42240, 21121, 1, 384] - - [319, 91.032] - - - [42624, 21120, 1, 384] - - [324, 92.034] - - - [42624, 21121, 1, 384] - - [324, 90.974] - - - [42624, 21504, 1, 384] - - [324, 92.03] - - - [43008, 21120, 1, 384] - - [319, 92.49] - - - [43008, 21121, 1, 384] - - [319, 91.371] - - - [43008, 21504, 1, 384] - - [324, 92.208] - - - [43008, 21505, 1, 384] - - [319, 88.593] - - - [43392, 21504, 1, 384] - - [319, 92.288] - - - [43392, 21505, 1, 384] - - [341, 89.017] - - - [43392, 21888, 1, 384] - - [365, 92.286] - - - [43776, 21504, 1, 384] - - [341, 92.123] - - - [43776, 21505, 1, 384] - - [319, 88.788] - - - [43776, 21888, 1, 384] - - [365, 92.276] - - - [43776, 21889, 1, 384] - - [319, 91.315] - - - [44160, 21888, 1, 384] - - [319, 92.357] - - - [44160, 21889, 1, 384] - - [319, 91.468] - - - [44160, 22272, 1, 384] - - [319, 92.553] - - - [44544, 21888, 1, 384] - - [365, 92.487] - - - [44544, 21889, 1, 384] - - [319, 91.611] - - - [44544, 22272, 1, 384] - - [319, 92.593] - - - [44544, 22273, 1, 384] - - [319, 91.635] - - - [44928, 384, 1, 384] - - [323, 80.407] - - - [44928, 22272, 1, 384] - - [365, 92.472] - - - [44928, 22273, 1, 384] - - [319, 91.371] - - - [44928, 22656, 1, 384] - - [319, 92.335] - - - [45312, 384, 1, 384] - - [321, 80.964] - - - [45312, 22272, 1, 384] - - [365, 92.386] - - - [45312, 22273, 1, 384] - - [319, 91.18] - - - [45312, 22656, 1, 384] - - [365, 92.306] - - - [45312, 22657, 1, 384] - - [319, 91.31] - - - [45696, 384, 1, 384] - - [319, 81.584] - - - [45696, 22656, 1, 384] - - [319, 92.496] - - - [45696, 22657, 1, 384] - - [319, 91.451] - - - [45696, 23040, 1, 384] - - [319, 92.56] - - - [46080, 384, 1, 384] - - [319, 81.939] - - - [46080, 22656, 1, 384] - - [319, 92.599] - - - [46080, 22657, 1, 384] - - [319, 91.433] - - - [46080, 23040, 1, 384] - - [319, 92.569] - - - [46080, 23041, 1, 384] - - [319, 91.54] - - - [46464, 384, 1, 384] - - [319, 82.292] - - - [46464, 23040, 1, 384] - - [318, 92.528] - - - [46464, 23041, 1, 384] - - [319, 91.536] - - - [46464, 23424, 1, 384] - - [319, 92.516] - - - [46848, 384, 1, 384] - - [319, 82.906] - - - [46848, 23040, 1, 384] - - [319, 92.505] - - - [46848, 23041, 1, 384] - - [319, 91.521] - - - [46848, 23424, 1, 384] - - [319, 92.525] - - - [46848, 23425, 1, 384] - - [319, 91.518] - - - [47232, 384, 1, 384] - - [304, 81.908] - - - [47232, 23424, 1, 384] - - [319, 92.586] - - - [47232, 23425, 1, 384] - - [319, 91.54] - - - [47232, 23808, 1, 384] - - [365, 92.452] - - - [47616, 384, 1, 384] - - [313, 82.205] - - - [47616, 23424, 1, 384] - - [319, 92.67] - - - [47616, 23425, 1, 384] - - [319, 91.574] - - - [47616, 23808, 1, 384] - - [365, 92.452] - - - [47616, 23809, 1, 384] - - [319, 91.579] - - - [48000, 384, 1, 384] - - [319, 79.761] - - - [48000, 23808, 1, 384] - - [365, 92.454] - - - [48000, 23809, 1, 384] - - [319, 91.429] - - - [48000, 24192, 1, 384] - - [319, 92.35] - - - [48384, 384, 1, 384] - - [321, 80.359] - - - [48384, 23808, 1, 384] - - [365, 92.42] - - - [48384, 23809, 1, 384] - - [365, 91.371] - - - [48384, 24192, 1, 384] - - [319, 92.487] - - - [48384, 24193, 1, 384] - - [365, 91.461] - - - [48768, 384, 1, 384] - - [318, 80.594] - - - [48768, 24192, 1, 384] - - [319, 92.609] - - - [48768, 24193, 1, 384] - - [319, 91.595] - - - [48768, 24576, 1, 384] - - [324, 92.391] - - - [49152, 384, 1, 384] - - [319, 80.409] - - - [49152, 24192, 1, 384] - - [319, 92.328] - - - [49152, 24193, 1, 384] - - [319, 90.984] - - - [49152, 24576, 1, 384] - - [324, 92.011] - - - [49152, 24577, 1, 384] - - [319, 88.798] - - - [49536, 384, 1, 384] - - [293, 78.411] - - - [49536, 24576, 1, 384] - - [324, 92.355] - - - [49536, 24577, 1, 384] - - [319, 89.514] - - - [49536, 24960, 1, 384] - - [319, 92.66] - - - [49920, 384, 1, 384] - - [324, 81.948] - - - [49920, 24576, 1, 384] - - [319, 92.293] - - - [49920, 24577, 1, 384] - - [319, 89.224] - - - [49920, 24960, 1, 384] - - [319, 92.67] - - - [49920, 24961, 1, 384] - - [319, 91.668] - - - [50304, 384, 1, 384] - - [425, 81.391] - - - [50304, 24960, 1, 384] - - [319, 92.698] - - - [50304, 24961, 1, 384] - - [324, 91.555] - - - [50304, 25344, 1, 384] - - [319, 92.687] - - - [50688, 384, 1, 384] - - [331, 82.698] - - - [50688, 24960, 1, 384] - - [365, 92.651] - - - [50688, 24961, 1, 384] - - [365, 91.598] - - - [50688, 25344, 1, 384] - - [365, 92.558] - - - [50688, 25345, 1, 384] - - [319, 91.685] - - - [51072, 384, 1, 384] - - [320, 82.018] - - - [51072, 25344, 1, 384] - - [365, 92.545] - - - [51072, 25345, 1, 384] - - [319, 91.638] - - - [51072, 25728, 1, 384] - - [319, 92.385] - - - [51456, 384, 1, 384] - - [422, 82.599] - - - [51456, 25344, 1, 384] - - [319, 92.525] - - - [51456, 25345, 1, 384] - - [319, 91.616] - - - [51456, 25728, 1, 384] - - [319, 92.472] - - - [51456, 25729, 1, 384] - - [319, 91.676] - - - [51840, 384, 1, 384] - - [320, 82.593] - - - [51840, 25728, 1, 384] - - [319, 92.539] - - - [51840, 25729, 1, 384] - - [319, 91.742] - - - [51840, 26112, 1, 384] - - [319, 92.777] - - - [52224, 384, 1, 384] - - [422, 83.393] - - - [52224, 25728, 1, 384] - - [365, 92.613] - - - [52224, 25729, 1, 384] - - [319, 91.752] - - - [52224, 26112, 1, 384] - - [365, 92.759] - - - [52224, 26113, 1, 384] - - [319, 91.818] - - - [52608, 384, 1, 384] - - [321, 80.513] - - - [52608, 26112, 1, 384] - - [319, 92.771] - - - [52608, 26113, 1, 384] - - [319, 91.81] - - - [52608, 26496, 1, 384] - - [319, 92.671] - - - [52992, 384, 1, 384] - - [319, 81.138] - - - [52992, 26112, 1, 384] - - [319, 92.72] - - - [52992, 26113, 1, 384] - - [319, 91.736] - - - [52992, 26496, 1, 384] - - [319, 92.717] - - - [52992, 26497, 1, 384] - - [319, 91.665] - - - [53376, 384, 1, 384] - - [323, 81.183] - - - [53376, 26496, 1, 384] - - [319, 92.609] - - - [53376, 26497, 1, 384] - - [319, 91.683] - - - [53376, 26880, 1, 384] - - [319, 92.648] - - - [53760, 384, 1, 384] - - [319, 81.618] - - - [53760, 26496, 1, 384] - - [365, 92.711] - - - [53760, 26497, 1, 384] - - [319, 91.796] - - - [53760, 26880, 1, 384] - - [365, 92.622] - - - [53760, 26881, 1, 384] - - [319, 91.796] - - - [54144, 384, 1, 384] - - [323, 82.239] - - - [54144, 26880, 1, 384] - - [319, 92.697] - - - [54144, 26881, 1, 384] - - [319, 91.715] - - - [54144, 27264, 1, 384] - - [319, 92.689] - - - [54528, 384, 1, 384] - - [323, 82.482] - - - [54528, 26880, 1, 384] - - [319, 92.68] - - - [54528, 26881, 1, 384] - - [319, 91.757] - - - [54528, 27264, 1, 384] - - [319, 92.762] - - - [54528, 27265, 1, 384] - - [319, 91.775] - - - [54912, 384, 1, 384] - - [323, 82.988] - - - [54912, 27264, 1, 384] - - [319, 92.772] - - - [54912, 27265, 1, 384] - - [319, 91.788] - - - [54912, 27648, 1, 384] - - [324, 92.571] - - - [55296, 384, 1, 384] - - [324, 83.359] - - - [55296, 27264, 1, 384] - - [319, 92.852] - - - [55296, 27265, 1, 384] - - [319, 91.898] - - - [55296, 27648, 1, 384] - - [324, 92.673] - - - [55296, 27649, 1, 384] - - [319, 88.969] - - - [55680, 384, 1, 384] - - [319, 83.654] - - - [55680, 27648, 1, 384] - - [324, 92.638] - - - [55680, 27649, 1, 384] - - [341, 89.602] - - - [55680, 28032, 1, 384] - - [319, 92.854] - - - [56064, 384, 1, 384] - - [425, 83.142] - - - [56064, 27648, 1, 384] - - [324, 92.58] - - - [56064, 27649, 1, 384] - - [319, 89.475] - - - [56064, 28032, 1, 384] - - [319, 92.714] - - - [56064, 28033, 1, 384] - - [319, 91.806] - - - [56448, 384, 1, 384] - - [321, 80.477] - - - [56448, 28032, 1, 384] - - [319, 92.57] - - - [56448, 28033, 1, 384] - - [324, 91.727] - - - [56448, 28416, 1, 384] - - [323, 92.598] - - - [56832, 384, 1, 384] - - [319, 80.99] - - - [56832, 28032, 1, 384] - - [319, 92.733] - - - [56832, 28033, 1, 384] - - [324, 91.868] - - - [56832, 28416, 1, 384] - - [319, 92.706] - - - [56832, 28417, 1, 384] - - [324, 91.918] - - - [57216, 384, 1, 384] - - [324, 81.224] - - - [57216, 28416, 1, 384] - - [319, 92.779] - - - [57216, 28417, 1, 384] - - [319, 91.848] - - - [57216, 28800, 1, 384] - - [319, 92.806] - - - [57600, 384, 1, 384] - - [323, 81.716] - - - [57600, 28416, 1, 384] - - [319, 92.698] - - - [57600, 28417, 1, 384] - - [319, 91.899] - - - [57600, 28800, 1, 384] - - [319, 92.843] - - - [57600, 28801, 1, 384] - - [319, 91.921] - - - [57984, 384, 1, 384] - - [323, 82.086] - - - [57984, 28800, 1, 384] - - [319, 92.752] - - - [57984, 28801, 1, 384] - - [341, 91.77] - - - [57984, 29184, 1, 384] - - [319, 92.813] - - - [58368, 384, 1, 384] - - [324, 82.314] - - - [58368, 28800, 1, 384] - - [319, 92.908] - - - [58368, 28801, 1, 384] - - [319, 91.822] - - - [58368, 29184, 1, 384] - - [319, 92.863] - - - [58368, 29185, 1, 384] - - [319, 91.929] - - - [58752, 384, 1, 384] - - [319, 82.767] - - - [58752, 29184, 1, 384] - - [319, 92.872] - - - [58752, 29185, 1, 384] - - [319, 91.906] - - - [58752, 29568, 1, 384] - - [324, 92.739] - - - [59136, 384, 1, 384] - - [321, 83.238] - - - [59136, 29184, 1, 384] - - [365, 92.768] - - - [59136, 29185, 1, 384] - - [324, 91.899] - - - [59136, 29568, 1, 384] - - [365, 92.79] - - - [59136, 29569, 1, 384] - - [319, 91.872] - - - [59520, 384, 1, 384] - - [324, 83.478] - - - [59520, 29568, 1, 384] - - [324, 92.759] - - - [59520, 29569, 1, 384] - - [319, 91.919] - - - [59520, 29952, 1, 384] - - [319, 92.773] - - - [59904, 384, 1, 384] - - [319, 83.945] - - - [59904, 29568, 1, 384] - - [319, 92.768] - - - [59904, 29569, 1, 384] - - [319, 91.981] - - - [59904, 29952, 1, 384] - - [319, 92.85] - - - [59904, 29953, 1, 384] - - [324, 91.949] - - - [60288, 384, 1, 384] - - [339, 84.058] - - - [60288, 29952, 1, 384] - - [319, 92.915] - - - [60288, 29953, 1, 384] - - [319, 92.018] - - - [60288, 30336, 1, 384] - - [319, 92.859] - - - [60672, 384, 1, 384] - - [341, 84.593] - - - [60672, 29952, 1, 384] - - [319, 92.807] - - - [60672, 29953, 1, 384] - - [319, 92.01] - - - [60672, 30336, 1, 384] - - [319, 92.866] - - - [60672, 30337, 1, 384] - - [319, 92.048] - - - [61056, 384, 1, 384] - - [341, 81.438] - - - [61056, 30336, 1, 384] - - [319, 92.909] - - - [61056, 30337, 1, 384] - - [319, 92.036] - - - [61056, 30720, 1, 384] - - [319, 92.921] - - - [61440, 384, 1, 384] - - [319, 81.168] - - - [61440, 30336, 1, 384] - - [319, 92.906] - - - [61440, 30337, 1, 384] - - [365, 91.899] - - - [61440, 30720, 1, 384] - - [324, 92.692] - - - [61440, 30721, 1, 384] - - [319, 89.242] - - - [61824, 384, 1, 384] - - [323, 81.871] - - - [61824, 30720, 1, 384] - - [319, 92.832] - - - [61824, 30721, 1, 384] - - [319, 90.084] - - - [61824, 31104, 1, 384] - - [319, 92.818] - - - [62208, 384, 1, 384] - - [319, 82.61] - - - [62208, 30720, 1, 384] - - [319, 92.81] - - - [62208, 30721, 1, 384] - - [319, 89.954] - - - [62208, 31104, 1, 384] - - [319, 92.853] - - - [62208, 31105, 1, 384] - - [319, 92.041] - - - [62592, 384, 1, 384] - - [339, 82.707] - - - [62592, 31104, 1, 384] - - [324, 92.81] - - - [62592, 31105, 1, 384] - - [324, 91.997] - - - [62592, 31488, 1, 384] - - [324, 92.834] - - - [62976, 384, 1, 384] - - [319, 83.032] - - - [62976, 31104, 1, 384] - - [319, 92.934] - - - [62976, 31105, 1, 384] - - [319, 92.072] - - - [62976, 31488, 1, 384] - - [324, 92.826] - - - [62976, 31489, 1, 384] - - [319, 92.105] - - - [63360, 384, 1, 384] - - [324, 83.357] - - - [63360, 31488, 1, 384] - - [324, 92.821] - - - [63360, 31489, 1, 384] - - [324, 91.972] - - - [63360, 31872, 1, 384] - - [319, 92.845] - - - [63744, 384, 1, 384] - - [319, 83.63] - - - [63744, 31488, 1, 384] - - [324, 92.771] - - - [63744, 31489, 1, 384] - - [319, 92.051] - - - [63744, 31872, 1, 384] - - [319, 92.955] - - - [63744, 31873, 1, 384] - - [319, 92.08] - - - [64128, 384, 1, 384] - - [425, 83.15] - - - [64128, 31872, 1, 384] - - [319, 92.987] - - - [64128, 31873, 1, 384] - - [324, 92.03] - - - [64128, 32256, 1, 384] - - [319, 92.994] - - - [64512, 384, 1, 384] - - [319, 84.404] - - - [64512, 31872, 1, 384] - - [319, 92.857] - - - [64512, 31873, 1, 384] - - [319, 91.947] - - - [64512, 32256, 1, 384] - - [324, 92.907] - - - [64512, 32257, 1, 384] - - [319, 92.007] - - - [64896, 384, 1, 384] - - [321, 84.524] - - - [64896, 32256, 1, 384] - - [365, 92.959] - - - [64896, 32257, 1, 384] - - [319, 92.202] - - - [64896, 32640, 1, 384] - - [319, 92.866] - - - [65280, 384, 1, 384] - - [318, 84.907] - - - [65280, 32256, 1, 384] - - [324, 92.867] - - - [65280, 32257, 1, 384] - - [319, 92.085] - - - [65280, 32640, 1, 384] - - [319, 92.86] - - - [65280, 32641, 1, 384] - - [319, 92.07] - - - [65664, 384, 1, 384] - - [282, 78.783] - - - [65664, 32640, 1, 384] - - [324, 92.877] - - - [65664, 32641, 1, 384] - - [324, 92.08] - - - [65664, 33024, 1, 384] - - [324, 92.901] - - - [66048, 384, 1, 384] - - [341, 82.052] - - - [66048, 32640, 1, 384] - - [319, 93.003] - - - [66048, 32641, 1, 384] - - [319, 92.147] - - - [66048, 33024, 1, 384] - - [319, 93.038] - - - [66048, 33025, 1, 384] - - [319, 92.168] - - - [66432, 384, 1, 384] - - [323, 82.606] - - - [66432, 33024, 1, 384] - - [319, 93.004] - - - [66432, 33025, 1, 384] - - [319, 92.12] - - - [66432, 33408, 1, 384] - - [319, 92.871] - - - [66816, 384, 1, 384] - - [319, 83.022] - - - [66816, 33024, 1, 384] - - [319, 92.906] - - - [66816, 33025, 1, 384] - - [324, 91.964] - - - [66816, 33408, 1, 384] - - [319, 92.896] - - - [66816, 33409, 1, 384] - - [324, 92.155] - - - [67200, 384, 1, 384] - - [319, 83.499] - - - [67200, 33408, 1, 384] - - [324, 92.959] - - - [67200, 33409, 1, 384] - - [324, 92.197] - - - [67200, 33792, 1, 384] - - [324, 93.01] - - - [67584, 384, 1, 384] - - [319, 83.772] - - - [67584, 33408, 1, 384] - - [324, 93.003] - - - [67584, 33409, 1, 384] - - [319, 92.116] - - - [67584, 33792, 1, 384] - - [324, 92.985] - - - [67584, 33793, 1, 384] - - [319, 89.609] - - - [67968, 384, 1, 384] - - [319, 83.81] - - - [67968, 33792, 1, 384] - - [319, 92.992] - - - [67968, 33793, 1, 384] - - [319, 90.356] - - - [67968, 34176, 1, 384] - - [319, 93.013] - - - [68352, 384, 1, 384] - - [356, 84.159] - - - [68352, 33792, 1, 384] - - [324, 92.961] - - - [68352, 33793, 1, 384] - - [319, 90.159] - - - [68352, 34176, 1, 384] - - [324, 92.99] - - - [68352, 34177, 1, 384] - - [324, 92.182] - - - [68736, 384, 1, 384] - - [323, 84.257] - - - [68736, 34176, 1, 384] - - [324, 93.293] - - - [68736, 34177, 1, 384] - - [324, 92.561] - - - [68736, 34560, 1, 384] - - [323, 93.319] - - - [69120, 384, 1, 384] - - [327, 84.759] - - - [69120, 34176, 1, 384] - - [324, 93.387] - - - [69120, 34177, 1, 384] - - [324, 92.573] - - - [69120, 34560, 1, 384] - - [324, 93.349] - - - [69120, 34561, 1, 384] - - [324, 92.538] - - - [69504, 384, 1, 384] - - [324, 84.73] - - - [69504, 34560, 1, 384] - - [324, 93.382] - - - [69504, 34561, 1, 384] - - [324, 92.476] - - - [69504, 34944, 1, 384] - - [319, 93.324] - - - [69888, 384, 1, 384] - - [324, 85.214] - - - [69888, 34560, 1, 384] - - [323, 93.25] - - - [69888, 34561, 1, 384] - - [324, 92.533] - - - [69888, 34944, 1, 384] - - [319, 93.22] - - - [69888, 34945, 1, 384] - - [324, 92.505] - - - [70272, 384, 1, 384] - - [326, 85.384] - - - [70272, 34944, 1, 384] - - [319, 93.257] - - - [70272, 34945, 1, 384] - - [324, 92.503] - - - [70272, 35328, 1, 384] - - [324, 93.388] - - - [70656, 384, 1, 384] - - [361, 82.691] - - - [70656, 34944, 1, 384] - - [319, 93.221] - - - [70656, 34945, 1, 384] - - [324, 92.384] - - - [70656, 35328, 1, 384] - - [324, 93.293] - - - [70656, 35329, 1, 384] - - [324, 92.389] - - - [71040, 384, 1, 384] - - [331, 82.994] - - - [71040, 35328, 1, 384] - - [324, 93.373] - - - [71040, 35329, 1, 384] - - [324, 92.496] - - - [71040, 35712, 1, 384] - - [324, 93.306] - - - [71424, 384, 1, 384] - - [323, 83.457] - - - [71424, 35328, 1, 384] - - [324, 93.26] - - - [71424, 35329, 1, 384] - - [319, 92.429] - - - [71424, 35712, 1, 384] - - [324, 93.263] - - - [71424, 35713, 1, 384] - - [324, 92.412] - - - [71808, 384, 1, 384] - - [374, 83.773] - - - [71808, 35712, 1, 384] - - [324, 93.277] - - - [71808, 35713, 1, 384] - - [324, 92.489] - - - [71808, 36096, 1, 384] - - [324, 93.323] - - - [72192, 384, 1, 384] - - [341, 83.809] - - - [72192, 35712, 1, 384] - - [319, 93.297] - - - [72192, 35713, 1, 384] - - [324, 92.505] - - - [72192, 36096, 1, 384] - - [324, 93.265] - - - [72192, 36097, 1, 384] - - [324, 92.522] - - - [72576, 384, 1, 384] - - [323, 84.225] - - - [72576, 36096, 1, 384] - - [324, 93.277] - - - [72576, 36097, 1, 384] - - [324, 92.487] - - - [72576, 36480, 1, 384] - - [324, 93.289] - - - [72960, 384, 1, 384] - - [319, 84.48] - - - [72960, 36096, 1, 384] - - [324, 93.261] - - - [72960, 36097, 1, 384] - - [319, 92.432] - - - [72960, 36480, 1, 384] - - [324, 93.296] - - - [72960, 36481, 1, 384] - - [319, 92.439] - - - [73344, 384, 1, 384] - - [319, 84.556] - - - [73344, 36480, 1, 384] - - [319, 93.225] - - - [73344, 36481, 1, 384] - - [324, 92.421] - - - [73344, 36864, 1, 384] - - [324, 93.251] - - - [73728, 384, 1, 384] - - [341, 83.95] - - - [73728, 36480, 1, 384] - - [324, 92.849] - - - [73728, 36481, 1, 384] - - [324, 91.857] - - - [73728, 36864, 1, 384] - - [324, 92.839] - - - [73728, 36865, 1, 384] - - [319, 89.489] - - - [74112, 384, 1, 384] - - [324, 85.264] - - - [74112, 36864, 1, 384] - - [324, 93.324] - - - [74112, 36865, 1, 384] - - [319, 90.592] - - - [74112, 37248, 1, 384] - - [324, 93.274] - - - [74496, 384, 1, 384] - - [319, 85.292] - - - [74496, 36864, 1, 384] - - [324, 93.29] - - - [74496, 36865, 1, 384] - - [319, 90.42] - - - [74496, 37248, 1, 384] - - [324, 93.251] - - - [74496, 37249, 1, 384] - - [324, 92.48] - - - [74880, 384, 1, 384] - - [361, 85.54] - - - [74880, 37248, 1, 384] - - [324, 93.236] - - - [74880, 37249, 1, 384] - - [324, 92.455] - - - [74880, 37632, 1, 384] - - [319, 93.276] - - - [75264, 384, 1, 384] - - [326, 83.093] - - - [75264, 37248, 1, 384] - - [324, 93.204] - - - [75264, 37249, 1, 384] - - [324, 92.412] - - - [75264, 37632, 1, 384] - - [324, 93.187] - - - [75264, 37633, 1, 384] - - [324, 92.394] - - - [75648, 384, 1, 384] - - [341, 83.601] - - - [75648, 37632, 1, 384] - - [319, 93.231] - - - [75648, 37633, 1, 384] - - [324, 92.489] - - - [75648, 38016, 1, 384] - - [319, 93.163] - - - [76032, 384, 1, 384] - - [321, 84.025] - - - [76032, 37632, 1, 384] - - [324, 93.2] - - - [76032, 37633, 1, 384] - - [324, 92.438] - - - [76032, 38016, 1, 384] - - [319, 93.076] - - - [76032, 38017, 1, 384] - - [324, 92.405] - - - [76416, 384, 1, 384] - - [323, 83.988] - - - [76416, 38016, 1, 384] - - [319, 93.185] - - - [76416, 38017, 1, 384] - - [319, 92.391] - - - [76416, 38400, 1, 384] - - [319, 93.239] - - - [76800, 384, 1, 384] - - [323, 84.326] - - - [76800, 38016, 1, 384] - - [319, 93.08] - - - [76800, 38017, 1, 384] - - [319, 92.255] - - - [76800, 38400, 1, 384] - - [324, 93.067] - - - [76800, 38401, 1, 384] - - [324, 92.265] - - - [77184, 384, 1, 384] - - [323, 84.734] - - - [77184, 38400, 1, 384] - - [324, 93.227] - - - [77184, 38401, 1, 384] - - [319, 92.449] - - - [77184, 38784, 1, 384] - - [324, 93.166] - - - [77568, 384, 1, 384] - - [324, 84.889] - - - [77568, 38400, 1, 384] - - [324, 93.161] - - - [77568, 38401, 1, 384] - - [324, 92.448] - - - [77568, 38784, 1, 384] - - [324, 93.108] - - - [77568, 38785, 1, 384] - - [324, 92.284] - - - [77952, 384, 1, 384] - - [324, 84.862] - - - [77952, 38784, 1, 384] - - [324, 93.14] - - - [77952, 38785, 1, 384] - - [319, 92.261] - - - [77952, 39168, 1, 384] - - [324, 93.115] - - - [78336, 384, 1, 384] - - [319, 85.099] - - - [78336, 38784, 1, 384] - - [319, 93.115] - - - [78336, 38785, 1, 384] - - [324, 92.364] - - - [78336, 39168, 1, 384] - - [324, 92.987] - - - [78336, 39169, 1, 384] - - [319, 92.292] - - - [78720, 384, 1, 384] - - [361, 85.113] - - - [78720, 39168, 1, 384] - - [324, 93.066] - - - [78720, 39169, 1, 384] - - [319, 92.341] - - - [78720, 39552, 1, 384] - - [319, 93.11] - - - [79104, 384, 1, 384] - - [323, 85.514] - - - [79104, 39168, 1, 384] - - [324, 93.026] - - - [79104, 39169, 1, 384] - - [324, 92.331] - - - [79104, 39552, 1, 384] - - [324, 93.017] - - - [79104, 39553, 1, 384] - - [324, 92.267] - - - [79488, 384, 1, 384] - - [323, 85.803] - - - [79488, 39552, 1, 384] - - [324, 93.086] - - - [79488, 39553, 1, 384] - - [324, 92.355] - - - [79488, 39936, 1, 384] - - [324, 93.132] - - - [79872, 384, 1, 384] - - [332, 83.822] - - - [79872, 39552, 1, 384] - - [324, 93.062] - - - [79872, 39553, 1, 384] - - [319, 92.254] - - - [79872, 39936, 1, 384] - - [324, 92.874] - - - [79872, 39937, 1, 384] - - [319, 89.714] - - - [80256, 384, 1, 384] - - [341, 84.158] - - - [80256, 39936, 1, 384] - - [324, 93.044] - - - [80256, 39937, 1, 384] - - [319, 90.42] - - - [80256, 40320, 1, 384] - - [324, 93.059] - - - [80640, 384, 1, 384] - - [331, 84.255] - - - [80640, 39936, 1, 384] - - [324, 93.019] - - - [80640, 39937, 1, 384] - - [319, 90.379] - - - [80640, 40320, 1, 384] - - [324, 92.949] - - - [80640, 40321, 1, 384] - - [324, 92.218] - - - [81024, 384, 1, 384] - - [318, 84.55] - - - [81024, 40320, 1, 384] - - [324, 93.055] - - - [81024, 40321, 1, 384] - - [324, 92.284] - - - [81024, 40704, 1, 384] - - [324, 93.043] - - - [81408, 384, 1, 384] - - [324, 84.924] - - - [81408, 40320, 1, 384] - - [324, 93.024] - - - [81408, 40321, 1, 384] - - [319, 92.24] - - - [81408, 40704, 1, 384] - - [324, 92.926] - - - [81408, 40705, 1, 384] - - [319, 92.158] - - - [81792, 384, 1, 384] - - [341, 84.885] - - - [81792, 40704, 1, 384] - - [319, 93.022] - - - [81792, 40705, 1, 384] - - [319, 92.249] - - - [81792, 41088, 1, 384] - - [319, 92.901] - - - [82176, 384, 1, 384] - - [331, 84.885] - - - [82176, 40704, 1, 384] - - [324, 92.882] - - - [82176, 40705, 1, 384] - - [324, 92.172] - - - [82176, 41088, 1, 384] - - [324, 92.857] - - - [82176, 41089, 1, 384] - - [324, 92.159] - - - [82560, 384, 1, 384] - - [319, 85.428] - - - [82560, 41088, 1, 384] - - [324, 92.896] - - - [82560, 41089, 1, 384] - - [324, 92.244] - - - [82560, 41472, 1, 384] - - [324, 93.001] - - - [82944, 384, 1, 384] - - [324, 85.268] - - - [82944, 41088, 1, 384] - - [324, 92.841] - - - [82944, 41089, 1, 384] - - [319, 92.056] - - - [82944, 41472, 1, 384] - - [324, 92.819] - - - [82944, 41473, 1, 384] - - [319, 92.008] - - - [83328, 384, 1, 384] - - [319, 85.57] - - - [83328, 41472, 1, 384] - - [324, 92.98] - - - [83328, 41473, 1, 384] - - [319, 92.201] - - - [83328, 41856, 1, 384] - - [324, 92.891] - - - [83712, 384, 1, 384] - - [324, 85.62] - - - [83712, 41472, 1, 384] - - [324, 92.896] - - - [83712, 41473, 1, 384] - - [324, 92.161] - - - [83712, 41856, 1, 384] - - [324, 92.813] - - - [83712, 41857, 1, 384] - - [324, 92.061] - - - [84096, 384, 1, 384] - - [341, 85.802] - - - [84096, 41856, 1, 384] - - [319, 92.879] - - - [84096, 41857, 1, 384] - - [324, 92.137] - - - [84096, 42240, 1, 384] - - [319, 92.852] - - - [84480, 384, 1, 384] - - [361, 85.785] - - - [84480, 41856, 1, 384] - - [324, 92.877] - - - [84480, 41857, 1, 384] - - [324, 92.105] - - - [84480, 42240, 1, 384] - - [324, 92.735] - - - [84480, 42241, 1, 384] - - [319, 92.007] - - - [84864, 384, 1, 384] - - [323, 84.383] - - - [84864, 42240, 1, 384] - - [319, 92.853] - - - [84864, 42241, 1, 384] - - [319, 92.101] - - - [84864, 42624, 1, 384] - - [319, 92.8] - - - [85248, 384, 1, 384] - - [332, 84.6] - - - [85248, 42240, 1, 384] - - [319, 92.729] - - - [85248, 42241, 1, 384] - - [319, 92.041] - - - [85248, 42624, 1, 384] - - [324, 92.705] - - - [85248, 42625, 1, 384] - - [324, 91.991] - - - [85632, 384, 1, 384] - - [323, 84.64] - - - [85632, 42624, 1, 384] - - [319, 92.799] - - - [85632, 42625, 1, 384] - - [324, 92.058] - - - [85632, 43008, 1, 384] - - [324, 92.754] - - - [86016, 384, 1, 384] - - [324, 84.981] - - - [86016, 42624, 1, 384] - - [319, 92.705] - - - [86016, 42625, 1, 384] - - [319, 91.928] - - - [86016, 43008, 1, 384] - - [324, 92.53] - - - [86016, 43009, 1, 384] - - [319, 89.555] - - - [86400, 384, 1, 384] - - [324, 85.335] - - - [86400, 43008, 1, 384] - - [324, 92.722] - - - [86400, 43009, 1, 384] - - [319, 90.219] - - - [86400, 43392, 1, 384] - - [324, 92.676] - - - [86784, 384, 1, 384] - - [324, 85.372] - - - [86784, 43008, 1, 384] - - [324, 92.684] - - - [86784, 43009, 1, 384] - - [319, 90.15] - - - [86784, 43392, 1, 384] - - [324, 92.57] - - - [86784, 43393, 1, 384] - - [324, 91.889] - - - [87168, 384, 1, 384] - - [341, 85.547] - - - [87168, 43392, 1, 384] - - [319, 92.659] - - - [87168, 43393, 1, 384] - - [319, 91.95] - - - [87168, 43776, 1, 384] - - [319, 92.562] - - - [87552, 384, 1, 384] - - [324, 85.558] - - - [87552, 43392, 1, 384] - - [319, 92.603] - - - [87552, 43393, 1, 384] - - [319, 91.913] - - - [87552, 43776, 1, 384] - - [324, 92.691] - - - [87552, 43777, 1, 384] - - [319, 92.01] - - - [87936, 384, 1, 384] - - [319, 85.631] - - - [87936, 43776, 1, 384] - - [324, 92.644] - - - [87936, 43777, 1, 384] - - [319, 91.946] - - - [87936, 44160, 1, 384] - - [319, 92.667] - - - [88320, 384, 1, 384] - - [361, 85.728] - - - [88320, 43776, 1, 384] - - [319, 92.599] - - - [88320, 43777, 1, 384] - - [324, 91.936] - - - [88320, 44160, 1, 384] - - [319, 92.554] - - - [88320, 44161, 1, 384] - - [324, 91.84] - - - [88704, 384, 1, 384] - - [371, 85.85] - - - [88704, 44160, 1, 384] - - [319, 92.674] - - - [88704, 44161, 1, 384] - - [319, 91.923] - - - [88704, 44544, 1, 384] - - [363, 92.604] - - - [89088, 384, 1, 384] - - [324, 85.71] - - - [89088, 44160, 1, 384] - - [319, 92.542] - - - [89088, 44161, 1, 384] - - [319, 91.767] - - - [89088, 44544, 1, 384] - - [324, 92.487] - - - [89088, 44545, 1, 384] - - [324, 91.739] - - - [89472, 384, 1, 384] - - [356, 84.73] - - - [89472, 44544, 1, 384] - - [324, 92.645] - - - [89472, 44545, 1, 384] - - [319, 91.956] - - - [89472, 44928, 1, 384] - - [324, 92.562] - - - [89856, 384, 1, 384] - - [361, 84.9] - - - [89856, 44544, 1, 384] - - [363, 92.527] - - - [89856, 44545, 1, 384] - - [324, 91.855] - - - [89856, 44928, 1, 384] - - [324, 92.445] - - - [89856, 44929, 1, 384] - - [324, 91.784] - - - [90240, 384, 1, 384] - - [331, 84.913] - - - [90240, 44928, 1, 384] - - [324, 92.532] - - - [90240, 44929, 1, 384] - - [324, 91.795] - - - [90240, 45312, 1, 384] - - [319, 92.524] - - - [90624, 384, 1, 384] - - [319, 85.444] - - - [90624, 44928, 1, 384] - - [324, 92.527] - - - [90624, 44929, 1, 384] - - [319, 91.83] - - - [90624, 45312, 1, 384] - - [319, 92.441] - - - [90624, 45313, 1, 384] - - [319, 91.775] - - - [91008, 384, 1, 384] - - [319, 85.696] - - - [91008, 45312, 1, 384] - - [319, 92.513] - - - [91008, 45313, 1, 384] - - [319, 91.795] - - - [91008, 45696, 1, 384] - - [324, 92.476] - - - [91392, 384, 1, 384] - - [341, 85.778] - - - [91392, 45312, 1, 384] - - [319, 92.421] - - - [91392, 45313, 1, 384] - - [319, 91.749] - - - [91392, 45696, 1, 384] - - [324, 92.37] - - - [91392, 45697, 1, 384] - - [319, 91.674] - - - [91776, 384, 1, 384] - - [326, 85.607] - - - [91776, 45696, 1, 384] - - [324, 92.379] - - - [91776, 45697, 1, 384] - - [324, 91.674] - - - [91776, 46080, 1, 384] - - [363, 92.42] - - - [92160, 384, 1, 384] - - [319, 85.856] - - - [92160, 45696, 1, 384] - - [324, 92.373] - - - [92160, 45697, 1, 384] - - [319, 91.624] - - - [92160, 46080, 1, 384] - - [324, 92.245] - - - [92160, 46081, 1, 384] - - [319, 89.53] - - - [92544, 384, 1, 384] - - [319, 85.793] - - - [92544, 46080, 1, 384] - - [363, 92.469] - - - [92544, 46081, 1, 384] - - [319, 90.234] - - - [92544, 46464, 1, 384] - - [319, 92.34] - - - [92928, 384, 1, 384] - - [332, 85.766] - - - [92928, 46080, 1, 384] - - [363, 92.435] - - - [92928, 46081, 1, 384] - - [319, 90.168] - - - [92928, 46464, 1, 384] - - [363, 92.313] - - - [92928, 46465, 1, 384] - - [319, 91.607] - - - [93312, 384, 1, 384] - - [319, 86.085] - - - [93312, 46464, 1, 384] - - [319, 92.345] - - - [93312, 46465, 1, 384] - - [324, 91.61] - - - [93312, 46848, 1, 384] - - [427, 92.327] - - - [93696, 384, 1, 384] - - [361, 86.104] - - - [93696, 46464, 1, 384] - - [363, 92.376] - - - [93696, 46465, 1, 384] - - [324, 91.604] - - - [93696, 46848, 1, 384] - - [363, 92.291] - - - [93696, 46849, 1, 384] - - [319, 91.547] - - - [94080, 384, 1, 384] - - [341, 85.057] - - - [94080, 46848, 1, 384] - - [363, 92.315] - - - [94080, 46849, 1, 384] - - [319, 91.594] - - - [94080, 47232, 1, 384] - - [373, 92.251] - - - [94464, 384, 1, 384] - - [319, 85.335] - - - [94464, 46848, 1, 384] - - [363, 92.291] - - - [94464, 46849, 1, 384] - - [319, 91.527] - - - [94464, 47232, 1, 384] - - [373, 92.204] - - - [94464, 47233, 1, 384] - - [319, 91.484] - - - [94848, 384, 1, 384] - - [341, 85.438] - - - [94848, 47232, 1, 384] - - [373, 92.265] - - - [94848, 47233, 1, 384] - - [319, 91.512] - - - [94848, 47616, 1, 384] - - [363, 92.368] - - - [95232, 384, 1, 384] - - [319, 85.517] - - - [95232, 47232, 1, 384] - - [319, 92.123] - - - [95232, 47233, 1, 384] - - [319, 91.376] - - - [95232, 47616, 1, 384] - - [363, 92.09] - - - [95232, 47617, 1, 384] - - [319, 91.374] - - - [95616, 384, 1, 384] - - [361, 85.688] - - - [95616, 47616, 1, 384] - - [363, 92.354] - - - [95616, 47617, 1, 384] - - [319, 91.487] - - - [95616, 48000, 1, 384] - - [363, 92.264] - - - [96000, 384, 1, 384] - - [326, 85.525] - - - [96000, 47616, 1, 384] - - [363, 92.296] - - - [96000, 47617, 1, 384] - - [363, 91.458] - - - [96000, 48000, 1, 384] - - [363, 92.25] - - - [96000, 48001, 1, 384] - - [363, 91.381] - - - [96384, 384, 1, 384] - - [318, 85.971] - - - [96384, 48000, 1, 384] - - [363, 92.173] - - - [96384, 48001, 1, 384] - - [324, 91.307] - - - [96384, 48384, 1, 384] - - [363, 92.169] - - - [96768, 384, 1, 384] - - [319, 86.094] - - - [96768, 48000, 1, 384] - - [363, 92.225] - - - [96768, 48001, 1, 384] - - [363, 91.378] - - - [96768, 48384, 1, 384] - - [363, 92.13] - - - [96768, 48385, 1, 384] - - [363, 91.359] - - - [97152, 384, 1, 384] - - [319, 86.067] - - - [97152, 48384, 1, 384] - - [363, 92.157] - - - [97152, 48385, 1, 384] - - [319, 91.313] - - - [97152, 48768, 1, 384] - - [363, 92.119] - - - [97536, 384, 1, 384] - - [319, 85.971] - - - [97536, 48384, 1, 384] - - [363, 92.149] - - - [97536, 48385, 1, 384] - - [363, 91.311] - - - [97536, 48768, 1, 384] - - [363, 92.152] - - - [97536, 48769, 1, 384] - - [363, 91.338] - - - [97920, 384, 1, 384] - - [326, 85.934] - - - [97920, 48768, 1, 384] - - [363, 91.965] - - - [97920, 48769, 1, 384] - - [319, 91.209] - - - [97920, 49152, 1, 384] - - [363, 91.911] - - - [98304, 384, 1, 384] - - [319, 83.088] - - - [98304, 48768, 1, 384] - - [324, 90.529] - - - [98304, 48769, 1, 384] - - [324, 89.38] - - - [98304, 49152, 1, 384] - - [324, 90.415] - - - [98304, 49153, 1, 384] - - [324, 88.259] - - - [98688, 384, 1, 384] - - [356, 85.423] - - - [98688, 49152, 1, 384] - - [363, 92.242] - - - [98688, 49153, 1, 384] - - [319, 89.64] - - - [98688, 49536, 1, 384] - - [363, 92.209] - - - [99072, 384, 1, 384] - - [319, 85.3] - - - [99072, 49152, 1, 384] - - [363, 92.029] - - - [99072, 49153, 1, 384] - - [319, 89.533] - - - [99072, 49536, 1, 384] - - [363, 92.069] - - - [99072, 49537, 1, 384] - - [363, 91.256] - - - [99456, 384, 1, 384] - - [318, 85.653] - - - [99456, 49536, 1, 384] - - [363, 92.083] - - - [99456, 49537, 1, 384] - - [363, 91.199] - - - [99456, 49920, 1, 384] - - [363, 92.093] - - - [99840, 384, 1, 384] - - [319, 85.95] - - - [99840, 49536, 1, 384] - - [363, 92.038] - - - [99840, 49537, 1, 384] - - [363, 91.205] - - - [99840, 49920, 1, 384] - - [363, 91.964] - - - [99840, 49921, 1, 384] - - [363, 91.191] - - - [100224, 384, 1, 384] - - [319, 86.158] - - - [100224, 49920, 1, 384] - - [363, 92.003] - - - [100224, 49921, 1, 384] - - [363, 91.196] - - - [100224, 50304, 1, 384] - - [373, 91.89] - - - [100608, 384, 1, 384] - - [319, 85.965] - - - [100608, 49920, 1, 384] - - [363, 91.931] - - - [100608, 49921, 1, 384] - - [363, 91.143] - - - [100608, 50304, 1, 384] - - [373, 91.88] - - - [100608, 50305, 1, 384] - - [363, 91.127] - - - [100992, 384, 1, 384] - - [319, 86.238] - - - [100992, 50304, 1, 384] - - [363, 91.849] - - - [100992, 50305, 1, 384] - - [363, 91.061] - - - [100992, 50688, 1, 384] - - [427, 91.964] - - - [101376, 384, 1, 384] - - [319, 85.991] - - - [101376, 50304, 1, 384] - - [363, 91.645] - - - [101376, 50305, 1, 384] - - [363, 90.706] - - - [101376, 50688, 1, 384] - - [363, 91.686] - - - [101376, 50689, 1, 384] - - [363, 90.729] - - - [101760, 384, 1, 384] - - [319, 86.229] - - - [101760, 50688, 1, 384] - - [363, 91.893] - - - [101760, 50689, 1, 384] - - [363, 91.067] - - - [101760, 51072, 1, 384] - - [363, 91.814] - - - [102144, 384, 1, 384] - - [324, 86.093] - - - [102144, 50688, 1, 384] - - [363, 91.848] - - - [102144, 50689, 1, 384] - - [363, 91.071] - - - [102144, 51072, 1, 384] - - [363, 91.826] - - - [102144, 51073, 1, 384] - - [363, 91.037] - - - [102528, 384, 1, 384] - - [319, 85.905] - - - [102528, 51072, 1, 384] - - [363, 91.777] - - - [102528, 51073, 1, 384] - - [363, 90.938] - - - [102528, 51456, 1, 384] - - [363, 91.773] - - - [102912, 384, 1, 384] - - [374, 86.074] - - - [102912, 51072, 1, 384] - - [363, 91.76] - - - [102912, 51073, 1, 384] - - [363, 90.948] - - - [102912, 51456, 1, 384] - - [363, 91.7] - - - [102912, 51457, 1, 384] - - [363, 90.918] - - - [103296, 384, 1, 384] - - [361, 85.545] - - - [103296, 51456, 1, 384] - - [363, 91.67] - - - [103296, 51457, 1, 384] - - [363, 90.841] - - - [103296, 51840, 1, 384] - - [363, 91.649] - - - [103680, 384, 1, 384] - - [326, 85.767] - - - [103680, 51456, 1, 384] - - [363, 91.615] - - - [103680, 51457, 1, 384] - - [363, 90.854] - - - [103680, 51840, 1, 384] - - [363, 91.681] - - - [103680, 51841, 1, 384] - - [363, 90.862] - - - [104064, 384, 1, 384] - - [324, 85.796] - - - [104064, 51840, 1, 384] - - [363, 91.625] - - - [104064, 51841, 1, 384] - - [363, 90.789] - - - [104064, 52224, 1, 384] - - [363, 91.638] - - - [104448, 384, 1, 384] - - [319, 85.91] - - - [104448, 51840, 1, 384] - - [363, 91.423] - - - [104448, 51841, 1, 384] - - [363, 90.442] - - - [104448, 52224, 1, 384] - - [363, 91.356] - - - [104448, 52225, 1, 384] - - [363, 89.319] - - - [104832, 384, 1, 384] - - [361, 86.237] - - - [104832, 52224, 1, 384] - - [363, 91.544] - - - [104832, 52225, 1, 384] - - [373, 89.134] - - - [104832, 52608, 1, 384] - - [363, 91.464] - - - [105216, 384, 1, 384] - - [324, 86.435] - - - [105216, 52224, 1, 384] - - [363, 91.546] - - - [105216, 52225, 1, 384] - - [373, 89.086] - - - [105216, 52608, 1, 384] - - [363, 91.48] - - - [105216, 52609, 1, 384] - - [363, 90.705] - - - [105600, 384, 1, 384] - - [361, 86.154] - - - [105600, 52608, 1, 384] - - [363, 91.435] - - - [105600, 52609, 1, 384] - - [363, 90.608] - - - [105600, 52992, 1, 384] - - [363, 91.462] - - - [105984, 384, 1, 384] - - [324, 86.234] - - - [105984, 52608, 1, 384] - - [363, 91.46] - - - [105984, 52609, 1, 384] - - [363, 90.635] - - - [105984, 52992, 1, 384] - - [363, 91.424] - - - [105984, 52993, 1, 384] - - [363, 90.635] - - - [106368, 384, 1, 384] - - [326, 86.031] - - - [106368, 52992, 1, 384] - - [427, 91.416] - - - [106368, 52993, 1, 384] - - [363, 90.619] - - - [106368, 53376, 1, 384] - - [363, 91.312] - - - [106752, 384, 1, 384] - - [323, 85.996] - - - [106752, 52992, 1, 384] - - [363, 91.346] - - - [106752, 52993, 1, 384] - - [363, 90.582] - - - [106752, 53376, 1, 384] - - [363, 91.264] - - - [106752, 53377, 1, 384] - - [363, 90.539] - - - [107136, 384, 1, 384] - - [374, 85.651] - - - [107136, 53376, 1, 384] - - [363, 91.229] - - - [107136, 53377, 1, 384] - - [363, 90.459] - - - [107136, 53760, 1, 384] - - [427, 91.374] - - - [107520, 384, 1, 384] - - [319, 85.71] - - - [107520, 53376, 1, 384] - - [363, 91.028] - - - [107520, 53377, 1, 384] - - [363, 90.13] - - - [107520, 53760, 1, 384] - - [363, 91.077] - - - [107520, 53761, 1, 384] - - [363, 90.139] - - - [107904, 384, 1, 384] - - [353, 85.985] - - - [107904, 53760, 1, 384] - - [427, 91.337] - - - [107904, 53761, 1, 384] - - [363, 90.502] - - - [107904, 54144, 1, 384] - - [363, 91.226] - - - [108288, 384, 1, 384] - - [323, 85.849] - - - [108288, 53760, 1, 384] - - [363, 91.24] - - - [108288, 53761, 1, 384] - - [363, 90.48] - - - [108288, 54144, 1, 384] - - [363, 91.2] - - - [108288, 54145, 1, 384] - - [363, 90.388] - - - [108672, 384, 1, 384] - - [318, 86.03] - - - [108672, 54144, 1, 384] - - [363, 91.138] - - - [108672, 54145, 1, 384] - - [363, 90.33] - - - [108672, 54528, 1, 384] - - [427, 91.222] - - - [109056, 384, 1, 384] - - [319, 86.283] - - - [109056, 54144, 1, 384] - - [363, 91.125] - - - [109056, 54145, 1, 384] - - [363, 90.304] - - - [109056, 54528, 1, 384] - - [363, 91.003] - - - [109056, 54529, 1, 384] - - [363, 90.259] - - - [109440, 384, 1, 384] - - [324, 86.279] - - - [109440, 54528, 1, 384] - - [427, 91.138] - - - [109440, 54529, 1, 384] - - [363, 90.211] - - - [109440, 54912, 1, 384] - - [363, 91.093] - - - [109824, 384, 1, 384] - - [319, 86.509] - - - [109824, 54528, 1, 384] - - [427, 91.06] - - - [109824, 54529, 1, 384] - - [363, 90.262] - - - [109824, 54912, 1, 384] - - [363, 90.972] - - - [109824, 54913, 1, 384] - - [363, 90.191] - - - [110208, 384, 1, 384] - - [319, 86.545] - - - [110208, 54912, 1, 384] - - [363, 90.96] - - - [110208, 54913, 1, 384] - - [363, 90.183] - - - [110208, 55296, 1, 384] - - [363, 90.979] - - - [110592, 384, 1, 384] - - [324, 86.144] - - - [110592, 54912, 1, 384] - - [363, 90.716] - - - [110592, 54913, 1, 384] - - [363, 89.759] - - - [110592, 55296, 1, 384] - - [363, 90.609] - - - [110592, 55297, 1, 384] - - [363, 88.908] - - - [110976, 384, 1, 384] - - [319, 86.413] - - - [110976, 55296, 1, 384] - - [363, 90.869] - - - [110976, 55297, 1, 384] - - [373, 88.698] - - - [110976, 55680, 1, 384] - - [363, 90.861] - - - [111360, 384, 1, 384] - - [318, 86.157] - - - [111360, 55296, 1, 384] - - [363, 90.804] - - - [111360, 55297, 1, 384] - - [373, 88.707] - - - [111360, 55680, 1, 384] - - [363, 90.783] - - - [111360, 55681, 1, 384] - - [363, 90.009] - - - [111744, 384, 1, 384] - - [326, 85.775] - - - [111744, 55680, 1, 384] - - [363, 90.744] - - - [111744, 55681, 1, 384] - - [363, 89.952] - - - [111744, 56064, 1, 384] - - [427, 90.793] - - - [112128, 384, 1, 384] - - [324, 85.91] - - - [112128, 55680, 1, 384] - - [363, 90.717] - - - [112128, 55681, 1, 384] - - [363, 89.943] - - - [112128, 56064, 1, 384] - - [427, 90.637] - - - [112128, 56065, 1, 384] - - [363, 89.85] - - - [112512, 384, 1, 384] - - [371, 85.898] - - - [112512, 56064, 1, 384] - - [427, 90.689] - - - [112512, 56065, 1, 384] - - [363, 89.837] - - - [112512, 56448, 1, 384] - - [363, 90.582] - - - [112896, 384, 1, 384] - - [319, 85.934] - - - [112896, 56064, 1, 384] - - [427, 90.592] - - - [112896, 56065, 1, 384] - - [363, 89.825] - - - [112896, 56448, 1, 384] - - [363, 90.518] - - - [112896, 56449, 1, 384] - - [363, 89.752] - - - [113280, 384, 1, 384] - - [331, 86.148] - - - [113280, 56448, 1, 384] - - [427, 90.474] - - - [113280, 56449, 1, 384] - - [363, 89.73] - - - [113280, 56832, 1, 384] - - [427, 90.603] - - - [113664, 384, 1, 384] - - [319, 86.301] - - - [113664, 56448, 1, 384] - - [363, 90.29] - - - [113664, 56449, 1, 384] - - [363, 89.419] - - - [113664, 56832, 1, 384] - - [427, 90.28] - - - [113664, 56833, 1, 384] - - [363, 89.377] - - - [114048, 384, 1, 384] - - [361, 86.415] - - - [114048, 56832, 1, 384] - - [427, 90.525] - - - [114048, 56833, 1, 384] - - [363, 89.653] - - - [114048, 57216, 1, 384] - - [428, 90.427] - - - [114432, 384, 1, 384] - - [324, 86.588] - - - [114432, 56832, 1, 384] - - [427, 90.441] - - - [114432, 56833, 1, 384] - - [363, 89.587] - - - [114432, 57216, 1, 384] - - [363, 90.348] - - - [114432, 57217, 1, 384] - - [363, 89.538] - - - [114816, 384, 1, 384] - - [326, 86.259] - - - [114816, 57216, 1, 384] - - [428, 90.473] - - - [114816, 57217, 1, 384] - - [363, 89.361] - - - [114816, 57600, 1, 384] - - [427, 90.242] - - - [115200, 384, 1, 384] - - [361, 86.304] - - - [115200, 57216, 1, 384] - - [363, 90.202] - - - [115200, 57217, 1, 384] - - [363, 89.376] - - - [115200, 57600, 1, 384] - - [427, 90.129] - - - [115200, 57601, 1, 384] - - [373, 89.271] - - - [115584, 384, 1, 384] - - [323, 86.126] - - - [115584, 57600, 1, 384] - - [427, 90.196] - - - [115584, 57601, 1, 384] - - [363, 89.331] - - - [115584, 57984, 1, 384] - - [428, 90.21] - - - [115968, 384, 1, 384] - - [319, 86.074] - - - [115968, 57600, 1, 384] - - [427, 90.1] - - - [115968, 57601, 1, 384] - - [363, 89.325] - - - [115968, 57984, 1, 384] - - [428, 90.093] - - - [115968, 57985, 1, 384] - - [373, 89.167] - - - [116352, 384, 1, 384] - - [371, 85.736] - - - [116352, 57984, 1, 384] - - [428, 90.236] - - - [116352, 57985, 1, 384] - - [363, 89.217] - - - [116352, 58368, 1, 384] - - [428, 90.124] - - - [116736, 384, 1, 384] - - [319, 85.64] - - - [116736, 57984, 1, 384] - - [363, 89.847] - - - [116736, 57985, 1, 384] - - [363, 88.975] - - - [116736, 58368, 1, 384] - - [363, 89.712] - - - [116736, 58369, 1, 384] - - [363, 88.209] - - - [117120, 384, 1, 384] - - [324, 85.95] - - - [117120, 58368, 1, 384] - - [428, 90.117] - - - [117120, 58369, 1, 384] - - [363, 88.082] - - - [117120, 58752, 1, 384] - - [428, 90.026] - - - [117504, 384, 1, 384] - - [319, 85.956] - - - [117504, 58368, 1, 384] - - [428, 89.982] - - - [117504, 58369, 1, 384] - - [373, 88.005] - - - [117504, 58752, 1, 384] - - [428, 89.953] - - - [117504, 58753, 1, 384] - - [373, 88.921] - - - [117888, 384, 1, 384] - - [319, 86.237] - - - [117888, 58752, 1, 384] - - [428, 90.067] - - - [117888, 58753, 1, 384] - - [373, 88.884] - - - [117888, 59136, 1, 384] - - [428, 89.937] - - - [118272, 384, 1, 384] - - [324, 86.366] - - - [118272, 58752, 1, 384] - - [428, 89.796] - - - [118272, 58753, 1, 384] - - [363, 88.931] - - - [118272, 59136, 1, 384] - - [428, 89.745] - - - [118272, 59137, 1, 384] - - [363, 88.873] - - - [118656, 384, 1, 384] - - [324, 86.58] - - - [118656, 59136, 1, 384] - - [428, 89.958] - - - [118656, 59137, 1, 384] - - [428, 88.873] - - - [118656, 59520, 1, 384] - - [428, 89.96] - - - [119040, 384, 1, 384] - - [319, 86.688] - - - [119040, 59136, 1, 384] - - [428, 89.935] - - - [119040, 59137, 1, 384] - - [427, 88.771] - - - [119040, 59520, 1, 384] - - [428, 89.886] - - - [119040, 59521, 1, 384] - - [428, 88.825] - - - [119424, 384, 1, 384] - - [319, 86.589] - - - [119424, 59520, 1, 384] - - [428, 89.932] - - - [119424, 59521, 1, 384] - - [428, 88.728] - - - [119424, 59904, 1, 384] - - [428, 89.941] - - - [119808, 384, 1, 384] - - [319, 86.424] - - - [119808, 59520, 1, 384] - - [427, 89.317] - - - [119808, 59521, 1, 384] - - [363, 88.467] - - - [119808, 59904, 1, 384] - - [427, 89.37] - - - [119808, 59905, 1, 384] - - [363, 88.456] - - - [120192, 384, 1, 384] - - [361, 86.028] - - - [120192, 59904, 1, 384] - - [428, 89.864] - - - [120192, 59905, 1, 384] - - [428, 88.689] - - - [120192, 60288, 1, 384] - - [428, 89.891] - - - [120576, 384, 1, 384] - - [324, 86.007] - - - [120576, 59904, 1, 384] - - [428, 89.859] - - - [120576, 59905, 1, 384] - - [363, 88.654] - - - [120576, 60288, 1, 384] - - [428, 89.833] - - - [120576, 60289, 1, 384] - - [428, 88.734] - - - [120960, 384, 1, 384] - - [324, 86.068] - - - [120960, 60288, 1, 384] - - [428, 89.911] - - - [120960, 60289, 1, 384] - - [428, 88.696] - - - [120960, 60672, 1, 384] - - [428, 89.712] - - - [121344, 384, 1, 384] - - [318, 85.718] - - - [121344, 60288, 1, 384] - - [428, 89.674] - - - [121344, 60289, 1, 384] - - [428, 88.514] - - - [121344, 60672, 1, 384] - - [428, 89.561] - - - [121344, 60673, 1, 384] - - [363, 88.345] - - - [121728, 384, 1, 384] - - [341, 85.97] - - - [121728, 60672, 1, 384] - - [428, 89.683] - - - [121728, 60673, 1, 384] - - [428, 88.674] - - - [121728, 61056, 1, 384] - - [428, 89.845] - - - [122112, 384, 1, 384] - - [324, 85.967] - - - [122112, 60672, 1, 384] - - [428, 89.619] - - - [122112, 60673, 1, 384] - - [428, 88.544] - - - [122112, 61056, 1, 384] - - [428, 89.747] - - - [122112, 61057, 1, 384] - - [428, 88.614] - - - [122496, 384, 1, 384] - - [324, 86.11] - - - [122496, 61056, 1, 384] - - [428, 89.801] - - - [122496, 61057, 1, 384] - - [428, 88.637] - - - [122496, 61440, 1, 384] - - [428, 89.814] - - - [122880, 384, 1, 384] - - [319, 85.61] - - - [122880, 61056, 1, 384] - - [427, 88.112] - - - [122880, 61057, 1, 384] - - [363, 87.07] - - - [122880, 61440, 1, 384] - - [363, 87.958] - - - [122880, 61441, 1, 384] - - [363, 86.846] - - - [123264, 384, 1, 384] - - [319, 86.604] - - - [123264, 61440, 1, 384] - - [428, 89.954] - - - [123264, 61441, 1, 384] - - [373, 87.219] - - - [123264, 61824, 1, 384] - - [428, 89.819] - - - [123648, 384, 1, 384] - - [319, 86.563] - - - [123648, 61440, 1, 384] - - [428, 89.723] - - - [123648, 61441, 1, 384] - - [373, 87.212] - - - [123648, 61824, 1, 384] - - [428, 89.78] - - - [123648, 61825, 1, 384] - - [428, 88.655] - - - [124032, 384, 1, 384] - - [319, 86.54] - - - [124032, 61824, 1, 384] - - [428, 89.751] - - - [124032, 61825, 1, 384] - - [428, 88.56] - - - [124032, 62208, 1, 384] - - [428, 89.665] - - - [124416, 384, 1, 384] - - [324, 86.59] - - - [124416, 61824, 1, 384] - - [428, 89.473] - - - [124416, 61825, 1, 384] - - [428, 88.332] - - - [124416, 62208, 1, 384] - - [428, 89.445] - - - [124416, 62209, 1, 384] - - [428, 88.124] - - - [124800, 384, 1, 384] - - [319, 86.117] - - - [124800, 62208, 1, 384] - - [428, 89.66] - - - [124800, 62209, 1, 384] - - [428, 88.568] - - - [124800, 62592, 1, 384] - - [428, 89.653] - - - [125184, 384, 1, 384] - - [319, 85.98] - - - [125184, 62208, 1, 384] - - [428, 89.659] - - - [125184, 62209, 1, 384] - - [428, 88.411] - - - [125184, 62592, 1, 384] - - [428, 89.58] - - - [125184, 62593, 1, 384] - - [428, 88.531] - - - [125568, 384, 1, 384] - - [356, 85.854] - - - [125568, 62592, 1, 384] - - [428, 89.625] - - - [125568, 62593, 1, 384] - - [428, 88.456] - - - [125568, 62976, 1, 384] - - [428, 89.539] - - - [125952, 384, 1, 384] - - [319, 85.737] - - - [125952, 62592, 1, 384] - - [428, 88.72] - - - [125952, 62593, 1, 384] - - [307, 87.36] - - - [125952, 62976, 1, 384] - - [428, 88.617] - - - [125952, 62977, 1, 384] - - [373, 87.21] - - - [126336, 384, 1, 384] - - [318, 85.814] - - - [126336, 62976, 1, 384] - - [428, 89.525] - - - [126336, 62977, 1, 384] - - [428, 88.393] - - - [126336, 63360, 1, 384] - - [428, 89.631] - - - [126720, 384, 1, 384] - - [319, 86.198] - - - [126720, 62976, 1, 384] - - [428, 89.554] - - - [126720, 62977, 1, 384] - - [428, 88.354] - - - [126720, 63360, 1, 384] - - [428, 89.587] - - - [126720, 63361, 1, 384] - - [428, 88.496] - - - [127104, 384, 1, 384] - - [319, 86.089] - - - [127104, 63360, 1, 384] - - [428, 89.608] - - - [127104, 63361, 1, 384] - - [428, 88.441] - - - [127104, 63744, 1, 384] - - [428, 89.373] - - - [127488, 384, 1, 384] - - [324, 86.384] - - - [127488, 63360, 1, 384] - - [428, 89.338] - - - [127488, 63361, 1, 384] - - [428, 88.193] - - - [127488, 63744, 1, 384] - - [428, 89.139] - - - [127488, 63745, 1, 384] - - [428, 87.803] - - - [127872, 384, 1, 384] - - [324, 86.187] - - - [127872, 63744, 1, 384] - - [428, 89.362] - - - [127872, 63745, 1, 384] - - [428, 88.345] - - - [127872, 64128, 1, 384] - - [428, 89.537] - - - [128256, 384, 1, 384] - - [324, 86.71] - - - [128256, 63744, 1, 384] - - [428, 89.328] - - - [128256, 63745, 1, 384] - - [428, 88.21] - - - [128256, 64128, 1, 384] - - [428, 89.482] - - - [768, 1, 1, 384] - - [342, 0.057] - - - [64128, 127489, 1, 384] - - [338, 87.906] - - - [63744, 126721, 1, 384] - - [338, 88.217] - - - [63744, 127105, 1, 384] - - [336, 88.569] - - - [63744, 127489, 1, 384] - - [338, 87.876] - - - [63360, 125953, 1, 384] - - [340, 86.829] - - - [63360, 126337, 1, 384] - - [336, 88.503] - - - [63360, 126721, 1, 384] - - [338, 88.199] - - - [62976, 125185, 1, 384] - - [338, 88.213] - - - [62976, 125569, 1, 384] - - [336, 88.472] - - - [62976, 125953, 1, 384] - - [340, 87.09] - - - [62592, 124417, 1, 384] - - [338, 87.862] - - - [62592, 124801, 1, 384] - - [336, 88.561] - - - [62592, 125185, 1, 384] - - [337, 88.243] - - - [62208, 123649, 1, 384] - - [338, 88.247] - - - [62208, 124033, 1, 384] - - [336, 88.527] - - - [62208, 124417, 1, 384] - - [338, 87.884] - - - [61824, 122881, 1, 384] - - [340, 86.013] - - - [61824, 123265, 1, 384] - - [336, 88.575] - - - [61824, 123649, 1, 384] - - [340, 88.261] - - - [61440, 122113, 1, 384] - - [338, 88.273] - - - [61440, 122497, 1, 384] - - [336, 88.604] - - - [61440, 122881, 1, 384] - - [340, 86.332] - - - [61056, 121345, 1, 384] - - [338, 87.916] - - - [61056, 121729, 1, 384] - - [336, 88.538] - - - [61056, 122113, 1, 384] - - [337, 88.222] - - - [60672, 120577, 1, 384] - - [338, 88.305] - - - [60672, 120961, 1, 384] - - [336, 88.537] - - - [60672, 121345, 1, 384] - - [338, 87.91] - - - [60288, 119809, 1, 384] - - [316, 86.905] - - - [60288, 120193, 1, 384] - - [336, 88.56] - - - [60288, 120577, 1, 384] - - [340, 88.274] - - - [59904, 119041, 1, 384] - - [338, 88.277] - - - [59904, 119425, 1, 384] - - [336, 88.599] - - - [59904, 119809, 1, 384] - - [335, 86.727] - - - [59520, 118273, 1, 384] - - [333, 88.073] - - - [59520, 118657, 1, 384] - - [336, 88.554] - - - [59520, 119041, 1, 384] - - [338, 88.258] - - - [59136, 117505, 1, 384] - - [338, 88.313] - - - [59136, 117889, 1, 384] - - [336, 88.564] - - - [59136, 118273, 1, 384] - - [333, 88.067] - - - [58752, 116737, 1, 384] - - [335, 87.014] - - - [58752, 117121, 1, 384] - - [336, 88.577] - - - [58752, 117505, 1, 384] - - [337, 88.295] - - - [58368, 115969, 1, 384] - - [333, 88.33] - - - [58368, 116353, 1, 384] - - [336, 88.55] - - - [58368, 116737, 1, 384] - - [335, 86.991] - - - [57984, 115201, 1, 384] - - [333, 88.424] - - - [57984, 115585, 1, 384] - - [336, 88.532] - - - [57984, 115969, 1, 384] - - [333, 88.3] - - - [57600, 114433, 1, 384] - - [333, 88.561] - - - [57600, 114817, 1, 384] - - [336, 88.556] - - - [57600, 115201, 1, 384] - - [333, 88.409] - - - [57216, 113665, 1, 384] - - [276, 87.566] - - - [57216, 114049, 1, 384] - - [334, 88.717] - - - [57216, 114433, 1, 384] - - [333, 88.521] - - - [56832, 112897, 1, 384] - - [333, 88.696] - - - [56832, 113281, 1, 384] - - [334, 88.789] - - - [56832, 113665, 1, 384] - - [335, 87.345] - - - [56448, 112129, 1, 384] - - [330, 88.902] - - - [56448, 112513, 1, 384] - - [333, 88.823] - - - [56448, 112897, 1, 384] - - [333, 88.698] - - - [56064, 111361, 1, 384] - - [333, 88.83] - - - [56064, 111745, 1, 384] - - [333, 88.895] - - - [56064, 112129, 1, 384] - - [333, 88.747] - - - [55680, 110593, 1, 384] - - [276, 88.423] - - - [55680, 110977, 1, 384] - - [333, 88.963] - - - [55680, 111361, 1, 384] - - [333, 88.892] - - - [55296, 109825, 1, 384] - - [333, 89.014] - - - [55296, 110209, 1, 384] - - [334, 89.093] - - - [55296, 110593, 1, 384] - - [276, 88.376] - - - [54912, 109057, 1, 384] - - [333, 89.052] - - - [54912, 109441, 1, 384] - - [334, 89.159] - - - [54912, 109825, 1, 384] - - [333, 89.018] - - - [54528, 108289, 1, 384] - - [330, 89.208] - - - [54528, 108673, 1, 384] - - [292, 89.184] - - - [54528, 109057, 1, 384] - - [333, 89.063] - - - [54144, 107521, 1, 384] - - [276, 88.661] - - - [54144, 107905, 1, 384] - - [330, 89.295] - - - [54144, 108289, 1, 384] - - [330, 89.425] - - - [53760, 106753, 1, 384] - - [330, 89.927] - - - [53760, 107137, 1, 384] - - [329, 89.437] - - - [53760, 107521, 1, 384] - - [276, 88.639] - - - [53376, 105985, 1, 384] - - [330, 89.716] - - - [53376, 106369, 1, 384] - - [333, 89.379] - - - [53376, 106753, 1, 384] - - [329, 89.402] - - - [52992, 105217, 1, 384] - - [330, 90.36] - - - [52992, 105601, 1, 384] - - [330, 89.771] - - - [52992, 105985, 1, 384] - - [330, 89.415] - - - [52608, 104449, 1, 384] - - [276, 88.823] - - - [52608, 104833, 1, 384] - - [330, 89.845] - - - [52608, 105217, 1, 384] - - [330, 90.402] - - - [52224, 103681, 1, 384] - - [330, 90.518] - - - [52224, 104065, 1, 384] - - [330, 90.477] - - - [52224, 104449, 1, 384] - - [276, 88.823] - - - [51840, 102913, 1, 384] - - [276, 90.027] - - - [51840, 103297, 1, 384] - - [276, 90.135] - - - [51840, 103681, 1, 384] - - [276, 90.59] - - - [51456, 102145, 1, 384] - - [276, 90.722] - - - [51456, 102529, 1, 384] - - [330, 90.57] - - - [51456, 102913, 1, 384] - - [330, 90.578] - - - [51072, 101377, 1, 384] - - [276, 89.038] - - - [51072, 101761, 1, 384] - - [276, 90.688] - - - [51072, 102145, 1, 384] - - [276, 90.767] - - - [50688, 100609, 1, 384] - - [276, 90.857] - - - [50688, 100993, 1, 384] - - [276, 90.842] - - - [50688, 101377, 1, 384] - - [276, 89.034] - - - [50304, 99841, 1, 384] - - [276, 90.898] - - - [50304, 100225, 1, 384] - - [276, 90.881] - - - [50304, 100609, 1, 384] - - [330, 90.763] - - - [49920, 99073, 1, 384] - - [276, 90.932] - - - [49920, 99457, 1, 384] - - [276, 90.987] - - - [49920, 99841, 1, 384] - - [276, 90.887] - - - [49536, 98305, 1, 384] - - [328, 88.163] - - - [49536, 98689, 1, 384] - - [276, 91.126] - - - [49536, 99073, 1, 384] - - [330, 90.812] - - - [49152, 97537, 1, 384] - - [276, 91.076] - - - [49152, 97921, 1, 384] - - [330, 90.675] - - - [49152, 98305, 1, 384] - - [328, 88.177] - - - [48768, 96769, 1, 384] - - [276, 91.084] - - - [48768, 97153, 1, 384] - - [276, 91.01] - - - [48768, 97537, 1, 384] - - [276, 91.053] - - - [48384, 96001, 1, 384] - - [329, 91.203] - - - [48384, 96385, 1, 384] - - [329, 91.253] - - - [48384, 96769, 1, 384] - - [329, 91.031] - - - [48000, 95233, 1, 384] - - [276, 89.293] - - - [48000, 95617, 1, 384] - - [329, 91.33] - - - [48000, 96001, 1, 384] - - [329, 91.211] - - - [47616, 94465, 1, 384] - - [276, 91.255] - - - [47616, 94849, 1, 384] - - [329, 91.271] - - - [47616, 95233, 1, 384] - - [276, 89.247] - - - [47232, 93697, 1, 384] - - [276, 91.254] - - - [47232, 94081, 1, 384] - - [329, 91.4] - - - [47232, 94465, 1, 384] - - [329, 91.328] - - - [46848, 92929, 1, 384] - - [276, 91.344] - - - [46848, 93313, 1, 384] - - [286, 91.348] - - - [46848, 93697, 1, 384] - - [276, 91.286] - - - [46464, 92161, 1, 384] - - [328, 89.524] - - - [46464, 92545, 1, 384] - - [286, 91.592] - - - [46464, 92929, 1, 384] - - [286, 91.483] - - - [46080, 91393, 1, 384] - - [286, 91.7] - - - [46080, 91777, 1, 384] - - [286, 91.495] - - - [46080, 92161, 1, 384] - - [328, 89.557] - - - [45696, 90625, 1, 384] - - [286, 91.59] - - - [45696, 91009, 1, 384] - - [286, 91.729] - - - [45696, 91393, 1, 384] - - [286, 91.753] - - - [45312, 89857, 1, 384] - - [286, 91.802] - - - [45312, 90241, 1, 384] - - [286, 91.805] - - - [45312, 90625, 1, 384] - - [286, 91.703] - - - [44928, 89089, 1, 384] - - [275, 89.762] - - - [44928, 89473, 1, 384] - - [286, 91.831] - - - [44928, 89857, 1, 384] - - [286, 91.841] - - - [44544, 88321, 1, 384] - - [286, 91.997] - - - [44544, 88705, 1, 384] - - [286, 91.915] - - - [44544, 89089, 1, 384] - - [275, 89.802] - - - [44160, 87553, 1, 384] - - [286, 92.037] - - - [44160, 87937, 1, 384] - - [286, 91.991] - - - [44160, 88321, 1, 384] - - [286, 91.997] - - - [43776, 86785, 1, 384] - - [286, 92.115] - - - [43776, 87169, 1, 384] - - [286, 92.101] - - - [43776, 87553, 1, 384] - - [286, 92.083] - - - [43392, 86017, 1, 384] - - [275, 90.273] - - - [43392, 86401, 1, 384] - - [286, 92.074] - - - [43392, 86785, 1, 384] - - [286, 92.127] - - - [43008, 85249, 1, 384] - - [286, 92.17] - - - [43008, 85633, 1, 384] - - [286, 92.15] - - - [43008, 86017, 1, 384] - - [275, 90.173] - - - [42624, 84481, 1, 384] - - [286, 92.233] - - - [42624, 84865, 1, 384] - - [286, 92.193] - - - [42624, 85249, 1, 384] - - [286, 92.168] - - - [42240, 83713, 1, 384] - - [286, 92.167] - - - [42240, 84097, 1, 384] - - [286, 92.151] - - - [42240, 84481, 1, 384] - - [286, 92.173] - - - [41856, 82945, 1, 384] - - [275, 90.565] - - - [41856, 83329, 1, 384] - - [286, 92.184] - - - [41856, 83713, 1, 384] - - [286, 92.176] - - - [41472, 82177, 1, 384] - - [286, 92.283] - - - [41472, 82561, 1, 384] - - [286, 92.305] - - - [41472, 82945, 1, 384] - - [275, 90.349] - - - [41088, 81409, 1, 384] - - [286, 92.421] - - - [41088, 81793, 1, 384] - - [286, 92.373] - - - [41088, 82177, 1, 384] - - [286, 92.307] - - - [40704, 80641, 1, 384] - - [286, 92.286] - - - [40704, 81025, 1, 384] - - [286, 92.24] - - - [40704, 81409, 1, 384] - - [286, 92.261] - - - [40320, 79873, 1, 384] - - [275, 90.704] - - - [40320, 80257, 1, 384] - - [286, 92.267] - - - [40320, 80641, 1, 384] - - [286, 92.264] - - - [39936, 79105, 1, 384] - - [286, 92.328] - - - [39936, 79489, 1, 384] - - [286, 92.253] - - - [39936, 79873, 1, 384] - - [275, 90.739] - - - [39552, 78337, 1, 384] - - [286, 92.33] - - - [39552, 78721, 1, 384] - - [286, 92.237] - - - [39552, 79105, 1, 384] - - [286, 92.263] - - - [39168, 77569, 1, 384] - - [286, 92.269] - - - [39168, 77953, 1, 384] - - [286, 92.191] - - - [39168, 78337, 1, 384] - - [286, 92.247] - - - [38784, 76801, 1, 384] - - [275, 90.727] - - - [38784, 77185, 1, 384] - - [286, 92.376] - - - [38784, 77569, 1, 384] - - [286, 92.402] - - - [38400, 76033, 1, 384] - - [286, 92.349] - - - [38400, 76417, 1, 384] - - [286, 92.308] - - - [38400, 76801, 1, 384] - - [275, 90.722] - - - [38016, 75265, 1, 384] - - [275, 92.346] - - - [38016, 75649, 1, 384] - - [275, 92.27] - - - [38016, 76033, 1, 384] - - [286, 92.286] - - - [37632, 74497, 1, 384] - - [275, 92.343] - - - [37632, 74881, 1, 384] - - [275, 92.312] - - - [37632, 75265, 1, 384] - - [275, 92.305] - - - [37248, 73729, 1, 384] - - [275, 90.613] - - - [37248, 74113, 1, 384] - - [275, 92.356] - - - [37248, 74497, 1, 384] - - [275, 92.352] - - - [36864, 72961, 1, 384] - - [275, 92.345] - - - [36864, 73345, 1, 384] - - [275, 92.302] - - - [36864, 73729, 1, 384] - - [275, 90.66] - - - [36480, 72193, 1, 384] - - [275, 92.295] - - - [36480, 72577, 1, 384] - - [275, 92.269] - - - [36480, 72961, 1, 384] - - [275, 92.266] - - - [36096, 71425, 1, 384] - - [275, 92.449] - - - [36096, 71809, 1, 384] - - [275, 92.443] - - - [36096, 72193, 1, 384] - - [275, 92.442] - - - [35712, 70657, 1, 384] - - [275, 90.934] - - - [35712, 71041, 1, 384] - - [275, 92.474] - - - [35712, 71425, 1, 384] - - [275, 92.48] - - - [35328, 69889, 1, 384] - - [275, 92.472] - - - [35328, 70273, 1, 384] - - [275, 92.462] - - - [35328, 70657, 1, 384] - - [275, 90.908] - - - [34944, 69121, 1, 384] - - [275, 92.457] - - - [34944, 69505, 1, 384] - - [275, 92.411] - - - [34944, 69889, 1, 384] - - [275, 92.417] - - - [34560, 68353, 1, 384] - - [275, 92.44] - - - [34560, 68737, 1, 384] - - [275, 92.403] - - - [34560, 69121, 1, 384] - - [275, 92.451] - - - [34176, 67585, 1, 384] - - [275, 90.857] - - - [34176, 67969, 1, 384] - - [275, 92.379] - - - [34176, 68353, 1, 384] - - [275, 92.445] - - - [33792, 66817, 1, 384] - - [274, 92.368] - - - [33792, 67201, 1, 384] - - [274, 92.353] - - - [33792, 67585, 1, 384] - - [275, 90.792] - - - [33408, 66049, 1, 384] - - [275, 92.542] - - - [33408, 66433, 1, 384] - - [275, 92.454] - - - [33408, 66817, 1, 384] - - [275, 92.52] - - - [33024, 65281, 1, 384] - - [275, 92.601] - - - [33024, 65665, 1, 384] - - [275, 92.527] - - - [33024, 66049, 1, 384] - - [275, 92.579] - - - [32640, 64513, 1, 384] - - [275, 91.045] - - - [32640, 64897, 1, 384] - - [275, 92.498] - - - [32640, 65281, 1, 384] - - [275, 92.558] - - - [32256, 63745, 1, 384] - - [275, 92.485] - - - [32256, 64129, 1, 384] - - [275, 92.465] - - - [32256, 64513, 1, 384] - - [275, 90.976] - - - [31872, 62977, 1, 384] - - [275, 92.407] - - - [31872, 63361, 1, 384] - - [275, 92.411] - - - [31872, 63745, 1, 384] - - [275, 92.405] - - - [31488, 62209, 1, 384] - - [275, 92.401] - - - [31488, 62593, 1, 384] - - [275, 92.34] - - - [31488, 62977, 1, 384] - - [275, 92.389] - - - [31104, 61441, 1, 384] - - [275, 90.783] - - - [31104, 61825, 1, 384] - - [274, 92.353] - - - [31104, 62209, 1, 384] - - [274, 92.344] - - - [30720, 60673, 1, 384] - - [274, 92.405] - - - [30720, 61057, 1, 384] - - [274, 92.375] - - - [30720, 61441, 1, 384] - - [275, 90.62] - - - [30336, 59905, 1, 384] - - [275, 92.573] - - - [30336, 60289, 1, 384] - - [275, 92.498] - - - [30336, 60673, 1, 384] - - [275, 92.502] - - - [29952, 59137, 1, 384] - - [275, 92.508] - - - [29952, 59521, 1, 384] - - [275, 92.485] - - - [29952, 59905, 1, 384] - - [275, 92.511] - - - [29568, 58369, 1, 384] - - [275, 90.828] - - - [29568, 58753, 1, 384] - - [275, 92.445] - - - [29568, 59137, 1, 384] - - [275, 92.41] - - - [29184, 57601, 1, 384] - - [275, 92.369] - - - [29184, 57985, 1, 384] - - [275, 92.371] - - - [29184, 58369, 1, 384] - - [275, 90.768] - - - [28800, 56833, 1, 384] - - [275, 92.255] - - - [28800, 57217, 1, 384] - - [275, 92.185] - - - [28800, 57601, 1, 384] - - [275, 92.21] - - - [28416, 56065, 1, 384] - - [316, 92.184] - - - [28416, 56449, 1, 384] - - [292, 92.169] - - - [28416, 56833, 1, 384] - - [275, 92.168] - - - [28032, 55297, 1, 384] - - [275, 90.529] - - - [28032, 55681, 1, 384] - - [292, 92.235] - - - [28032, 56065, 1, 384] - - [292, 92.161] - - - [27648, 54529, 1, 384] - - [275, 92.451] - - - [27648, 54913, 1, 384] - - [275, 92.45] - - - [27648, 55297, 1, 384] - - [275, 90.69] - - - [27264, 53761, 1, 384] - - [275, 92.424] - - - [27264, 54145, 1, 384] - - [275, 92.392] - - - [27264, 54529, 1, 384] - - [275, 92.434] - - - [26880, 52993, 1, 384] - - [275, 92.398] - - - [26880, 53377, 1, 384] - - [275, 92.365] - - - [26880, 53761, 1, 384] - - [275, 92.41] - - - [26496, 52225, 1, 384] - - [275, 90.707] - - - [26496, 52609, 1, 384] - - [275, 92.304] - - - [26496, 52993, 1, 384] - - [275, 92.352] - - - [26112, 51457, 1, 384] - - [275, 92.162] - - - [26112, 51841, 1, 384] - - [288, 92.33] - - - [26112, 52225, 1, 384] - - [275, 90.68] - - - [25728, 50689, 1, 384] - - [275, 92.071] - - - [25728, 51073, 1, 384] - - [288, 92.193] - - - [25728, 51457, 1, 384] - - [274, 92.103] - - - [25344, 49921, 1, 384] - - [288, 92.271] - - - [25344, 50305, 1, 384] - - [288, 92.341] - - - [25344, 50689, 1, 384] - - [274, 91.781] - - - [24960, 49153, 1, 384] - - [275, 90.127] - - - [24960, 49537, 1, 384] - - [275, 92.344] - - - [24960, 49921, 1, 384] - - [275, 92.382] - - - [24576, 48385, 1, 384] - - [275, 92.354] - - - [24576, 48769, 1, 384] - - [275, 92.298] - - - [24576, 49153, 1, 384] - - [275, 90.599] - - - [24192, 47617, 1, 384] - - [275, 92.288] - - - [24192, 48001, 1, 384] - - [275, 92.26] - - - [24192, 48385, 1, 384] - - [275, 92.308] - - - [23808, 46849, 1, 384] - - [274, 92.225] - - - [23808, 47233, 1, 384] - - [288, 92.235] - - - [23808, 47617, 1, 384] - - [274, 92.197] - - - [23424, 46081, 1, 384] - - [275, 90.613] - - - [23424, 46465, 1, 384] - - [288, 92.124] - - - [23424, 46849, 1, 384] - - [275, 92.114] - - - [23040, 45313, 1, 384] - - [288, 92.046] - - - [23040, 45697, 1, 384] - - [288, 92.096] - - - [23040, 46081, 1, 384] - - [275, 90.361] - - - [22656, 44545, 1, 384] - - [274, 91.887] - - - [22656, 44929, 1, 384] - - [288, 92.062] - - - [22656, 45313, 1, 384] - - [288, 92.088] - - - [22272, 43777, 1, 384] - - [275, 92.231] - - - [22272, 44161, 1, 384] - - [275, 92.158] - - - [22272, 44545, 1, 384] - - [275, 92.183] - - - [21888, 43009, 1, 384] - - [275, 90.509] - - - [21888, 43393, 1, 384] - - [275, 92.143] - - - [21888, 43777, 1, 384] - - [275, 92.204] - - - [21504, 42241, 1, 384] - - [275, 92.069] - - - [21504, 42625, 1, 384] - - [275, 92.031] - - - [21504, 43009, 1, 384] - - [275, 90.457] - - - [21120, 41473, 1, 384] - - [281, 91.82] - - - [21120, 41857, 1, 384] - - [275, 91.749] - - - [21120, 42241, 1, 384] - - [288, 91.798] - - - [20736, 40705, 1, 384] - - [274, 92.079] - - - [20736, 41089, 1, 384] - - [274, 92.017] - - - [20736, 41473, 1, 384] - - [274, 92.031] - - - [20352, 39937, 1, 384] - - [275, 90.3] - - - [20352, 40321, 1, 384] - - [274, 91.997] - - - [20352, 40705, 1, 384] - - [274, 91.947] - - - [19968, 39169, 1, 384] - - [288, 91.859] - - - [19968, 39553, 1, 384] - - [288, 91.829] - - - [19968, 39937, 1, 384] - - [275, 89.971] - - - [19584, 38401, 1, 384] - - [275, 91.756] - - - [19584, 38785, 1, 384] - - [275, 91.79] - - - [19584, 39169, 1, 384] - - [275, 91.893] - - - [19200, 37633, 1, 384] - - [275, 92.021] - - - [19200, 38017, 1, 384] - - [275, 91.964] - - - [19200, 38401, 1, 384] - - [275, 91.996] - - - [18816, 36865, 1, 384] - - [275, 90.208] - - - [18816, 37249, 1, 384] - - [275, 91.826] - - - [18816, 37633, 1, 384] - - [275, 91.92] - - - [18432, 36097, 1, 384] - - [275, 91.745] - - - [18432, 36481, 1, 384] - - [275, 91.7] - - - [18432, 36865, 1, 384] - - [275, 90.171] - - - [18048, 35329, 1, 384] - - [275, 91.71] - - - [18048, 35713, 1, 384] - - [275, 91.78] - - - [18048, 36097, 1, 384] - - [275, 91.763] - - - [17664, 34561, 1, 384] - - [275, 91.516] - - - [17664, 34945, 1, 384] - - [275, 91.496] - - - [17664, 35329, 1, 384] - - [275, 91.537] - - - [17280, 33793, 1, 384] - - [275, 89.78] - - - [17280, 34177, 1, 384] - - [274, 91.699] - - - [17280, 34561, 1, 384] - - [274, 91.589] - - - [16896, 33025, 1, 384] - - [274, 91.515] - - - [16896, 33409, 1, 384] - - [274, 91.488] - - - [16896, 33793, 1, 384] - - [276, 88.965] - - - [16512, 32257, 1, 384] - - [275, 91.746] - - - [16512, 32641, 1, 384] - - [275, 91.652] - - - [16512, 33025, 1, 384] - - [275, 91.637] - - - [16128, 31489, 1, 384] - - [275, 91.455] - - - [16128, 31873, 1, 384] - - [275, 91.455] - - - [16128, 32257, 1, 384] - - [275, 91.474] - - - [15744, 30721, 1, 384] - - [275, 89.905] - - - [15744, 31105, 1, 384] - - [275, 91.293] - - - [15744, 31489, 1, 384] - - [275, 91.334] - - - [15360, 29953, 1, 384] - - [281, 91.275] - - - [15360, 30337, 1, 384] - - [281, 91.209] - - - [15360, 30721, 1, 384] - - [275, 89.767] - - - [14976, 29185, 1, 384] - - [275, 91.141] - - - [14976, 29569, 1, 384] - - [288, 91.07] - - - [14976, 29953, 1, 384] - - [275, 91.176] - - - [14592, 28417, 1, 384] - - [275, 90.799] - - - [14592, 28801, 1, 384] - - [275, 90.745] - - - [14592, 29185, 1, 384] - - [288, 90.825] - - - [14208, 27649, 1, 384] - - [276, 88.273] - - - [14208, 28033, 1, 384] - - [288, 90.529] - - - [14208, 28417, 1, 384] - - [288, 90.527] - - - [13824, 26881, 1, 384] - - [275, 91.192] - - - [13824, 27265, 1, 384] - - [275, 91.19] - - - [13824, 27649, 1, 384] - - [281, 89.333] - - - [13440, 26113, 1, 384] - - [274, 91.029] - - - [13440, 26497, 1, 384] - - [274, 90.923] - - - [13440, 26881, 1, 384] - - [274, 90.944] - - - [13056, 25345, 1, 384] - - [275, 90.78] - - - [13056, 25729, 1, 384] - - [275, 90.728] - - - [13056, 26113, 1, 384] - - [275, 90.836] - - - [12672, 24577, 1, 384] - - [275, 88.815] - - - [12672, 24961, 1, 384] - - [274, 90.484] - - - [12672, 25345, 1, 384] - - [274, 90.465] - - - [12288, 23809, 1, 384] - - [275, 90.521] - - - [12288, 24193, 1, 384] - - [275, 90.523] - - - [12288, 24577, 1, 384] - - [275, 89.249] - - - [11904, 23041, 1, 384] - - [274, 90.548] - - - [11904, 23425, 1, 384] - - [275, 90.459] - - - [11904, 23809, 1, 384] - - [274, 90.505] - - - [11520, 22273, 1, 384] - - [274, 90.629] - - - [11520, 22657, 1, 384] - - [274, 90.558] - - - [11520, 23041, 1, 384] - - [274, 90.457] - - - [11136, 21505, 1, 384] - - [281, 88.68] - - - [11136, 21889, 1, 384] - - [275, 90.601] - - - [11136, 22273, 1, 384] - - [275, 90.885] - - - [10752, 20737, 1, 384] - - [275, 90.592] - - - [10752, 21121, 1, 384] - - [275, 90.587] - - - [10752, 21505, 1, 384] - - [275, 89.124] - - - [10368, 19969, 1, 384] - - [288, 90.75] - - - [10368, 20353, 1, 384] - - [274, 90.789] - - - [10368, 20737, 1, 384] - - [275, 90.426] - - - [9984, 19201, 1, 384] - - [288, 90.553] - - - [9984, 19585, 1, 384] - - [275, 90.613] - - - [9984, 19969, 1, 384] - - [274, 90.613] - - - [9600, 18433, 1, 384] - - [275, 89.325] - - - [9600, 18817, 1, 384] - - [288, 90.628] - - - [9600, 19201, 1, 384] - - [288, 90.656] - - - [9216, 17665, 1, 384] - - [275, 90.672] - - - [9216, 18049, 1, 384] - - [275, 90.502] - - - [9216, 18433, 1, 384] - - [275, 89.1] - - - [8832, 16897, 1, 384] - - [275, 90.805] - - - [8832, 17281, 1, 384] - - [275, 90.554] - - - [8832, 17665, 1, 384] - - [275, 90.539] - - - [8448, 16129, 1, 384] - - [288, 90.54] - - - [8448, 16513, 1, 384] - - [288, 90.768] - - - [8448, 16897, 1, 384] - - [292, 90.557] - - - [8064, 15361, 1, 384] - - [275, 89.079] - - - [8064, 15745, 1, 384] - - [292, 90.445] - - - [8064, 16129, 1, 384] - - [288, 90.472] - - - [7680, 14593, 1, 384] - - [275, 90.144] - - - [7680, 14977, 1, 384] - - [288, 90.612] - - - [7680, 15361, 1, 384] - - [275, 88.932] - - - [7296, 13825, 1, 384] - - [275, 89.674] - - - [7296, 14209, 1, 384] - - [275, 89.682] - - - [7296, 14593, 1, 384] - - [275, 89.842] - - - [6912, 13057, 1, 384] - - [288, 89.608] - - - [6912, 13441, 1, 384] - - [275, 89.826] - - - [6912, 13825, 1, 384] - - [275, 89.872] - - - [6528, 12289, 1, 384] - - [275, 87.482] - - - [6528, 12673, 1, 384] - - [275, 89.165] - - - [6528, 13057, 1, 384] - - [275, 89.319] - - - [6144, 11521, 1, 384] - - [275, 88.629] - - - [6144, 11905, 1, 384] - - [286, 88.133] - - - [6144, 12289, 1, 384] - - [275, 87.106] - - - [5760, 10753, 1, 384] - - [274, 88.042] - - - [5760, 11137, 1, 384] - - [275, 88.117] - - - [5760, 11521, 1, 384] - - [275, 88.432] - - - [5376, 9985, 1, 384] - - [276, 86.702] - - - [5376, 10369, 1, 384] - - [281, 87.062] - - - [5376, 10753, 1, 384] - - [283, 87.59] - - - [4992, 9217, 1, 384] - - [275, 85.149] - - - [4992, 9601, 1, 384] - - [275, 87.416] - - - [4992, 9985, 1, 384] - - [279, 86.83] - - - [4608, 8449, 1, 384] - - [275, 85.804] - - - [4608, 8833, 1, 384] - - [280, 86.023] - - - [4608, 9217, 1, 384] - - [281, 84.861] - - - [4224, 7681, 1, 384] - - [275, 84.569] - - - [4224, 8065, 1, 384] - - [275, 83.93] - - - [4224, 8449, 1, 384] - - [279, 83.752] - - - [3840, 6913, 1, 384] - - [275, 83.92] - - - [3840, 7297, 1, 384] - - [275, 83.602] - - - [3840, 7681, 1, 384] - - [275, 84.06] - - - [3456, 6145, 1, 384] - - [320, 83.459] - - - [3456, 6529, 1, 384] - - [275, 81.809] - - - [3456, 6913, 1, 384] - - [274, 81.482] - - - [3072, 5377, 1, 384] - - [274, 76.944] - - - [3072, 5761, 1, 384] - - [275, 77.299] - - - [3072, 6145, 1, 384] - - [275, 79.328] - - - [2688, 4609, 1, 384] - - [272, 74.133] - - - [2688, 4993, 1, 384] - - [273, 77.161] - - - [2688, 5377, 1, 384] - - [350, 79.51] - - - [2304, 3841, 1, 384] - - [270, 73.591] - - - [2304, 4225, 1, 384] - - [266, 73.981] - - - [2304, 4609, 1, 384] - - [270, 74.009] - - - [1920, 3073, 1, 384] - - [421, 66.238] - - - [1920, 3457, 1, 384] - - [417, 65.98] - - - [1920, 3841, 1, 384] - - [270, 70.159] - - - [1536, 2305, 1, 384] - - [419, 59.47] - - - [1536, 2689, 1, 384] - - [266, 65.936] - - - [1536, 3073, 1, 384] - - [306, 64.205] - - - [1152, 1537, 1, 384] - - [260, 48.178] - - - [1152, 1921, 1, 384] - - [262, 54.787] - - - [1152, 2305, 1, 384] - - [263, 55.296] - - - [768, 1153, 1, 384] - - [261, 34.205] - - - [768, 1537, 1, 384] - - [260, 44.797] - - - [384, 769, 1, 384] - - [258, 20.523] - - - [512, 1025, 1, 512] - - [344, 32.377] - - - [1024, 2049, 1, 512] - - [284, 55.968] - - - [1536, 3073, 1, 512] - - [350, 69.017] - - - [2048, 4097, 1, 512] - - [420, 75.057] - - - [2560, 5121, 1, 512] - - [326, 79.143] - - - [3072, 6145, 1, 512] - - [319, 84.135] - - - [3584, 7169, 1, 512] - - [451, 86.716] - - - [1024, 1024, 8, 1024] - - [353, 87.182] - - - [2048, 2048, 4, 2048] - - [362, 90.467] - - - [4096, 4096, 2, 4096] - - [194, 96.942] - - - [8192, 8192, 1, 8192] - - [194, 97.575] - - - [16384, 16384, 1, 16384] - - [194, 99.795] - - - [768, 768, 1, 768] - - [195, 43.067] - - - [1152, 1152, 1, 1152] - - [385, 57.789] - - - [1536, 1536, 1, 1536] - - [262, 76.268] - - - [1920, 1920, 1, 1920] - - [417, 85.094] - - - [2304, 2304, 1, 2304] - - [355, 82.706] - - - [2688, 2688, 1, 2688] - - [198, 87.132] - - - [3072, 3072, 1, 3072] - - [199, 95.558] - - - [3456, 3456, 1, 3456] - - [200, 95.674] - - - [3840, 3840, 1, 3840] - - [201, 97.134] - - - [4224, 4224, 1, 4224] - - [202, 97.539] - - - [4992, 4992, 1, 4992] - - [197, 97.764] - - - [5376, 5376, 1, 5376] - - [196, 96.198] - - - [6144, 6144, 1, 6144] - - [197, 99.046] - - - [6528, 6528, 1, 6528] - - [203, 97.946] - - - [6912, 6912, 1, 6912] - - [202, 97.667] - - - [7296, 7296, 1, 7296] - - [203, 97.983] - - - [7680, 7680, 1, 7680] - - [202, 98.739] - - - [1024, 1024, 1, 2048] - - [204, 64.821] - - - [1024, 1024, 1, 3072] - - [205, 68.93] - - - [1024, 2048, 1, 11264] - - [206, 76.197] - - - [1024, 2048, 1, 15360] - - [207, 76.452] - - - [1024, 2048, 1, 3072] - - [206, 74.102] - - - [1024, 2048, 1, 7168] - - [206, 75.65] - - - [1024, 4096, 1, 13312] - - [208, 91.746] - - - [1024, 4096, 1, 5120] - - [209, 90.65] - - - [1024, 8192, 1, 9216] - - [210, 92.392] - - - [2048, 2048, 1, 4096] - - [211, 83.954] - - - [2048, 2048, 1, 5120] - - [212, 86.155] - - - [2048, 2048, 1, 6144] - - [213, 84.986] - - - [2048, 2048, 1, 7168] - - [230, 94.47] - - - [2048, 4096, 1, 14336] - - [214, 92.588] - - - [2048, 4096, 1, 6144] - - [215, 92.096] - - - [2048, 8192, 1, 10240] - - [204, 94.987] - - - [256, 256, 1, 512] - - [216, 6.691] - - - [3072, 4096, 1, 15360] - - [217, 99.323] - - - [3072, 4096, 1, 7168] - - [217, 98.993] - - - [3072, 8192, 1, 11264] - - [215, 99.396] - - - [4096, 4096, 1, 10240] - - [218, 94.88] - - - [4096, 4096, 1, 11264] - - [219, 94.974] - - - [4096, 4096, 1, 12288] - - [220, 94.911] - - - [4096, 4096, 1, 13312] - - [219, 95.034] - - - [4096, 4096, 1, 14336] - - [220, 95.025] - - - [4096, 4096, 1, 15360] - - [219, 95.07] - - - [4096, 4096, 1, 8192] - - [221, 94.473] - - - [4096, 4096, 1, 9216] - - [219, 94.893] - - - [4096, 8192, 1, 12288] - - [222, 97.713] - - - [512, 512, 1, 1024] - - [223, 31.259] - - - [5120, 8192, 1, 13312] - - [215, 96.746] - - - [6144, 8192, 1, 14336] - - [215, 99.532] - - - [7168, 8192, 1, 15360] - - [215, 98.548] - - - [8192, 8192, 1, 16384] - - [224, 97.775] - - - [1024, 1024, 2, 4096] - - [204, 74.316] - - - [1024, 1024, 2, 5120] - - [206, 75.126] - - - [128, 128, 2, 512] - - [183, 1.78] - - - [2048, 2048, 2, 10240] - - [217, 92.451] - - - [2048, 2048, 2, 11264] - - [225, 92.491] - - - [2048, 2048, 2, 8192] - - [226, 91.95] - - - [2048, 2048, 2, 9216] - - [217, 92.394] - - - [256, 256, 2, 1024] - - [216, 15.589] - - - [4096, 4096, 2, 16384] - - [224, 97.694] - - - [512, 512, 2, 2048] - - [351, 49.578] - - - [1024, 1024, 3, 6144] - - [210, 85.459] - - - [1024, 1024, 3, 7168] - - [210, 85.68] - - - [2048, 2048, 3, 12288] - - [225, 99.208] - - - [2048, 2048, 3, 13312] - - [215, 99.259] - - - [2048, 2048, 3, 14336] - - [228, 99.225] - - - [2048, 2048, 3, 15360] - - [229, 99.235] - - - [512, 512, 3, 3072] - - [119, 69.949] - - - [1024, 1024, 4, 8192] - - [242, 91.695] - - - [1024, 1024, 4, 9216] - - [230, 91.434] - - - [128, 128, 4, 1024] - - [231, 7.695] - - - [2048, 2048, 4, 16384] - - [232, 44.087] - - - [256, 256, 4, 2048] - - [231, 28.602] - - - [512, 512, 4, 4096] - - [233, 67.207] - - - [64, 64, 4, 512] - - [234, 1.575] - - - [1024, 1024, 5, 10240] - - [230, 85.793] - - - [1024, 1024, 5, 11264] - - [230, 85.905] - - - [512, 512, 5, 5120] - - [93, 69.988] - - - [1024, 1024, 6, 12288] - - [353, 90.902] - - - [1024, 1024, 6, 13312] - - [296, 91.006] - - - [256, 256, 6, 3072] - - [223, 47.307] - - - [512, 512, 6, 6144] - - [222, 80.948] - - - [1024, 1024, 7, 14336] - - [221, 88.813] - - - [1024, 1024, 7, 15360] - - [236, 88.875] - - - [512, 512, 7, 7168] - - [394, 68.431] - - - [1024, 1024, 8, 16384] - - [238, 53.337] - - - [128, 128, 8, 2048] - - [254, 14.333] - - - [256, 256, 8, 4096] - - [354, 57.235] - - - [32, 32, 8, 512] - - [223, 0.67] - - - [512, 512, 8, 8192] - - [394, 78.57] - - - [64, 64, 8, 1024] - - [223, 2.24] - - - [512, 512, 9, 9216] - - [209, 88.467] - - - [256, 256, 10, 5120] - - [208, 63.029] - - - [512, 512, 10, 10240] - - [239, 67.597] - - - [512, 512, 11, 11264] - - [240, 71.195] - - - [128, 128, 12, 3072] - - [253, 23.444] - - - [256, 256, 12, 6144] - - [227, 77.865] - - - [512, 512, 12, 12288] - - [241, 54.771] - - - [512, 512, 13, 13312] - - [227, 78.421] - - - [256, 256, 14, 7168] - - [367, 64.422] - - - [512, 512, 14, 14336] - - [209, 84.572] - - - [512, 512, 15, 15360] - - [209, 90.576] - - - [128, 128, 16, 4096] - - [253, 34.17] - - - [256, 256, 16, 8192] - - [362, 74.973] - - - [32, 32, 16, 1024] - - [252, 1.393] - - - [512, 512, 16, 16384] - - [242, 91.972] - - - [64, 64, 16, 2048] - - [234, 7.353] - - - [256, 256, 18, 9216] - - [243, 20.658] - - - [128, 128, 20, 5120] - - [253, 44.0] - - - [256, 256, 20, 10240] - - [244, 18.355] - - - [256, 256, 22, 11264] - - [217, 19.873] - - - [128, 128, 24, 6144] - - [231, 52.706] - - - [256, 256, 24, 12288] - - [245, 13.467] - - - [64, 64, 24, 3072] - - [231, 12.232] - - - [256, 256, 26, 13312] - - [390, 95.133] - - - [128, 128, 28, 7168] - - [246, 46.149] - - - [256, 256, 28, 14336] - - [235, 66.832] - - - [256, 256, 30, 15360] - - [247, 71.605] - - - [128, 128, 32, 8192] - - [246, 53.609] - - - [256, 256, 32, 16384] - - [233, 76.406] - - - [32, 32, 32, 2048] - - [251, 3.639] - - - [64, 64, 32, 4096] - - [216, 16.609] - - - [128, 128, 36, 9216] - - [207, 60.213] - - - [128, 128, 40, 10240] - - [208, 67.681] - - - [64, 64, 40, 5120] - - [254, 21.564] - - - [128, 128, 44, 11264] - - [230, 73.615] - - - [128, 128, 48, 12288] - - [206, 77.025] - - - [32, 32, 48, 3072] - - [251, 6.013] - - - [64, 64, 48, 6144] - - [251, 26.857] - - - [128, 128, 52, 13312] - - [206, 77.699] - - - [128, 128, 56, 14336] - - [248, 49.512] - - - [64, 64, 56, 7168] - - [251, 32.103] - - - [128, 128, 60, 15360] - - [249, 53.128] - - - [128, 128, 64, 16384] - - [250, 56.728] - - - [32, 32, 64, 4096] - - [234, 8.454] - - - [64, 64, 64, 8192] - - [251, 36.465] - - - [64, 64, 72, 9216] - - [223, 37.434] - - - [32, 32, 80, 5120] - - [253, 11.069] - - - [64, 64, 80, 10240] - - [223, 37.946] - - - [64, 64, 88, 11264] - - [223, 38.19] - - - [32, 32, 96, 6144] - - [251, 13.39] - - - [64, 64, 96, 12288] - - [223, 38.343] - - - [64, 64, 104, 13312] - - [223, 38.234] - - - [32, 32, 112, 7168] - - [128, 10.321] - - - [64, 64, 112, 14336] - - [124, 33.174] - - - [64, 64, 120, 15360] - - [167, 34.887] - - - [32, 32, 128, 8192] - - [137, 11.619] - - - [64, 64, 128, 16384] - - [130, 37.247] - - - [32, 32, 144, 9216] - - [128, 13.05] - - - [32, 32, 160, 10240] - - [97, 14.145] - - - [32, 32, 176, 11264] - - [128, 15.519] - - - [32, 32, 192, 12288] - - [137, 16.488] - - - [32, 32, 208, 13312] - - [124, 14.88] - - - [32, 32, 224, 14336] - - [158, 15.623] - - - [32, 32, 240, 15360] - - [158, 16.459] - - - [32, 32, 256, 16384] - - [158, 16.901] - - - [512, 512, 11, 512] - - [299, 59.312] - - - [512, 512, 21, 512] - - [418, 70.271] - - - [512, 512, 31, 512] - - [314, 75.499] - - - [512, 512, 41, 512] - - [255, 78.047] - - - [512, 512, 51, 512] - - [339, 82.719] - - - [512, 512, 61, 512] - - [331, 81.697] - - - [512, 512, 71, 512] - - [255, 79.37] - - - [512, 512, 81, 512] - - [255, 83.325] - - - [512, 512, 91, 512] - - [255, 81.232] - - - [3840, 4223, 1, 4096] - - [429, 97.857] - - - [3840, 4225, 1, 4096] - - [430, 92.088] - - - [3840, 4223, 1, 4224] - - [429, 97.953] - - - [3840, 4225, 1, 4224] - - [430, 92.221] - - - [3840, 4223, 1, 4320] - - [431, 97.97] - - - [3840, 4225, 1, 4320] - - [430, 92.232] - - - [7680, 8447, 1, 8192] - - [432, 99.471] - - - [7680, 8449, 1, 8192] - - [429, 97.449] - - - [7680, 8447, 1, 8448] - - [432, 99.483] - - - [7680, 8449, 1, 8448] - - [429, 97.466] - - - [7680, 8447, 1, 8640] - - [432, 99.497] - - - [7680, 8449, 1, 8640] - - [429, 97.475] - - - [3840, 4224, 1, 4095] - - [433, 97.732] - - - [3840, 4224, 1, 4097] - - [434, 97.732] - - - [3840, 4224, 1, 4223] - - [432, 97.78] - - - [3840, 4224, 1, 4225] - - [435, 97.792] - - - [3840, 4224, 1, 4319] - - [432, 97.852] - - - [3840, 4224, 1, 4321] - - [436, 97.9] - - - [7680, 8448, 1, 8191] - - [434, 99.431] - - - [7680, 8448, 1, 8193] - - [434, 99.413] - - - [7680, 8448, 1, 8447] - - [433, 99.439] - - - [7680, 8448, 1, 8449] - - [433, 99.447] - - - [7680, 8448, 1, 8639] - - [432, 99.455] - - - [7680, 8448, 1, 8641] - - [433, 99.455] - - - [3839, 4224, 1, 4096] - - [437, 97.689] - - - [3841, 4224, 1, 4096] - - [438, 91.004] - - - [3839, 4224, 1, 4224] - - [437, 97.728] - - - [3841, 4224, 1, 4224] - - [191, 92.464] - - - [3839, 4224, 1, 4320] - - [437, 97.806] - - - [3841, 4224, 1, 4320] - - [438, 91.18] - - - [7679, 8448, 1, 8192] - - [439, 99.402] - - - [7681, 8448, 1, 8192] - - [437, 97.351] - - - [7679, 8448, 1, 8448] - - [439, 99.422] - - - [7681, 8448, 1, 8448] - - [437, 97.373] - - - [7679, 8448, 1, 8640] - - [439, 99.436] - - - [7681, 8448, 1, 8640] - - [437, 97.395] - - - [100, 50, 1, 11776] - - [440, 6.404] - - - [100, 50, 1, 5888] - - [441, 3.446] - - - [50, 25, 1, 11776] - - [442, 1.702] - - - [50, 25, 1, 5888] - - [443, 0.948] - - - [5939, 5939, 1, 1009] - - [444, 83.29] - - - [10789, 10789, 1, 2211] - - [445, 91.365] - - - [15957, 15957, 1, 1382] - - [445, 91.576] - - - [20613, 20613, 1, 2189] - - [446, 92.613] - - - [25429, 25429, 1, 2404] - - [447, 93.043] - - - [31985, 31985, 1, 1573] - - [447, 92.685] - - - [37053, 37053, 1, 3873] - - [445, 93.873] - - - [43909, 43909, 1, 1995] - - [448, 92.982] - - - [56549, 56549, 1, 2278] - - [445, 93.264] - - - [62002, 62002, 1, 2408] - - [449, 93.472] - - - [127488, 38400, 1, 512] - - [450, 94.593] - - - [128000, 38400, 1, 512] - - [451, 94.637] - - - [127488, 25088, 1, 512] - - [451, 94.585] - - - [128000, 25088, 1, 512] - - [451, 94.625] - - - [127488, 25600, 1, 512] - - [451, 94.615] - - - [128000, 25600, 1, 512] - - [451, 94.617] - - - [127488, 25089, 1, 512] - - [451, 94.113] - - - [128000, 25089, 1, 512] - - [451, 94.14] - - - [126976, 38400, 1, 512] - - [451, 94.635] - - - [126976, 25088, 1, 512] - - [451, 94.609] - - - [126976, 25089, 1, 512] - - [451, 94.144] - - - [126976, 24577, 1, 512] - - [451, 94.053] - - - [126464, 38400, 1, 512] - - [451, 94.632] - - - [126464, 24576, 1, 512] - - [451, 94.618] - - - [126976, 24576, 1, 512] - - [451, 94.606] - - - [126464, 25088, 1, 512] - - [451, 94.635] - - - [126464, 24577, 1, 512] - - [454, 94.091] - - - [125952, 38400, 1, 512] - - [451, 94.626] - - - [125952, 24576, 1, 512] - - [451, 94.592] - - - [125952, 24577, 1, 512] - - [455, 94.076] - - - [125952, 24065, 1, 512] - - [451, 94.108] - - - [125440, 38400, 1, 512] - - [451, 94.62] - - - [125440, 24064, 1, 512] - - [451, 94.626] - - - [125952, 24064, 1, 512] - - [451, 94.613] - - - [125440, 24576, 1, 512] - - [451, 94.585] - - - [125440, 24065, 1, 512] - - [451, 94.099] - - - [124928, 38400, 1, 512] - - [451, 94.611] - - - [124928, 24064, 1, 512] - - [451, 94.612] - - - [124928, 24065, 1, 512] - - [451, 94.11] - - - [124928, 23553, 1, 512] - - [451, 94.076] - - - [124416, 38400, 1, 512] - - [451, 94.637] - - - [124416, 23552, 1, 512] - - [451, 94.605] - - - [124928, 23552, 1, 512] - - [451, 94.596] - - - [124416, 24064, 1, 512] - - [451, 94.624] - - - [124416, 23553, 1, 512] - - [451, 94.092] - - - [123904, 38400, 1, 512] - - [451, 94.64] - - - [123904, 23552, 1, 512] - - [451, 94.589] - - - [123904, 23553, 1, 512] - - [451, 94.079] - - - [123904, 23041, 1, 512] - - [453, 94.076] - - - [123392, 38400, 1, 512] - - [451, 94.647] - - - [123904, 23040, 1, 512] - - [451, 94.609] - - - [123392, 23040, 1, 512] - - [451, 94.617] - - - [123392, 23552, 1, 512] - - [451, 94.626] - - - [123392, 23041, 1, 512] - - [453, 94.075] - - - [122880, 38400, 1, 512] - - [451, 94.624] - - - [122880, 23040, 1, 512] - - [451, 94.605] - - - [122880, 23041, 1, 512] - - [455, 94.05] - - - [122880, 22529, 1, 512] - - [454, 94.021] - - - [122368, 38400, 1, 512] - - [455, 94.577] - - - [122368, 22528, 1, 512] - - [451, 94.603] - - - [122880, 22528, 1, 512] - - [453, 94.571] - - - [122368, 23040, 1, 512] - - [451, 94.63] - - - [122368, 22529, 1, 512] - - [458, 94.059] - - - [121856, 38400, 1, 512] - - [451, 94.629] - - - [121856, 22528, 1, 512] - - [451, 94.604] - - - [121856, 22529, 1, 512] - - [454, 94.046] - - - [121856, 22017, 1, 512] - - [451, 94.058] - - - [121344, 38400, 1, 512] - - [451, 94.651] - - - [121344, 22016, 1, 512] - - [451, 94.618] - - - [121856, 22016, 1, 512] - - [451, 94.59] - - - [121344, 22528, 1, 512] - - [451, 94.611] - - - [121344, 22017, 1, 512] - - [451, 94.047] - - - [120832, 38400, 1, 512] - - [451, 94.632] - - - [120832, 22016, 1, 512] - - [451, 94.607] - - - [120832, 22017, 1, 512] - - [451, 94.048] - - - [120832, 21505, 1, 512] - - [451, 94.058] - - - [120320, 38400, 1, 512] - - [451, 94.631] - - - [120832, 21504, 1, 512] - - [451, 94.61] - - - [120320, 21504, 1, 512] - - [454, 94.612] - - - [120320, 22016, 1, 512] - - [458, 94.596] - - - [120320, 21505, 1, 512] - - [451, 94.057] - - - [119808, 38400, 1, 512] - - [451, 94.64] - - - [119808, 21504, 1, 512] - - [451, 94.619] - - - [119808, 21505, 1, 512] - - [451, 94.064] - - - [119808, 20993, 1, 512] - - [453, 93.976] - - - [119296, 38400, 1, 512] - - [451, 94.641] - - - [119296, 20992, 1, 512] - - [451, 94.584] - - - [119808, 20992, 1, 512] - - [451, 94.595] - - - [119296, 21504, 1, 512] - - [451, 94.606] - - - [119296, 20993, 1, 512] - - [451, 93.987] - - - [118784, 38400, 1, 512] - - [451, 94.635] - - - [118784, 20992, 1, 512] - - [451, 94.572] - - - [118784, 20993, 1, 512] - - [453, 93.986] - - - [118784, 20481, 1, 512] - - [455, 93.983] - - - [118272, 38400, 1, 512] - - [451, 94.634] - - - [118272, 20480, 1, 512] - - [451, 94.613] - - - [118784, 20480, 1, 512] - - [454, 94.574] - - - [118272, 20992, 1, 512] - - [451, 94.603] - - - [118272, 20481, 1, 512] - - [451, 94.013] - - - [117760, 38400, 1, 512] - - [451, 94.634] - - - [117760, 20480, 1, 512] - - [453, 94.603] - - - [117760, 20481, 1, 512] - - [454, 94.01] - - - [117760, 19969, 1, 512] - - [451, 93.975] - - - [117248, 38400, 1, 512] - - [451, 94.641] - - - [117760, 19968, 1, 512] - - [451, 94.584] - - - [117248, 19968, 1, 512] - - [451, 94.597] - - - [117248, 20480, 1, 512] - - [451, 94.614] - - - [117248, 19969, 1, 512] - - [451, 93.991] - - - [116736, 38400, 1, 512] - - [451, 94.634] - - - [116736, 19968, 1, 512] - - [451, 94.586] - - - [116736, 19969, 1, 512] - - [451, 93.992] - - - [116736, 19457, 1, 512] - - [453, 93.937] - - - [116224, 38400, 1, 512] - - [451, 94.637] - - - [116224, 19456, 1, 512] - - [451, 94.614] - - - [116736, 19456, 1, 512] - - [453, 94.6] - - - [116224, 19968, 1, 512] - - [451, 94.6] - - - [116224, 19457, 1, 512] - - [454, 93.945] - - - [115712, 38400, 1, 512] - - [451, 94.632] - - - [115712, 19456, 1, 512] - - [451, 94.6] - - - [115712, 19457, 1, 512] - - [454, 93.949] - - - [115712, 18945, 1, 512] - - [453, 93.927] - - - [115200, 38400, 1, 512] - - [451, 94.638] - - - [115200, 18944, 1, 512] - - [458, 94.571] - - - [115712, 18944, 1, 512] - - [455, 94.583] - - - [115200, 19456, 1, 512] - - [451, 94.6] - - - [115200, 18945, 1, 512] - - [454, 93.908] - - - [114688, 38400, 1, 512] - - [453, 94.619] - - - [114688, 18944, 1, 512] - - [453, 94.566] - - - [114688, 18945, 1, 512] - - [453, 93.922] - - - [114688, 18433, 1, 512] - - [454, 93.902] - - - [114176, 38400, 1, 512] - - [451, 94.645] - - - [114176, 18432, 1, 512] - - [451, 94.592] - - - [114688, 18432, 1, 512] - - [453, 94.587] - - - [114176, 18944, 1, 512] - - [455, 94.556] - - - [114176, 18433, 1, 512] - - [453, 93.951] - - - [113664, 38400, 1, 512] - - [451, 94.632] - - - [113664, 18432, 1, 512] - - [453, 94.6] - - - [113664, 18433, 1, 512] - - [454, 93.951] - - - [113664, 17921, 1, 512] - - [451, 93.938] - - - [113152, 38400, 1, 512] - - [451, 94.643] - - - [113152, 17920, 1, 512] - - [451, 94.602] - - - [113664, 17920, 1, 512] - - [451, 94.606] - - - [113152, 18432, 1, 512] - - [451, 94.597] - - - [113152, 17921, 1, 512] - - [451, 93.922] - - - [112640, 38400, 1, 512] - - [451, 94.64] - - - [112640, 17920, 1, 512] - - [451, 94.59] - - - [112640, 17921, 1, 512] - - [451, 93.928] - - - [112640, 17409, 1, 512] - - [454, 93.845] - - - [112128, 38400, 1, 512] - - [451, 94.639] - - - [112128, 17408, 1, 512] - - [451, 94.577] - - - [112640, 17408, 1, 512] - - [451, 94.578] - - - [112128, 17409, 1, 512] - - [454, 93.864] - - - [112128, 17920, 1, 512] - - [451, 94.593] - - - [111616, 38400, 1, 512] - - [453, 94.613] - - - [111616, 17408, 1, 512] - - [451, 94.582] - - - [111616, 17409, 1, 512] - - [454, 93.901] - - - [111616, 16897, 1, 512] - - [451, 93.886] - - - [111104, 38400, 1, 512] - - [451, 94.624] - - - [111104, 16896, 1, 512] - - [451, 94.589] - - - [111616, 16896, 1, 512] - - [451, 94.573] - - - [111104, 16897, 1, 512] - - [451, 93.866] - - - [111104, 17408, 1, 512] - - [454, 94.565] - - - [110592, 38400, 1, 512] - - [451, 94.63] - - - [110592, 16896, 1, 512] - - [451, 94.582] - - - [110592, 16897, 1, 512] - - [451, 93.848] - - - [110592, 16385, 1, 512] - - [451, 93.812] - - - [110080, 38400, 1, 512] - - [451, 94.636] - - - [110080, 16384, 1, 512] - - [451, 94.575] - - - [110592, 16384, 1, 512] - - [451, 94.562] - - - [110080, 16896, 1, 512] - - [451, 94.573] - - - [110080, 16385, 1, 512] - - [454, 93.84] - - - [109568, 38400, 1, 512] - - [451, 94.615] - - - [109568, 16384, 1, 512] - - [458, 94.55] - - - [109568, 16385, 1, 512] - - [454, 93.798] - - - [109568, 15873, 1, 512] - - [454, 93.732] - - - [109056, 38400, 1, 512] - - [451, 94.622] - - - [109056, 15872, 1, 512] - - [451, 94.566] - - - [109568, 15872, 1, 512] - - [451, 94.563] - - - [109056, 16384, 1, 512] - - [455, 94.553] - - - [109056, 15873, 1, 512] - - [451, 93.768] - - - [108544, 38400, 1, 512] - - [451, 94.613] - - - [108544, 15872, 1, 512] - - [451, 94.55] - - - [108544, 15873, 1, 512] - - [451, 93.754] - - - [108544, 15361, 1, 512] - - [455, 93.761] - - - [108032, 38400, 1, 512] - - [451, 94.639] - - - [108032, 15360, 1, 512] - - [451, 94.546] - - - [108544, 15360, 1, 512] - - [451, 94.542] - - - [108032, 15361, 1, 512] - - [455, 93.728] - - - [108032, 15872, 1, 512] - - [451, 94.57] - - - [107520, 38400, 1, 512] - - [451, 94.616] - - - [107520, 15360, 1, 512] - - [451, 94.547] - - - [107520, 15361, 1, 512] - - [455, 93.758] - - - [107520, 14849, 1, 512] - - [458, 93.684] - - - [107008, 38400, 1, 512] - - [451, 94.626] - - - [107008, 14848, 1, 512] - - [453, 94.54] - - - [107520, 14848, 1, 512] - - [451, 94.564] - - - [107008, 15360, 1, 512] - - [453, 94.533] - - - [107008, 14849, 1, 512] - - [451, 93.727] - - - [106496, 38400, 1, 512] - - [451, 94.617] - - - [106496, 14848, 1, 512] - - [451, 94.537] - - - [106496, 14849, 1, 512] - - [458, 93.718] - - - [106496, 14337, 1, 512] - - [451, 93.719] - - - [105984, 38400, 1, 512] - - [451, 94.627] - - - [105984, 14336, 1, 512] - - [451, 94.59] - - - [106496, 14336, 1, 512] - - [451, 94.56] - - - [105984, 14848, 1, 512] - - [451, 94.55] - - - [105984, 14337, 1, 512] - - [451, 93.735] - - - [105472, 38400, 1, 512] - - [451, 94.641] - - - [105472, 14336, 1, 512] - - [451, 94.565] - - - [105472, 14337, 1, 512] - - [451, 93.72] - - - [105472, 13825, 1, 512] - - [455, 93.641] - - - [104960, 38400, 1, 512] - - [451, 94.618] - - - [104960, 13824, 1, 512] - - [453, 94.563] - - - [105472, 13824, 1, 512] - - [455, 94.528] - - - [104960, 14336, 1, 512] - - [451, 94.557] - - - [104960, 13825, 1, 512] - - [453, 93.7] - - - [104448, 38400, 1, 512] - - [451, 94.635] - - - [104448, 13824, 1, 512] - - [453, 94.544] - - - [104448, 13825, 1, 512] - - [453, 93.673] - - - [104448, 13313, 1, 512] - - [454, 93.67] - - - [103936, 38400, 1, 512] - - [451, 94.634] - - - [103936, 13312, 1, 512] - - [451, 94.556] - - - [104448, 13312, 1, 512] - - [451, 94.557] - - - [103936, 13313, 1, 512] - - [454, 93.667] - - - [103936, 13824, 1, 512] - - [453, 94.553] - - - [103424, 38400, 1, 512] - - [451, 94.631] - - - [103424, 13312, 1, 512] - - [451, 94.56] - - - [103424, 13313, 1, 512] - - [454, 93.653] - - - [103424, 12801, 1, 512] - - [455, 93.573] - - - [102912, 38400, 1, 512] - - [451, 94.622] - - - [102912, 12800, 1, 512] - - [455, 94.494] - - - [103424, 12800, 1, 512] - - [455, 94.486] - - - [102912, 13312, 1, 512] - - [451, 94.559] - - - [102912, 12801, 1, 512] - - [451, 93.582] - - - [102400, 38400, 1, 512] - - [451, 94.609] - - - [102400, 12800, 1, 512] - - [451, 94.524] - - - [102400, 12801, 1, 512] - - [455, 93.602] - - - [102400, 12289, 1, 512] - - [454, 93.542] - - - [101888, 38400, 1, 512] - - [451, 94.627] - - - [101888, 12288, 1, 512] - - [451, 94.551] - - - [102400, 12288, 1, 512] - - [454, 94.516] - - - [101888, 12800, 1, 512] - - [451, 94.534] - - - [101888, 12289, 1, 512] - - [454, 93.567] - - - [101376, 38400, 1, 512] - - [451, 94.638] - - - [101376, 12288, 1, 512] - - [453, 94.533] - - - [101376, 12289, 1, 512] - - [454, 93.537] - - - [101376, 11777, 1, 512] - - [453, 93.462] - - - [100864, 38400, 1, 512] - - [451, 94.613] - - - [100864, 11776, 1, 512] - - [451, 94.495] - - - [101376, 11776, 1, 512] - - [454, 94.515] - - - [100864, 12288, 1, 512] - - [451, 94.542] - - - [100864, 11777, 1, 512] - - [458, 93.444] - - - [100352, 38400, 1, 512] - - [451, 94.642] - - - [100352, 11776, 1, 512] - - [454, 94.508] - - - [100352, 11777, 1, 512] - - [454, 93.472] - - - [100352, 11265, 1, 512] - - [454, 93.452] - - - [99840, 38400, 1, 512] - - [451, 94.636] - - - [99840, 11264, 1, 512] - - [453, 94.526] - - - [100352, 11264, 1, 512] - - [454, 94.528] - - - [99840, 11265, 1, 512] - - [454, 93.46] - - - [99840, 11776, 1, 512] - - [455, 94.515] - - - [99328, 38400, 1, 512] - - [451, 94.633] - - - [99328, 11264, 1, 512] - - [454, 94.552] - - - [99328, 11265, 1, 512] - - [454, 93.507] - - - [99328, 10753, 1, 512] - - [451, 93.426] - - - [98816, 38400, 1, 512] - - [451, 94.632] - - - [98816, 10752, 1, 512] - - [451, 94.544] - - - [99328, 10752, 1, 512] - - [458, 94.526] - - - [98816, 10753, 1, 512] - - [451, 93.42] - - - [98816, 11264, 1, 512] - - [454, 94.522] - - - [98304, 38400, 1, 512] - - [451, 94.565] - - - [98304, 10752, 1, 512] - - [451, 94.471] - - - [98304, 10753, 1, 512] - - [451, 93.345] - - - [98304, 10241, 1, 512] - - [455, 93.274] - - - [97792, 38400, 1, 512] - - [451, 94.627] - - - [97792, 10240, 1, 512] - - [453, 94.531] - - - [98304, 10240, 1, 512] - - [453, 94.476] - - - [97792, 10752, 1, 512] - - [451, 94.485] - - - [97792, 10241, 1, 512] - - [455, 93.372] - - - [97280, 38400, 1, 512] - - [451, 94.637] - - - [97280, 10240, 1, 512] - - [455, 94.534] - - - [97280, 10241, 1, 512] - - [454, 93.375] - - - [97280, 9729, 1, 512] - - [451, 93.282] - - - [96768, 38400, 1, 512] - - [451, 94.637] - - - [96768, 9728, 1, 512] - - [453, 94.491] - - - [97280, 9728, 1, 512] - - [451, 94.492] - - - [96768, 10240, 1, 512] - - [453, 94.525] - - - [96768, 9729, 1, 512] - - [451, 93.278] - - - [96256, 38400, 1, 512] - - [451, 94.64] - - - [96256, 9728, 1, 512] - - [451, 94.508] - - - [96256, 9729, 1, 512] - - [455, 93.238] - - - [96256, 9217, 1, 512] - - [454, 93.233] - - - [95744, 38400, 1, 512] - - [451, 94.648] - - - [95744, 9216, 1, 512] - - [453, 94.5] - - - [96256, 9216, 1, 512] - - [453, 94.535] - - - [95744, 9217, 1, 512] - - [454, 93.24] - - - [95744, 9728, 1, 512] - - [455, 94.482] - - - [95232, 38400, 1, 512] - - [451, 94.627] - - - [95232, 9216, 1, 512] - - [454, 94.514] - - - [95232, 9217, 1, 512] - - [453, 93.269] - - - [95232, 8705, 1, 512] - - [455, 93.139] - - - [94720, 38400, 1, 512] - - [451, 94.64] - - - [94720, 8704, 1, 512] - - [451, 94.491] - - - [95232, 8704, 1, 512] - - [455, 94.545] - - - [94720, 8705, 1, 512] - - [455, 93.151] - - - [94720, 9216, 1, 512] - - [454, 94.53] - - - [94208, 38400, 1, 512] - - [454, 94.6] - - - [94208, 8704, 1, 512] - - [451, 94.524] - - - [94208, 8705, 1, 512] - - [455, 93.126] - - - [94208, 8193, 1, 512] - - [455, 93.054] - - - [93696, 38400, 1, 512] - - [451, 94.638] - - - [93696, 8192, 1, 512] - - [454, 94.517] - - - [94208, 8192, 1, 512] - - [458, 94.523] - - - [93696, 8704, 1, 512] - - [455, 94.518] - - - [93696, 8193, 1, 512] - - [454, 93.056] - - - [93184, 38400, 1, 512] - - [451, 94.632] - - - [93184, 8192, 1, 512] - - [458, 94.516] - - - [93184, 8193, 1, 512] - - [454, 93.109] - - - [93184, 7681, 1, 512] - - [455, 92.922] - - - [92672, 38400, 1, 512] - - [451, 94.635] - - - [92672, 7680, 1, 512] - - [455, 94.526] - - - [93184, 7680, 1, 512] - - [458, 94.479] - - - [92672, 8192, 1, 512] - - [458, 94.494] - - - [92672, 7681, 1, 512] - - [453, 92.94] - - - [92160, 38400, 1, 512] - - [451, 94.635] - - - [92160, 7680, 1, 512] - - [453, 94.488] - - - [92160, 7681, 1, 512] - - [453, 92.923] - - - [92160, 7169, 1, 512] - - [454, 92.808] - - - [91648, 38400, 1, 512] - - [451, 94.619] - - - [91648, 7168, 1, 512] - - [451, 94.495] - - - [92160, 7168, 1, 512] - - [454, 94.475] - - - [91648, 7680, 1, 512] - - [458, 94.52] - - - [91648, 7169, 1, 512] - - [451, 92.895] - - - [91136, 38400, 1, 512] - - [451, 94.632] - - - [91136, 7168, 1, 512] - - [451, 94.504] - - - [91136, 7169, 1, 512] - - [454, 92.871] - - - [91136, 6657, 1, 512] - - [458, 92.641] - - - [90624, 38400, 1, 512] - - [451, 94.632] - - - [90624, 6656, 1, 512] - - [458, 94.505] - - - [91136, 6656, 1, 512] - - [458, 94.506] - - - [90624, 7168, 1, 512] - - [451, 94.503] - - - [90624, 6657, 1, 512] - - [452, 92.672] - - - [90112, 38400, 1, 512] - - [451, 94.628] - - - [90112, 6656, 1, 512] - - [453, 94.477] - - - [90112, 6657, 1, 512] - - [452, 92.665] - - - [90112, 6145, 1, 512] - - [457, 92.475] - - - [89600, 38400, 1, 512] - - [451, 94.631] - - - [89600, 6144, 1, 512] - - [454, 94.421] - - - [90112, 6144, 1, 512] - - [452, 94.442] - - - [89600, 6656, 1, 512] - - [458, 94.49] - - - [89600, 6145, 1, 512] - - [454, 92.511] - - - [89088, 38400, 1, 512] - - [451, 94.626] - - - [89088, 6144, 1, 512] - - [454, 94.443] - - - [89088, 6145, 1, 512] - - [454, 92.488] - - - [89088, 5633, 1, 512] - - [458, 92.38] - - - [88576, 38400, 1, 512] - - [451, 94.638] - - - [88576, 5632, 1, 512] - - [453, 94.389] - - - [89088, 5632, 1, 512] - - [458, 94.493] - - - [88576, 6144, 1, 512] - - [454, 94.441] - - - [88576, 5633, 1, 512] - - [458, 92.348] - - - [88064, 38400, 1, 512] - - [451, 94.641] - - - [88064, 5632, 1, 512] - - [458, 94.407] - - - [88064, 5633, 1, 512] - - [458, 92.274] - - - [88064, 5121, 1, 512] - - [454, 92.095] - - - [87552, 38400, 1, 512] - - [451, 94.618] - - - [88064, 5120, 1, 512] - - [458, 94.392] - - - [87552, 5120, 1, 512] - - [454, 94.292] - - - [87552, 5121, 1, 512] - - [454, 92.133] - - - [87552, 5632, 1, 512] - - [458, 94.487] - - - [87040, 38400, 1, 512] - - [451, 94.634] - - - [87040, 5120, 1, 512] - - [454, 94.37] - - - [87040, 5121, 1, 512] - - [454, 92.035] - - - [87040, 4609, 1, 512] - - [453, 91.824] - - - [86528, 38400, 1, 512] - - [451, 94.623] - - - [86528, 4608, 1, 512] - - [455, 94.299] - - - [87040, 4608, 1, 512] - - [455, 94.383] - - - [86528, 5120, 1, 512] - - [455, 94.438] - - - [86528, 4609, 1, 512] - - [457, 91.732] - - - [86016, 38400, 1, 512] - - [451, 94.64] - - - [86016, 4608, 1, 512] - - [453, 94.394] - - - [86016, 4609, 1, 512] - - [453, 91.79] - - - [86016, 4097, 1, 512] - - [454, 91.416] - - - [85504, 38400, 1, 512] - - [451, 94.637] - - - [85504, 4096, 1, 512] - - [454, 94.236] - - - [86016, 4096, 1, 512] - - [457, 94.298] - - - [85504, 4608, 1, 512] - - [453, 94.344] - - - [85504, 4097, 1, 512] - - [454, 91.453] - - - [84992, 38400, 1, 512] - - [451, 94.642] - - - [84992, 4096, 1, 512] - - [453, 94.185] - - - [84992, 4097, 1, 512] - - [454, 91.502] - - - [84992, 3585, 1, 512] - - [451, 90.995] - - - [84480, 38400, 1, 512] - - [451, 94.632] - - - [84480, 3584, 1, 512] - - [458, 94.16] - - - [84992, 3584, 1, 512] - - [455, 94.045] - - - [84480, 4096, 1, 512] - - [458, 94.319] - - - [84480, 3585, 1, 512] - - [451, 90.961] - - - [83968, 38400, 1, 512] - - [451, 94.63] - - - [83968, 3584, 1, 512] - - [455, 94.272] - - - [83968, 3585, 1, 512] - - [451, 90.958] - - - [83968, 3073, 1, 512] - - [454, 90.431] - - - [83456, 38400, 1, 512] - - [451, 94.647] - - - [83456, 3072, 1, 512] - - [454, 94.002] - - - [83968, 3072, 1, 512] - - [454, 93.942] - - - [83456, 3584, 1, 512] - - [458, 94.25] - - - [83456, 3073, 1, 512] - - [454, 90.428] - - - [82944, 38400, 1, 512] - - [451, 94.639] - - - [82944, 3072, 1, 512] - - [454, 94.031] - - - [82944, 3073, 1, 512] - - [454, 90.434] - - - [82944, 2561, 1, 512] - - [455, 89.378] - - - [82432, 38400, 1, 512] - - [451, 94.624] - - - [82432, 2560, 1, 512] - - [455, 93.692] - - - [82944, 2560, 1, 512] - - [455, 94.044] - - - [82432, 2561, 1, 512] - - [455, 89.46] - - - [82432, 3072, 1, 512] - - [454, 94.018] - - - [81920, 38400, 1, 512] - - [451, 94.59] - - - [81920, 2560, 1, 512] - - [455, 93.826] - - - [81920, 2561, 1, 512] - - [455, 89.507] - - - [81920, 2049, 1, 512] - - [454, 88.229] - - - [81408, 38400, 1, 512] - - [451, 94.64] - - - [81408, 2048, 1, 512] - - [454, 93.75] - - - [81920, 2048, 1, 512] - - [454, 93.515] - - - [81408, 2049, 1, 512] - - [454, 88.407] - - - [81408, 2560, 1, 512] - - [458, 93.933] - - - [80896, 38400, 1, 512] - - [451, 94.637] - - - [80896, 2048, 1, 512] - - [454, 93.836] - - - [80896, 2049, 1, 512] - - [454, 88.058] - - - [80896, 1537, 1, 512] - - [450, 86.143] - - - [80384, 38400, 1, 512] - - [451, 94.643] - - - [80384, 1536, 1, 512] - - [459, 93.445] - - - [80896, 1536, 1, 512] - - [451, 93.033] - - - [80384, 2048, 1, 512] - - [454, 93.566] - - - [80384, 1537, 1, 512] - - [457, 86.556] - - - [79872, 38400, 1, 512] - - [451, 94.642] - - - [79872, 1536, 1, 512] - - [450, 93.031] - - - [79872, 1537, 1, 512] - - [457, 86.141] - - - [79872, 1025, 1, 512] - - [457, 82.691] - - - [79360, 38400, 1, 512] - - [451, 94.637] - - - [79360, 1024, 1, 512] - - [458, 92.51] - - - [79872, 1024, 1, 512] - - [458, 92.899] - - - [79360, 1536, 1, 512] - - [454, 93.482] - - - [79360, 1025, 1, 512] - - [454, 82.625] - - - [78848, 38400, 1, 512] - - [451, 94.644] - - - [78848, 1024, 1, 512] - - [458, 92.072] - - - [78848, 1025, 1, 512] - - [454, 82.236] - - - [78848, 513, 1, 512] - - [460, 72.143] - - - [78336, 38400, 1, 512] - - [451, 94.627] - - - [78336, 513, 1, 512] - - [450, 73.146] - - - [78336, 1024, 1, 512] - - [457, 92.964] - - - [77824, 513, 1, 512] - - [451, 72.656] - - - [90880, 512, 1, 512] - - [452, 91.158] - - - [90880, 54272, 1, 512] - - [451, 94.647] - - - [90624, 54272, 1, 512] - - [451, 94.658] - - - [90624, 36352, 1, 512] - - [451, 94.653] - - - [90880, 36352, 1, 512] - - [451, 94.636] - - - [90624, 36097, 1, 512] - - [451, 94.303] - - - [90880, 36097, 1, 512] - - [451, 94.286] - - - [90368, 512, 1, 512] - - [455, 90.659] - - - [90368, 54272, 1, 512] - - [451, 94.639] - - - [90624, 35840, 1, 512] - - [451, 94.648] - - - [90368, 35840, 1, 512] - - [451, 94.65] - - - [90368, 36097, 1, 512] - - [451, 94.293] - - - [90112, 54272, 1, 512] - - [451, 94.647] - - - [90112, 35840, 1, 512] - - [451, 94.647] - - - [90112, 35585, 1, 512] - - [453, 94.314] - - - [90368, 35585, 1, 512] - - [451, 94.271] - - - [89856, 54272, 1, 512] - - [451, 94.65] - - - [89856, 512, 1, 512] - - [456, 90.29] - - - [90112, 35328, 1, 512] - - [451, 94.637] - - - [89856, 35328, 1, 512] - - [451, 94.622] - - - [89856, 35585, 1, 512] - - [453, 94.306] - - - [89600, 54272, 1, 512] - - [451, 94.658] - - - [89600, 35328, 1, 512] - - [451, 94.643] - - - [89856, 35073, 1, 512] - - [453, 94.26] - - - [89600, 35073, 1, 512] - - [453, 94.276] - - - [89344, 512, 1, 512] - - [458, 89.935] - - - [89344, 54272, 1, 512] - - [451, 94.662] - - - [89344, 34816, 1, 512] - - [451, 94.624] - - - [89600, 34816, 1, 512] - - [451, 94.642] - - - [89344, 35073, 1, 512] - - [458, 94.261] - - - [89088, 54272, 1, 512] - - [451, 94.655] - - - [89088, 34816, 1, 512] - - [451, 94.654] - - - [89344, 34561, 1, 512] - - [453, 94.292] - - - [89088, 34561, 1, 512] - - [451, 94.299] - - - [88832, 54272, 1, 512] - - [451, 94.642] - - - [88832, 512, 1, 512] - - [451, 89.673] - - - [88832, 34304, 1, 512] - - [453, 94.634] - - - [89088, 34304, 1, 512] - - [451, 94.649] - - - [88832, 34561, 1, 512] - - [451, 94.287] - - - [88576, 54272, 1, 512] - - [451, 94.655] - - - [88576, 34304, 1, 512] - - [453, 94.645] - - - [88576, 34049, 1, 512] - - [451, 94.299] - - - [88832, 34049, 1, 512] - - [451, 94.293] - - - [88320, 54272, 1, 512] - - [451, 94.642] - - - [88320, 512, 1, 512] - - [451, 91.226] - - - [88576, 33792, 1, 512] - - [451, 94.65] - - - [88320, 33792, 1, 512] - - [451, 94.636] - - - [88320, 34049, 1, 512] - - [451, 94.295] - - - [88064, 54272, 1, 512] - - [453, 94.656] - - - [88064, 33792, 1, 512] - - [451, 94.653] - - - [88320, 33537, 1, 512] - - [451, 94.232] - - - [88064, 33537, 1, 512] - - [453, 94.254] - - - [87808, 54272, 1, 512] - - [451, 94.643] - - - [87808, 512, 1, 512] - - [460, 91.089] - - - [87808, 33280, 1, 512] - - [453, 94.619] - - - [88064, 33280, 1, 512] - - [451, 94.629] - - - [87808, 33537, 1, 512] - - [451, 94.248] - - - [87552, 54272, 1, 512] - - [451, 94.651] - - - [87552, 33280, 1, 512] - - [453, 94.627] - - - [87808, 33025, 1, 512] - - [451, 94.251] - - - [87552, 33025, 1, 512] - - [453, 94.283] - - - [87296, 54272, 1, 512] - - [451, 94.65] - - - [87296, 512, 1, 512] - - [456, 90.596] - - - [87552, 32768, 1, 512] - - [451, 94.633] - - - [87296, 32768, 1, 512] - - [453, 94.623] - - - [87296, 33025, 1, 512] - - [451, 94.253] - - - [87040, 54272, 1, 512] - - [451, 94.652] - - - [87040, 32768, 1, 512] - - [451, 94.624] - - - [87040, 32513, 1, 512] - - [453, 94.242] - - - [87296, 32513, 1, 512] - - [451, 94.226] - - - [86784, 54272, 1, 512] - - [451, 94.656] - - - [86784, 512, 1, 512] - - [458, 90.267] - - - [87040, 32256, 1, 512] - - [451, 94.654] - - - [86784, 32256, 1, 512] - - [451, 94.635] - - - [86784, 32513, 1, 512] - - [453, 94.256] - - - [86528, 54272, 1, 512] - - [451, 94.661] - - - [86528, 32256, 1, 512] - - [453, 94.65] - - - [86784, 32001, 1, 512] - - [451, 94.214] - - - [86528, 32001, 1, 512] - - [451, 94.237] - - - [86272, 54272, 1, 512] - - [451, 94.654] - - - [86272, 512, 1, 512] - - [455, 89.914] - - - [86528, 31744, 1, 512] - - [451, 94.64] - - - [86272, 31744, 1, 512] - - [451, 94.643] - - - [86272, 32001, 1, 512] - - [451, 94.233] - - - [86016, 54272, 1, 512] - - [451, 94.654] - - - [86016, 31744, 1, 512] - - [451, 94.64] - - - [86016, 31489, 1, 512] - - [458, 94.22] - - - [86272, 31489, 1, 512] - - [451, 94.222] - - - [85760, 54272, 1, 512] - - [451, 94.656] - - - [85760, 512, 1, 512] - - [450, 89.589] - - - [86016, 31232, 1, 512] - - [451, 94.648] - - - [85760, 31232, 1, 512] - - [451, 94.629] - - - [85760, 31489, 1, 512] - - [453, 94.217] - - - [85504, 54272, 1, 512] - - [451, 94.659] - - - [85504, 31232, 1, 512] - - [451, 94.64] - - - [85504, 30977, 1, 512] - - [451, 94.237] - - - [85760, 30977, 1, 512] - - [451, 94.243] - - - [85248, 54272, 1, 512] - - [451, 94.659] - - - [85248, 512, 1, 512] - - [456, 89.136] - - - [85504, 30720, 1, 512] - - [451, 94.638] - - - [85248, 30720, 1, 512] - - [451, 94.623] - - - [85248, 30977, 1, 512] - - [451, 94.236] - - - [84992, 54272, 1, 512] - - [451, 94.653] - - - [84992, 30720, 1, 512] - - [451, 94.623] - - - [84992, 30465, 1, 512] - - [451, 94.265] - - - [85248, 30465, 1, 512] - - [451, 94.249] - - - [84736, 54272, 1, 512] - - [451, 94.656] - - - [84736, 512, 1, 512] - - [459, 91.086] - - - [84992, 30208, 1, 512] - - [451, 94.639] - - - [84736, 30208, 1, 512] - - [451, 94.642] - - - [84736, 30465, 1, 512] - - [451, 94.241] - - - [84480, 54272, 1, 512] - - [451, 94.655] - - - [84480, 30208, 1, 512] - - [451, 94.643] - - - [84480, 29953, 1, 512] - - [453, 94.237] - - - [84736, 29953, 1, 512] - - [453, 94.235] - - - [84224, 54272, 1, 512] - - [451, 94.64] - - - [84224, 512, 1, 512] - - [452, 90.637] - - - [84480, 29696, 1, 512] - - [453, 94.633] - - - [84224, 29696, 1, 512] - - [453, 94.637] - - - [84224, 29953, 1, 512] - - [453, 94.216] - - - [83968, 54272, 1, 512] - - [451, 94.654] - - - [83968, 29696, 1, 512] - - [453, 94.648] - - - [83968, 29441, 1, 512] - - [451, 94.222] - - - [84224, 29441, 1, 512] - - [451, 94.201] - - - [83712, 54272, 1, 512] - - [451, 94.657] - - - [83712, 512, 1, 512] - - [457, 90.236] - - - [83968, 29184, 1, 512] - - [451, 94.654] - - - [83712, 29184, 1, 512] - - [451, 94.635] - - - [83712, 29441, 1, 512] - - [451, 94.203] - - - [83456, 54272, 1, 512] - - [451, 94.65] - - - [83456, 29184, 1, 512] - - [451, 94.63] - - - [83456, 28929, 1, 512] - - [451, 94.197] - - - [83712, 28929, 1, 512] - - [451, 94.191] - - - [83200, 54272, 1, 512] - - [451, 94.651] - - - [83200, 512, 1, 512] - - [450, 89.86] - - - [83456, 28672, 1, 512] - - [451, 94.64] - - - [83200, 28672, 1, 512] - - [451, 94.637] - - - [83200, 28929, 1, 512] - - [451, 94.214] - - - [82944, 54272, 1, 512] - - [451, 94.678] - - - [82944, 28417, 1, 512] - - [453, 94.206] - - - [83200, 28417, 1, 512] - - [453, 94.18] - - - [82944, 28672, 1, 512] - - [453, 94.649] - - - [82688, 54272, 1, 512] - - [451, 94.654] - - - [82688, 512, 1, 512] - - [459, 89.501] - - - [82944, 28160, 1, 512] - - [451, 94.608] - - - [82688, 28160, 1, 512] - - [451, 94.611] - - - [82688, 28417, 1, 512] - - [453, 94.172] - - - [82432, 54272, 1, 512] - - [451, 94.657] - - - [82432, 27905, 1, 512] - - [453, 94.175] - - - [82688, 27905, 1, 512] - - [453, 94.174] - - - [82432, 28160, 1, 512] - - [451, 94.621] - - - [82176, 54272, 1, 512] - - [451, 94.65] - - - [82176, 512, 1, 512] - - [452, 89.139] - - - [82432, 27648, 1, 512] - - [451, 94.649] - - - [82176, 27648, 1, 512] - - [451, 94.619] - - - [82176, 27905, 1, 512] - - [453, 94.165] - - - [81920, 54272, 1, 512] - - [458, 94.642] - - - [82176, 27393, 1, 512] - - [451, 94.168] - - - [81920, 27393, 1, 512] - - [451, 94.132] - - - [81920, 27648, 1, 512] - - [458, 94.617] - - - [81664, 54272, 1, 512] - - [451, 94.661] - - - [81664, 512, 1, 512] - - [459, 90.888] - - - [81920, 27136, 1, 512] - - [458, 94.599] - - - [81664, 27136, 1, 512] - - [451, 94.626] - - - [81664, 27393, 1, 512] - - [453, 94.179] - - - [81408, 54272, 1, 512] - - [451, 94.659] - - - [81408, 26881, 1, 512] - - [451, 94.19] - - - [81664, 26881, 1, 512] - - [451, 94.192] - - - [81408, 27136, 1, 512] - - [451, 94.624] - - - [81152, 54272, 1, 512] - - [451, 94.655] - - - [81152, 512, 1, 512] - - [450, 90.73] - - - [81408, 26624, 1, 512] - - [451, 94.636] - - - [81152, 26624, 1, 512] - - [451, 94.634] - - - [81152, 26881, 1, 512] - - [451, 94.188] - - - [80896, 54272, 1, 512] - - [451, 94.665] - - - [81152, 26369, 1, 512] - - [453, 94.168] - - - [80896, 26369, 1, 512] - - [454, 94.15] - - - [80896, 26624, 1, 512] - - [451, 94.631] - - - [80640, 54272, 1, 512] - - [451, 94.658] - - - [80640, 512, 1, 512] - - [452, 90.36] - - - [80896, 26112, 1, 512] - - [451, 94.618] - - - [80640, 26112, 1, 512] - - [453, 94.613] - - - [80640, 26369, 1, 512] - - [453, 94.137] - - - [80384, 54272, 1, 512] - - [451, 94.667] - - - [80384, 25857, 1, 512] - - [451, 94.166] - - - [80640, 25857, 1, 512] - - [451, 94.135] - - - [80384, 26112, 1, 512] - - [453, 94.62] - - - [80128, 54272, 1, 512] - - [451, 94.643] - - - [80128, 512, 1, 512] - - [456, 89.811] - - - [80128, 25600, 1, 512] - - [451, 94.601] - - - [80384, 25600, 1, 512] - - [451, 94.617] - - - [80128, 25857, 1, 512] - - [451, 94.155] - - - [79872, 54272, 1, 512] - - [451, 94.656] - - - [79872, 25345, 1, 512] - - [453, 94.136] - - - [80128, 25345, 1, 512] - - [451, 94.129] - - - [79872, 25600, 1, 512] - - [451, 94.613] - - - [79616, 54272, 1, 512] - - [451, 94.662] - - - [79616, 512, 1, 512] - - [453, 89.445] - - - [79872, 25088, 1, 512] - - [451, 94.642] - - - [79616, 25088, 1, 512] - - [451, 94.624] - - - [79616, 25345, 1, 512] - - [453, 94.156] - - - [79360, 54272, 1, 512] - - [451, 94.667] - - - [79360, 24833, 1, 512] - - [451, 94.109] - - - [79616, 24833, 1, 512] - - [451, 94.105] - - - [79360, 25088, 1, 512] - - [451, 94.632] - - - [79104, 54272, 1, 512] - - [451, 94.645] - - - [79104, 512, 1, 512] - - [457, 89.134] - - - [79360, 24576, 1, 512] - - [451, 94.621] - - - [79104, 24576, 1, 512] - - [454, 94.602] - - - [79104, 24833, 1, 512] - - [451, 94.101] - - - [78848, 54272, 1, 512] - - [451, 94.669] - - - [79104, 24321, 1, 512] - - [454, 94.072] - - - [78848, 24321, 1, 512] - - [451, 94.095] - - - [78848, 24576, 1, 512] - - [453, 94.615] - - - [78592, 54272, 1, 512] - - [451, 94.655] - - - [78592, 512, 1, 512] - - [455, 88.653] - - - [78848, 24064, 1, 512] - - [451, 94.632] - - - [78592, 24064, 1, 512] - - [451, 94.608] - - - [78592, 24321, 1, 512] - - [451, 94.087] - - - [78336, 54272, 1, 512] - - [451, 94.66] - - - [78592, 23809, 1, 512] - - [451, 94.097] - - - [78336, 23809, 1, 512] - - [451, 94.134] - - - [78336, 24064, 1, 512] - - [451, 94.637] - - - [78080, 54272, 1, 512] - - [451, 94.659] - - - [78080, 512, 1, 512] - - [451, 90.62] - - - [78336, 23552, 1, 512] - - [451, 94.621] - - - [78080, 23552, 1, 512] - - [451, 94.602] - - - [78080, 23809, 1, 512] - - [451, 94.094] - - - [77824, 54272, 1, 512] - - [451, 94.668] - - - [77824, 23297, 1, 512] - - [451, 94.12] - - - [78080, 23297, 1, 512] - - [451, 94.097] - - - [77824, 23552, 1, 512] - - [453, 94.603] - - - [77568, 54272, 1, 512] - - [451, 94.649] - - - [77568, 512, 1, 512] - - [458, 90.32] - - - [77824, 23040, 1, 512] - - [453, 94.602] - - - [77568, 23040, 1, 512] - - [453, 94.602] - - - [77568, 23297, 1, 512] - - [451, 94.122] - - - [77312, 54272, 1, 512] - - [451, 94.654] - - - [77312, 22785, 1, 512] - - [451, 94.024] - - - [77568, 22785, 1, 512] - - [451, 94.055] - - - [77312, 23040, 1, 512] - - [453, 94.63] - - - [77056, 54272, 1, 512] - - [451, 94.653] - - - [77056, 512, 1, 512] - - [459, 89.732] - - - [77056, 22528, 1, 512] - - [453, 94.597] - - - [77312, 22528, 1, 512] - - [453, 94.598] - - - [77056, 22785, 1, 512] - - [453, 94.05] - - - [76800, 54272, 1, 512] - - [451, 94.662] - - - [76800, 22273, 1, 512] - - [451, 94.07] - - - [77056, 22273, 1, 512] - - [451, 94.057] - - - [76800, 22528, 1, 512] - - [451, 94.599] - - - [76544, 54272, 1, 512] - - [451, 94.664] - - - [76544, 512, 1, 512] - - [458, 89.344] - - - [76800, 22016, 1, 512] - - [453, 94.593] - - - [76544, 22016, 1, 512] - - [453, 94.59] - - - [76544, 22273, 1, 512] - - [451, 94.055] - - - [76288, 54272, 1, 512] - - [451, 94.655] - - - [76288, 21761, 1, 512] - - [453, 94.039] - - - [76544, 21761, 1, 512] - - [453, 94.058] - - - [76288, 22016, 1, 512] - - [451, 94.615] - - - [76032, 54272, 1, 512] - - [453, 94.619] - - - [76032, 512, 1, 512] - - [450, 89.062] - - - [76288, 21504, 1, 512] - - [453, 94.594] - - - [76032, 21504, 1, 512] - - [451, 94.621] - - - [76032, 21761, 1, 512] - - [451, 94.03] - - - [75776, 54272, 1, 512] - - [451, 94.66] - - - [75776, 21249, 1, 512] - - [453, 94.005] - - - [76032, 21249, 1, 512] - - [454, 93.988] - - - [75776, 21504, 1, 512] - - [453, 94.61] - - - [75520, 54272, 1, 512] - - [451, 94.649] - - - [75520, 512, 1, 512] - - [451, 88.529] - - - [75776, 20992, 1, 512] - - [451, 94.576] - - - [75520, 20992, 1, 512] - - [451, 94.57] - - - [75520, 21249, 1, 512] - - [453, 94.034] - - - [75264, 54272, 1, 512] - - [451, 94.658] - - - [75264, 20737, 1, 512] - - [453, 94.008] - - - [75520, 20737, 1, 512] - - [453, 94.007] - - - [75264, 20992, 1, 512] - - [453, 94.598] - - - [75008, 54272, 1, 512] - - [451, 94.663] - - - [75008, 512, 1, 512] - - [455, 90.729] - - - [75264, 20480, 1, 512] - - [453, 94.612] - - - [75008, 20480, 1, 512] - - [451, 94.618] - - - [75008, 20737, 1, 512] - - [453, 94.022] - - - [74752, 54272, 1, 512] - - [451, 94.661] - - - [74752, 20225, 1, 512] - - [451, 93.995] - - - [75008, 20225, 1, 512] - - [454, 93.977] - - - [74752, 20480, 1, 512] - - [451, 94.61] - - - [74496, 54272, 1, 512] - - [451, 94.659] - - - [74496, 512, 1, 512] - - [454, 90.407] - - - [74752, 19968, 1, 512] - - [451, 94.58] - - - [74496, 19968, 1, 512] - - [451, 94.6] - - - [74496, 20225, 1, 512] - - [451, 93.978] - - - [74240, 54272, 1, 512] - - [451, 94.656] - - - [74240, 19713, 1, 512] - - [451, 94.02] - - - [74496, 19713, 1, 512] - - [451, 93.999] - - - [74240, 19968, 1, 512] - - [454, 94.586] - - - [73984, 54272, 1, 512] - - [451, 94.657] - - - [73984, 512, 1, 512] - - [454, 89.742] - - - [74240, 19456, 1, 512] - - [453, 94.593] - - - [73984, 19456, 1, 512] - - [453, 94.597] - - - [73984, 19713, 1, 512] - - [451, 93.982] - - - [73728, 54272, 1, 512] - - [451, 94.652] - - - [73984, 19201, 1, 512] - - [455, 93.948] - - - [73728, 19201, 1, 512] - - [455, 93.949] - - - [73728, 19456, 1, 512] - - [451, 94.599] - - - [73472, 54272, 1, 512] - - [451, 94.653] - - - [73472, 512, 1, 512] - - [450, 89.33] - - - [73472, 18944, 1, 512] - - [451, 94.556] - - - [73728, 18944, 1, 512] - - [453, 94.559] - - - [73472, 19201, 1, 512] - - [455, 93.944] - - - [73216, 54272, 1, 512] - - [451, 94.654] - - - [73216, 18689, 1, 512] - - [451, 93.953] - - - [73472, 18689, 1, 512] - - [451, 93.939] - - - [73216, 18944, 1, 512] - - [453, 94.557] - - - [72960, 54272, 1, 512] - - [451, 94.665] - - - [72960, 512, 1, 512] - - [453, 88.889] - - - [72960, 18432, 1, 512] - - [451, 94.581] - - - [73216, 18432, 1, 512] - - [453, 94.622] - - - [72960, 18689, 1, 512] - - [451, 93.927] - - - [72704, 54272, 1, 512] - - [451, 94.664] - - - [72960, 18177, 1, 512] - - [451, 93.896] - - - [72704, 18177, 1, 512] - - [454, 93.889] - - - [72704, 18432, 1, 512] - - [454, 94.579] - - - [72448, 54272, 1, 512] - - [451, 94.661] - - - [72448, 512, 1, 512] - - [456, 88.515] - - - [72704, 17920, 1, 512] - - [451, 94.595] - - - [72448, 17920, 1, 512] - - [451, 94.579] - - - [72448, 18177, 1, 512] - - [451, 93.928] - - - [72192, 54272, 1, 512] - - [451, 94.666] - - - [72192, 17665, 1, 512] - - [450, 93.863] - - - [72448, 17665, 1, 512] - - [451, 93.884] - - - [72192, 17920, 1, 512] - - [453, 94.591] - - - [71936, 54272, 1, 512] - - [451, 94.671] - - - [71936, 512, 1, 512] - - [453, 90.564] - - - [71936, 17408, 1, 512] - - [451, 94.569] - - - [72192, 17408, 1, 512] - - [451, 94.555] - - - [71936, 17665, 1, 512] - - [451, 93.868] - - - [71680, 54272, 1, 512] - - [451, 94.656] - - - [71680, 17153, 1, 512] - - [453, 93.861] - - - [71936, 17153, 1, 512] - - [451, 93.842] - - - [71680, 17408, 1, 512] - - [455, 94.568] - - - [71424, 54272, 1, 512] - - [451, 94.663] - - - [71424, 512, 1, 512] - - [450, 90.192] - - - [71680, 16896, 1, 512] - - [453, 94.589] - - - [71424, 16896, 1, 512] - - [453, 94.577] - - - [71424, 17153, 1, 512] - - [453, 93.847] - - - [71168, 54272, 1, 512] - - [453, 94.657] - - - [71424, 16641, 1, 512] - - [451, 93.837] - - - [71168, 16641, 1, 512] - - [455, 93.811] - - - [71168, 16896, 1, 512] - - [453, 94.574] - - - [70912, 54272, 1, 512] - - [451, 94.667] - - - [70912, 512, 1, 512] - - [456, 89.763] - - - [71168, 16384, 1, 512] - - [454, 94.553] - - - [70912, 16384, 1, 512] - - [454, 94.58] - - - [70912, 16641, 1, 512] - - [453, 93.83] - - - [70656, 54272, 1, 512] - - [451, 94.668] - - - [70656, 16129, 1, 512] - - [453, 93.841] - - - [70912, 16129, 1, 512] - - [451, 93.825] - - - [70656, 16384, 1, 512] - - [454, 94.58] - - - [70400, 54272, 1, 512] - - [451, 94.672] - - - [70400, 512, 1, 512] - - [453, 89.278] - - - [70656, 15872, 1, 512] - - [451, 94.545] - - - [70400, 15872, 1, 512] - - [453, 94.574] - - - [70400, 16129, 1, 512] - - [451, 93.862] - - - [70144, 54272, 1, 512] - - [451, 94.668] - - - [70144, 15617, 1, 512] - - [454, 93.693] - - - [70400, 15617, 1, 512] - - [451, 93.761] - - - [70144, 15872, 1, 512] - - [451, 94.563] - - - [69888, 54272, 1, 512] - - [451, 94.669] - - - [69888, 512, 1, 512] - - [458, 88.935] - - - [69888, 15360, 1, 512] - - [453, 94.561] - - - [70144, 15360, 1, 512] - - [454, 94.593] - - - [69888, 15617, 1, 512] - - [451, 93.728] - - - [69632, 54272, 1, 512] - - [451, 94.659] - - - [69632, 15105, 1, 512] - - [451, 93.731] - - - [69888, 15105, 1, 512] - - [451, 93.756] - - - [69632, 15360, 1, 512] - - [453, 94.54] - - - [69376, 54272, 1, 512] - - [451, 94.664] - - - [69376, 512, 1, 512] - - [454, 88.473] - - - [69376, 14848, 1, 512] - - [451, 94.556] - - - [69632, 14848, 1, 512] - - [453, 94.563] - - - [69376, 15105, 1, 512] - - [451, 93.769] - - - [69120, 54272, 1, 512] - - [451, 94.659] - - - [69120, 14593, 1, 512] - - [455, 93.718] - - - [69376, 14593, 1, 512] - - [450, 93.697] - - - [69120, 14848, 1, 512] - - [451, 94.574] - - - [68864, 54272, 1, 512] - - [451, 94.676] - - - [68864, 512, 1, 512] - - [451, 87.937] - - - [68864, 14336, 1, 512] - - [451, 94.593] - - - [69120, 14336, 1, 512] - - [451, 94.595] - - - [68864, 14593, 1, 512] - - [453, 93.721] - - - [68608, 54272, 1, 512] - - [451, 94.657] - - - [68608, 14081, 1, 512] - - [455, 93.702] - - - [68864, 14081, 1, 512] - - [455, 93.71] - - - [68608, 14336, 1, 512] - - [451, 94.564] - - - [68352, 54272, 1, 512] - - [451, 94.679] - - - [68352, 512, 1, 512] - - [452, 90.186] - - - [68352, 13824, 1, 512] - - [458, 94.569] - - - [68608, 13824, 1, 512] - - [453, 94.54] - - - [68352, 14081, 1, 512] - - [455, 93.712] - - - [68096, 54272, 1, 512] - - [451, 94.666] - - - [68096, 13569, 1, 512] - - [453, 93.633] - - - [68352, 13569, 1, 512] - - [453, 93.601] - - - [68096, 13824, 1, 512] - - [453, 94.558] - - - [67840, 54272, 1, 512] - - [451, 94.672] - - - [67840, 512, 1, 512] - - [452, 89.729] - - - [68096, 13312, 1, 512] - - [454, 94.553] - - - [67840, 13312, 1, 512] - - [453, 94.557] - - - [67840, 13569, 1, 512] - - [455, 93.612] - - - [67584, 54272, 1, 512] - - [453, 94.656] - - - [67584, 13057, 1, 512] - - [451, 93.616] - - - [67840, 13057, 1, 512] - - [454, 93.596] - - - [67584, 13312, 1, 512] - - [454, 94.566] - - - [67328, 54272, 1, 512] - - [451, 94.678] - - - [67328, 512, 1, 512] - - [454, 89.241] - - - [67328, 12800, 1, 512] - - [455, 94.557] - - - [67584, 12800, 1, 512] - - [455, 94.511] - - - [67328, 13057, 1, 512] - - [451, 93.637] - - - [67072, 54272, 1, 512] - - [451, 94.674] - - - [67072, 12545, 1, 512] - - [451, 93.596] - - - [67328, 12545, 1, 512] - - [451, 93.558] - - - [67072, 12800, 1, 512] - - [455, 94.541] - - - [66816, 54272, 1, 512] - - [451, 94.669] - - - [66816, 512, 1, 512] - - [456, 88.778] - - - [66816, 12288, 1, 512] - - [454, 94.569] - - - [67072, 12288, 1, 512] - - [451, 94.585] - - - [66816, 12545, 1, 512] - - [451, 93.622] - - - [66560, 54272, 1, 512] - - [451, 94.678] - - - [66560, 12033, 1, 512] - - [458, 93.519] - - - [66816, 12033, 1, 512] - - [454, 93.543] - - - [66560, 12288, 1, 512] - - [453, 94.532] - - - [66304, 54272, 1, 512] - - [451, 94.67] - - - [66304, 512, 1, 512] - - [458, 88.323] - - - [66304, 11776, 1, 512] - - [457, 94.534] - - - [66560, 11776, 1, 512] - - [451, 94.543] - - - [66304, 12033, 1, 512] - - [453, 93.521] - - - [66048, 54272, 1, 512] - - [451, 94.655] - - - [66048, 11521, 1, 512] - - [455, 93.484] - - - [66304, 11521, 1, 512] - - [455, 93.557] - - - [66048, 11776, 1, 512] - - [454, 94.498] - - - [65792, 54272, 1, 512] - - [451, 94.678] - - - [65792, 512, 1, 512] - - [450, 87.866] - - - [65792, 11264, 1, 512] - - [454, 94.55] - - - [66048, 11264, 1, 512] - - [455, 94.532] - - - [65792, 11521, 1, 512] - - [453, 93.528] - - - [65536, 54272, 1, 512] - - [458, 94.56] - - - [65536, 11009, 1, 512] - - [454, 93.334] - - - [65792, 11009, 1, 512] - - [455, 93.419] - - - [65536, 11264, 1, 512] - - [458, 94.456] - - - [65280, 54272, 1, 512] - - [451, 94.674] - - - [65280, 512, 1, 512] - - [457, 90.071] - - - [65536, 10752, 1, 512] - - [455, 94.44] - - - [65280, 10752, 1, 512] - - [451, 94.544] - - - [65280, 11009, 1, 512] - - [455, 93.401] - - - [65024, 54272, 1, 512] - - [451, 94.68] - - - [65280, 10497, 1, 512] - - [457, 93.299] - - - [65024, 10497, 1, 512] - - [453, 93.318] - - - [65024, 10752, 1, 512] - - [455, 94.515] - - - [64768, 54272, 1, 512] - - [451, 94.672] - - - [64768, 512, 1, 512] - - [457, 89.681] - - - [65024, 10240, 1, 512] - - [455, 94.49] - - - [64768, 10240, 1, 512] - - [453, 94.531] - - - [64768, 10497, 1, 512] - - [455, 93.344] - - - [64512, 54272, 1, 512] - - [451, 94.673] - - - [64512, 9985, 1, 512] - - [453, 93.297] - - - [64768, 9985, 1, 512] - - [453, 93.291] - - - [64512, 10240, 1, 512] - - [455, 94.526] - - - [64256, 54272, 1, 512] - - [451, 94.674] - - - [64256, 512, 1, 512] - - [453, 89.239] - - - [64256, 9728, 1, 512] - - [451, 94.552] - - - [64512, 9728, 1, 512] - - [453, 94.488] - - - [64256, 9985, 1, 512] - - [451, 93.305] - - - [64000, 54272, 1, 512] - - [451, 94.664] - - - [64000, 9473, 1, 512] - - [451, 93.164] - - - [64256, 9473, 1, 512] - - [455, 93.225] - - - [64000, 9728, 1, 512] - - [451, 94.48] - - - [63744, 54272, 1, 512] - - [451, 94.678] - - - [63744, 512, 1, 512] - - [454, 88.706] - - - [63744, 9216, 1, 512] - - [454, 94.476] - - - [64000, 9216, 1, 512] - - [454, 94.508] - - - [63744, 9473, 1, 512] - - [458, 93.255] - - - [63488, 54272, 1, 512] - - [451, 94.671] - - - [63488, 8961, 1, 512] - - [455, 93.154] - - - [63744, 8961, 1, 512] - - [455, 93.179] - - - [63488, 9216, 1, 512] - - [454, 94.522] - - - [63232, 54272, 1, 512] - - [451, 94.675] - - - [63232, 512, 1, 512] - - [451, 88.121] - - - [63488, 8704, 1, 512] - - [455, 94.471] - - - [63232, 8704, 1, 512] - - [458, 94.433] - - - [63232, 8961, 1, 512] - - [455, 93.186] - - - [62976, 54272, 1, 512] - - [451, 94.671] - - - [62976, 8449, 1, 512] - - [458, 93.069] - - - [63232, 8449, 1, 512] - - [458, 92.991] - - - [62976, 8704, 1, 512] - - [451, 94.482] - - - [62720, 54272, 1, 512] - - [451, 94.675] - - - [62720, 512, 1, 512] - - [454, 87.681] - - - [62720, 8192, 1, 512] - - [455, 94.4] - - - [62976, 8192, 1, 512] - - [454, 94.463] - - - [62720, 8449, 1, 512] - - [458, 93.085] - - - [62464, 54272, 1, 512] - - [451, 94.675] - - - [62464, 7937, 1, 512] - - [457, 92.876] - - - [62720, 7937, 1, 512] - - [454, 92.93] - - - [62464, 8192, 1, 512] - - [458, 94.484] - - - [62208, 54272, 1, 512] - - [451, 94.677] - - - [62208, 512, 1, 512] - - [453, 88.389] - - - [62208, 7680, 1, 512] - - [455, 94.426] - - - [62464, 7680, 1, 512] - - [451, 94.46] - - - [62208, 7937, 1, 512] - - [451, 92.955] - - - [61952, 54272, 1, 512] - - [451, 94.672] - - - [61952, 7425, 1, 512] - - [455, 92.804] - - - [62208, 7425, 1, 512] - - [455, 92.812] - - - [61952, 7680, 1, 512] - - [453, 94.405] - - - [61696, 54272, 1, 512] - - [451, 94.679] - - - [61696, 512, 1, 512] - - [456, 89.754] - - - [61696, 7168, 1, 512] - - [454, 94.409] - - - [61952, 7168, 1, 512] - - [454, 94.403] - - - [61696, 7425, 1, 512] - - [454, 92.793] - - - [61440, 54272, 1, 512] - - [451, 94.674] - - - [61440, 6913, 1, 512] - - [453, 92.711] - - - [61696, 6913, 1, 512] - - [453, 92.728] - - - [61440, 7168, 1, 512] - - [454, 94.38] - - - [61184, 54272, 1, 512] - - [451, 94.667] - - - [61184, 512, 1, 512] - - [458, 89.217] - - - [61184, 6656, 1, 512] - - [453, 94.362] - - - [61440, 6656, 1, 512] - - [453, 94.357] - - - [61184, 6913, 1, 512] - - [453, 92.715] - - - [60928, 54272, 1, 512] - - [451, 94.676] - - - [60928, 6401, 1, 512] - - [455, 92.548] - - - [61184, 6401, 1, 512] - - [455, 92.588] - - - [60928, 6656, 1, 512] - - [458, 94.327] - - - [60672, 54272, 1, 512] - - [451, 94.672] - - - [60672, 512, 1, 512] - - [456, 88.625] - - - [60672, 6144, 1, 512] - - [455, 94.375] - - - [60928, 6144, 1, 512] - - [454, 94.382] - - - [60672, 6401, 1, 512] - - [455, 92.588] - - - [60416, 54272, 1, 512] - - [451, 94.669] - - - [60416, 5889, 1, 512] - - [453, 92.179] - - - [60672, 5889, 1, 512] - - [450, 92.153] - - - [60416, 6144, 1, 512] - - [455, 94.369] - - - [60160, 54272, 1, 512] - - [451, 94.671] - - - [60160, 512, 1, 512] - - [451, 88.114] - - - [60160, 5632, 1, 512] - - [453, 94.275] - - - [60416, 5632, 1, 512] - - [453, 94.24] - - - [60160, 5889, 1, 512] - - [454, 92.208] - - - [59904, 54272, 1, 512] - - [451, 94.664] - - - [60160, 5377, 1, 512] - - [451, 92.227] - - - [59904, 5377, 1, 512] - - [451, 91.976] - - - [59904, 5632, 1, 512] - - [453, 94.314] - - - [59648, 54272, 1, 512] - - [451, 94.674] - - - [59648, 512, 1, 512] - - [458, 87.642] - - - [59648, 5120, 1, 512] - - [454, 94.244] - - - [59904, 5120, 1, 512] - - [453, 94.156] - - - [59648, 5377, 1, 512] - - [451, 92.057] - - - [59392, 54272, 1, 512] - - [451, 94.669] - - - [59392, 4865, 1, 512] - - [453, 91.624] - - - [59648, 4865, 1, 512] - - [455, 91.765] - - - [59392, 5120, 1, 512] - - [454, 94.277] - - - [59136, 54272, 1, 512] - - [451, 94.672] - - - [59136, 512, 1, 512] - - [458, 87.043] - - - [59136, 4608, 1, 512] - - [453, 94.211] - - - [59392, 4608, 1, 512] - - [453, 94.159] - - - [59136, 4865, 1, 512] - - [452, 91.669] - - - [58880, 54272, 1, 512] - - [451, 94.663] - - - [58880, 4353, 1, 512] - - [458, 91.446] - - - [59136, 4353, 1, 512] - - [454, 91.249] - - - [58880, 4608, 1, 512] - - [453, 94.018] - - - [58624, 54272, 1, 512] - - [451, 94.68] - - - [58624, 512, 1, 512] - - [454, 89.747] - - - [58880, 4096, 1, 512] - - [453, 93.862] - - - [58624, 4096, 1, 512] - - [454, 94.055] - - - [58624, 4353, 1, 512] - - [451, 91.5] - - - [58368, 54272, 1, 512] - - [453, 94.667] - - - [58368, 3841, 1, 512] - - [455, 90.983] - - - [58624, 3841, 1, 512] - - [458, 90.971] - - - [58368, 4096, 1, 512] - - [454, 94.179] - - - [58112, 54272, 1, 512] - - [451, 94.673] - - - [58112, 512, 1, 512] - - [453, 89.094] - - - [58112, 3584, 1, 512] - - [455, 93.898] - - - [58368, 3584, 1, 512] - - [451, 93.883] - - - [58112, 3841, 1, 512] - - [455, 91.083] - - - [57856, 54272, 1, 512] - - [451, 94.68] - - - [58112, 3329, 1, 512] - - [455, 90.442] - - - [57856, 3329, 1, 512] - - [452, 90.374] - - - [57856, 3584, 1, 512] - - [451, 94.081] - - - [57600, 54272, 1, 512] - - [451, 94.68] - - - [57600, 512, 1, 512] - - [453, 88.417] - - - [57856, 3072, 1, 512] - - [453, 93.58] - - - [57600, 3072, 1, 512] - - [454, 93.897] - - - [57600, 3329, 1, 512] - - [455, 90.539] - - - [57344, 54272, 1, 512] - - [451, 94.667] - - - [57344, 2817, 1, 512] - - [451, 89.776] - - - [57600, 2817, 1, 512] - - [458, 89.547] - - - [57344, 3072, 1, 512] - - [453, 93.626] - - - [57088, 54272, 1, 512] - - [451, 94.663] - - - [57088, 512, 1, 512] - - [452, 87.967] - - - [57088, 2560, 1, 512] - - [454, 93.398] - - - [57344, 2560, 1, 512] - - [458, 93.668] - - - [57088, 2817, 1, 512] - - [458, 89.6] - - - [56832, 54272, 1, 512] - - [451, 94.675] - - - [56832, 2305, 1, 512] - - [453, 88.705] - - - [57088, 2305, 1, 512] - - [450, 88.836] - - - [56832, 2560, 1, 512] - - [454, 93.812] - - - [56576, 54272, 1, 512] - - [451, 94.672] - - - [56576, 512, 1, 512] - - [456, 87.339] - - - [56576, 2048, 1, 512] - - [454, 92.906] - - - [56832, 2048, 1, 512] - - [453, 93.173] - - - [56576, 2305, 1, 512] - - [453, 88.401] - - - [56320, 54272, 1, 512] - - [451, 94.671] - - - [56576, 1793, 1, 512] - - [451, 87.332] - - - [56320, 1793, 1, 512] - - [459, 87.002] - - - [56320, 2048, 1, 512] - - [453, 93.511] - - - [56064, 54272, 1, 512] - - [451, 94.675] - - - [56064, 512, 1, 512] - - [451, 86.989] - - - [56064, 1536, 1, 512] - - [459, 92.446] - - - [56320, 1536, 1, 512] - - [451, 92.791] - - - [56064, 1793, 1, 512] - - [459, 86.634] - - - [55808, 54272, 1, 512] - - [451, 94.678] - - - [55808, 1281, 1, 512] - - [455, 84.755] - - - [56064, 1281, 1, 512] - - [455, 83.624] - - - [55808, 1536, 1, 512] - - [454, 92.17] - - - [55552, 54272, 1, 512] - - [451, 94.681] - - - [55552, 512, 1, 512] - - [451, 89.669] - - - [55808, 1024, 1, 512] - - [458, 90.754] - - - [55552, 1024, 1, 512] - - [455, 92.224] - - - [55552, 1281, 1, 512] - - [455, 84.432] - - - [55296, 54272, 1, 512] - - [451, 94.662] - - - [55296, 769, 1, 512] - - [450, 78.344] - - - [55552, 769, 1, 512] - - [450, 78.588] - - - [55296, 1024, 1, 512] - - [455, 91.995] - - - [55040, 54272, 1, 512] - - [451, 94.689] - - - [55040, 512, 1, 512] - - [452, 89.119] - - - [55040, 769, 1, 512] - - [450, 77.922] - - - [54784, 54272, 1, 512] - - [451, 94.674] - - - [54784, 257, 1, 512] - - [453, 57.105] - - - [55040, 257, 1, 512] - - [460, 57.324] - - - [54528, 54272, 1, 512] - - [451, 94.67] - - - [54528, 512, 1, 512] - - [459, 88.531] - - - [54528, 257, 1, 512] - - [453, 56.697] - - - [54528, 54017, 1, 512] - - [451, 94.455] - - - [54272, 54017, 1, 512] - - [451, 94.446] - - - [54272, 54272, 1, 512] - - [451, 94.676] - - - [54016, 54017, 1, 512] - - [451, 94.453] - - - [54016, 512, 1, 512] - - [458, 87.779] - - - [54016, 53760, 1, 512] - - [451, 94.69] - - - [54272, 53760, 1, 512] - - [451, 94.685] - - - [53760, 53505, 1, 512] - - [453, 94.444] - - - [54016, 53505, 1, 512] - - [451, 94.461] - - - [53760, 53760, 1, 512] - - [451, 94.691] - - - [53504, 53505, 1, 512] - - [451, 94.454] - - - [53504, 512, 1, 512] - - [452, 87.151] - - - [53504, 53248, 1, 512] - - [451, 94.686] - - - [53760, 53248, 1, 512] - - [451, 94.683] - - - [53248, 52993, 1, 512] - - [453, 94.446] - - - [53504, 52993, 1, 512] - - [451, 94.459] - - - [53248, 53248, 1, 512] - - [451, 94.674] - - - [52992, 52993, 1, 512] - - [453, 94.439] - - - [52992, 512, 1, 512] - - [452, 86.611] - - - [52992, 52736, 1, 512] - - [451, 94.691] - - - [53248, 52736, 1, 512] - - [451, 94.682] - - - [52992, 52481, 1, 512] - - [451, 94.45] - - - [52736, 52481, 1, 512] - - [451, 94.447] - - - [52736, 52736, 1, 512] - - [451, 94.687] - - - [52480, 52481, 1, 512] - - [451, 94.458] - - - [52480, 512, 1, 512] - - [452, 86.157] - - - [52480, 52224, 1, 512] - - [453, 94.682] - - - [52736, 52224, 1, 512] - - [451, 94.687] - - - [52480, 51969, 1, 512] - - [451, 94.464] - - - [52224, 51969, 1, 512] - - [451, 94.485] - - - [52224, 52224, 1, 512] - - [451, 94.691] - - - [51968, 51969, 1, 512] - - [451, 94.472] - - - [51968, 512, 1, 512] - - [452, 89.037] - - - [52224, 51712, 1, 512] - - [451, 94.702] - - - [51968, 51712, 1, 512] - - [451, 94.693] - - - [51968, 51457, 1, 512] - - [451, 94.441] - - - [51712, 51457, 1, 512] - - [451, 94.445] - - - [51712, 51712, 1, 512] - - [451, 94.692] - - - [51456, 51457, 1, 512] - - [453, 94.438] - - - [51456, 512, 1, 512] - - [451, 88.354] - - - [51712, 51200, 1, 512] - - [451, 94.685] - - - [51456, 51200, 1, 512] - - [453, 94.661] - - - [51200, 50945, 1, 512] - - [451, 94.458] - - - [51456, 50945, 1, 512] - - [451, 94.432] - - - [51200, 51200, 1, 512] - - [451, 94.687] - - - [50944, 50945, 1, 512] - - [451, 94.447] - - - [50944, 512, 1, 512] - - [456, 87.667] - - - [50944, 50688, 1, 512] - - [451, 94.682] - - - [51200, 50688, 1, 512] - - [451, 94.686] - - - [50944, 50433, 1, 512] - - [451, 94.452] - - - [50688, 50433, 1, 512] - - [451, 94.444] - - - [50688, 50688, 1, 512] - - [451, 94.69] - - - [50432, 50433, 1, 512] - - [451, 94.434] - - - [50432, 512, 1, 512] - - [451, 87.002] - - - [50432, 50176, 1, 512] - - [453, 94.684] - - - [50688, 50176, 1, 512] - - [451, 94.691] - - - [50176, 49921, 1, 512] - - [451, 94.442] - - - [50432, 49921, 1, 512] - - [453, 94.424] - - - [50176, 50176, 1, 512] - - [451, 94.702] - - - [49920, 49921, 1, 512] - - [451, 94.43] - - - [49920, 512, 1, 512] - - [458, 86.369] - - - [49920, 49664, 1, 512] - - [451, 94.679] - - - [50176, 49664, 1, 512] - - [451, 94.682] - - - [49664, 49409, 1, 512] - - [451, 94.436] - - - [49920, 49409, 1, 512] - - [451, 94.434] - - - [49664, 49664, 1, 512] - - [451, 94.695] - - - [49408, 49409, 1, 512] - - [451, 94.43] - - - [49408, 512, 1, 512] - - [456, 85.685] - - - [49408, 49152, 1, 512] - - [451, 94.678] - - - [49664, 49152, 1, 512] - - [451, 94.693] - - - [49408, 48897, 1, 512] - - [451, 94.41] - - - [49152, 48897, 1, 512] - - [453, 94.409] - - - [49152, 49152, 1, 512] - - [453, 94.65] - - - [48896, 48897, 1, 512] - - [451, 94.446] - - - [48896, 512, 1, 512] - - [457, 88.996] - - - [48896, 48640, 1, 512] - - [451, 94.677] - - - [49152, 48640, 1, 512] - - [453, 94.65] - - - [48640, 48385, 1, 512] - - [451, 94.454] - - - [48896, 48385, 1, 512] - - [451, 94.465] - - - [48640, 48640, 1, 512] - - [451, 94.679] - - - [48384, 48385, 1, 512] - - [451, 94.456] - - - [48384, 512, 1, 512] - - [456, 88.296] - - - [48384, 48128, 1, 512] - - [451, 94.695] - - - [48640, 48128, 1, 512] - - [451, 94.707] - - - [48128, 47873, 1, 512] - - [451, 94.43] - - - [48384, 47873, 1, 512] - - [453, 94.417] - - - [48128, 48128, 1, 512] - - [451, 94.695] - - - [47872, 47873, 1, 512] - - [451, 94.429] - - - [47872, 512, 1, 512] - - [454, 87.451] - - - [47872, 47616, 1, 512] - - [451, 94.695] - - - [48128, 47616, 1, 512] - - [451, 94.686] - - - [47616, 47361, 1, 512] - - [451, 94.431] - - - [47872, 47361, 1, 512] - - [451, 94.435] - - - [47616, 47616, 1, 512] - - [451, 94.709] - - - [47360, 47361, 1, 512] - - [451, 94.439] - - - [47360, 512, 1, 512] - - [453, 86.757] - - - [47360, 47104, 1, 512] - - [451, 94.705] - - - [47616, 47104, 1, 512] - - [451, 94.696] - - - [47104, 46849, 1, 512] - - [451, 94.434] - - - [47360, 46849, 1, 512] - - [451, 94.437] - - - [47104, 47104, 1, 512] - - [451, 94.711] - - - [46848, 46849, 1, 512] - - [451, 94.433] - - - [46848, 512, 1, 512] - - [454, 86.097] - - - [46848, 46592, 1, 512] - - [451, 94.687] - - - [47104, 46592, 1, 512] - - [451, 94.711] - - - [46848, 46337, 1, 512] - - [451, 94.431] - - - [46592, 46337, 1, 512] - - [451, 94.438] - - - [46592, 46592, 1, 512] - - [451, 94.713] - - - [46336, 46337, 1, 512] - - [453, 94.418] - - - [46336, 512, 1, 512] - - [459, 85.52] - - - [46336, 46080, 1, 512] - - [451, 94.689] - - - [46592, 46080, 1, 512] - - [451, 94.711] - - - [46336, 45825, 1, 512] - - [454, 94.41] - - - [46080, 45825, 1, 512] - - [451, 94.424] - - - [46080, 46080, 1, 512] - - [451, 94.698] - - - [45824, 45825, 1, 512] - - [451, 94.406] - - - [45824, 512, 1, 512] - - [459, 84.964] - - - [45824, 45568, 1, 512] - - [451, 94.688] - - - [46080, 45568, 1, 512] - - [451, 94.7] - - - [45568, 45313, 1, 512] - - [451, 94.419] - - - [45824, 45313, 1, 512] - - [451, 94.402] - - - [45568, 45568, 1, 512] - - [451, 94.701] - - - [45312, 45313, 1, 512] - - [451, 94.41] - - - [45312, 512, 1, 512] - - [450, 88.259] - - - [45312, 45056, 1, 512] - - [451, 94.688] - - - [45568, 45056, 1, 512] - - [451, 94.693] - - - [45056, 44801, 1, 512] - - [451, 94.431] - - - [45312, 44801, 1, 512] - - [451, 94.447] - - - [45056, 45056, 1, 512] - - [451, 94.683] - - - [44800, 44801, 1, 512] - - [451, 94.461] - - - [44800, 512, 1, 512] - - [454, 87.289] - - - [44800, 44544, 1, 512] - - [451, 94.706] - - - [45056, 44544, 1, 512] - - [451, 94.719] - - - [44544, 44289, 1, 512] - - [454, 94.396] - - - [44800, 44289, 1, 512] - - [451, 94.413] - - - [44544, 44544, 1, 512] - - [451, 94.708] - - - [44288, 44289, 1, 512] - - [451, 94.41] - - - [44288, 512, 1, 512] - - [453, 86.52] - - - [44288, 44032, 1, 512] - - [451, 94.691] - - - [44544, 44032, 1, 512] - - [451, 94.698] - - - [44032, 43777, 1, 512] - - [453, 94.416] - - - [44288, 43777, 1, 512] - - [453, 94.428] - - - [44032, 44032, 1, 512] - - [451, 94.694] - - - [43776, 43777, 1, 512] - - [453, 94.414] - - - [43776, 512, 1, 512] - - [458, 85.761] - - - [43776, 43520, 1, 512] - - [451, 94.69] - - - [44032, 43520, 1, 512] - - [451, 94.697] - - - [43520, 43265, 1, 512] - - [453, 94.392] - - - [43776, 43265, 1, 512] - - [451, 94.407] - - - [43520, 43520, 1, 512] - - [451, 94.695] - - - [43264, 43265, 1, 512] - - [451, 94.409] - - - [43264, 512, 1, 512] - - [459, 84.985] - - - [43264, 43008, 1, 512] - - [451, 94.689] - - - [43520, 43008, 1, 512] - - [451, 94.703] - - - [43008, 42753, 1, 512] - - [453, 94.39] - - - [43264, 42753, 1, 512] - - [451, 94.397] - - - [43008, 43008, 1, 512] - - [453, 94.68] - - - [42752, 42753, 1, 512] - - [453, 94.385] - - - [42752, 512, 1, 512] - - [451, 84.324] - - - [42752, 42496, 1, 512] - - [451, 94.698] - - - [43008, 42496, 1, 512] - - [453, 94.675] - - - [42496, 42241, 1, 512] - - [451, 94.409] - - - [42752, 42241, 1, 512] - - [451, 94.413] - - - [42496, 42496, 1, 512] - - [451, 94.685] - - - [42240, 42241, 1, 512] - - [451, 94.409] - - - [42240, 512, 1, 512] - - [457, 88.156] - - - [42240, 41984, 1, 512] - - [451, 94.691] - - - [42496, 41984, 1, 512] - - [451, 94.699] - - - [41984, 41729, 1, 512] - - [451, 94.394] - - - [42240, 41729, 1, 512] - - [451, 94.38] - - - [41984, 41984, 1, 512] - - [451, 94.701] - - - [41728, 41729, 1, 512] - - [451, 94.381] - - - [41728, 512, 1, 512] - - [456, 87.102] - - - [41728, 41472, 1, 512] - - [451, 94.677] - - - [41984, 41472, 1, 512] - - [451, 94.693] - - - [41472, 41217, 1, 512] - - [451, 94.408] - - - [41728, 41217, 1, 512] - - [451, 94.398] - - - [41472, 41472, 1, 512] - - [451, 94.686] - - - [41216, 41217, 1, 512] - - [451, 94.383] - - - [41216, 512, 1, 512] - - [460, 86.149] - - - [41216, 40960, 1, 512] - - [451, 94.666] - - - [41472, 40960, 1, 512] - - [451, 94.68] - - - [40960, 40705, 1, 512] - - [455, 94.334] - - - [41216, 40705, 1, 512] - - [451, 94.334] - - - [40960, 40960, 1, 512] - - [451, 94.668] - - - [40704, 40705, 1, 512] - - [453, 94.368] - - - [40704, 512, 1, 512] - - [452, 85.451] - - - [40704, 40448, 1, 512] - - [451, 94.677] - - - [40960, 40448, 1, 512] - - [451, 94.658] - - - [40448, 40193, 1, 512] - - [451, 94.385] - - - [40704, 40193, 1, 512] - - [451, 94.386] - - - [40448, 40448, 1, 512] - - [451, 94.688] - - - [40192, 40193, 1, 512] - - [451, 94.364] - - - [40192, 512, 1, 512] - - [454, 84.574] - - - [40192, 39936, 1, 512] - - [451, 94.668] - - - [40448, 39936, 1, 512] - - [451, 94.679] - - - [39936, 39936, 1, 512] - - [451, 94.685] - - - [40192, 39681, 1, 512] - - [451, 94.377] - - - [39936, 39681, 1, 512] - - [455, 94.362] - - - [39680, 39681, 1, 512] - - [451, 94.378] - - - [39680, 512, 1, 512] - - [451, 84.021] - - - [39680, 39424, 1, 512] - - [451, 94.69] - - - [39936, 39424, 1, 512] - - [451, 94.685] - - - [39424, 39424, 1, 512] - - [451, 94.689] - - - [39680, 39169, 1, 512] - - [453, 94.378] - - - [39424, 39169, 1, 512] - - [453, 94.383] - - - [39168, 39169, 1, 512] - - [453, 94.392] - - - [39168, 512, 1, 512] - - [455, 87.849] - - - [39168, 38912, 1, 512] - - [451, 94.703] - - - [39424, 38912, 1, 512] - - [451, 94.678] - - - [38912, 38912, 1, 512] - - [451, 94.703] - - - [38912, 38657, 1, 512] - - [451, 94.375] - - - [39168, 38657, 1, 512] - - [451, 94.389] - - - [38656, 38657, 1, 512] - - [451, 94.386] - - - [38656, 512, 1, 512] - - [459, 86.904] - - - [38656, 38400, 1, 512] - - [451, 94.686] - - - [38912, 38400, 1, 512] - - [451, 94.688] - - - [38400, 38400, 1, 512] - - [451, 94.68] - - - [38400, 38145, 1, 512] - - [451, 94.349] - - - [38656, 38145, 1, 512] - - [451, 94.371] - - - [38144, 38145, 1, 512] - - [451, 94.356] - - - [38144, 512, 1, 512] - - [455, 86.047] - - - [38144, 37888, 1, 512] - - [451, 94.67] - - - [38400, 37888, 1, 512] - - [453, 94.665] - - - [37888, 37888, 1, 512] - - [451, 94.669] - - - [38144, 37633, 1, 512] - - [451, 94.386] - - - [37888, 37633, 1, 512] - - [451, 94.379] - - - [37632, 37633, 1, 512] - - [451, 94.362] - - - [37632, 512, 1, 512] - - [459, 85.014] - - - [37632, 37376, 1, 512] - - [451, 94.68] - - - [37888, 37376, 1, 512] - - [451, 94.667] - - - [37376, 37376, 1, 512] - - [451, 94.669] - - - [37376, 37121, 1, 512] - - [451, 94.336] - - - [37632, 37121, 1, 512] - - [455, 94.332] - - - [37120, 37121, 1, 512] - - [454, 94.308] - - - [37120, 512, 1, 512] - - [454, 84.315] - - - [37120, 36864, 1, 512] - - [451, 94.671] - - - [37376, 36864, 1, 512] - - [453, 94.663] - - - [36864, 36864, 1, 512] - - [451, 94.659] - - - [36864, 36609, 1, 512] - - [451, 94.344] - - - [37120, 36609, 1, 512] - - [451, 94.325] - - - [36608, 36609, 1, 512] - - [451, 94.336] - - - [36608, 512, 1, 512] - - [456, 83.288] - - - [36608, 36352, 1, 512] - - [451, 94.666] - - - [36864, 36352, 1, 512] - - [451, 94.657] - - - [36352, 36352, 1, 512] - - [451, 94.674] - - - [36352, 36097, 1, 512] - - [451, 94.318] - - - [36608, 36097, 1, 512] - - [451, 94.331] - - - [36096, 36097, 1, 512] - - [451, 94.344] - - - [36096, 512, 1, 512] - - [454, 82.514] - - - [36096, 35840, 1, 512] - - [451, 94.672] - - - [36352, 35840, 1, 512] - - [451, 94.668] - - - [35840, 35840, 1, 512] - - [451, 94.669] - - - [35840, 35585, 1, 512] - - [451, 94.295] - - - [36096, 35585, 1, 512] - - [451, 94.313] - - - [35584, 35585, 1, 512] - - [451, 94.309] - - - [35584, 512, 1, 512] - - [456, 86.503] - - - [35584, 35328, 1, 512] - - [451, 94.66] - - - [35840, 35328, 1, 512] - - [451, 94.666] - - - [35328, 35328, 1, 512] - - [451, 94.67] - - - [35328, 35073, 1, 512] - - [451, 94.336] - - - [35584, 35073, 1, 512] - - [451, 94.319] - - - [35072, 35073, 1, 512] - - [453, 94.3] - - - [35072, 512, 1, 512] - - [457, 85.725] - - - [35072, 34816, 1, 512] - - [451, 94.655] - - - [35328, 34816, 1, 512] - - [451, 94.701] - - - [34816, 34816, 1, 512] - - [451, 94.683] - - - [34816, 34561, 1, 512] - - [453, 94.337] - - - [35072, 34561, 1, 512] - - [451, 94.311] - - - [34560, 34561, 1, 512] - - [453, 94.318] - - - [34560, 512, 1, 512] - - [456, 84.573] - - - [34560, 34304, 1, 512] - - [451, 94.67] - - - [34816, 34304, 1, 512] - - [451, 94.665] - - - [34304, 34304, 1, 512] - - [451, 94.669] - - - [34304, 34049, 1, 512] - - [451, 94.319] - - - [34560, 34049, 1, 512] - - [451, 94.299] - - - [34048, 34049, 1, 512] - - [451, 94.332] - - - [34048, 512, 1, 512] - - [456, 83.631] - - - [34048, 33792, 1, 512] - - [451, 94.649] - - - [34304, 33792, 1, 512] - - [454, 94.641] - - - [33792, 33792, 1, 512] - - [451, 94.682] - - - [33792, 33537, 1, 512] - - [451, 94.292] - - - [34048, 33537, 1, 512] - - [453, 94.296] - - - [33536, 33537, 1, 512] - - [453, 94.29] - - - [33536, 512, 1, 512] - - [457, 82.863] - - - [33536, 33280, 1, 512] - - [451, 94.659] - - - [33792, 33280, 1, 512] - - [451, 94.642] - - - [33280, 33280, 1, 512] - - [451, 94.632] - - - [33536, 33025, 1, 512] - - [451, 94.294] - - - [33280, 33025, 1, 512] - - [451, 94.277] - - - [33024, 33025, 1, 512] - - [450, 94.295] - - - [33024, 512, 1, 512] - - [452, 81.921] - - - [33024, 32768, 1, 512] - - [451, 94.616] - - - [33280, 32768, 1, 512] - - [451, 94.626] - - - [32768, 32768, 1, 512] - - [451, 94.541] - - - [32768, 32513, 1, 512] - - [451, 94.152] - - - [33024, 32513, 1, 512] - - [451, 94.253] - - - [32512, 32513, 1, 512] - - [451, 94.136] - - - [32512, 512, 1, 512] - - [458, 86.122] - - - [32512, 32256, 1, 512] - - [451, 94.558] - - - [32768, 32256, 1, 512] - - [451, 94.559] - - - [32256, 32256, 1, 512] - - [451, 94.633] - - - [32256, 32001, 1, 512] - - [451, 94.235] - - - [32512, 32001, 1, 512] - - [458, 94.097] - - - [32000, 32001, 1, 512] - - [455, 94.223] - - - [32000, 512, 1, 512] - - [450, 85.309] - - - [32000, 31744, 1, 512] - - [451, 94.616] - - - [32256, 31744, 1, 512] - - [451, 94.624] - - - [31744, 31744, 1, 512] - - [451, 94.614] - - - [31744, 31489, 1, 512] - - [451, 94.309] - - - [32000, 31489, 1, 512] - - [451, 94.227] - - - [31488, 31489, 1, 512] - - [451, 94.242] - - - [31488, 512, 1, 512] - - [457, 84.124] - - - [31488, 31232, 1, 512] - - [451, 94.628] - - - [31744, 31232, 1, 512] - - [451, 94.652] - - - [31232, 31232, 1, 512] - - [451, 94.626] - - - [31488, 30977, 1, 512] - - [451, 94.2] - - - [31232, 30977, 1, 512] - - [450, 94.222] - - - [30976, 30977, 1, 512] - - [451, 94.204] - - - [30976, 512, 1, 512] - - [454, 83.075] - - - [30976, 30720, 1, 512] - - [454, 94.595] - - - [31232, 30720, 1, 512] - - [450, 94.617] - - - [30720, 30720, 1, 512] - - [451, 94.629] - - - [30976, 30465, 1, 512] - - [451, 94.25] - - - [30720, 30465, 1, 512] - - [451, 94.212] - - - [30464, 30465, 1, 512] - - [451, 94.221] - - - [30464, 512, 1, 512] - - [459, 82.019] - - - [30464, 30208, 1, 512] - - [451, 94.606] - - - [30720, 30208, 1, 512] - - [451, 94.608] - - - [30208, 30208, 1, 512] - - [451, 94.62] - - - [30208, 29953, 1, 512] - - [453, 94.216] - - - [30464, 29953, 1, 512] - - [453, 94.193] - - - [29952, 29953, 1, 512] - - [450, 94.23] - - - [29952, 512, 1, 512] - - [453, 81.218] - - - [29952, 29696, 1, 512] - - [450, 94.594] - - - [30208, 29696, 1, 512] - - [451, 94.601] - - - [29696, 29696, 1, 512] - - [451, 94.589] - - - [29696, 29441, 1, 512] - - [451, 94.2] - - - [29952, 29441, 1, 512] - - [455, 94.166] - - - [29440, 29441, 1, 512] - - [451, 94.19] - - - [29440, 512, 1, 512] - - [451, 86.048] - - - [29440, 29184, 1, 512] - - [451, 94.608] - - - [29696, 29184, 1, 512] - - [451, 94.597] - - - [29184, 29184, 1, 512] - - [451, 94.59] - - - [29184, 28929, 1, 512] - - [451, 94.153] - - - [29440, 28929, 1, 512] - - [451, 94.163] - - - [28928, 28929, 1, 512] - - [451, 94.179] - - - [28928, 512, 1, 512] - - [453, 84.889] - - - [28928, 28672, 1, 512] - - [451, 94.602] - - - [29184, 28672, 1, 512] - - [451, 94.593] - - - [28672, 28672, 1, 512] - - [451, 94.615] - - - [28928, 28417, 1, 512] - - [450, 94.179] - - - [28672, 28417, 1, 512] - - [451, 94.155] - - - [28416, 28417, 1, 512] - - [451, 94.132] - - - [28416, 512, 1, 512] - - [457, 83.603] - - - [28416, 28160, 1, 512] - - [451, 94.594] - - - [28672, 28160, 1, 512] - - [451, 94.604] - - - [28160, 28160, 1, 512] - - [451, 94.587] - - - [28160, 27905, 1, 512] - - [451, 94.16] - - - [28416, 27905, 1, 512] - - [451, 94.113] - - - [27904, 27905, 1, 512] - - [451, 94.144] - - - [27904, 512, 1, 512] - - [454, 82.349] - - - [27904, 27648, 1, 512] - - [451, 94.577] - - - [28160, 27648, 1, 512] - - [451, 94.585] - - - [27648, 27648, 1, 512] - - [451, 94.59] - - - [27648, 27393, 1, 512] - - [451, 94.136] - - - [27904, 27393, 1, 512] - - [451, 94.097] - - - [27392, 27393, 1, 512] - - [451, 94.105] - - - [27392, 512, 1, 512] - - [459, 81.24] - - - [27392, 27136, 1, 512] - - [451, 94.545] - - - [27648, 27136, 1, 512] - - [451, 94.548] - - - [27136, 27136, 1, 512] - - [451, 94.538] - - - [27392, 26881, 1, 512] - - [451, 94.137] - - - [27136, 26881, 1, 512] - - [450, 94.097] - - - [26880, 26881, 1, 512] - - [451, 94.158] - - - [26880, 512, 1, 512] - - [459, 80.121] - - - [26880, 26624, 1, 512] - - [451, 94.549] - - - [27136, 26624, 1, 512] - - [451, 94.57] - - - [26624, 26624, 1, 512] - - [451, 94.558] - - - [26624, 26369, 1, 512] - - [451, 94.076] - - - [26880, 26369, 1, 512] - - [451, 94.056] - - - [26368, 26369, 1, 512] - - [450, 94.047] - - - [26368, 512, 1, 512] - - [458, 78.934] - - - [26368, 26112, 1, 512] - - [453, 94.521] - - - [26624, 26112, 1, 512] - - [453, 94.542] - - - [26112, 26112, 1, 512] - - [453, 94.578] - - - [26112, 25857, 1, 512] - - [451, 94.12] - - - [26368, 25857, 1, 512] - - [454, 94.035] - - - [25856, 25857, 1, 512] - - [451, 94.066] - - - [25856, 512, 1, 512] - - [450, 84.441] - - - [25856, 25600, 1, 512] - - [451, 94.532] - - - [26112, 25600, 1, 512] - - [454, 94.571] - - - [25600, 25345, 1, 512] - - [451, 94.06] - - - [25856, 25345, 1, 512] - - [453, 94.097] - - - [25344, 25345, 1, 512] - - [450, 94.07] - - - [25344, 512, 1, 512] - - [457, 82.921] - - - [25344, 25088, 1, 512] - - [453, 94.493] - - - [25600, 25088, 1, 512] - - [451, 94.524] - - - [25088, 25088, 1, 512] - - [451, 94.52] - - - [25088, 24833, 1, 512] - - [451, 93.988] - - - [25344, 24833, 1, 512] - - [451, 93.963] - - - [24832, 24833, 1, 512] - - [451, 94.007] - - - [24832, 512, 1, 512] - - [454, 81.674] - - - [24832, 24576, 1, 512] - - [451, 94.457] - - - [25088, 24576, 1, 512] - - [450, 94.477] - - - [24576, 24576, 1, 512] - - [456, 94.49] - - - [24576, 24321, 1, 512] - - [451, 94.006] - - - [24832, 24321, 1, 512] - - [455, 94.033] - - - [24320, 24321, 1, 512] - - [456, 94.031] - - - [24320, 512, 1, 512] - - [458, 80.137] - - - [24320, 24064, 1, 512] - - [453, 94.478] - - - [24576, 24064, 1, 512] - - [451, 94.507] - - - [24064, 24064, 1, 512] - - [453, 94.456] - - - [24064, 23809, 1, 512] - - [450, 93.987] - - - [24320, 23809, 1, 512] - - [450, 93.97] - - - [23808, 23809, 1, 512] - - [450, 94.02] - - - [23808, 512, 1, 512] - - [455, 78.929] - - - [23808, 23552, 1, 512] - - [453, 94.43] - - - [24064, 23552, 1, 512] - - [453, 94.458] - - - [23552, 23552, 1, 512] - - [453, 94.479] - - - [23552, 23297, 1, 512] - - [451, 94.004] - - - [23808, 23297, 1, 512] - - [451, 93.996] - - - [23296, 23297, 1, 512] - - [451, 93.954] - - - [23296, 512, 1, 512] - - [457, 77.727] - - - [23296, 23040, 1, 512] - - [451, 94.438] - - - [23552, 23040, 1, 512] - - [453, 94.433] - - - [23040, 23040, 1, 512] - - [451, 94.434] - - - [23296, 22785, 1, 512] - - [451, 93.89] - - - [23040, 22785, 1, 512] - - [455, 93.882] - - - [22784, 22785, 1, 512] - - [458, 93.833] - - - [22784, 512, 1, 512] - - [453, 83.584] - - - [22784, 22528, 1, 512] - - [454, 94.424] - - - [23040, 22528, 1, 512] - - [451, 94.393] - - - [22528, 22528, 1, 512] - - [451, 94.41] - - - [22528, 22273, 1, 512] - - [450, 93.932] - - - [22784, 22273, 1, 512] - - [450, 93.927] - - - [22272, 22273, 1, 512] - - [451, 93.819] - - - [22272, 512, 1, 512] - - [459, 82.067] - - - [22272, 22016, 1, 512] - - [450, 94.369] - - - [22528, 22016, 1, 512] - - [454, 94.392] - - - [22016, 22016, 1, 512] - - [450, 94.413] - - - [22016, 21761, 1, 512] - - [456, 93.889] - - - [22272, 21761, 1, 512] - - [455, 93.867] - - - [21760, 21761, 1, 512] - - [453, 93.881] - - - [21760, 512, 1, 512] - - [459, 80.442] - - - [21760, 21504, 1, 512] - - [453, 94.445] - - - [22016, 21504, 1, 512] - - [451, 94.41] - - - [21504, 21504, 1, 512] - - [453, 94.391] - - - [21504, 21249, 1, 512] - - [450, 93.805] - - - [21760, 21249, 1, 512] - - [451, 93.784] - - - [21248, 21249, 1, 512] - - [456, 93.778] - - - [21248, 512, 1, 512] - - [455, 78.782] - - - [21248, 20992, 1, 512] - - [456, 94.35] - - - [21504, 20992, 1, 512] - - [451, 94.375] - - - [20992, 20992, 1, 512] - - [456, 94.353] - - - [20992, 20737, 1, 512] - - [450, 93.849] - - - [21248, 20737, 1, 512] - - [450, 93.845] - - - [20736, 20737, 1, 512] - - [450, 93.822] - - - [20736, 512, 1, 512] - - [451, 77.182] - - - [20736, 20480, 1, 512] - - [451, 94.369] - - - [20992, 20480, 1, 512] - - [450, 94.34] - - - [20480, 20480, 1, 512] - - [451, 94.369] - - - [20736, 20225, 1, 512] - - [451, 93.708] - - - [20480, 20225, 1, 512] - - [451, 93.686] - - - [20224, 20225, 1, 512] - - [460, 93.736] - - - [20224, 512, 1, 512] - - [456, 75.727] - - - [20224, 19968, 1, 512] - - [453, 94.267] - - - [20480, 19968, 1, 512] - - [450, 94.335] - - - [19968, 19968, 1, 512] - - [450, 94.256] - - - [19968, 19713, 1, 512] - - [451, 93.808] - - - [20224, 19713, 1, 512] - - [451, 93.809] - - - [19712, 19713, 1, 512] - - [451, 93.789] - - - [19712, 512, 1, 512] - - [453, 76.114] - - - [19712, 19456, 1, 512] - - [451, 94.238] - - - [19968, 19456, 1, 512] - - [451, 94.263] - - - [19456, 19456, 1, 512] - - [453, 94.26] - - - [19456, 19201, 1, 512] - - [456, 93.767] - - - [19712, 19201, 1, 512] - - [450, 93.738] - - - [19200, 19201, 1, 512] - - [456, 93.755] - - - [19200, 512, 1, 512] - - [456, 80.722] - - - [19200, 18944, 1, 512] - - [453, 94.218] - - - [19456, 18944, 1, 512] - - [450, 94.235] - - - [18944, 18944, 1, 512] - - [451, 94.232] - - - [18944, 18689, 1, 512] - - [460, 93.688] - - - [19200, 18689, 1, 512] - - [456, 93.626] - - - [18688, 18689, 1, 512] - - [453, 93.485] - - - [18688, 512, 1, 512] - - [455, 78.968] - - - [18688, 18432, 1, 512] - - [451, 94.304] - - - [18944, 18432, 1, 512] - - [450, 94.236] - - - [18432, 18432, 1, 512] - - [460, 94.288] - - - [18432, 18177, 1, 512] - - [451, 93.584] - - - [18688, 18177, 1, 512] - - [455, 93.509] - - - [18176, 18177, 1, 512] - - [450, 93.562] - - - [18176, 512, 1, 512] - - [451, 77.065] - - - [18176, 17920, 1, 512] - - [451, 94.222] - - - [18432, 17920, 1, 512] - - [451, 94.194] - - - [17920, 17920, 1, 512] - - [456, 94.289] - - - [18176, 17665, 1, 512] - - [450, 93.533] - - - [17920, 17665, 1, 512] - - [453, 93.452] - - - [17664, 17665, 1, 512] - - [450, 93.641] - - - [17664, 512, 1, 512] - - [454, 75.38] - - - [17664, 17408, 1, 512] - - [451, 94.192] - - - [17920, 17408, 1, 512] - - [454, 94.184] - - - [17408, 17408, 1, 512] - - [451, 94.066] - - - [17664, 17153, 1, 512] - - [451, 93.433] - - - [17408, 17153, 1, 512] - - [450, 93.551] - - - [17152, 17153, 1, 512] - - [451, 93.351] - - - [17152, 512, 1, 512] - - [459, 73.623] - - - [17152, 16896, 1, 512] - - [451, 94.027] - - - [17408, 16896, 1, 512] - - [453, 94.264] - - - [16896, 16896, 1, 512] - - [451, 94.124] - - - [16896, 16641, 1, 512] - - [456, 93.435] - - - [17152, 16641, 1, 512] - - [456, 93.606] - - - [16640, 16641, 1, 512] - - [456, 93.594] - - - [16640, 512, 1, 512] - - [456, 72.219] - - - [16640, 16384, 1, 512] - - [460, 94.213] - - - [16896, 16384, 1, 512] - - [450, 94.04] - - - [16384, 16384, 1, 512] - - [456, 94.033] - - - [16384, 16129, 1, 512] - - [451, 93.307] - - - [16640, 16129, 1, 512] - - [450, 93.517] - - - [16128, 16129, 1, 512] - - [451, 93.448] - - - [16128, 512, 1, 512] - - [459, 71.946] - - - [16128, 15872, 1, 512] - - [450, 94.054] - - - [16384, 15872, 1, 512] - - [453, 94.019] - - - [15872, 15872, 1, 512] - - [456, 94.016] - - - [15872, 15617, 1, 512] - - [455, 93.108] - - - [16128, 15617, 1, 512] - - [456, 93.284] - - - [15616, 15617, 1, 512] - - [451, 93.302] - - - [15616, 512, 1, 512] - - [459, 77.185] - - - [15616, 15360, 1, 512] - - [451, 93.913] - - - [15872, 15360, 1, 512] - - [450, 94.042] - - - [15360, 15360, 1, 512] - - [460, 94.159] - - - [15360, 15105, 1, 512] - - [453, 93.29] - - - [15616, 15105, 1, 512] - - [453, 93.014] - - - [15104, 15105, 1, 512] - - [455, 93.072] - - - [15104, 512, 1, 512] - - [459, 75.22] - - - [15104, 14848, 1, 512] - - [450, 94.01] - - - [15360, 14848, 1, 512] - - [453, 93.841] - - - [14848, 14848, 1, 512] - - [456, 93.994] - - - [14848, 14593, 1, 512] - - [450, 93.277] - - - [15104, 14593, 1, 512] - - [450, 93.373] - - - [14592, 14593, 1, 512] - - [450, 93.13] - - - [14592, 512, 1, 512] - - [454, 72.931] - - - [14592, 14336, 1, 512] - - [456, 94.124] - - - [14848, 14336, 1, 512] - - [451, 93.728] - - - [14336, 14336, 1, 512] - - [460, 93.975] - - - [14336, 14081, 1, 512] - - [455, 93.257] - - - [14592, 14081, 1, 512] - - [458, 93.279] - - - [14080, 14081, 1, 512] - - [455, 93.147] - - - [14080, 512, 1, 512] - - [458, 70.826] - - - [14080, 13824, 1, 512] - - [453, 93.673] - - - [14336, 13824, 1, 512] - - [455, 93.805] - - - [13824, 13824, 1, 512] - - [451, 93.597] - - - [14080, 13569, 1, 512] - - [451, 92.783] - - - [13824, 13569, 1, 512] - - [456, 92.684] - - - [13568, 13569, 1, 512] - - [451, 92.652] - - - [13568, 512, 1, 512] - - [451, 68.748] - - - [13568, 13312, 1, 512] - - [460, 93.975] - - - [13824, 13312, 1, 512] - - [454, 94.008] - - - [13312, 13312, 1, 512] - - [451, 93.936] - - - [13568, 13057, 1, 512] - - [450, 93.266] - - - [13312, 13057, 1, 512] - - [450, 93.191] - - - [13056, 13057, 1, 512] - - [450, 93.176] - - - [13056, 512, 1, 512] - - [453, 77.279] - - - [13056, 12800, 1, 512] - - [451, 93.897] - - - [13312, 12800, 1, 512] - - [455, 93.853] - - - [12800, 12800, 1, 512] - - [453, 93.873] - - - [13056, 12545, 1, 512] - - [451, 93.209] - - - [12800, 12545, 1, 512] - - [459, 93.16] - - - [12544, 12545, 1, 512] - - [451, 93.225] - - - [12544, 512, 1, 512] - - [450, 74.496] - - - [12544, 12288, 1, 512] - - [451, 93.322] - - - [12800, 12288, 1, 512] - - [451, 93.943] - - - [12288, 12288, 1, 512] - - [453, 93.344] - - - [12544, 12033, 1, 512] - - [451, 92.26] - - - [12288, 12033, 1, 512] - - [456, 92.315] - - - [12032, 12033, 1, 512] - - [454, 92.415] - - - [12032, 512, 1, 512] - - [451, 71.762] - - - [12032, 11776, 1, 512] - - [451, 93.632] - - - [12288, 11776, 1, 512] - - [454, 93.508] - - - [11776, 11776, 1, 512] - - [451, 93.724] - - - [11776, 11521, 1, 512] - - [453, 93.042] - - - [12032, 11521, 1, 512] - - [453, 92.913] - - - [11520, 11521, 1, 512] - - [456, 92.337] - - - [11520, 512, 1, 512] - - [454, 69.204] - - - [11520, 11264, 1, 512] - - [453, 93.387] - - - [11776, 11264, 1, 512] - - [453, 93.19] - - - [11264, 11264, 1, 512] - - [451, 93.634] - - - [11264, 11009, 1, 512] - - [456, 92.655] - - - [11520, 11009, 1, 512] - - [453, 92.36] - - - [11008, 11009, 1, 512] - - [456, 91.888] - - - [11008, 512, 1, 512] - - [451, 66.447] - - - [11008, 10752, 1, 512] - - [453, 93.451] - - - [11264, 10752, 1, 512] - - [454, 93.137] - - - [10752, 10752, 1, 512] - - [454, 92.763] - - - [11008, 10497, 1, 512] - - [453, 92.452] - - - [10752, 10497, 1, 512] - - [453, 91.785] - - - [10496, 10497, 1, 512] - - [453, 92.107] - - - [10496, 512, 1, 512] - - [457, 63.85] - - - [10496, 10240, 1, 512] - - [454, 92.743] - - - [10752, 10240, 1, 512] - - [454, 93.381] - - - [10240, 10240, 1, 512] - - [454, 93.136] - - - [10496, 9985, 1, 512] - - [450, 92.048] - - - [10240, 9985, 1, 512] - - [450, 92.428] - - - [9984, 9985, 1, 512] - - [450, 91.845] - - - [9984, 512, 1, 512] - - [451, 61.636] - - - [9984, 9728, 1, 512] - - [455, 93.424] - - - [10240, 9728, 1, 512] - - [451, 92.923] - - - [9728, 9728, 1, 512] - - [454, 92.851] - - - [9728, 9473, 1, 512] - - [451, 91.809] - - - [9984, 9473, 1, 512] - - [453, 91.209] - - - [9472, 9473, 1, 512] - - [456, 91.104] - - - [9472, 512, 1, 512] - - [454, 70.49] - - - [9472, 9216, 1, 512] - - [450, 92.188] - - - [9728, 9216, 1, 512] - - [454, 92.76] - - - [9216, 9216, 1, 512] - - [454, 92.879] - - - [9472, 8961, 1, 512] - - [455, 91.579] - - - [9216, 8961, 1, 512] - - [456, 90.9] - - - [8960, 8961, 1, 512] - - [451, 91.833] - - - [8960, 512, 1, 512] - - [459, 67.08] - - - [8960, 8704, 1, 512] - - [451, 92.49] - - - [9216, 8704, 1, 512] - - [455, 93.052] - - - [8704, 8704, 1, 512] - - [455, 91.892] - - - [8704, 8449, 1, 512] - - [458, 91.459] - - - [8960, 8449, 1, 512] - - [458, 91.796] - - - [8448, 8449, 1, 512] - - [450, 90.829] - - - [8448, 512, 1, 512] - - [458, 63.782] - - - [8448, 8192, 1, 512] - - [452, 91.795] - - - [8704, 8192, 1, 512] - - [450, 92.245] - - - [8192, 8192, 1, 512] - - [450, 91.24] - - - [8192, 7937, 1, 512] - - [458, 90.286] - - - [8448, 7937, 1, 512] - - [459, 90.748] - - - [7936, 7937, 1, 512] - - [458, 89.947] - - - [7936, 512, 1, 512] - - [457, 60.28] - - - [7936, 7680, 1, 512] - - [450, 91.483] - - - [8192, 7680, 1, 512] - - [455, 91.946] - - - [7680, 7680, 1, 512] - - [458, 91.0] - - - [7936, 7425, 1, 512] - - [458, 90.653] - - - [7680, 7425, 1, 512] - - [454, 90.15] - - - [7424, 7425, 1, 512] - - [458, 89.713] - - - [7424, 512, 1, 512] - - [460, 57.085] - - - [7424, 7168, 1, 512] - - [452, 91.8] - - - [7680, 7168, 1, 512] - - [457, 92.101] - - - [7168, 7168, 1, 512] - - [451, 91.486] - - - [7168, 6913, 1, 512] - - [450, 89.149] - - - [7424, 6913, 1, 512] - - [450, 89.476] - - - [6912, 6913, 1, 512] - - [450, 88.786] - - - [6912, 512, 1, 512] - - [454, 53.71] - - - [6912, 6656, 1, 512] - - [456, 90.582] - - - [7168, 6656, 1, 512] - - [455, 90.716] - - - [6656, 6656, 1, 512] - - [453, 90.339] - - - [6912, 6401, 1, 512] - - [455, 90.174] - - - [6656, 6401, 1, 512] - - [455, 90.045] - - - [6400, 6401, 1, 512] - - [455, 89.955] - - - [6400, 512, 1, 512] - - [450, 64.491] - - - [6400, 6144, 1, 512] - - [452, 89.954] - - - [6656, 6144, 1, 512] - - [455, 89.943] - - - [6144, 6144, 1, 512] - - [451, 90.026] - - - [6144, 5889, 1, 512] - - [456, 86.586] - - - [6400, 5889, 1, 512] - - [453, 89.492] - - - [5888, 5889, 1, 512] - - [452, 86.392] - - - [5888, 512, 1, 512] - - [457, 59.38] - - - [5888, 5632, 1, 512] - - [458, 90.191] - - - [6144, 5632, 1, 512] - - [455, 89.78] - - - [5632, 5632, 1, 512] - - [452, 90.379] - - - [5632, 5377, 1, 512] - - [451, 87.09] - - - [5888, 5377, 1, 512] - - [459, 87.541] - - - [5376, 5377, 1, 512] - - [450, 87.33] - - - [5376, 512, 1, 512] - - [459, 54.48] - - - [5376, 5120, 1, 512] - - [452, 87.786] - - - [5632, 5120, 1, 512] - - [453, 87.181] - - - [5120, 5120, 1, 512] - - [450, 88.217] - - - [5120, 4865, 1, 512] - - [460, 84.206] - - - [5376, 4865, 1, 512] - - [453, 87.592] - - - [4864, 4865, 1, 512] - - [453, 84.85] - - - [4864, 512, 1, 512] - - [455, 49.776] - - - [4864, 4608, 1, 512] - - [455, 85.828] - - - [5120, 4608, 1, 512] - - [455, 85.161] - - - [4608, 4608, 1, 512] - - [450, 86.829] - - - [4608, 4353, 1, 512] - - [454, 82.328] - - - [4864, 4353, 1, 512] - - [458, 86.538] - - - [4352, 4353, 1, 512] - - [453, 83.512] - - - [4352, 512, 1, 512] - - [450, 44.941] - - - [4352, 4096, 1, 512] - - [458, 85.25] - - - [4608, 4096, 1, 512] - - [453, 83.871] - - - [4096, 4096, 1, 512] - - [453, 87.059] - - - [4096, 3841, 1, 512] - - [455, 82.392] - - - [4352, 3841, 1, 512] - - [450, 82.131] - - - [3840, 3841, 1, 512] - - [450, 79.6] - - - [3840, 512, 1, 512] - - [455, 40.219] - - - [3840, 3584, 1, 512] - - [457, 80.14] - - - [4096, 3584, 1, 512] - - [455, 84.435] - - - [3584, 3584, 1, 512] - - [451, 82.278] - - - [3840, 3329, 1, 512] - - [452, 81.732] - - - [3584, 3329, 1, 512] - - [452, 78.083] - - - [3328, 3329, 1, 512] - - [452, 79.874] - - - [3328, 512, 1, 512] - - [454, 49.505] - - - [3328, 3072, 1, 512] - - [452, 75.355] - - - [3584, 3072, 1, 512] - - [451, 79.779] - - - [3072, 3072, 1, 512] - - [457, 78.257] - - - [63488, 76800, 1, 512] - - [451, 94.691] - - - [64000, 76800, 1, 512] - - [451, 94.695] - - - [64000, 50177, 1, 512] - - [451, 94.455] - - - [63488, 50177, 1, 512] - - [451, 94.454] - - - [63488, 49665, 1, 512] - - [453, 94.417] - - - [62976, 76800, 1, 512] - - [451, 94.699] - - - [62976, 49153, 1, 512] - - [451, 94.431] - - - [63488, 49153, 1, 512] - - [451, 94.426] - - - [62976, 48641, 1, 512] - - [453, 94.42] - - - [62464, 76800, 1, 512] - - [451, 94.694] - - - [62464, 48129, 1, 512] - - [451, 94.418] - - - [62976, 48129, 1, 512] - - [453, 94.418] - - - [62464, 47617, 1, 512] - - [451, 94.422] - - - [61952, 76800, 1, 512] - - [451, 94.691] - - - [61952, 47105, 1, 512] - - [451, 94.428] - - - [62464, 47105, 1, 512] - - [451, 94.417] - - - [61952, 46593, 1, 512] - - [451, 94.423] - - - [61440, 76800, 1, 512] - - [453, 94.689] - - - [61440, 46081, 1, 512] - - [453, 94.414] - - - [61952, 46081, 1, 512] - - [453, 94.395] - - - [61440, 45569, 1, 512] - - [451, 94.408] - - - [60928, 76800, 1, 512] - - [451, 94.7] - - - [60928, 45057, 1, 512] - - [458, 94.389] - - - [61440, 45057, 1, 512] - - [451, 94.405] - - - [60928, 44545, 1, 512] - - [451, 94.378] - - - [60416, 76800, 1, 512] - - [451, 94.684] - - - [60416, 44033, 1, 512] - - [451, 94.388] - - - [60928, 44033, 1, 512] - - [451, 94.393] - - - [60416, 43521, 1, 512] - - [451, 94.397] - - - [59904, 76800, 1, 512] - - [451, 94.693] - - - [59904, 43009, 1, 512] - - [451, 94.379] - - - [60416, 43009, 1, 512] - - [451, 94.375] - - - [59904, 42497, 1, 512] - - [451, 94.344] - - - [59392, 76800, 1, 512] - - [451, 94.691] - - - [59392, 41985, 1, 512] - - [451, 94.377] - - - [59904, 41985, 1, 512] - - [451, 94.381] - - - [59392, 41473, 1, 512] - - [453, 94.363] - - - [58880, 76800, 1, 512] - - [451, 94.695] - - - [58880, 40961, 1, 512] - - [451, 94.346] - - - [59392, 40961, 1, 512] - - [454, 94.352] - - - [58880, 40449, 1, 512] - - [451, 94.349] - - - [58368, 76800, 1, 512] - - [451, 94.701] - - - [58368, 39937, 1, 512] - - [453, 94.355] - - - [58880, 39937, 1, 512] - - [451, 94.356] - - - [58368, 39425, 1, 512] - - [451, 94.38] - - - [57856, 76800, 1, 512] - - [451, 94.689] - - - [58368, 38913, 1, 512] - - [451, 94.332] - - - [57856, 38913, 1, 512] - - [451, 94.368] - - - [57856, 38401, 1, 512] - - [451, 94.355] - - - [57344, 76800, 1, 512] - - [451, 94.687] - - - [57856, 37889, 1, 512] - - [451, 94.326] - - - [57344, 37889, 1, 512] - - [451, 94.339] - - - [57344, 37377, 1, 512] - - [451, 94.319] - - - [56832, 76800, 1, 512] - - [451, 94.69] - - - [57344, 36865, 1, 512] - - [453, 94.345] - - - [56832, 36865, 1, 512] - - [453, 94.354] - - - [56832, 36353, 1, 512] - - [451, 94.325] - - - [56320, 76800, 1, 512] - - [453, 94.663] - - - [56320, 35841, 1, 512] - - [454, 94.288] - - - [56832, 35841, 1, 512] - - [451, 94.347] - - - [56320, 35329, 1, 512] - - [451, 94.293] - - - [55808, 76800, 1, 512] - - [453, 94.696] - - - [55808, 34817, 1, 512] - - [451, 94.285] - - - [56320, 34817, 1, 512] - - [451, 94.303] - - - [55808, 34305, 1, 512] - - [455, 94.285] - - - [55296, 76800, 1, 512] - - [451, 94.689] - - - [55808, 33793, 1, 512] - - [451, 94.272] - - - [55296, 33793, 1, 512] - - [454, 94.274] - - - [55296, 33281, 1, 512] - - [453, 94.289] - - - [54784, 76800, 1, 512] - - [453, 94.69] - - - [55296, 32769, 1, 512] - - [451, 94.259] - - - [54784, 32769, 1, 512] - - [453, 94.259] - - - [54784, 32257, 1, 512] - - [451, 94.304] - - - [54272, 76800, 1, 512] - - [453, 94.689] - - - [54784, 31745, 1, 512] - - [454, 94.246] - - - [54272, 31745, 1, 512] - - [454, 94.25] - - - [54272, 31233, 1, 512] - - [451, 94.255] - - - [53760, 76800, 1, 512] - - [451, 94.691] - - - [54272, 30721, 1, 512] - - [451, 94.24] - - - [53760, 30721, 1, 512] - - [453, 94.239] - - - [53760, 30209, 1, 512] - - [453, 94.206] - - - [53248, 76800, 1, 512] - - [453, 94.682] - - - [53760, 29697, 1, 512] - - [454, 94.237] - - - [53248, 29697, 1, 512] - - [451, 94.213] - - - [53248, 29185, 1, 512] - - [451, 94.208] - - - [52736, 76800, 1, 512] - - [451, 94.691] - - - [53248, 28673, 1, 512] - - [451, 94.23] - - - [52736, 28673, 1, 512] - - [451, 94.232] - - - [52736, 28161, 1, 512] - - [455, 94.206] - - - [52224, 76800, 1, 512] - - [451, 94.685] - - - [52736, 27649, 1, 512] - - [453, 94.217] - - - [52224, 27649, 1, 512] - - [453, 94.203] - - - [52224, 27137, 1, 512] - - [453, 94.154] - - - [51712, 76800, 1, 512] - - [453, 94.68] - - - [52224, 26625, 1, 512] - - [454, 94.179] - - - [51712, 26625, 1, 512] - - [454, 94.16] - - - [51712, 26113, 1, 512] - - [458, 94.107] - - - [51200, 76800, 1, 512] - - [451, 94.677] - - - [50688, 76800, 1, 512] - - [451, 94.679] - - - [50688, 24577, 1, 512] - - [454, 94.107] - - - [51200, 24577, 1, 512] - - [454, 94.071] - - - [50688, 24065, 1, 512] - - [453, 94.11] - - - [50176, 76800, 1, 512] - - [451, 94.678] - - - [50688, 23553, 1, 512] - - [454, 94.091] - - - [50176, 23553, 1, 512] - - [451, 94.065] - - - [50176, 23041, 1, 512] - - [453, 94.094] - - - [49664, 76800, 1, 512] - - [451, 94.683] - - - [49664, 22529, 1, 512] - - [454, 94.063] - - - [50176, 22529, 1, 512] - - [454, 94.05] - - - [49664, 22017, 1, 512] - - [451, 94.001] - - - [49152, 76800, 1, 512] - - [451, 94.681] - - - [49664, 21505, 1, 512] - - [451, 94.053] - - - [49152, 21505, 1, 512] - - [451, 94.027] - - - [49152, 20993, 1, 512] - - [451, 93.978] - - - [48640, 76800, 1, 512] - - [453, 94.673] - - - [49152, 20481, 1, 512] - - [455, 93.999] - - - [48640, 20481, 1, 512] - - [455, 94.011] - - - [48640, 19969, 1, 512] - - [451, 93.969] - - - [48128, 76800, 1, 512] - - [451, 94.682] - - - [48128, 19457, 1, 512] - - [454, 93.945] - - - [48640, 19457, 1, 512] - - [454, 93.948] - - - [48128, 18945, 1, 512] - - [451, 93.924] - - - [47616, 76800, 1, 512] - - [451, 94.678] - - - [48128, 18433, 1, 512] - - [454, 93.931] - - - [47616, 18433, 1, 512] - - [453, 93.933] - - - [47616, 17921, 1, 512] - - [451, 93.91] - - - [47104, 76800, 1, 512] - - [451, 94.674] - - - [47616, 17409, 1, 512] - - [454, 93.891] - - - [47104, 17409, 1, 512] - - [455, 93.848] - - - [47104, 16897, 1, 512] - - [451, 93.816] - - - [46592, 76800, 1, 512] - - [453, 94.667] - - - [46592, 16385, 1, 512] - - [454, 93.794] - - - [47104, 16385, 1, 512] - - [454, 93.805] - - - [46592, 15873, 1, 512] - - [458, 93.752] - - - [46080, 76800, 1, 512] - - [451, 94.68] - - - [46592, 15361, 1, 512] - - [454, 93.788] - - - [46080, 15361, 1, 512] - - [455, 93.799] - - - [46080, 14849, 1, 512] - - [451, 93.697] - - - [45568, 76800, 1, 512] - - [451, 94.679] - - - [46080, 14337, 1, 512] - - [451, 93.708] - - - [45568, 14337, 1, 512] - - [454, 93.67] - - - [45568, 13825, 1, 512] - - [453, 93.666] - - - [45056, 76800, 1, 512] - - [451, 94.673] - - - [45568, 13313, 1, 512] - - [454, 93.625] - - - [45056, 13313, 1, 512] - - [454, 93.58] - - - [45056, 12801, 1, 512] - - [455, 93.524] - - - [44544, 76800, 1, 512] - - [451, 94.67] - - - [45056, 12289, 1, 512] - - [454, 93.524] - - - [44544, 12289, 1, 512] - - [454, 93.496] - - - [44544, 11777, 1, 512] - - [455, 93.391] - - - [44032, 76800, 1, 512] - - [451, 94.674] - - - [44544, 11265, 1, 512] - - [454, 93.408] - - - [44032, 11265, 1, 512] - - [458, 93.443] - - - [44032, 10753, 1, 512] - - [451, 93.305] - - - [43520, 76800, 1, 512] - - [453, 94.673] - - - [44032, 10241, 1, 512] - - [455, 93.233] - - - [43520, 10241, 1, 512] - - [454, 93.316] - - - [43520, 9729, 1, 512] - - [453, 93.112] - - - [43008, 76800, 1, 512] - - [451, 94.676] - - - [43520, 9217, 1, 512] - - [457, 93.136] - - - [43008, 9217, 1, 512] - - [453, 93.069] - - - [43008, 8705, 1, 512] - - [458, 92.936] - - - [42496, 76800, 1, 512] - - [451, 94.675] - - - [43008, 8193, 1, 512] - - [451, 92.926] - - - [42496, 8193, 1, 512] - - [454, 92.848] - - - [42496, 7681, 1, 512] - - [455, 92.706] - - - [41984, 76800, 1, 512] - - [451, 94.666] - - - [42496, 7169, 1, 512] - - [451, 92.553] - - - [41984, 7169, 1, 512] - - [454, 92.493] - - - [41984, 6657, 1, 512] - - [452, 92.242] - - - [41472, 76800, 1, 512] - - [451, 94.674] - - - [41984, 6145, 1, 512] - - [454, 92.232] - - - [41472, 6145, 1, 512] - - [454, 92.255] - - - [41472, 5633, 1, 512] - - [458, 92.14] - - - [40960, 76800, 1, 512] - - [453, 94.664] - - - [41472, 5121, 1, 512] - - [454, 91.575] - - - [40960, 5121, 1, 512] - - [454, 91.75] - - - [40960, 4609, 1, 512] - - [457, 91.493] - - - [40448, 76800, 1, 512] - - [453, 94.671] - - - [40960, 4097, 1, 512] - - [454, 90.9] - - - [40448, 4097, 1, 512] - - [454, 90.742] - - - [40448, 3585, 1, 512] - - [451, 90.585] - - - [39936, 76800, 1, 512] - - [451, 94.662] - - - [40448, 3073, 1, 512] - - [454, 89.528] - - - [39936, 3073, 1, 512] - - [454, 89.59] - - - [39936, 2561, 1, 512] - - [455, 88.347] - - - [39424, 76800, 1, 512] - - [451, 94.645] - - - [39424, 2049, 1, 512] - - [454, 86.903] - - - [39936, 2049, 1, 512] - - [454, 87.892] - - - [39424, 1537, 1, 512] - - [460, 84.4] - - - [38912, 76800, 1, 512] - - [453, 94.661] - - - [39424, 1025, 1, 512] - - [454, 80.04] - - - [38912, 1025, 1, 512] - - [454, 81.535] - - - [38912, 513, 1, 512] - - [455, 70.685] - - - [38400, 76800, 1, 512] - - [453, 94.655] - - - [89600, 89089, 1, 512] - - [453, 94.385] - - - [89088, 88577, 1, 512] - - [453, 94.403] - - - [88576, 88065, 1, 512] - - [453, 94.405] - - - [88064, 87553, 1, 512] - - [453, 94.404] - - - [87552, 87041, 1, 512] - - [453, 94.39] - - - [87040, 86529, 1, 512] - - [453, 94.391] - - - [86528, 86017, 1, 512] - - [453, 94.385] - - - [86016, 85505, 1, 512] - - [458, 94.406] - - - [85504, 84993, 1, 512] - - [453, 94.399] - - - [84992, 84481, 1, 512] - - [453, 94.388] - - - [84480, 83969, 1, 512] - - [453, 94.382] - - - [83968, 83457, 1, 512] - - [453, 94.377] - - - [83456, 82945, 1, 512] - - [453, 94.413] - - - [82944, 82433, 1, 512] - - [458, 94.381] - - - [82432, 81921, 1, 512] - - [453, 94.38] - - - [81920, 81409, 1, 512] - - [453, 94.394] - - - [81408, 80897, 1, 512] - - [453, 94.375] - - - [80896, 80385, 1, 512] - - [453, 94.357] - - - [80384, 79873, 1, 512] - - [453, 94.38] - - - [79872, 79361, 1, 512] - - [455, 94.346] - - - [79360, 78849, 1, 512] - - [453, 94.378] - - - [78848, 78337, 1, 512] - - [453, 94.381] - - - [78336, 77825, 1, 512] - - [453, 94.36] - - - [77824, 77313, 1, 512] - - [453, 94.366] - - - [77312, 76801, 1, 512] - - [453, 94.379] - - - [76800, 76289, 1, 512] - - [453, 94.364] - - - [76288, 75777, 1, 512] - - [458, 94.354] - - - [75776, 75265, 1, 512] - - [453, 94.364] - - - [75264, 74753, 1, 512] - - [453, 94.378] - - - [74752, 74241, 1, 512] - - [453, 94.351] - - - [74240, 73729, 1, 512] - - [453, 94.38] - - - [73728, 73217, 1, 512] - - [458, 94.362] - - - [73216, 72705, 1, 512] - - [453, 94.34] - - - [72704, 72193, 1, 512] - - [453, 94.365] - - - [72192, 71681, 1, 512] - - [453, 94.338] - - - [71680, 71169, 1, 512] - - [458, 94.376] - - - [71168, 70657, 1, 512] - - [458, 94.341] - - - [70656, 70145, 1, 512] - - [458, 94.336] - - - [70144, 69633, 1, 512] - - [453, 94.334] - - - [69632, 69121, 1, 512] - - [453, 94.35] - - - [69120, 68609, 1, 512] - - [458, 94.344] - - - [68608, 68097, 1, 512] - - [458, 94.351] - - - [68096, 67585, 1, 512] - - [458, 94.346] - - - [67584, 67073, 1, 512] - - [453, 94.329] - - - [67072, 66561, 1, 512] - - [458, 94.306] - - - [66560, 66049, 1, 512] - - [458, 94.305] - - - [66048, 65537, 1, 512] - - [453, 94.321] - - - [65536, 65025, 1, 512] - - [453, 94.328] - - - [65024, 64513, 1, 512] - - [453, 94.332] - - - [64512, 64001, 1, 512] - - [453, 94.319] - - - [64000, 63489, 1, 512] - - [453, 94.302] - - - [63488, 62977, 1, 512] - - [458, 94.337] - - - [62976, 62465, 1, 512] - - [458, 94.328] - - - [62464, 61953, 1, 512] - - [458, 94.339] - - - [61952, 61441, 1, 512] - - [458, 94.295] - - - [61440, 60929, 1, 512] - - [453, 94.284] - - - [60928, 60417, 1, 512] - - [453, 94.309] - - - [60416, 59905, 1, 512] - - [453, 94.319] - - - [59904, 59393, 1, 512] - - [458, 94.277] - - - [59392, 58881, 1, 512] - - [458, 94.275] - - - [58880, 58369, 1, 512] - - [453, 94.303] - - - [58368, 57857, 1, 512] - - [453, 94.293] - - - [57856, 57345, 1, 512] - - [458, 94.298] - - - [57344, 56833, 1, 512] - - [458, 94.261] - - - [56832, 56321, 1, 512] - - [453, 94.282] - - - [56320, 55809, 1, 512] - - [453, 94.286] - - - [55808, 55297, 1, 512] - - [453, 94.29] - - - [55296, 54785, 1, 512] - - [458, 94.267] - - - [54784, 54273, 1, 512] - - [458, 94.229] - - - [54272, 53761, 1, 512] - - [453, 94.262] - - - [53760, 53249, 1, 512] - - [458, 94.275] - - - [53248, 52737, 1, 512] - - [453, 94.24] - - - [52736, 52225, 1, 512] - - [458, 94.251] - - - [52224, 51713, 1, 512] - - [458, 94.258] - - - [51712, 51201, 1, 512] - - [453, 94.242] - - - [51200, 50689, 1, 512] - - [453, 94.25] - - - [50688, 50177, 1, 512] - - [458, 94.225] - - - [50176, 49665, 1, 512] - - [453, 94.234] - - - [49664, 49153, 1, 512] - - [453, 94.228] - - - [49152, 48641, 1, 512] - - [453, 94.213] - - - [48640, 48129, 1, 512] - - [453, 94.214] - - - [48128, 47617, 1, 512] - - [453, 94.194] - - - [47616, 47105, 1, 512] - - [453, 94.224] - - - [47104, 46593, 1, 512] - - [458, 94.21] - - - [46592, 46081, 1, 512] - - [458, 94.2] - - - [46080, 45569, 1, 512] - - [453, 94.182] - - - [45568, 45057, 1, 512] - - [458, 94.22] - - - [45056, 44545, 1, 512] - - [453, 94.169] - - - [44544, 44033, 1, 512] - - [453, 94.14] - - - [44032, 43521, 1, 512] - - [458, 94.168] - - - [43520, 43009, 1, 512] - - [458, 94.152] - - - [43008, 42497, 1, 512] - - [458, 94.156] - - - [42496, 41985, 1, 512] - - [458, 94.1] - - - [41984, 41473, 1, 512] - - [453, 94.121] - - - [41472, 40961, 1, 512] - - [458, 94.096] - - - [40960, 40449, 1, 512] - - [453, 94.081] - - - [40448, 39937, 1, 512] - - [453, 94.091] - - - [39936, 39425, 1, 512] - - [458, 94.082] - - - [39424, 38913, 1, 512] - - [453, 94.035] - - - [38912, 38401, 1, 512] - - [455, 94.023] - - - [38400, 37889, 1, 512] - - [458, 94.088] - - - [37888, 37377, 1, 512] - - [458, 94.073] - - - [37376, 36865, 1, 512] - - [453, 94.086] - - - [36864, 36353, 1, 512] - - [453, 94.014] - - - [36352, 35841, 1, 512] - - [458, 94.032] - - - [35840, 35329, 1, 512] - - [453, 94.02] - - - [35328, 34817, 1, 512] - - [458, 93.985] - - - [34816, 34305, 1, 512] - - [458, 93.962] - - - [34304, 33793, 1, 512] - - [458, 94.035] - - - [33792, 33281, 1, 512] - - [458, 93.969] - - - [33280, 32769, 1, 512] - - [455, 93.912] - - - [32768, 32257, 1, 512] - - [458, 93.944] - - - [32256, 31745, 1, 512] - - [458, 93.939] - - - [31744, 31233, 1, 512] - - [452, 93.912] - - - [31232, 30721, 1, 512] - - [453, 93.895] - - - [30720, 30209, 1, 512] - - [458, 93.874] - - - [30208, 29697, 1, 512] - - [458, 93.857] - - - [29696, 29185, 1, 512] - - [458, 93.82] - - - [29184, 28673, 1, 512] - - [453, 93.834] - - - [28672, 28161, 1, 512] - - [455, 93.794] - - - [28160, 27649, 1, 512] - - [458, 93.767] - - - [27648, 27137, 1, 512] - - [453, 93.713] - - - [27136, 26625, 1, 512] - - [458, 93.722] - - - [26624, 26113, 1, 512] - - [458, 93.71] - - - [26112, 25601, 1, 512] - - [452, 93.649] - - - [25600, 25089, 1, 512] - - [459, 93.648] - - - [25088, 24577, 1, 512] - - [454, 93.596] - - - [24576, 24065, 1, 512] - - [453, 93.588] - - - [24064, 23553, 1, 512] - - [454, 93.598] - - - [23552, 23041, 1, 512] - - [458, 93.471] - - - [23040, 22529, 1, 512] - - [458, 93.583] - - - [22528, 22017, 1, 512] - - [458, 93.458] - - - [22016, 21505, 1, 512] - - [451, 93.446] - - - [21504, 20993, 1, 512] - - [452, 93.374] - - - [20992, 20481, 1, 512] - - [455, 93.405] - - - [20480, 19969, 1, 512] - - [452, 93.367] - - - [19968, 19457, 1, 512] - - [453, 93.233] - - - [19456, 18945, 1, 512] - - [455, 93.202] - - - [18944, 18433, 1, 512] - - [453, 93.201] - - - [18432, 17921, 1, 512] - - [455, 93.115] - - - [17920, 17409, 1, 512] - - [458, 92.994] - - - [17408, 16897, 1, 512] - - [458, 93.041] - - - [16896, 16385, 1, 512] - - [457, 92.921] - - - [16384, 15873, 1, 512] - - [453, 92.91] - - - [15872, 15361, 1, 512] - - [455, 92.787] - - - [15360, 14849, 1, 512] - - [452, 92.693] - - - [14848, 14337, 1, 512] - - [454, 92.662] - - - [14336, 13825, 1, 512] - - [453, 92.681] - - - [13824, 13313, 1, 512] - - [454, 92.402] - - - [13312, 12801, 1, 512] - - [455, 92.384] - - - [12800, 12289, 1, 512] - - [457, 92.292] - - - [12288, 11777, 1, 512] - - [452, 92.046] - - - [11776, 11265, 1, 512] - - [458, 92.001] - - - [11264, 10753, 1, 512] - - [457, 91.91] - - - [10752, 10241, 1, 512] - - [455, 91.883] - - - [10240, 9729, 1, 512] - - [452, 91.138] - - - [9728, 9217, 1, 512] - - [454, 91.205] - - - [9216, 8705, 1, 512] - - [457, 90.541] - - - [8704, 8193, 1, 512] - - [454, 90.392] - - - [8192, 7681, 1, 512] - - [455, 89.827] - - - [7680, 7169, 1, 512] - - [460, 89.697] - - - [7168, 6657, 1, 512] - - [452, 88.368] - - - [6656, 6145, 1, 512] - - [457, 88.868] - - - [6144, 5633, 1, 512] - - [460, 86.592] - - - [5632, 5121, 1, 512] - - [460, 86.576] - - - [5120, 4609, 1, 512] - - [460, 84.11] - - - [4608, 4097, 1, 512] - - [454, 83.015] - - - [4096, 3585, 1, 512] - - [460, 78.76] - - - [3584, 3073, 1, 512] - - [460, 78.362] -- null diff --git a/library/src/blas3/Tensile/Logic/nonMFMA_legacy/aldebaran_Cijk_Ailk_Bjlk_DB.yaml b/library/src/blas3/Tensile/Logic/nonMFMA_legacy/aldebaran_Cijk_Ailk_Bjlk_DB.yaml deleted file mode 100644 index ba3eb1e..0000000 --- a/library/src/blas3/Tensile/Logic/nonMFMA_legacy/aldebaran_Cijk_Ailk_Bjlk_DB.yaml +++ /dev/null @@ -1,144526 +0,0 @@ -- {MinimumRequiredVersion: 4.8.1} -- aldebaran -- gfx90a -- [Device 0050, Device 0051, Device 0052, Device 0054, Device 0062, Device 7400, Device - 740c] -- AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false -- - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 0 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x064x08_PLR1_TT04_04_WG16_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 1 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x064x04_PLR1_TT04_04_WG16_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: 1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 2 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x048x04_PLR0_TT04_06_WG16_08_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 3 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x048x04_PLR0_TT04_06_WG16_08_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 4 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x048x04_PLR0_TT04_06_WG16_08_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 5 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x048x04_PLR0_TT04_06_WG16_08_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 6 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x064x04_PLR1_TT04_08_WG16_08_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 4 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 7 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x064x04_PLR1_TT04_08_WG16_08_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 8 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x064x04_PLR1_TT04_08_WG16_08_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 9 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT048x064x04_PLR0_TT06_04_WG08_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 4 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 10 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x064x04_PLR1_TT08_04_WG08_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 4] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 96 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 48 - LVPA: 2 - LVPB: 2 - LdsNumElements: 1792 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 384 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 96 - MacroTileA: 96 - MacroTileB: 96 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 36 - NumGlobalWriteVectorsPerThread: 18 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 2 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 11 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT096x096x04_PLR0_TT06_06_WG16_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 6] - ThreadTile0: 6 - ThreadTile1: 6 - ThreadTileA: 6 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 4 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 12 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x128x04_PLR1_TT04_08_WG16_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 13 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x064x08_PLR0_TT04_04_WG16_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 14 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x064x08_PLR0_TT04_04_WG16_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 2 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 2 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: 1 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 32 - LSPA: 2 - LSPB: 4 - LVCA: 48 - LVCB: 16 - LVPA: 1 - LVPB: 2 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 384 - LdsOffsetB_Blk: 896 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 32 - MacroTileA: 96 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 15 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT096x032x04_PLR1_TT06_04_WG16_08_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: 1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 2 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 2 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: 1 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 16 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x048x04_PLR1_TT04_06_WG16_08_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: 1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 2 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 2 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: 1 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdsNumElements: 896 - LdsNumElementsAlignedA: 128 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 128 - LdsOffsetB_Blk: 640 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 17 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT032x064x04_PLR0_TT04_04_WG08_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: 1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 2 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 2 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: 1 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 18 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT048x064x04_PLR0_TT06_04_WG08_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: 1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 2 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 2 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: 1 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdsNumElements: 1664 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 19 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT096x064x04_PLR0_TT06_04_WG16_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: 1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 2 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 2 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: 1 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 20 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x032x08_PLR1_TT04_04_WG16_08_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: 1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 2 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 2 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: 1 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 21 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x064x08_PLR1_TT04_04_WG16_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: 1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 2 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 2 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: 1 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 22 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x064x08_PLR0_TT04_04_WG16_16_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: 1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 2 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 2 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 12 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: 1 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdsNumElements: 3200 - LdsNumElementsAlignedA: 768 - LdsNumElementsAlignedB: 384 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 768 - LdsOffsetB_Blk: 2816 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 12 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 3 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 23 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x032x12_PLR1_TT04_04_WG16_08_01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: 1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 1 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 4 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 24 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x048x04_PK00_PLR0_SNLL0_TT04_06_WG16_08_01_WGM01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 25 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x032x04_PK00_PLR0_SNLL0_TT04_04_WG16_08_01_WGM08 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 26 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x048x04_PK00_PLR0_SNLL1_TT04_06_WG16_08_01_WGM08 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 27 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x048x04_PK00_PLR0_SNLL0_TT04_06_WG16_08_01_WGM08 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 28 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT048x064x04_PK00_PLR0_SNLL1_TT06_04_WG08_16_01_WGM08 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 29 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT048x064x04_PK00_PLR0_SNLL0_TT06_04_WG08_16_01_WGM08 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdsNumElements: 1664 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 30 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT096x064x04_PK00_PLR0_SNLL1_TT06_04_WG16_16_01_WGM08 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 31 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x032x08_PK00_PLR1_SNLL0_TT04_04_WG16_08_01_WGM01 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 32 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x064x08_PK00_PLR1_SNLL1_TT04_04_WG16_16_01_WGM08 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentLDA: 5 - IndexAssignmentLDB: 6 - IndexAssignmentLDC: 4 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 33 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT064x064x08_PK00_PLR1_SNLL0_TT04_04_WG16_16_01_WGM08 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 34 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR0_SNLL0_TT4_4_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 35 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x48x4_SE_PLR1_SNLL0_TT4_6_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 36 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x48x4_SE_PLR0_SNLL0_TT4_6_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 4 - LVCA: 64 - LVCB: 16 - LVPA: 1 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1664 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 128 - MacroTile1: 32 - MacroTileA: 128 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 37 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x32x4_SE_PLR0_SNLL0_TT8_4_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [8, 4] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 38 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x4_SE_PLR1_SNLL1_TT4_8_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 39 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x4_SE_PLR1_SNLL0_TT4_8_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 40 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x4_SE_PLR1_SNLL0_TT8_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 4] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 41 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x4_SE_PLR1_SNLL1_TT4_8_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 42 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x4_SE_PLR1_SNLL0_TT4_8_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 32 - LSPA: 2 - LSPB: 4 - LVCA: 48 - LVCB: 16 - LVPA: 1 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 384 - LdsOffsetB_Blk: 896 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 32 - MacroTileA: 96 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 43 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x32x4_SE_PLR1_SNLL1_TT6_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 44 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x48x4_SE_PLR0_SNLL1_TT4_6_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 45 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x48x4_SE_PLR0_SNLL0_TT4_6_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 32 - LSPA: 2 - LSPB: 4 - LVCA: 64 - LVCB: 16 - LVPA: 1 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1664 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 128 - MacroTile1: 32 - MacroTileA: 128 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 46 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x32x4_SE_PLR0_SNLL1_TT8_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [8, 4] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 47 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT48x64x4_SE_PLR0_SNLL0_TT6_4_WG8_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 48 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR0_SNLL0_TT4_4_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 49 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x48x4_SE_PLR0_SNLL1_TT4_6_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 50 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x48x4_SE_PLR0_SNLL0_TT4_6_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 128 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 128 - LdsOffsetB_Blk: 640 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 51 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x64x4_SE_PLR0_SNLL0_TT4_4_WG8_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 52 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT48x64x4_SE_PLR1_SNLL1_TT6_4_WG8_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 53 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT48x64x4_SE_PLR0_SNLL0_TT6_4_WG8_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 54 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_SE_PLR1_SNLL1_TT4_4_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 55 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_SE_PLR1_SNLL0_TT4_4_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 56 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR1_SNLL1_TT4_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 57 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR0_SNLL1_TT4_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 58 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR1_SNLL0_TT4_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 59 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR0_SNLL0_TT4_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 60 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_SE_PLR1_SNLL1_TT4_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 48 - LVCB: 16 - LVPA: 1 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 768 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 768 - LdsOffsetB_Blk: 1792 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 96 - MacroTile1: 32 - MacroTileA: 96 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 61 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x32x8_SE_PLR1_SNLL1_TT6_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 62 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR1_SNLL1_TT4_4_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 63 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR0_SNLL1_TT4_4_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 64 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR0_SNLL0_TT4_4_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 65 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_SE_PLR1_SNLL0_TT4_4_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 66 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR1_SNLL1_TT4_4_WG16_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 67 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR0_SNLL1_TT4_4_WG16_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 68 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR1_SNLL0_TT4_4_WG16_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 69 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR0_SNLL0_TT4_4_WG16_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 70 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR1_SNLL1_TT4_4_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 71 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR0_SNLL1_TT4_4_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 72 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR1_SNLL0_TT4_4_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 73 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR0_SNLL0_TT4_4_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 32 - LSPA: 2 - LSPB: 4 - LVCA: 48 - LVCB: 16 - LVPA: 1 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 384 - LdsOffsetB_Blk: 896 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 32 - MacroTileA: 96 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 74 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x32x4_SE_PLR1_SNLL1_TT6_4_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 75 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x4_SE_PLR1_SNLL1_TT4_8_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 76 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT48x64x4_SE_PLR0_SNLL1_TT6_4_WG8_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 77 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT48x64x4_SE_PLR0_SNLL0_TT6_4_WG8_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1664 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 78 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x4_SE_PLR0_SNLL1_TT6_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1664 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 79 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x4_SE_PLR0_SNLL0_TT6_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 80 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x4_SE_PLR1_SNLL1_TT8_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [8, 4] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 81 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x4_SE_PLR1_SNLL0_TT8_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 4] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 82 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x4_SE_PLR1_SNLL1_TT4_8_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 83 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x4_SE_PLR1_SNLL0_TT4_8_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 84 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR0_SNLL1_TT4_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 85 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR0_SNLL0_TT4_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 32 - LSPA: 2 - LSPB: 4 - LVCA: 48 - LVCB: 16 - LVPA: 1 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 384 - LdsOffsetB_Blk: 896 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 32 - MacroTileA: 96 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 86 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x32x4_SE_PLR1_SNLL0_TT6_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 87 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT48x64x4_SE_PLR0_SNLL1_TT6_4_WG8_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 88 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT48x64x4_SE_PLR0_SNLL0_TT6_4_WG8_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1664 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 89 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x4_SE_PLR0_SNLL1_TT6_4_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1664 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 90 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x4_SE_PLR0_SNLL0_TT6_4_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 91 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x4_SE_PLR1_SNLL0_TT4_8_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 92 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR0_SNLL1_TT4_4_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 93 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR0_SNLL0_TT4_4_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 32 - LSPA: 2 - LSPB: 4 - LVCA: 48 - LVCB: 16 - LVPA: 1 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 384 - LdsOffsetB_Blk: 896 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 32 - MacroTileA: 96 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 94 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x32x4_SE_PLR1_SNLL1_TT6_4_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 95 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT48x64x4_SE_PLR0_SNLL1_TT6_4_WG8_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 96 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT48x64x4_SE_PLR0_SNLL0_TT6_4_WG8_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1664 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 97 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x4_SE_PLR0_SNLL0_TT6_4_WG16_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 512 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 98 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_SE_PLR0_SNLL0_TT4_4_WG32_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 32 - SubGroup1: 16 - SubGroupA: 32 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [32, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 99 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_SE_PLR1_SNLL1_TT4_4_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 100 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_SE_PLR1_SNLL0_TT4_4_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 8 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 101 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x64x8_SE_PLR1_SNLL1_TT4_4_WG8_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 102 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR1_SNLL1_TT4_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 103 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR0_SNLL1_TT4_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 104 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR1_SNLL0_TT4_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 105 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR0_SNLL0_TT4_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 106 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_SE_PLR1_SNLL1_TT4_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 107 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_SE_PLR1_SNLL0_TT4_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 32 - LSPA: 2 - LSPB: 8 - LVCA: 48 - LVCB: 16 - LVPA: 1 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 768 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 768 - LdsOffsetB_Blk: 1792 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 96 - MacroTile1: 32 - MacroTileA: 96 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 108 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x32x8_SE_PLR1_SNLL0_TT6_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 8 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 109 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x64x8_SE_PLR1_SNLL1_TT4_4_WG8_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 110 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR1_SNLL1_TT4_4_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 111 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR0_SNLL1_TT4_4_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 112 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR1_SNLL0_TT4_4_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 113 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR0_SNLL0_TT4_4_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 114 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_SE_PLR1_SNLL1_TT4_4_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 115 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_SE_PLR1_SNLL0_TT4_4_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 8 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 116 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x64x8_SE_PLR1_SNLL1_TT4_4_WG8_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 117 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR1_SNLL1_TT4_4_WG16_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 118 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR0_SNLL1_TT4_4_WG16_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 119 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR1_SNLL0_TT4_4_WG16_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 120 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR0_SNLL0_TT4_4_WG16_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 121 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x48x4_SE_PLR0_SNLL1_TT4_6_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 128 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 128 - LdsOffsetB_Blk: 640 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 122 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x64x4_SE_PLR0_SNLL1_TT4_4_WG8_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 128 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 128 - LdsOffsetB_Blk: 640 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 123 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x64x4_SE_PLR0_SNLL0_TT4_4_WG8_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 124 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x48x4_SE_PLR0_SNLL1_TT4_6_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 125 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x48x4_SE_PLR0_SNLL0_TT4_6_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 128 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 128 - LdsOffsetB_Blk: 640 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 126 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x64x4_SE_PLR0_SNLL0_TT4_4_WG8_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 127 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR0_SNLL1_TT4_4_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 32 - LSPA: 2 - LSPB: 4 - LVCA: 48 - LVCB: 16 - LVPA: 1 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 384 - LdsOffsetB_Blk: 896 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 32 - MacroTileA: 96 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 128 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x32x4_SE_PLR1_SNLL0_TT6_4_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 129 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x48x4_SE_PLR0_SNLL1_TT4_6_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1664 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 130 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x4_SE_PLR0_SNLL1_TT6_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1664 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 131 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x4_SE_PLR0_SNLL0_TT6_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 132 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x4_SE_PLR1_SNLL1_TT8_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [8, 4] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 133 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_SE_PLR1_SNLL0_TT4_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 2 - LSPB: 4 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 134 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_SE_ - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: 1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 1 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 135 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT48x64x4_SE_PLR0_SNLL1_TT6_4_WG8_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1664 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 136 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x4_SE_PLR1_SNLL0_TT6_4_WG16_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 137 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR1_SNLL1_TT4_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 138 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR0_SNLL1_TT4_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 128 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 128 - LdsOffsetB_Blk: 640 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 139 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x64x4_SE_PLR0_SNLL1_TT4_4_WG8_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1664 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 140 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x4_SE_PLR1_SNLL1_TT6_4_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1664 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 141 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x4_SE_PLR1_SNLL0_TT6_4_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 8 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 142 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x64x8_SE_PLR1_SNLL1_TT4_4_WG8_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 8 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 143 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x64x8_SE_PLR1_SNLL1_TT4_4_WG8_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 144 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_SE_PLR1_SNLL0_TT4_4_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 145 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_SE_PLR1_SNLL1_TT4_4_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 8 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 146 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x64x8_SE_PLR1_SNLL1_TT4_4_WG8_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 8 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 147 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x64x8_SE_PLR1_SNLL0_TT4_4_WG8_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 148 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT48x64x4_SE_PLR1_SNLL0_TT6_4_WG8_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 149 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR0_SNLL0_TT4_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 150 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x48x4_SE_PLR1_SNLL0_TT4_6_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 151 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x4_SE_PLR1_SNLL1_TT4_8_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 152 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x48x4_SE_PLR1_SNLL1_TT4_6_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: false - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 153 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT48x64x4_SE_PLR0_SNLL1_TT6_4_WG8_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1664 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 154 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x4_SE_PLR1_SNLL1_TT6_4_WG16_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 155 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x4_SE_PLR1_SNLL0_TT8_4_WG16_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 4] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 8 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 156 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x64x8_SE_PLR1_SNLL0_TT4_4_WG8_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 5 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 3 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1920 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 157 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT48x64x8_SE_PLR1_SNLL0_TT6_4_WG8_16_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 3 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 64 - LSPA: 8 - LSPB: 4 - LVCA: 16 - LVCB: 32 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 64 - MacroTileA: 32 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 158 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT32x64x8_SE_PLR1_SNLL0_TT4_4_WG8_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 159 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x48x4_SE_PLR1_SNLL1_TT4_6_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 48 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 24 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 192 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 48 - MacroTileA: 64 - MacroTileB: 48 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 160 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x48x4_SE_PLR1_SNLL0_TT4_6_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 6] - ThreadTile0: 4 - ThreadTile1: 6 - ThreadTileA: 4 - ThreadTileB: 6 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 161 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x4_SE_PLR1_SNLL0_TT4_8_WG16_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 162 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x4_SE_PLR1_SNLL0_TT4_8_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 16 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 896 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 163 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x4_SE_PLR1_SNLL0_TT4_4_WG16_8_1_WGM1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 4] - ThreadTile0: 4 - ThreadTile1: 4 - ThreadTileA: 4 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 1 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 32 - LSPA: 2 - LSPB: 4 - LVCA: 48 - LVCB: 16 - LVPA: 1 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 384 - LdsOffsetB_Blk: 896 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 32 - MacroTileA: 96 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 164 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x32x4_SE_PLR1_SNLL0_TT6_4_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 48 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 24 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 960 - LdsNumElementsAlignedA: 192 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 192 - LdsOffsetB_Blk: 704 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 48 - MacroTile1: 64 - MacroTileA: 48 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 165 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT48x64x4_SE_PLR1_SNLL1_TT6_4_WG8_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 8 - SubGroup1: 16 - SubGroupA: 8 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [8, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 166 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x4_SE_PLR1_SNLL1_TT8_4_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [8, 4] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1792 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 256 - LdsOffsetB_Blk: 1280 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 167 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x4_SE_PLR1_SNLL1_TT4_8_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 32 - LSPA: 2 - LSPB: 4 - LVCA: 48 - LVCB: 16 - LVPA: 1 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 384 - LdsOffsetB_Blk: 896 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 32 - MacroTileA: 96 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 168 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x32x4_SE_PLR1_SNLL0_TT6_4_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 169 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x4_SE_PLR1_SNLL1_TT4_8_WG16_8_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: true - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 48 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1664 - LdsNumElementsAlignedA: 384 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 384 - LdsOffsetB_Blk: 1408 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 170 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x4_SE_PLR1_SNLL0_TT6_4_WG16_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 5 - LSPB: 8 - LVCA: 48 - LVCB: 32 - LVPA: 3 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 3328 - LdsNumElementsAlignedA: 768 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 768 - LdsOffsetB_Blk: 2816 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 171 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x8_SE_PLR1_SNLL1_TT6_4_WG16_16_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: true - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 3 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 96 - LSCB: 64 - LSPA: 5 - LSPB: 8 - LVCA: 48 - LVCB: 32 - LVPA: 3 - LVPB: 4 - LdcEqualsLdd: true - LdsNumElements: 3328 - LdsNumElementsAlignedA: 768 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 768 - LdsOffsetB_Blk: 2816 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 8 - MacroTile0: 96 - MacroTile1: 64 - MacroTileA: 96 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 24 - NumGlobalWriteVectorsPerThread: 12 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 172 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x64x8_SE_PLR1_SNLL0_TT6_4_WG16_16_1_WGM8 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [6, 4] - ThreadTile0: 6 - ThreadTile1: 4 - ThreadTileA: 6 - ThreadTileB: 4 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _staggerStrideShift: 2 - fractionalPerpOverhangA: 3 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 1 - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableKernelPieces: 0 - EdgeType: ShiftPtr - ExpandPointerSwap: true - FractionalLoad: true - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - InnerUnroll: 1 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 32 - LVCB: 32 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsNumElements: 1024 - LdsNumElementsAlignedA: 256 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 512 - LdsOffsetB: 256 - LdsOffsetB_Blk: 768 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopTail: true - LoopUnroll: 4 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MaxOccupancy: 40 - MinGlobalWriteVectorWidth: 1 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackedC0Indices: [I] - PackedC1Indices: [J] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStrides: false - ReplacementKernel: true - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 173 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x4_SE_PLR1_SNLL0_TT4_8_WG16_8_1_WGM4 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [4, 8] - ThreadTile0: 4 - ThreadTile1: 8 - ThreadTileA: 4 - ThreadTileB: 8 - UnrollMemFence: false - UseSgprForGRO: false - Valid: true - VectorAtomicWidth: 1 - VectorStore: true - VectorWidth: 2 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _staggerStrideShift: 3 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 8] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 512 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 0 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 174 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 1 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 8] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 8 - LVPB: 8 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 2 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 4 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 4 - NumGlobalWriteVectorsPerThread: 2 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 4 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 175 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 2] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: 1 - _VectorStore: 1 - _staggerStrideShift: 1 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 16 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 8] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsOffsetA: 0 - LdsOffsetB: 512 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 16 - LoopTail: true - LoopUnroll: 16 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 176 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 16 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 8] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsOffsetA: 0 - LdsOffsetB: 512 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 4 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 177 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 16 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 8] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsOffsetA: 0 - LdsOffsetB: 512 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 178 - StaggerU: 16 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 0 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 8] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 512 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 179 - StaggerU: 32 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 1 - - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: false - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: false - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 8] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 512 - LdsOffsetA: 0 - LdsOffsetB: 256 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: -1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 4 - MacroTile0: 32 - MacroTile1: 32 - MacroTileA: 32 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 128 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 64 - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: false - PrefetchLocalRead: 0 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 180 - StaggerU: 32 - StaggerUMapping: 1 - StaggerUStride: 128 - StoreVectorWidth: 4 - SubGroup0: 16 - SubGroup1: 8 - SubGroupA: 16 - SubGroupB: 8 - SuppressNoLoadLoop: false - ThreadTile: [2, 4] - ThreadTile0: 2 - ThreadTile1: 4 - ThreadTileA: 2 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _UseSgprForGRO: false - _VectorStore: 1 - _staggerStrideShift: 1 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 181 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_SE_FL0_WGM11 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 182 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_SE_FL0_WGM8 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 8 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - FractionalLoad: 1 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 8 - LoopTail: true - LoopUnroll: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 183 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_SE_FL1_WGM8 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - fractionalPerpOverhangA: 0 - fractionalPerpOverhangB: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - DepthU: 4 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: false - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 2 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstruction: [] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 184 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x4_SE_FL0_WGM11 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [8, 8] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 3 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 185 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB0_EPS1_IU1_NLCA1_PGR1_SIA3_TT8_32_WG8_32_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 186 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB0_EPS1_IU1_NLCA1_PGR1_SIA3_TT8_32_WG16_16_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 187 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB0_EPS1_IU1_NLCA1_PGR1_SIA3_TT8_32_WG8_32_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 188 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB0_EPS1_IU1_NLCA1_PGR1_SIA3_TT8_32_WG16_16_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 189 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG16_16_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 190 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG8_32_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 191 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB0_EPS1_IU1_NLCA1_PGR1_SIA3_TT8_32_WG16_16_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 192 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB0_EPS1_IU1_NLCA1_PGR1_SIA3_TT8_32_WG8_32_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 193 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG8_32_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 194 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG16_16_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 195 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB0_EPS1_IU1_NLCA1_PGR1_SIA3_TT8_32_WG8_32_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 196 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB0_EPS1_IU1_NLCA1_PGR1_SIA3_TT8_32_WG16_16_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 197 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG8_32_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 198 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG16_16_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 199 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB0_EPS1_IU1_NLCA1_PGR1_SIA3_TT8_32_WG16_16_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 4 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 200 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SE_1LDSB1_EPS1_IU4_NLCA1_PGR1_SIA2_TT8_32_WG16_16_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 201 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB0_EPS1_IU1_NLCA1_PGR1_SIA3_TT8_32_WG8_32_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 202 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB0_EPS1_IU1_NLCA1_PGR1_SIA3_TT8_32_WG8_32_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: true - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 203 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB0_EPS1_IU1_NLCA1_PGR1_SIA3_TT8_32_WG16_16_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 4 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 204 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SE_1LDSB1_EPS1_IU4_NLCA1_PGR1_SIA2_TT8_32_WG16_16_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 205 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG16_16_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 4 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 206 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SE_1LDSB1_EPS1_IU4_NLCA1_PGR1_SIA2_TT8_32_WG8_32_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 207 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG8_32_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 208 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG8_32_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [8, 32, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 209 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG16_16_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: true - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 1 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: false - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 256 - LSCB: 64 - LSPA: 2 - LSPB: 8 - LVCA: 128 - LVCB: 32 - LVPA: 1 - LVPB: 4 - LdcEqualsLdd: true - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2560 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 4 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 256 - MacroTile1: 64 - MacroTileA: 256 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: true - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 2 - ScheduleLocalWrite: 1 - SolutionIndex: 210 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT256x64x8_MI16x16x4x1_SE_1LDSB1_EPS1_IU2_NLCA1_PGR1_SIA2_TT8_32_WG32_8_1_WGM8 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreVectorWidth: 1 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 211 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 1 - LSPB: 2 - LVCA: 64 - LVCB: 32 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 212 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_64_WG64_4_1_WGM11 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 213 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 214 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 215 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM10 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 10 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 216 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 217 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR5_SU32_SUS128_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 218 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR5_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 219 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 220 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR5_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 221 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 222 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS256_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 223 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 224 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS256_TT2_128_WG64_4_1_WGM10 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 10 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 1 - LSPB: 2 - LVCA: 64 - LVCB: 32 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 225 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_64_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 1 - LSPB: 2 - LVCA: 64 - LVCB: 32 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 226 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_64_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 1 - LSPB: 2 - LVCA: 64 - LVCB: 32 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 227 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_64_WG64_4_1_WGM11 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 228 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 2 - LSPB: 1 - LVCA: 32 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1536 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 512 - LdsOffsetB_Blk: 2560 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 1 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 229 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x8_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 230 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_TT2_128_WG64_4_1_WGM10 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 10 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 231 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM4 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 232 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 233 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 32 - LSCB: 128 - LSPA: 4 - LSPB: 1 - LVCA: 16 - LVCB: 64 - LVPA: 2 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3584 - LdsNumElementsAlignedA: 1536 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1536 - LdsOffsetB_Blk: 5632 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [6, 2] - MIWaveTileA: 6 - MIWaveTileB: 2 - MacroTile0: 96 - MacroTile1: 128 - MacroTileA: 96 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 48 - NumGlobalWriteVectorsPerThread: 24 - NumLoadsA: 3 - NumLoadsB: 4 - NumLoadsCoalescedA: 3 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 234 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT96x128x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA3_NLCB1_PLR5_SU0_SUS256_TT6_32_WG16_16_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [6, 32] - ThreadTile0: 24 - ThreadTile1: 2 - ThreadTileA: 24 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 235 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 236 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_GRVW2_NEPBS1_NLCA1_NLCB1_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM10 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 10 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 2 - LSPB: 2 - LVCA: 32 - LVCB: 32 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 237 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS256_TT2_32_WG32_8_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 1 - LSPB: 2 - LVCA: 64 - LVCB: 32 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 238 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS256_TT2_64_WG64_4_1_WGM11 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 8 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 239 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW1_NEPBS2_NLCA2_NLCB1_PLR5_SU0_SUS256_TT2_64_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 8 - NumLoadsB: 4 - NumLoadsCoalescedA: 2 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 240 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW1_NEPBS2_NLCA2_NLCB1_PLR5_SU0_SUS256_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 1 - LSPB: 2 - LVCA: 64 - LVCB: 32 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 241 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS256_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 1 - LSPB: 2 - LVCA: 64 - LVCB: 32 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1536 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 242 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x8_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR3_SU0_SUS256_TT2_64_WG64_4_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 1 - LSPB: 2 - LVCA: 64 - LVCB: 32 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 243 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_TT2_64_WG64_4_1_WGM8 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 2 - LSPB: 2 - LVCA: 32 - LVCB: 32 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 244 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 2 - LSPB: 2 - LVCA: 32 - LVCB: 32 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 245 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW2_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalReadWarmup: false - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 1 - LSPB: 1 - LVCA: 64 - LVCB: 64 - LVPA: 1 - LVPB: 1 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 2 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 5 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 246 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_GRVW1_NEPBS2_NLCA1_NLCB1_PLR5_SU32_SUS128_TT2_32_WG32_8_1_WGM4 - SourceSwap: 1 - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 1 - WaveSeparateGlobalReadB: 1 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 4 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: -1 - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {3: 512} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {0: 128, 1: 128} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: DGEMM_Aldebaran_PKFixedAtomic512Latest - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: Branch - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 1 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 247 - SolutionNameMin: DGEMM_Aldebaran_PKFixedAtomic512Latest - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 0 - AggressivePerfMode: 1 - AssertAlphaValue: -1 - AssertBetaValue: 1 - AssertCEqualsD: true - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {3: 512} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {0: 128, 1: 128} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: DGEMM_Aldebaran_PKFixedAtomic512_104 - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: Branch - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: false - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 0 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 64 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 1 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 1 - ScheduleLocalWrite: 1 - SolutionIndex: 248 - SolutionNameMin: DGEMM_Aldebaran_PKFixedAtomic512_104 - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: false - StoreRemapVectorWidth: 0 - StoreSyncOpt: 0 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 8 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 249 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 250 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 251 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 252 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 253 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 254 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 255 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 256 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 257 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 258 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 259 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 260 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 261 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 262 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 263 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 264 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 265 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 266 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 267 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 268 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 269 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 270 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 271 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 272 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 273 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 274 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 275 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 276 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 277 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 278 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 279 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 280 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 281 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 282 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 283 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 284 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 285 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 286 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 287 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 288 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 289 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 290 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 291 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 292 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 293 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 294 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 295 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 296 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 297 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 298 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 299 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 300 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 301 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 302 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 303 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 304 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 305 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 306 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 307 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 308 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 309 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 310 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 311 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 312 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 313 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 314 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 315 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 316 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 317 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 318 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 319 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 320 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 321 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 322 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 323 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 324 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 325 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 326 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU0_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 327 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS256_TT8_32_WG16_16_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 328 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS128_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 329 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 330 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 4] - MIWaveTile: [8, 2] - MIWaveTileA: 8 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 331 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT8_32_WG16_16_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 4 - SubGroup1: 64 - SubGroupA: 4 - SubGroupB: 64 - SuppressNoLoadLoop: false - ThreadTile: [8, 32] - ThreadTile0: 32 - ThreadTile1: 2 - ThreadTileA: 32 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 16, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 332 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 333 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 334 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU0_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 335 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 336 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 337 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 338 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 339 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 340 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 341 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 342 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 343 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 344 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 345 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 346 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 347 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 348 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 349 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 350 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 351 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 352 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 353 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 354 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 355 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 356 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 357 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 358 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 359 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 360 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 361 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 362 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS128_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 363 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 364 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 365 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 366 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 367 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 368 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 369 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 370 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 371 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 372 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 373 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 374 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 375 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 376 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 377 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 378 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 379 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 380 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 381 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 382 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 383 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 384 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU32_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 385 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 386 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 387 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 388 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 389 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 390 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 391 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 392 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 393 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 394 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 395 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 396 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 397 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 398 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 399 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 400 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 401 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 402 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 403 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 404 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 405 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 406 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 407 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW1_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 408 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 409 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 410 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 411 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 412 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 413 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 414 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 415 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 416 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 417 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 418 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 419 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 420 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 421 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 422 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS128_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 423 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT2_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 424 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 128 - LSPA: 8 - LSPB: 4 - LVCA: 32 - LVCB: 64 - LVPA: 4 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 5120 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 4] - MIWaveTileA: 2 - MIWaveTileB: 4 - MacroTile0: 64 - MacroTile1: 128 - MacroTileA: 64 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 2 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 425 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x128x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU0_SUS256_TT2_64_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 64] - ThreadTile0: 8 - ThreadTile1: 4 - ThreadTileA: 8 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 426 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT4_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 427 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 428 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU32_SUS256_TT4_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 429 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS128_TT4_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 430 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM1_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 431 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x16_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 64 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 2 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 3072 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 2] - MIWaveTileA: 4 - MIWaveTileB: 2 - MacroTile0: 128 - MacroTile1: 64 - MacroTileA: 128 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 32 - NumGlobalWriteVectorsPerThread: 16 - NumLoadsA: 4 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 432 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x64x16_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS128_TT4_32_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 32] - ThreadTile0: 16 - ThreadTile1: 2 - ThreadTileA: 16 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: true - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 8 - LSPB: 8 - LVCA: 32 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 1 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 1 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 1 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 433 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS1_GRVW2_IU2_PGR1_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 32 - LSPA: 4 - LSPB: 8 - LVCA: 64 - LVCB: 32 - LVPA: 4 - LVPB: 8 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 768 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 256 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 1] - MIWaveTileA: 2 - MIWaveTileB: 1 - MacroTile0: 64 - MacroTile1: 32 - MacroTileA: 64 - MacroTileB: 32 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 8 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 1 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 1 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 434 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x32x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS128_TT2_16_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 16] - ThreadTile0: 8 - ThreadTile1: 1 - ThreadTileA: 8 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 2 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 64 - LSCB: 64 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 1024 - LdsNumElementsAlignedA: 512 - LdsNumElementsAlignedB: 512 - LdsOffsetA: 0 - LdsOffsetA_Blk: 1024 - LdsOffsetB: 512 - LdsOffsetB_Blk: 1536 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [2, 2] - MIWaveTileA: 2 - MIWaveTileB: 2 - MacroTile0: 64 - MacroTile1: 64 - MacroTileA: 64 - MacroTileB: 64 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 16 - NumGlobalWriteVectorsPerThread: 8 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 435 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT64x64x8_MI16x16x4x1_SN_AF0EM2_EPS0_GRVW1_IU2_PGR2_PLR1_SU0_SUS256_TT2_32_WG32_8_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [2, 32] - ThreadTile0: 8 - ThreadTile1: 2 - ThreadTileA: 8 - ThreadTileB: 2 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 0 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 1 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 16 - LSCB: 16 - LSPA: 4 - LSPB: 4 - LVCA: 16 - LVCB: 16 - LVPA: 4 - LVPB: 4 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 256 - LdsNumElementsAlignedA: 128 - LdsNumElementsAlignedB: 128 - LdsOffsetA: 0 - LdsOffsetA_Blk: 256 - LdsOffsetB: 128 - LdsOffsetB_Blk: 384 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 1 - LoopTail: true - LoopUnroll: 4 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [1, 1] - MIWaveTile: [1, 1] - MIWaveTileA: 1 - MIWaveTileB: 1 - MacroTile0: 16 - MacroTile1: 16 - MacroTileA: 16 - MacroTileB: 16 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 4 - NumGlobalWriteVectorsPerThread: 4 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 64 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 436 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT16x16x8_MI16x16x4x1_SN_TT1_16_WG16_4_1 - SourceSwap: false - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 1 - SubGroup0: 4 - SubGroup1: 16 - SubGroupA: 4 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [1, 16] - ThreadTile0: 4 - ThreadTile1: 1 - ThreadTileA: 4 - ThreadTileB: 1 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 1 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [16, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 437 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 438 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 439 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 440 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 441 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 442 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR3_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 443 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 444 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 445 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 446 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS256_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 447 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 2 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 448 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR3_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 2 - NumLoadsB: 2 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 2 - NumLoadsPerpendicularB: 2 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 449 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 8 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 2048 - LdsNumElementsAlignedA: 1024 - LdsNumElementsAlignedB: 1024 - LdsOffsetA: 0 - LdsOffsetA_Blk: 2048 - LdsOffsetB: 1024 - LdsOffsetB_Blk: 3072 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 450 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x8_MI16x16x4x1_SN_EPS0_GRVW1_IU1_PGR2_PLR1_SU0_SUS128_TT2_128_WG64_4_1_WGM11 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 8 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 1 - GlobalLoadVectorWidthB: 1 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 1 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: true - GuaranteeNoPartialB: true - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 2 - LSPB: 2 - LVCA: 128 - LVCB: 128 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 8 - NumLoadsB: 8 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 8 - NumLoadsPerpendicularB: 8 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 451 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW1_IU2_PGR2_PLR3_SU0_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 0 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: 1 - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [2, 2] - MIWaveTile: [4, 4] - MIWaveTileA: 4 - MIWaveTileB: 4 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 452 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS256_TT4_64_WG32_8_1_WGM11 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 8 - SubGroup1: 32 - SubGroupA: 8 - SubGroupB: 32 - SuppressNoLoadLoop: false - ThreadTile: [4, 64] - ThreadTile0: 16 - ThreadTile1: 4 - ThreadTileA: 16 - ThreadTileB: 4 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [32, 8, 1] - WorkGroupMapping: 11 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 2 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 2 - LoopTail: true - LoopUnroll: 8 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 3 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 453 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU2_PGR2_PLR3_SU32_SUS256_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 256 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 1 - - 1LDSBuffer: 1 - AggressivePerfMode: 1 - AssertAlphaValue: false - AssertBetaValue: false - AssertCEqualsD: false - AssertFree0ElementMultiple: 1 - AssertFree1ElementMultiple: 1 - AssertMinApproxSize: 3 - AssertSizeEqual: {} - AssertSizeGreaterThan: {} - AssertSizeLessThan: {} - AssertSizeMultiple: {} - AssertStrideAEqual: {0: 1} - AssertStrideBEqual: {0: 1} - AssertStrideCEqual: {0: 1} - AssertStrideDEqual: {0: 1} - AssertSummationElementMultiple: 1 - AssignedDerivedParameters: true - AssignedProblemIndependentDerivedParameters: true - AtomicAddC: false - BufferLoad: true - BufferStore: true - CheckDimOverflow: 0 - CheckTensorDimAsserts: false - CodeObjectVersion: default - CustomKernelName: '' - DepthU: 16 - DepthULdsDivisor: 1 - DirectToLds: false - DirectToLdsA: false - DirectToLdsB: false - DisableAtomicFail: 0 - DisableKernelPieces: 0 - DisableVgprOverlapping: false - EdgeType: ShiftPtr - EnableMatrixInstruction: true - ExpandPointerSwap: 0 - FractionalLoad: 0 - GlobalLoadVectorWidthA: 2 - GlobalLoadVectorWidthB: 2 - GlobalRead2A: true - GlobalRead2B: true - GlobalReadCoalesceGroupA: true - GlobalReadCoalesceGroupB: true - GlobalReadCoalesceVectorA: true - GlobalReadCoalesceVectorB: true - GlobalReadPerMfma: 1 - GlobalReadVectorWidth: 2 - GlobalSplitU: 1 - GlobalSplitUAlgorithm: MultipleBuffer - GlobalSplitUSummationAssignmentRoundRobin: true - GlobalSplitUWorkGroupMappingRoundRobin: false - GlobalWriteVectorWidth: 2 - GroupLoadStore: 1 - GuaranteeNoPartialA: false - GuaranteeNoPartialB: false - ISA: [9, 0, 10] - InnerUnroll: 1 - InterleaveAlpha: 0 - KernelLanguage: Assembly - LSCA: 128 - LSCB: 128 - LSPA: 4 - LSPB: 4 - LVCA: 64 - LVCB: 64 - LVPA: 2 - LVPB: 2 - LdcEqualsLdd: false - LdsBlockSizePerPad: 0 - LdsBlockSizePerPadA: 0 - LdsBlockSizePerPadB: 0 - LdsNumElements: 4096 - LdsNumElementsAlignedA: 2048 - LdsNumElementsAlignedB: 2048 - LdsOffsetA: 0 - LdsOffsetA_Blk: 4096 - LdsOffsetB: 2048 - LdsOffsetB_Blk: 6144 - LdsPadA: 0 - LdsPadB: 0 - LocalDotLayout: 1 - LocalRead2A: true - LocalRead2B: true - LocalReadVectorWidth: 1 - LocalSplitU: 1 - LocalWrite2A: true - LocalWrite2B: true - LocalWritePerMfma: -1 - LocalWriteUseSgprA: false - LocalWriteUseSgprB: false - LoopDoWhile: false - LoopIters: 4 - LoopTail: true - LoopUnroll: 16 - MACInstruction: MAC - MFMA_BF16_1K: false - MIBlock: [16, 16, 4, 1, 1, 1] - MIInputPerThread: 1 - MIOutputVectorWidth: 1 - MIRegPerOut: 2 - MIUseAccVgpr: true - MIWaveGroup: [4, 1] - MIWaveTile: [2, 8] - MIWaveTileA: 2 - MIWaveTileB: 8 - MacroTile0: 128 - MacroTile1: 128 - MacroTileA: 128 - MacroTileB: 128 - MacroTileShapeMax: 64 - MacroTileShapeMin: 1 - MagicDivAlg: 2 - MatrixInstB: 1 - MatrixInstBM: 1 - MatrixInstBN: 1 - MatrixInstK: 4 - MatrixInstM: 16 - MatrixInstN: 16 - MatrixInstruction: [16, 16, 4, 1] - MaxOccupancy: 40 - MaxVgprNumber: 256 - MinVgprNumber: 0 - NoReject: false - NonTemporalA: 0 - NonTemporalB: 0 - NonTemporalC: 0 - NumElementsPerBatchStore: 1 - NumElementsPerThread: 64 - NumGlobalWriteVectorsPerThread: 32 - NumLoadsA: 4 - NumLoadsB: 4 - NumLoadsCoalescedA: 1 - NumLoadsCoalescedB: 1 - NumLoadsPerpendicularA: 4 - NumLoadsPerpendicularB: 4 - NumThreads: 256 - OptNoLoadLoop: 1 - OptPreLoopVmcnt: 0 - PackBatchDims: 0 - PackFreeDims: 1 - PackGranularity: 2 - PackSummationDims: 0 - PackedC0IdxChars: [I] - PackedC0IndicesX: [0] - PackedC1IdxChars: [J] - PackedC1IndicesX: [1] - PerformanceSyncLocation: -1 - PerformanceWaitCount: -1 - PerformanceWaitLocation: -1 - PersistentKernel: 0 - PersistentKernelAlongBatch: false - PrefetchAcrossPersistent: 0 - PrefetchGlobalRead: 2 - PrefetchLocalRead: 1 - ProblemType: - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - ConvolutionConfig: [] - DataType: 1 - DestDataType: 1 - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseInitialStridesAB: false - UseInitialStridesCD: false - ZeroPadA: [] - ZeroPadB: [] - ReplacementKernel: false - ScheduleGlobalRead: 1 - ScheduleIterAlg: 3 - ScheduleLocalWrite: 1 - SolutionIndex: 454 - SolutionNameMin: Cijk_Ailk_Bjlk_DB_MT128x128x16_MI16x16x4x1_SN_EPS0_GRVW2_IU1_PGR2_PLR1_SU32_SUS128_TT2_128_WG64_4_1_WGM5 - SourceSwap: true - StaggerU: 32 - StaggerUMapping: 0 - StaggerUStride: 128 - StorePriorityOpt: 1 - StoreRemapVectorWidth: 0 - StoreSyncOpt: 6 - StoreVectorWidth: 2 - SubGroup0: 16 - SubGroup1: 16 - SubGroupA: 16 - SubGroupB: 16 - SuppressNoLoadLoop: false - ThreadTile: [2, 128] - ThreadTile0: 8 - ThreadTile1: 8 - ThreadTileA: 8 - ThreadTileB: 8 - TransposeLDS: 0 - UnrollIncIsDepthU: 0 - UnrollMajorLDSA: 0 - UnrollMajorLDSB: 0 - UnrollMemFence: false - Use64bShadowLimit: 1 - UseInstOffsetForGRO: 0 - UseSgprForGRO: -1 - Valid: true - VectorAtomicWidth: 1 - VectorStore: -1 - VectorWidth: 2 - WaveSeparateGlobalReadA: 0 - WaveSeparateGlobalReadB: 0 - WavefrontSize: 64 - WorkGroup: [64, 4, 1] - WorkGroupMapping: 5 - WorkGroupMappingType: B - _DepthULds: 16 - _GlobalAccumulation: null - _UseSgprForGRO: false - _VectorStore: 1 - _WorkspaceSizePerElemC: 0 - _staggerStrideShift: 0 -- [2, 3, 0, 1] -- - - [38144, 38144, 1, 256] - - [9, 0.0] - - - [29568, 128, 1, 384] - - [21, 0.0] - - - [30848, 128, 1, 256] - - [21, 0.0] - - - [25728, 128, 1, 384] - - [21, 0.0] - - - [32256, 32256, 1, 256] - - [9, 0.0] - - - [7680, 7680, 1, 256] - - [5, 0.0] - - - [41984, 41984, 1, 256] - - [8, 0.0] - - - [40448, 40448, 1, 256] - - [9, 0.0] - - - [25728, 128, 1, 256] - - [21, 0.0] - - - [64, 64, 1, 64] - - [0, 0.0] - - - [15104, 15104, 1, 256] - - [5, 0.0] - - - [17280, 17280, 1, 384] - - [5, 0.0] - - - [34688, 128, 1, 384] - - [21, 0.0] - - - [27392, 27392, 1, 256] - - [4, 0.0] - - - [6528, 128, 1, 256] - - [17, 0.0] - - - [35328, 35328, 1, 256] - - [5, 0.0] - - - [18432, 18432, 1, 384] - - [9, 0.0] - - - [31232, 31232, 1, 256] - - [9, 0.0] - - - [7808, 128, 1, 256] - - [21, 0.0] - - - [38400, 38400, 1, 384] - - [8, 0.0] - - - [16128, 16128, 1, 256] - - [5, 0.0] - - - [9472, 9472, 1, 256] - - [5, 0.0] - - - [21888, 21888, 1, 384] - - [8, 0.0] - - - [38656, 38656, 1, 256] - - [5, 0.0] - - - [20224, 20224, 1, 256] - - [5, 0.0] - - - [8960, 8960, 1, 256] - - [5, 0.0] - - - [29952, 29952, 1, 384] - - [8, 0.0] - - - [36864, 36864, 1, 384] - - [8, 0.0] - - - [33408, 33408, 1, 384] - - [8, 0.0] - - - [20608, 128, 1, 384] - - [21, 0.0] - - - [23424, 23424, 1, 384] - - [8, 0.0] - - - [4864, 4864, 1, 256] - - [5, 0.0] - - - [21504, 21504, 1, 384] - - [5, 0.0] - - - [25600, 25600, 1, 256] - - [9, 0.0] - - - [40960, 40960, 1, 256] - - [6, 0.0] - - - [19200, 19200, 1, 384] - - [5, 0.0] - - - [64, 1, 1, 64] - - [1, 0.0] - - - [25088, 25088, 1, 256] - - [9, 0.0] - - - [41728, 41728, 1, 256] - - [5, 0.0] - - - [35840, 35840, 1, 256] - - [8, 0.0] - - - [34560, 34560, 1, 256] - - [5, 0.0] - - - [26368, 26368, 1, 256] - - [5, 0.0] - - - [5888, 5888, 1, 256] - - [5, 0.0] - - - [28032, 28032, 1, 384] - - [4, 0.0] - - - [42496, 42496, 1, 256] - - [9, 0.0] - - - [27008, 128, 1, 256] - - [21, 0.0] - - - [38400, 38400, 1, 256] - - [9, 0.0] - - - [11008, 11008, 1, 256] - - [5, 0.0] - - - [32000, 32000, 1, 256] - - [9, 0.0] - - - [37248, 37248, 1, 384] - - [8, 0.0] - - - [10496, 10496, 1, 256] - - [5, 0.0] - - - [16640, 16640, 1, 256] - - [5, 0.0] - - - [24960, 24960, 1, 384] - - [5, 0.0] - - - [18688, 18688, 1, 256] - - [5, 0.0] - - - [22272, 22272, 1, 384] - - [8, 0.0] - - - [15488, 128, 1, 256] - - [21, 0.0] - - - [28416, 28416, 1, 384] - - [6, 0.0] - - - [3840, 3840, 1, 256] - - [4, 0.0] - - - [19968, 19968, 1, 384] - - [5, 0.0] - - - [43776, 43776, 1, 256] - - [5, 0.0] - - - [35072, 35072, 1, 256] - - [8, 0.0] - - - [20736, 20736, 1, 256] - - [5, 0.0] - - - [7168, 7168, 1, 256] - - [9, 0.0] - - - [18432, 18432, 1, 256] - - [9, 0.0] - - - [38016, 38016, 1, 384] - - [8, 0.0] - - - [35328, 35328, 1, 384] - - [8, 0.0] - - - [38784, 38784, 1, 384] - - [8, 0.0] - - - [26112, 26112, 1, 384] - - [6, 0.0] - - - [27264, 27264, 1, 384] - - [5, 0.0] - - - [44928, 44928, 1, 384] - - [8, 0.0] - - - [41088, 128, 1, 384] - - [18, 0.0] - - - [42368, 128, 1, 256] - - [21, 0.0] - - - [10752, 10752, 1, 256] - - [5, 0.0] - - - [9088, 128, 1, 384] - - [18, 0.0] - - - [17152, 17152, 1, 256] - - [5, 0.0] - - - [44928, 128, 1, 384] - - [18, 0.0] - - - [7808, 128, 1, 384] - - [21, 0.0] - - - [29184, 29184, 1, 256] - - [9, 0.0] - - - [11776, 11776, 1, 256] - - [9, 0.0] - - - [1, 64, 1, 64] - - [1, 0.0] - - - [27136, 27136, 1, 256] - - [8, 0.0] - - - [33408, 128, 1, 256] - - [21, 0.0] - - - [33792, 33792, 1, 384] - - [8, 0.0] - - - [43520, 43520, 1, 256] - - [8, 0.0] - - - [14592, 14592, 1, 384] - - [5, 0.0] - - - [41472, 41472, 1, 256] - - [6, 0.0] - - - [14080, 14080, 1, 256] - - [5, 0.0] - - - [34688, 128, 1, 256] - - [21, 0.0] - - - [16896, 16896, 1, 256] - - [9, 0.0] - - - [15744, 15744, 1, 384] - - [5, 0.0] - - - [28416, 28416, 1, 256] - - [5, 0.0] - - - [23808, 23808, 1, 256] - - [5, 0.0] - - - [27648, 27648, 1, 256] - - [9, 0.0] - - - [1152, 3072, 1, 384] - - [9, 0.0] - - - [21888, 128, 1, 256] - - [21, 0.0] - - - [34816, 34816, 1, 256] - - [9, 0.0] - - - [43776, 43776, 1, 384] - - [8, 0.0] - - - [36096, 36096, 1, 256] - - [5, 0.0] - - - [24320, 24320, 1, 256] - - [5, 0.0] - - - [12544, 12544, 1, 256] - - [5, 0.0] - - - [29184, 29184, 1, 384] - - [8, 0.0] - - - [29568, 29568, 1, 384] - - [8, 0.0] - - - [12928, 128, 1, 384] - - [21, 0.0] - - - [36480, 36480, 1, 384] - - [8, 0.0] - - - [30720, 30720, 1, 256] - - [9, 0.0] - - - [25728, 25728, 1, 384] - - [5, 0.0] - - - [34048, 34048, 1, 256] - - [5, 0.0] - - - [12928, 128, 1, 256] - - [21, 0.0] - - - [9728, 9728, 1, 256] - - [9, 0.0] - - - [128, 128, 1, 256] - - [20, 0.0] - - - [33024, 33024, 1, 256] - - [5, 0.0] - - - [15488, 128, 1, 384] - - [21, 0.0] - - - [39808, 128, 1, 384] - - [21, 0.0] - - - [18176, 18176, 1, 256] - - [5, 0.0] - - - [21504, 21504, 1, 256] - - [5, 0.0] - - - [16384, 16384, 1, 256] - - [5, 0.0] - - - [27008, 128, 1, 384] - - [18, 0.0] - - - [27904, 27904, 1, 256] - - [4, 0.0] - - - [24448, 128, 1, 384] - - [21, 0.0] - - - [35968, 128, 1, 384] - - [18, 0.0] - - - [37632, 37632, 1, 256] - - [5, 0.0] - - - [14848, 14848, 1, 256] - - [5, 0.0] - - - [23552, 23552, 1, 256] - - [9, 0.0] - - - [4608, 4608, 1, 50000] - - [11, 0.0] - - - [13056, 13056, 1, 256] - - [5, 0.0] - - - [38528, 128, 1, 256] - - [21, 0.0] - - - [19584, 19584, 1, 384] - - [5, 0.0] - - - [16768, 128, 1, 384] - - [18, 0.0] - - - [22784, 22784, 1, 256] - - [9, 0.0] - - - [44160, 44160, 1, 384] - - [8, 0.0] - - - [28160, 28160, 1, 256] - - [8, 0.0] - - - [14592, 14592, 1, 256] - - [5, 0.0] - - - [20992, 20992, 1, 256] - - [9, 0.0] - - - [41216, 41216, 1, 256] - - [5, 0.0] - - - [21760, 21760, 1, 256] - - [5, 0.0] - - - [25344, 25344, 1, 256] - - [9, 0.0] - - - [4608, 4608, 1, 256] - - [9, 0.0] - - - [2560, 2048, 1, 256] - - [13, 0.0] - - - [30464, 30464, 1, 256] - - [5, 0.0] - - - [19200, 19200, 1, 256] - - [5, 0.0] - - - [22272, 22272, 1, 256] - - [9, 0.0] - - - [29952, 29952, 1, 256] - - [5, 0.0] - - - [20480, 20480, 1, 256] - - [9, 0.0] - - - [17408, 17408, 1, 256] - - [9, 0.0] - - - [32768, 32768, 1, 256] - - [6, 0.0] - - - [18816, 18816, 1, 384] - - [6, 0.0] - - - [34944, 34944, 1, 384] - - [8, 0.0] - - - [18048, 18048, 1, 384] - - [7, 0.0] - - - [34560, 34560, 1, 384] - - [8, 0.0] - - - [9088, 128, 1, 256] - - [18, 0.0] - - - [24576, 24576, 1, 256] - - [9, 0.0] - - - [32128, 128, 1, 384] - - [21, 0.0] - - - [8448, 8448, 1, 256] - - [5, 0.0] - - - [42752, 42752, 1, 256] - - [4, 0.0] - - - [5376, 5376, 1, 256] - - [5, 0.0] - - - [18048, 128, 1, 256] - - [21, 0.0] - - - [3584, 3584, 1, 256] - - [5, 0.0] - - - [37120, 37120, 1, 256] - - [5, 0.0] - - - [39936, 39936, 1, 384] - - [8, 0.0] - - - [20736, 20736, 1, 384] - - [5, 0.0] - - - [35584, 35584, 1, 256] - - [5, 0.0] - - - [26112, 26112, 1, 256] - - [5, 0.0] - - - [16896, 16896, 1, 384] - - [9, 0.0] - - - [40704, 40704, 1, 384] - - [8, 0.0] - - - [33280, 33280, 1, 256] - - [9, 0.0] - - - [5632, 5632, 1, 256] - - [5, 0.0] - - - [19456, 19456, 1, 256] - - [9, 0.0] - - - [22016, 22016, 1, 256] - - [9, 0.0] - - - [14208, 128, 1, 256] - - [21, 0.0] - - - [13568, 13568, 1, 256] - - [5, 0.0] - - - [30848, 128, 1, 384] - - [21, 0.0] - - - [1408, 128, 1, 384] - - [21, 0.0] - - - [5760, 5760, 1, 5760] - - [12, 0.0] - - - [39936, 39936, 1, 256] - - [9, 0.0] - - - [1920, 3072, 1, 384] - - [3, 0.0] - - - [9984, 9984, 1, 256] - - [5, 0.0] - - - [2816, 2048, 1, 256] - - [14, 0.0] - - - [23168, 128, 1, 256] - - [21, 0.0] - - - [19968, 19968, 1, 256] - - [5, 0.0] - - - [44800, 44800, 1, 256] - - [5, 0.0] - - - [14976, 14976, 1, 384] - - [5, 0.0] - - - [35712, 35712, 1, 384] - - [8, 0.0] - - - [43008, 43008, 1, 384] - - [8, 0.0] - - - [41088, 41088, 1, 384] - - [8, 0.0] - - - [16128, 16128, 1, 384] - - [7, 0.0] - - - [5120, 5120, 1, 256] - - [9, 0.0] - - - [25856, 25856, 1, 256] - - [5, 0.0] - - - [12288, 12288, 1, 256] - - [5, 0.0] - - - [6400, 6400, 1, 256] - - [5, 0.0] - - - [2688, 128, 1, 256] - - [21, 0.0] - - - [11648, 128, 1, 256] - - [21, 0.0] - - - [43264, 43264, 1, 256] - - [9, 0.0] - - - [19712, 19712, 1, 256] - - [5, 0.0] - - - [34176, 34176, 1, 384] - - [8, 0.0] - - - [31104, 31104, 1, 384] - - [8, 0.0] - - - [36608, 36608, 1, 256] - - [5, 0.0] - - - [39808, 128, 1, 256] - - [21, 0.0] - - - [13824, 13824, 1, 384] - - [9, 0.0] - - - [42624, 42624, 1, 384] - - [8, 0.0] - - - [21120, 21120, 1, 384] - - [8, 0.0] - - - [23296, 23296, 1, 256] - - [5, 0.0] - - - [42240, 42240, 1, 256] - - [5, 0.0] - - - [33408, 128, 1, 384] - - [19, 0.0] - - - [43648, 128, 1, 256] - - [21, 0.0] - - - [19328, 128, 1, 384] - - [21, 0.0] - - - [33792, 33792, 1, 256] - - [5, 0.0] - - - [31488, 31488, 1, 256] - - [4, 0.0] - - - [768, 3072, 1, 384] - - [2, 0.0] - - - [6144, 6144, 1, 256] - - [9, 0.0] - - - [20352, 20352, 1, 384] - - [5, 0.0] - - - [23168, 128, 1, 384] - - [18, 0.0] - - - [33536, 33536, 1, 256] - - [4, 0.0] - - - [32640, 32640, 1, 384] - - [8, 0.0] - - - [1536, 3072, 1, 384] - - [9, 0.0] - - - [19328, 128, 1, 256] - - [21, 0.0] - - - [2688, 3072, 1, 384] - - [9, 0.0] - - - [24192, 24192, 1, 384] - - [8, 0.0] - - - [6912, 6912, 1, 256] - - [5, 0.0] - - - [15360, 15360, 1, 256] - - [5, 0.0] - - - [18944, 18944, 1, 256] - - [9, 0.0] - - - [37376, 37376, 1, 256] - - [8, 0.0] - - - [31488, 31488, 1, 384] - - [8, 0.0] - - - [26880, 26880, 1, 256] - - [5, 0.0] - - - [44928, 128, 1, 128] - - [22, 0.0] - - - [24448, 128, 1, 256] - - [21, 0.0] - - - [31872, 31872, 1, 384] - - [8, 0.0] - - - [1408, 128, 1, 256] - - [21, 0.0] - - - [38528, 128, 1, 384] - - [21, 0.0] - - - [15616, 15616, 1, 256] - - [5, 0.0] - - - [39552, 39552, 1, 384] - - [8, 0.0] - - - [4352, 4352, 1, 256] - - [5, 0.0] - - - [28288, 128, 1, 384] - - [21, 0.0] - - - [10368, 128, 1, 256] - - [22, 0.0] - - - [32128, 128, 1, 256] - - [21, 0.0] - - - [4608, 4608, 1, 4608] - - [11, 0.0] - - - [8704, 8704, 1, 256] - - [9, 0.0] - - - [17664, 17664, 1, 256] - - [5, 0.0] - - - [24576, 24576, 1, 384] - - [8, 0.0] - - - [37248, 128, 1, 384] - - [21, 0.0] - - - [34304, 34304, 1, 256] - - [6, 0.0] - - - [42368, 128, 1, 384] - - [19, 0.0] - - - [17664, 17664, 1, 384] - - [5, 0.0] - - - [12800, 12800, 1, 256] - - [9, 0.0] - - - [26624, 26624, 1, 256] - - [9, 0.0] - - - [36864, 36864, 1, 256] - - [8, 0.0] - - - [40704, 40704, 1, 256] - - [5, 0.0] - - - [12032, 12032, 1, 256] - - [5, 0.0] - - - [33024, 33024, 1, 384] - - [8, 0.0] - - - [28800, 28800, 1, 384] - - [5, 0.0] - - - [22656, 22656, 1, 384] - - [5, 0.0] - - - [41472, 41472, 1, 384] - - [6, 0.0] - - - [39680, 39680, 1, 256] - - [5, 0.0] - - - [44032, 44032, 1, 256] - - [8, 0.0] - - - [43392, 43392, 1, 384] - - [8, 0.0] - - - [42240, 42240, 1, 384] - - [8, 0.0] - - - [38912, 38912, 1, 256] - - [9, 0.0] - - - [23040, 23040, 1, 384] - - [8, 0.0] - - - [13312, 13312, 1, 256] - - [9, 0.0] - - - [128, 128, 1, 384] - - [23, 0.0] - - - [39168, 39168, 1, 256] - - [5, 0.0] - - - [25344, 25344, 1, 384] - - [5, 0.0] - - - [5248, 128, 1, 256] - - [16, 0.0] - - - [30208, 30208, 1, 256] - - [9, 0.0] - - - [40192, 40192, 1, 256] - - [5, 0.0] - - - [15872, 15872, 1, 256] - - [9, 0.0] - - - [44544, 44544, 1, 256] - - [9, 0.0] - - - [11520, 11520, 1, 256] - - [5, 0.0] - - - [15360, 15360, 1, 384] - - [9, 0.0] - - - [23040, 23040, 1, 256] - - [5, 0.0] - - - [26496, 26496, 1, 384] - - [8, 0.0] - - - [11264, 11264, 1, 256] - - [9, 0.0] - - - [18048, 128, 1, 384] - - [18, 0.0] - - - [30976, 30976, 1, 256] - - [4, 0.0] - - - [11648, 128, 1, 384] - - [15, 0.0] - - - [2304, 3072, 1, 384] - - [3, 0.0] - - - [28928, 28928, 1, 256] - - [9, 0.0] - - - [43008, 43008, 1, 256] - - [5, 0.0] - - - [29440, 29440, 1, 256] - - [5, 0.0] - - - [36352, 36352, 1, 256] - - [5, 0.0] - - - [32256, 32256, 1, 384] - - [8, 0.0] - - - [23808, 23808, 1, 384] - - [8, 0.0] - - - [37248, 128, 1, 256] - - [21, 0.0] - - - [1, 1, 1, 64] - - [1, 0.0] - - - [37888, 37888, 1, 256] - - [6, 0.0] - - - [35968, 128, 1, 256] - - [21, 0.0] - - - [13824, 13824, 1, 256] - - [9, 0.0] - - - [39168, 39168, 1, 384] - - [8, 0.0] - - - [37632, 37632, 1, 384] - - [8, 0.0] - - - [29568, 128, 1, 256] - - [21, 0.0] - - - [14336, 14336, 1, 256] - - [9, 0.0] - - - [28288, 128, 1, 256] - - [21, 0.0] - - - [16512, 16512, 1, 384] - - [5, 0.0] - - - [30720, 30720, 1, 384] - - [8, 0.0] - - - [21248, 21248, 1, 256] - - [5, 0.0] - - - [29696, 29696, 1, 256] - - [9, 0.0] - - - [384, 3072, 1, 384] - - [3, 0.0] - - - [28672, 28672, 1, 256] - - [5, 0.0] - - - [32512, 32512, 1, 256] - - [5, 0.0] - - - [9216, 9216, 1, 256] - - [9, 0.0] - - - [6656, 6656, 1, 256] - - [9, 0.0] - - - [30336, 30336, 1, 384] - - [5, 0.0] - - - [20608, 128, 1, 256] - - [21, 0.0] - - - [7936, 7936, 1, 256] - - [5, 0.0] - - - [41856, 41856, 1, 384] - - [8, 0.0] - - - [44288, 44288, 1, 256] - - [5, 0.0] - - - [7744, 7744, 1, 7744] - - [10, 0.0] - - - [7424, 7424, 1, 256] - - [5, 0.0] - - - [39424, 39424, 1, 256] - - [6, 0.0] - - - [43648, 128, 1, 384] - - [21, 0.0] - - - [14208, 14208, 1, 384] - - [5, 0.0] - - - [36096, 36096, 1, 384] - - [8, 0.0] - - - [44544, 44544, 1, 384] - - [8, 0.0] - - - [22528, 22528, 1, 256] - - [9, 0.0] - - - [4096, 4096, 1, 256] - - [5, 0.0] - - - [31744, 31744, 1, 256] - - [9, 0.0] - - - [3968, 128, 1, 384] - - [21, 0.0] - - - [17920, 17920, 1, 256] - - [9, 0.0] - - - [5248, 128, 1, 384] - - [16, 0.0] - - - [26880, 26880, 1, 384] - - [8, 0.0] - - - [8192, 8192, 1, 256] - - [9, 0.0] - - - [3968, 128, 1, 256] - - [21, 0.0] - - - [41088, 128, 1, 256] - - [21, 0.0] - - - [21888, 128, 1, 384] - - [21, 0.0] - - - [16768, 128, 1, 256] - - [21, 0.0] - - - [24064, 24064, 1, 256] - - [9, 0.0] - - - [44928, 128, 1, 256] - - [21, 0.0] - - - [27648, 27648, 1, 384] - - [6, 0.0] - - - [24832, 24832, 1, 256] - - [5, 0.0] - - - [10240, 10240, 1, 256] - - [9, 0.0] - - - [40320, 40320, 1, 384] - - [8, 0.0] - - - [18432, 2688, 1, 384] - - [27, 0.0] - - - [43008, 2304, 1, 384] - - [29, 0.0] - - - [3840, 3072, 1, 384] - - [27, 0.0] - - - [33408, 1920, 1, 384] - - [27, 0.0] - - - [33792, 2688, 1, 384] - - [27, 0.0] - - - [8064, 2688, 1, 384] - - [27, 0.0] - - - [33408, 2304, 1, 384] - - [27, 0.0] - - - [31872, 1536, 1, 384] - - [26, 0.0] - - - [41088, 1920, 1, 384] - - [27, 0.0] - - - [41088, 2304, 1, 384] - - [27, 0.0] - - - [5376, 1536, 1, 384] - - [30, 0.0] - - - [16128, 1536, 1, 384] - - [29, 0.0] - - - [36480, 2688, 1, 384] - - [27, 0.0] - - - [15360, 768, 1, 384] - - [27, 0.0] - - - [42624, 768, 1, 384] - - [29, 0.0] - - - [4992, 1536, 1, 384] - - [29, 0.0] - - - [29952, 1536, 1, 384] - - [29, 0.0] - - - [10752, 2688, 1, 384] - - [26, 0.0] - - - [42240, 2688, 1, 384] - - [27, 0.0] - - - [36096, 1536, 1, 384] - - [29, 0.0] - - - [26496, 1536, 1, 384] - - [27, 0.0] - - - [42624, 2688, 1, 384] - - [27, 0.0] - - - [17664, 2688, 1, 384] - - [27, 0.0] - - - [37248, 1536, 1, 384] - - [27, 0.0] - - - [16896, 2304, 1, 384] - - [27, 0.0] - - - [22272, 1920, 1, 384] - - [27, 0.0] - - - [26880, 2688, 1, 384] - - [27, 0.0] - - - [384, 768, 1, 384] - - [31, 0.0] - - - [16896, 1920, 1, 384] - - [27, 0.0] - - - [32640, 2304, 1, 384] - - [29, 0.0] - - - [5760, 2304, 1, 384] - - [29, 0.0] - - - [11904, 2304, 1, 384] - - [29, 0.0] - - - [24576, 2304, 1, 384] - - [29, 0.0] - - - [33024, 1536, 1, 384] - - [27, 0.0] - - - [36096, 2304, 1, 384] - - [29, 0.0] - - - [20352, 2688, 1, 384] - - [27, 0.0] - - - [14592, 2304, 1, 384] - - [29, 0.0] - - - [16128, 1920, 1, 384] - - [27, 0.0] - - - [16512, 1920, 1, 384] - - [27, 0.0] - - - [35712, 1920, 1, 384] - - [27, 0.0] - - - [9216, 2688, 1, 384] - - [29, 0.0] - - - [23808, 2688, 1, 384] - - [27, 0.0] - - - [18048, 768, 1, 384] - - [29, 0.0] - - - [14592, 2688, 1, 384] - - [27, 0.0] - - - [14208, 1920, 1, 384] - - [27, 0.0] - - - [14976, 2688, 1, 384] - - [27, 0.0] - - - [17280, 2304, 1, 384] - - [29, 0.0] - - - [11520, 2304, 1, 384] - - [27, 0.0] - - - [18432, 768, 1, 384] - - [26, 0.0] - - - [4608, 768, 1, 384] - - [29, 0.0] - - - [34944, 1920, 1, 384] - - [27, 0.0] - - - [13824, 2688, 1, 384] - - [27, 0.0] - - - [39936, 2304, 1, 384] - - [29, 0.0] - - - [7680, 2688, 1, 384] - - [26, 0.0] - - - [19968, 2304, 1, 384] - - [27, 0.0] - - - [27648, 2688, 1, 384] - - [27, 0.0] - - - [4224, 768, 1, 384] - - [32, 0.0] - - - [24192, 1920, 1, 384] - - [27, 0.0] - - - [32640, 1920, 1, 384] - - [26, 0.0] - - - [34176, 2688, 1, 384] - - [27, 0.0] - - - [35328, 1536, 1, 384] - - [27, 0.0] - - - [8832, 2688, 1, 384] - - [27, 0.0] - - - [18048, 1920, 1, 384] - - [27, 0.0] - - - [31488, 768, 1, 384] - - [29, 0.0] - - - [21504, 2304, 1, 384] - - [27, 0.0] - - - [11136, 2688, 1, 384] - - [27, 0.0] - - - [768, 1152, 1, 384] - - [25, 0.0] - - - [29184, 2688, 1, 384] - - [27, 0.0] - - - [4608, 2688, 1, 384] - - [27, 0.0] - - - [21504, 2688, 1, 384] - - [27, 0.0] - - - [34176, 768, 1, 384] - - [27, 0.0] - - - [23808, 1536, 1, 384] - - [27, 0.0] - - - [43392, 1536, 1, 384] - - [27, 0.0] - - - [13824, 768, 1, 384] - - [29, 0.0] - - - [38016, 1536, 1, 384] - - [29, 0.0] - - - [20736, 2688, 1, 384] - - [27, 0.0] - - - [15744, 1536, 1, 384] - - [27, 0.0] - - - [16512, 1536, 1, 384] - - [29, 0.0] - - - [3072, 2304, 1, 384] - - [28, 0.0] - - - [5760, 2688, 1, 384] - - [26, 0.0] - - - [38400, 2304, 1, 384] - - [27, 0.0] - - - [15360, 2688, 1, 384] - - [27, 0.0] - - - [29952, 2688, 1, 384] - - [27, 0.0] - - - [43008, 2688, 1, 384] - - [27, 0.0] - - - [13440, 1920, 1, 384] - - [27, 0.0] - - - [6528, 2688, 1, 384] - - [27, 0.0] - - - [2304, 1536, 1, 384] - - [24, 0.0] - - - [40320, 1536, 1, 384] - - [29, 0.0] - - - [13440, 1536, 1, 384] - - [29, 0.0] - - - [40320, 2688, 1, 384] - - [29, 0.0] - - - [30336, 2304, 1, 384] - - [27, 0.0] - - - [24192, 2688, 1, 384] - - [27, 0.0] - - - [35328, 768, 1, 384] - - [29, 0.0] - - - [23040, 768, 1, 384] - - [29, 0.0] - - - [29952, 2304, 1, 384] - - [27, 0.0] - - - [33024, 1920, 1, 384] - - [27, 0.0] - - - [14976, 768, 1, 384] - - [26, 0.0] - - - [42624, 1920, 1, 384] - - [26, 0.0] - - - [32640, 2688, 1, 384] - - [28, 0.0] - - - [11520, 1536, 1, 384] - - [29, 0.0] - - - [6912, 768, 1, 384] - - [29, 0.0] - - - [39552, 1920, 1, 384] - - [27, 0.0] - - - [32256, 1920, 1, 384] - - [29, 0.0] - - - [10752, 1536, 1, 384] - - [28, 0.0] - - - [24576, 2688, 1, 384] - - [27, 0.0] - - - [12672, 2688, 1, 384] - - [27, 0.0] - - - [10752, 1920, 1, 384] - - [27, 0.0] - - - [40704, 1536, 1, 384] - - [29, 0.0] - - - [32256, 768, 1, 384] - - [28, 0.0] - - - [18816, 2688, 1, 384] - - [29, 0.0] - - - [11520, 2688, 1, 384] - - [27, 0.0] - - - [35712, 2688, 1, 384] - - [27, 0.0] - - - [29952, 1920, 1, 384] - - [27, 0.0] - - - [26880, 1920, 1, 384] - - [27, 0.0] - - - [33408, 2688, 1, 384] - - [27, 0.0] - - - [35328, 2688, 1, 384] - - [27, 0.0] - - - [21120, 2688, 1, 384] - - [27, 0.0] - - - [19584, 1920, 1, 384] - - [27, 0.0] - - - [17664, 1536, 1, 384] - - [27, 0.0] - - - [36864, 768, 1, 384] - - [29, 0.0] - - - [14592, 1536, 1, 384] - - [28, 0.0] - - - [11136, 2304, 1, 384] - - [27, 0.0] - - - [9600, 2688, 1, 384] - - [27, 0.0] - - - [9216, 2304, 1, 384] - - [29, 0.0] - - - [21120, 768, 1, 384] - - [29, 0.0] - - - [4992, 2688, 1, 384] - - [27, 0.0] - - - [41472, 768, 1, 384] - - [29, 0.0] - - - [37632, 1536, 1, 384] - - [26, 0.0] - - - [38784, 2304, 1, 384] - - [27, 0.0] - - - [8448, 2688, 1, 384] - - [27, 0.0] - - - [36864, 2304, 1, 384] - - [27, 0.0] - - - [40704, 1920, 1, 384] - - [27, 0.0] - - - [39552, 2688, 1, 384] - - [27, 0.0] - - - [26112, 768, 1, 384] - - [27, 0.0] - - - [29184, 1536, 1, 384] - - [29, 0.0] - - - [32640, 1536, 1, 384] - - [27, 0.0] - - - [5376, 2688, 1, 384] - - [27, 0.0] - - - [13056, 768, 1, 384] - - [29, 0.0] - - - [13824, 2304, 1, 384] - - [29, 0.0] - - - [16896, 768, 1, 384] - - [29, 0.0] - - - [30336, 1920, 1, 384] - - [27, 0.0] - - - [27264, 2304, 1, 384] - - [27, 0.0] - - - [7680, 1536, 1, 384] - - [28, 0.0] - - - [30720, 2688, 1, 384] - - [27, 0.0] - - - [36096, 2688, 1, 384] - - [27, 0.0] - - - [5760, 1920, 1, 384] - - [27, 0.0] - - - [42240, 1536, 1, 384] - - [29, 0.0] - - - [8448, 1920, 1, 384] - - [27, 0.0] - - - [32256, 1536, 1, 384] - - [29, 0.0] - - - [44160, 2304, 1, 384] - - [29, 0.0] - - - [30336, 2688, 1, 384] - - [27, 0.0] - - - [6144, 2688, 1, 384] - - [27, 0.0] - - - [39168, 1536, 1, 384] - - [29, 0.0] - - - [11904, 1920, 1, 384] - - [27, 0.0] - - - [8064, 1536, 1, 384] - - [28, 0.0] - - - [21120, 1920, 1, 384] - - [27, 0.0] - - - [22656, 2304, 1, 384] - - [27, 0.0] - - - [19968, 2688, 1, 384] - - [27, 0.0] - - - [10752, 768, 1, 384] - - [28, 0.0] - - - [18432, 2304, 1, 384] - - [29, 0.0] - - - [14976, 1920, 1, 384] - - [27, 0.0] - - - [33024, 2688, 1, 384] - - [27, 0.0] - - - [1536, 768, 1, 384] - - [29, 0.0] - - - [33024, 2304, 1, 384] - - [27, 0.0] - - - [14208, 2688, 1, 384] - - [27, 0.0] - - - [38016, 2304, 1, 384] - - [27, 0.0] - - - [16896, 2688, 1, 384] - - [27, 0.0] - - - [31104, 768, 1, 384] - - [27, 0.0] - - - [41472, 2304, 1, 384] - - [27, 0.0] - - - [23424, 2688, 1, 384] - - [27, 0.0] - - - [26496, 2688, 1, 384] - - [27, 0.0] - - - [16512, 2304, 1, 384] - - [27, 0.0] - - - [11520, 1920, 1, 384] - - [27, 0.0] - - - [39552, 768, 1, 384] - - [27, 0.0] - - - [6144, 2304, 1, 384] - - [26, 0.0] - - - [14208, 2304, 1, 384] - - [27, 0.0] - - - [19584, 2304, 1, 384] - - [27, 0.0] - - - [36480, 768, 1, 384] - - [26, 0.0] - - - [15744, 2688, 1, 384] - - [27, 0.0] - - - [34560, 1536, 1, 384] - - [27, 0.0] - - - [8448, 2304, 1, 384] - - [27, 0.0] - - - [26112, 2688, 1, 384] - - [27, 0.0] - - - [39936, 768, 1, 384] - - [27, 0.0] - - - [19200, 1920, 1, 384] - - [29, 0.0] - - - [38400, 768, 1, 384] - - [29, 0.0] - - - [8448, 1536, 1, 384] - - [29, 0.0] - - - [13824, 1536, 1, 384] - - [29, 0.0] - - - [9600, 768, 1, 384] - - [29, 0.0] - - - [10368, 768, 1, 384] - - [27, 0.0] - - - [20736, 1536, 1, 384] - - [29, 0.0] - - - [28800, 768, 1, 384] - - [27, 0.0] - - - [10368, 1536, 1, 384] - - [26, 0.0] - - - [21888, 1536, 1, 384] - - [26, 0.0] - - - [38784, 2688, 1, 384] - - [26, 0.0] - - - [27648, 2304, 1, 384] - - [29, 0.0] - - - [11136, 1920, 1, 384] - - [27, 0.0] - - - [37248, 768, 1, 384] - - [29, 0.0] - - - [23040, 2688, 1, 384] - - [27, 0.0] - - - [37632, 1920, 1, 384] - - [27, 0.0] - - - [7680, 768, 1, 384] - - [27, 0.0] - - - [38016, 1920, 1, 384] - - [27, 0.0] - - - [35712, 2304, 1, 384] - - [27, 0.0] - - - [37248, 2688, 1, 384] - - [27, 0.0] - - - [29568, 1920, 1, 384] - - [27, 0.0] - - - [38400, 2688, 1, 384] - - [27, 0.0] - - - [25728, 768, 1, 384] - - [26, 0.0] - - - [8832, 1920, 1, 384] - - [26, 0.0] - - - [43776, 1920, 1, 384] - - [27, 0.0] - - - [15744, 768, 1, 384] - - [29, 0.0] - - - [27264, 1920, 1, 384] - - [27, 0.0] - - - [33792, 2304, 1, 384] - - [27, 0.0] - - - [8832, 2304, 1, 384] - - [29, 0.0] - - - [39168, 2688, 1, 384] - - [27, 0.0] - - - [35328, 1920, 1, 384] - - [27, 0.0] - - - [35328, 2304, 1, 384] - - [29, 0.0] - - - [29184, 768, 1, 384] - - [27, 0.0] - - - [18048, 2688, 1, 384] - - [27, 0.0] - - - [32256, 2688, 1, 384] - - [27, 0.0] - - - [18816, 1536, 1, 384] - - [26, 0.0] - - - [13056, 1536, 1, 384] - - [29, 0.0] - - - [34944, 1536, 1, 384] - - [27, 0.0] - - - [38400, 1920, 1, 384] - - [27, 0.0] - - - [15360, 2304, 1, 384] - - [27, 0.0] - - - [27264, 2688, 1, 384] - - [27, 0.0] - - - [11136, 1536, 1, 384] - - [29, 0.0] - - - [30720, 2304, 1, 384] - - [27, 0.0] - - - [24960, 2688, 1, 384] - - [27, 0.0] - - - [13824, 1920, 1, 384] - - [27, 0.0] - - - [17280, 2688, 1, 384] - - [27, 0.0] - - - [31872, 768, 1, 384] - - [27, 0.0] - - - [11904, 2688, 1, 384] - - [27, 0.0] - - - [7296, 768, 1, 384] - - [26, 0.0] - - - [19200, 1536, 1, 384] - - [29, 0.0] - - - [12288, 768, 1, 384] - - [29, 0.0] - - - [33792, 768, 1, 384] - - [27, 0.0] - - - [21888, 2688, 1, 384] - - [27, 0.0] - - - [2688, 1920, 1, 384] - - [33, 0.0] - - - [19968, 768, 1, 384] - - [29, 0.0] - - - [12288, 2688, 1, 384] - - [26, 0.0] - - - [12288, 2304, 1, 384] - - [26, 0.0] - - - [28416, 768, 1, 384] - - [27, 0.0] - - - [34560, 768, 1, 384] - - [27, 0.0] - - - [39936, 2688, 1, 384] - - [27, 0.0] - - - [8064, 1920, 1, 384] - - [26, 0.0] - - - [26880, 1536, 1, 384] - - [27, 0.0] - - - [28032, 2688, 1, 384] - - [27, 0.0] - - - [41472, 2688, 1, 384] - - [27, 0.0] - - - [29568, 2688, 1, 384] - - [27, 0.0] - - - [31104, 2688, 1, 384] - - [27, 0.0] - - - [5376, 1920, 1, 384] - - [26, 0.0] - - - [41856, 2688, 1, 384] - - [27, 0.0] - - - [9984, 768, 1, 384] - - [29, 0.0] - - - [3456, 2688, 1, 384] - - [29, 0.0] - - - [43392, 2688, 1, 384] - - [27, 0.0] - - - [36480, 1920, 1, 384] - - [27, 0.0] - - - [29568, 1536, 1, 384] - - [29, 0.0] - - - [36864, 2688, 1, 384] - - [27, 0.0] - - - [12672, 768, 1, 384] - - [29, 0.0] - - - [24064, 3072, 1, 256] - - [53, 0.0] - - - [256, 512, 1, 256] - - [54, 0.0] - - - [40960, 27648, 1, 256] - - [53, 0.0] - - - [31744, 3072, 1, 256] - - [53, 0.0] - - - [13056, 1792, 1, 256] - - [47, 0.0] - - - [35328, 22785, 1, 256] - - [47, 0.0] - - - [28160, 15872, 1, 256] - - [53, 0.0] - - - [39168, 1792, 1, 256] - - [47, 0.0] - - - [23808, 11265, 1, 256] - - [53, 0.0] - - - [16640, 4353, 1, 256] - - [47, 0.0] - - - [38912, 26624, 1, 256] - - [53, 0.0] - - - [6912, 3585, 1, 256] - - [53, 0.0] - - - [32768, 1792, 1, 256] - - [47, 0.0] - - - [30976, 18688, 1, 256] - - [47, 0.0] - - - [512, 2048, 1, 256] - - [41, 0.0] - - - [15872, 3584, 1, 256] - - [53, 0.0] - - - [6400, 1792, 1, 256] - - [53, 0.0] - - - [39680, 27393, 1, 256] - - [47, 0.0] - - - [36864, 24577, 1, 256] - - [53, 0.0] - - - [26112, 1536, 1, 256] - - [53, 0.0] - - - [26368, 1536, 1, 256] - - [53, 0.0] - - - [16896, 4353, 1, 256] - - [47, 0.0] - - - [14336, 1793, 1, 256] - - [47, 0.0] - - - [3840, 3072, 1, 256] - - [53, 0.0] - - - [2560, 3072, 1, 256] - - [59, 0.0] - - - [6656, 1536, 1, 256] - - [69, 0.0] - - - [27136, 1792, 1, 256] - - [47, 0.0] - - - [43776, 3072, 1, 256] - - [53, 0.0] - - - [23296, 1792, 1, 256] - - [47, 0.0] - - - [11264, 7937, 1, 256] - - [47, 0.0] - - - [768, 3072, 1, 256] - - [47, 0.0] - - - [6912, 3841, 1, 256] - - [47, 0.0] - - - [40960, 769, 1, 256] - - [47, 0.0] - - - [40448, 9216, 1, 256] - - [53, 0.0] - - - [7680, 4353, 1, 256] - - [47, 0.0] - - - [23296, 3072, 1, 256] - - [53, 0.0] - - - [7936, 4609, 1, 256] - - [53, 0.0] - - - [20736, 8448, 1, 256] - - [47, 0.0] - - - [768, 1024, 1, 256] - - [61, 0.0] - - - [38656, 3072, 1, 256] - - [53, 0.0] - - - [28160, 1792, 1, 256] - - [47, 0.0] - - - [13824, 3072, 1, 256] - - [53, 0.0] - - - [42752, 1792, 1, 256] - - [47, 0.0] - - - [35584, 23041, 1, 256] - - [53, 0.0] - - - [13056, 3072, 1, 256] - - [53, 0.0] - - - [37888, 768, 1, 256] - - [47, 0.0] - - - [19456, 3072, 1, 256] - - [53, 0.0] - - - [15872, 9216, 1, 256] - - [53, 0.0] - - - [30976, 1792, 1, 256] - - [47, 0.0] - - - [26368, 14081, 1, 256] - - [47, 0.0] - - - [35328, 23041, 1, 256] - - [53, 0.0] - - - [27648, 15105, 1, 256] - - [47, 0.0] - - - [25856, 13568, 1, 256] - - [47, 0.0] - - - [23296, 9216, 1, 256] - - [53, 0.0] - - - [2048, 1024, 1, 256] - - [39, 0.0] - - - [12032, 1792, 1, 256] - - [47, 0.0] - - - [11520, 1536, 1, 256] - - [53, 0.0] - - - [16128, 768, 1, 256] - - [47, 0.0] - - - [15360, 3072, 1, 256] - - [53, 0.0] - - - [38912, 26369, 1, 256] - - [47, 0.0] - - - [25344, 13056, 1, 256] - - [47, 0.0] - - - [39168, 26880, 1, 256] - - [47, 0.0] - - - [39424, 768, 1, 256] - - [47, 0.0] - - - [10496, 1792, 1, 256] - - [47, 0.0] - - - [28672, 3072, 1, 256] - - [53, 0.0] - - - [27392, 768, 1, 256] - - [47, 0.0] - - - [39680, 768, 1, 256] - - [47, 0.0] - - - [11520, 8193, 1, 256] - - [53, 0.0] - - - [17408, 4865, 1, 256] - - [47, 0.0] - - - [14080, 1537, 1, 256] - - [53, 0.0] - - - [29184, 768, 1, 256] - - [53, 0.0] - - - [19200, 6913, 1, 256] - - [47, 0.0] - - - [33536, 9216, 1, 256] - - [53, 0.0] - - - [5632, 3072, 1, 256] - - [53, 0.0] - - - [32768, 20480, 1, 256] - - [53, 0.0] - - - [29440, 9216, 1, 256] - - [53, 0.0] - - - [40960, 1792, 1, 256] - - [47, 0.0] - - - [10240, 3072, 1, 256] - - [53, 0.0] - - - [20992, 1792, 1, 256] - - [47, 0.0] - - - [42240, 9216, 1, 256] - - [53, 0.0] - - - [19200, 6912, 1, 256] - - [47, 0.0] - - - [27392, 1792, 1, 256] - - [47, 0.0] - - - [42496, 1536, 1, 256] - - [53, 0.0] - - - [29440, 16897, 1, 256] - - [53, 0.0] - - - [20480, 8192, 1, 256] - - [53, 0.0] - - - [11264, 8193, 1, 256] - - [53, 0.0] - - - [26880, 14337, 1, 256] - - [53, 0.0] - - - [28928, 16641, 1, 256] - - [47, 0.0] - - - [15360, 2817, 1, 256] - - [47, 0.0] - - - [44288, 1536, 1, 256] - - [53, 0.0] - - - [7936, 1536, 1, 256] - - [53, 0.0] - - - [18176, 5633, 1, 256] - - [53, 0.0] - - - [8448, 3072, 1, 256] - - [53, 0.0] - - - [17920, 5632, 1, 256] - - [53, 0.0] - - - [1792, 2048, 1, 256] - - [64, 0.0] - - - [39936, 3072, 1, 256] - - [53, 0.0] - - - [20480, 3072, 1, 256] - - [53, 0.0] - - - [24832, 1792, 1, 256] - - [47, 0.0] - - - [37376, 25088, 1, 256] - - [53, 0.0] - - - [7168, 4097, 1, 256] - - [53, 0.0] - - - [21504, 768, 1, 256] - - [47, 0.0] - - - [13312, 3072, 1, 256] - - [53, 0.0] - - - [40960, 1025, 1, 256] - - [50, 0.0] - - - [12032, 1536, 1, 256] - - [53, 0.0] - - - [9216, 768, 1, 256] - - [53, 0.0] - - - [44288, 27648, 1, 256] - - [53, 0.0] - - - [32512, 1792, 1, 256] - - [47, 0.0] - - - [23808, 11520, 1, 256] - - [47, 0.0] - - - [25600, 13057, 1, 256] - - [47, 0.0] - - - [40448, 1792, 1, 256] - - [47, 0.0] - - - [25088, 12800, 1, 256] - - [53, 0.0] - - - [22784, 10496, 1, 256] - - [47, 0.0] - - - [38400, 26113, 1, 256] - - [53, 0.0] - - - [9728, 3072, 1, 256] - - [53, 0.0] - - - [20736, 1792, 1, 256] - - [47, 0.0] - - - [7680, 3072, 1, 256] - - [53, 0.0] - - - [5376, 2305, 1, 256] - - [47, 0.0] - - - [12800, 3072, 1, 256] - - [53, 0.0] - - - [43520, 3584, 1, 256] - - [53, 0.0] - - - [12288, 3072, 1, 256] - - [53, 0.0] - - - [12800, 1536, 1, 256] - - [53, 0.0] - - - [21504, 8961, 1, 256] - - [47, 0.0] - - - [39680, 9216, 1, 256] - - [53, 0.0] - - - [3584, 513, 1, 256] - - [39, 0.0] - - - [1280, 3072, 1, 256] - - [59, 0.0] - - - [13056, 9216, 1, 256] - - [53, 0.0] - - - [22016, 768, 1, 256] - - [53, 0.0] - - - [33024, 1536, 1, 256] - - [53, 0.0] - - - [26880, 9216, 1, 256] - - [53, 0.0] - - - [44032, 27648, 1, 256] - - [53, 0.0] - - - [7680, 768, 1, 256] - - [47, 0.0] - - - [32000, 19712, 1, 256] - - [47, 0.0] - - - [26880, 14593, 1, 256] - - [47, 0.0] - - - [24064, 9216, 1, 256] - - [53, 0.0] - - - [39424, 26881, 1, 256] - - [47, 0.0] - - - [27392, 3072, 1, 256] - - [53, 0.0] - - - [10752, 1792, 1, 256] - - [47, 0.0] - - - [8960, 5633, 1, 256] - - [53, 0.0] - - - [34560, 3072, 1, 256] - - [53, 0.0] - - - [23808, 9216, 1, 256] - - [53, 0.0] - - - [29696, 17153, 1, 256] - - [47, 0.0] - - - [11776, 1536, 1, 256] - - [53, 0.0] - - - [13568, 1536, 1, 256] - - [53, 0.0] - - - [30208, 9216, 1, 256] - - [53, 0.0] - - - [36608, 1536, 1, 256] - - [53, 0.0] - - - [12800, 513, 1, 256] - - [50, 0.0] - - - [7680, 1792, 1, 256] - - [47, 0.0] - - - [42496, 2305, 1, 256] - - [47, 0.0] - - - [37376, 1536, 1, 256] - - [53, 0.0] - - - [20224, 1792, 1, 256] - - [47, 0.0] - - - [43520, 1536, 1, 256] - - [53, 0.0] - - - [26368, 768, 1, 256] - - [53, 0.0] - - - [18176, 3072, 1, 256] - - [53, 0.0] - - - [24320, 12033, 1, 256] - - [47, 0.0] - - - [17408, 9216, 1, 256] - - [53, 0.0] - - - [36352, 1792, 1, 256] - - [47, 0.0] - - - [20992, 8705, 1, 256] - - [53, 0.0] - - - [19712, 7424, 1, 256] - - [47, 0.0] - - - [38144, 768, 1, 256] - - [47, 0.0] - - - [10752, 1536, 1, 256] - - [53, 0.0] - - - [4096, 3072, 1, 256] - - [53, 0.0] - - - [29696, 17409, 1, 256] - - [53, 0.0] - - - [10240, 6913, 1, 256] - - [47, 0.0] - - - [18944, 1536, 1, 256] - - [53, 0.0] - - - [38656, 26113, 1, 256] - - [53, 0.0] - - - [37376, 25089, 1, 256] - - [53, 0.0] - - - [38400, 1536, 1, 256] - - [53, 0.0] - - - [8448, 1792, 1, 256] - - [47, 0.0] - - - [13056, 769, 1, 256] - - [53, 0.0] - - - [24320, 11777, 1, 256] - - [53, 0.0] - - - [17664, 9216, 1, 256] - - [53, 0.0] - - - [8192, 4865, 1, 256] - - [47, 0.0] - - - [17920, 1792, 1, 256] - - [47, 0.0] - - - [32000, 19713, 1, 256] - - [47, 0.0] - - - [8960, 768, 1, 256] - - [47, 0.0] - - - [31232, 3072, 1, 256] - - [53, 0.0] - - - [12544, 257, 1, 256] - - [49, 0.0] - - - [43776, 3585, 1, 256] - - [53, 0.0] - - - [11008, 1792, 1, 256] - - [47, 0.0] - - - [29696, 17408, 1, 256] - - [53, 0.0] - - - [34560, 22272, 1, 256] - - [47, 0.0] - - - [256, 2048, 1, 256] - - [56, 0.0] - - - [32768, 20481, 1, 256] - - [53, 0.0] - - - [14336, 3072, 1, 256] - - [53, 0.0] - - - [19456, 7168, 1, 256] - - [53, 0.0] - - - [13312, 9216, 1, 256] - - [53, 0.0] - - - [22272, 768, 1, 256] - - [47, 0.0] - - - [24064, 1792, 1, 256] - - [47, 0.0] - - - [16896, 1792, 1, 256] - - [47, 0.0] - - - [27904, 15616, 1, 256] - - [47, 0.0] - - - [37888, 3072, 1, 256] - - [53, 0.0] - - - [13056, 513, 1, 256] - - [49, 0.0] - - - [36608, 24065, 1, 256] - - [53, 0.0] - - - [40704, 3072, 1, 256] - - [53, 0.0] - - - [28928, 16640, 1, 256] - - [47, 0.0] - - - [24576, 12288, 1, 256] - - [53, 0.0] - - - [17152, 3072, 1, 256] - - [53, 0.0] - - - [17152, 4864, 1, 256] - - [47, 0.0] - - - [42496, 9216, 1, 256] - - [53, 0.0] - - - [32256, 768, 1, 256] - - [47, 0.0] - - - [4352, 1792, 1, 256] - - [47, 0.0] - - - [5632, 768, 1, 256] - - [69, 0.0] - - - [40704, 513, 1, 256] - - [49, 0.0] - - - [19712, 768, 1, 256] - - [47, 0.0] - - - [33536, 20993, 1, 256] - - [53, 0.0] - - - [2816, 3072, 1, 256] - - [59, 0.0] - - - [3584, 3072, 1, 256] - - [53, 0.0] - - - [4608, 1537, 1, 256] - - [53, 0.0] - - - [44032, 9216, 1, 256] - - [53, 0.0] - - - [33792, 21249, 1, 256] - - [47, 0.0] - - - [32512, 20225, 1, 256] - - [47, 0.0] - - - [38656, 9216, 1, 256] - - [53, 0.0] - - - [17664, 5377, 1, 256] - - [47, 0.0] - - - [19456, 7169, 1, 256] - - [53, 0.0] - - - [8448, 5121, 1, 256] - - [53, 0.0] - - - [29440, 17152, 1, 256] - - [47, 0.0] - - - [40448, 513, 1, 256] - - [50, 0.0] - - - [41472, 1792, 1, 256] - - [47, 0.0] - - - [17920, 3072, 1, 256] - - [53, 0.0] - - - [35072, 9216, 1, 256] - - [53, 0.0] - - - [34816, 22273, 1, 256] - - [47, 0.0] - - - [35072, 22785, 1, 256] - - [47, 0.0] - - - [39168, 9216, 1, 256] - - [53, 0.0] - - - [42752, 2817, 1, 256] - - [47, 0.0] - - - [11776, 3072, 1, 256] - - [53, 0.0] - - - [24832, 12289, 1, 256] - - [53, 0.0] - - - [24576, 12033, 1, 256] - - [47, 0.0] - - - [6400, 1536, 1, 256] - - [53, 0.0] - - - [32512, 3072, 1, 256] - - [53, 0.0] - - - [30976, 3072, 1, 256] - - [53, 0.0] - - - [22016, 9473, 1, 256] - - [47, 0.0] - - - [19968, 1792, 1, 256] - - [47, 0.0] - - - [29440, 3072, 1, 256] - - [53, 0.0] - - - [43776, 3840, 1, 256] - - [47, 0.0] - - - [41472, 768, 1, 256] - - [47, 0.0] - - - [8192, 1792, 1, 256] - - [47, 0.0] - - - [35840, 3072, 1, 256] - - [53, 0.0] - - - [8704, 3072, 1, 256] - - [53, 0.0] - - - [9728, 1792, 1, 256] - - [47, 0.0] - - - [22272, 9729, 1, 256] - - [53, 0.0] - - - [32768, 3072, 1, 256] - - [53, 0.0] - - - [3072, 2048, 1, 256] - - [53, 0.0] - - - [36864, 24576, 1, 256] - - [53, 0.0] - - - [9984, 1536, 1, 256] - - [53, 0.0] - - - [12032, 8961, 1, 256] - - [47, 0.0] - - - [38400, 25857, 1, 256] - - [47, 0.0] - - - [20224, 7937, 1, 256] - - [47, 0.0] - - - [34304, 21761, 1, 256] - - [47, 0.0] - - - [30720, 18432, 1, 256] - - [53, 0.0] - - - [31744, 9216, 1, 256] - - [53, 0.0] - - - [27136, 14848, 1, 256] - - [53, 0.0] - - - [34048, 9216, 1, 256] - - [53, 0.0] - - - [3584, 257, 1, 256] - - [37, 0.0] - - - [18688, 6145, 1, 256] - - [53, 0.0] - - - [36096, 768, 1, 256] - - [47, 0.0] - - - [36608, 9216, 1, 256] - - [53, 0.0] - - - [35584, 9216, 1, 256] - - [53, 0.0] - - - [29952, 17664, 1, 256] - - [47, 0.0] - - - [34816, 1792, 1, 256] - - [47, 0.0] - - - [24064, 11776, 1, 256] - - [53, 0.0] - - - [40448, 3072, 1, 256] - - [53, 0.0] - - - [18688, 6401, 1, 256] - - [47, 0.0] - - - [20480, 1536, 1, 256] - - [53, 0.0] - - - [18432, 3072, 1, 256] - - [53, 0.0] - - - [20224, 768, 1, 256] - - [53, 0.0] - - - [25344, 768, 1, 256] - - [47, 0.0] - - - [36608, 24320, 1, 256] - - [47, 0.0] - - - [34816, 9216, 1, 256] - - [53, 0.0] - - - [41216, 27648, 1, 256] - - [53, 0.0] - - - [30464, 9216, 1, 256] - - [53, 0.0] - - - [7424, 3072, 1, 256] - - [53, 0.0] - - - [20480, 1792, 1, 256] - - [47, 0.0] - - - [41984, 1793, 1, 256] - - [47, 0.0] - - - [18688, 1792, 1, 256] - - [47, 0.0] - - - [13824, 1792, 1, 256] - - [47, 0.0] - - - [38144, 3072, 1, 256] - - [53, 0.0] - - - [33280, 3072, 1, 256] - - [53, 0.0] - - - [35584, 23296, 1, 256] - - [47, 0.0] - - - [43520, 768, 1, 256] - - [47, 0.0] - - - [40704, 1536, 1, 256] - - [53, 0.0] - - - [29696, 3072, 1, 256] - - [53, 0.0] - - - [32256, 19969, 1, 256] - - [53, 0.0] - - - [40960, 9216, 1, 256] - - [53, 0.0] - - - [37632, 9216, 1, 256] - - [53, 0.0] - - - [42240, 2305, 1, 256] - - [47, 0.0] - - - [17920, 5377, 1, 256] - - [47, 0.0] - - - [27904, 9216, 1, 256] - - [53, 0.0] - - - [34304, 22016, 1, 256] - - [53, 0.0] - - - [11776, 8705, 1, 256] - - [53, 0.0] - - - [22272, 1536, 1, 256] - - [53, 0.0] - - - [25856, 9216, 1, 256] - - [53, 0.0] - - - [19712, 3072, 1, 256] - - [53, 0.0] - - - [41472, 9216, 1, 256] - - [53, 0.0] - - - [42496, 27648, 1, 256] - - [53, 0.0] - - - [44288, 4352, 1, 256] - - [47, 0.0] - - - [42496, 2561, 1, 256] - - [53, 0.0] - - - [9984, 6657, 1, 256] - - [53, 0.0] - - - [43008, 3073, 1, 256] - - [53, 0.0] - - - [36352, 24065, 1, 256] - - [53, 0.0] - - - [24832, 3072, 1, 256] - - [53, 0.0] - - - [29184, 16641, 1, 256] - - [47, 0.0] - - - [1024, 2048, 1, 256] - - [57, 0.0] - - - [42240, 27648, 1, 256] - - [53, 0.0] - - - [9984, 1792, 1, 256] - - [47, 0.0] - - - [44288, 3072, 1, 256] - - [53, 0.0] - - - [11008, 768, 1, 256] - - [47, 0.0] - - - [28672, 16129, 1, 256] - - [47, 0.0] - - - [17920, 9216, 1, 256] - - [53, 0.0] - - - [25088, 12801, 1, 256] - - [53, 0.0] - - - [19712, 9216, 1, 256] - - [53, 0.0] - - - [31744, 19457, 1, 256] - - [53, 0.0] - - - [36864, 1792, 1, 256] - - [47, 0.0] - - - [42496, 1792, 1, 256] - - [47, 0.0] - - - [39936, 9216, 1, 256] - - [53, 0.0] - - - [8960, 1792, 1, 256] - - [47, 0.0] - - - [17664, 5121, 1, 256] - - [53, 0.0] - - - [38144, 25601, 1, 256] - - [53, 0.0] - - - [27136, 14849, 1, 256] - - [53, 0.0] - - - [31744, 19456, 1, 256] - - [53, 0.0] - - - [33024, 3072, 1, 256] - - [53, 0.0] - - - [37888, 9216, 1, 256] - - [53, 0.0] - - - [6912, 1792, 1, 256] - - [47, 0.0] - - - [42240, 2049, 1, 256] - - [53, 0.0] - - - [34048, 3072, 1, 256] - - [53, 0.0] - - - [37120, 9216, 1, 256] - - [53, 0.0] - - - [14080, 9216, 1, 256] - - [53, 0.0] - - - [38400, 1792, 1, 256] - - [47, 0.0] - - - [43776, 9216, 1, 256] - - [53, 0.0] - - - [14336, 2049, 1, 256] - - [53, 0.0] - - - [37120, 24577, 1, 256] - - [53, 0.0] - - - [30976, 18433, 1, 256] - - [53, 0.0] - - - [37632, 3072, 1, 256] - - [53, 0.0] - - - [34560, 1792, 1, 256] - - [47, 0.0] - - - [5120, 3072, 1, 256] - - [53, 0.0] - - - [21760, 9217, 1, 256] - - [53, 0.0] - - - [24064, 11521, 1, 256] - - [47, 0.0] - - - [7936, 3072, 1, 256] - - [53, 0.0] - - - [21760, 9472, 1, 256] - - [47, 0.0] - - - [9216, 6145, 1, 256] - - [53, 0.0] - - - [8192, 1536, 1, 256] - - [53, 0.0] - - - [39936, 27648, 1, 256] - - [53, 0.0] - - - [21248, 9216, 1, 256] - - [53, 0.0] - - - [5376, 2049, 1, 256] - - [53, 0.0] - - - [35072, 22529, 1, 256] - - [53, 0.0] - - - [13312, 769, 1, 256] - - [50, 0.0] - - - [35840, 9216, 1, 256] - - [53, 0.0] - - - [39424, 27136, 1, 256] - - [53, 0.0] - - - [26368, 9216, 1, 256] - - [53, 0.0] - - - [34048, 21505, 1, 256] - - [53, 0.0] - - - [26112, 1792, 1, 256] - - [47, 0.0] - - - [23296, 768, 1, 256] - - [47, 0.0] - - - [43264, 27648, 1, 256] - - [53, 0.0] - - - [18432, 9216, 1, 256] - - [53, 0.0] - - - [38912, 3072, 1, 256] - - [53, 0.0] - - - [30464, 17921, 1, 256] - - [53, 0.0] - - - [37376, 9216, 1, 256] - - [53, 0.0] - - - [256, 3072, 1, 256] - - [57, 0.0] - - - [9472, 3072, 1, 256] - - [53, 0.0] - - - [35840, 23552, 1, 256] - - [53, 0.0] - - - [8960, 3072, 1, 256] - - [53, 0.0] - - - [34816, 3072, 1, 256] - - [53, 0.0] - - - [11008, 3072, 1, 256] - - [53, 0.0] - - - [36864, 1536, 1, 256] - - [53, 0.0] - - - [23552, 9216, 1, 256] - - [53, 0.0] - - - [31232, 18945, 1, 256] - - [53, 0.0] - - - [27136, 9216, 1, 256] - - [53, 0.0] - - - [19968, 7681, 1, 256] - - [53, 0.0] - - - [31488, 18945, 1, 256] - - [53, 0.0] - - - [33280, 1792, 1, 256] - - [47, 0.0] - - - [14592, 3072, 1, 256] - - [53, 0.0] - - - [30976, 18689, 1, 256] - - [47, 0.0] - - - [4096, 769, 1, 256] - - [47, 0.0] - - - [31488, 3072, 1, 256] - - [53, 0.0] - - - [33024, 1792, 1, 256] - - [47, 0.0] - - - [11520, 8449, 1, 256] - - [47, 0.0] - - - [44544, 4353, 1, 256] - - [47, 0.0] - - - [18176, 5889, 1, 256] - - [47, 0.0] - - - [5632, 2305, 1, 256] - - [47, 0.0] - - - [39936, 27393, 1, 256] - - [47, 0.0] - - - [10240, 7169, 1, 256] - - [53, 0.0] - - - [39168, 26625, 1, 256] - - [53, 0.0] - - - [10752, 7681, 1, 256] - - [53, 0.0] - - - [13824, 1536, 1, 256] - - [53, 0.0] - - - [14336, 9216, 1, 256] - - [53, 0.0] - - - [37632, 25345, 1, 256] - - [47, 0.0] - - - [35840, 23553, 1, 256] - - [53, 0.0] - - - [23552, 3072, 1, 256] - - [53, 0.0] - - - [19712, 7169, 1, 256] - - [53, 0.0] - - - [5888, 2561, 1, 256] - - [53, 0.0] - - - [27136, 768, 1, 256] - - [47, 0.0] - - - [22272, 1792, 1, 256] - - [47, 0.0] - - - [15616, 1536, 1, 256] - - [53, 0.0] - - - [3840, 769, 1, 256] - - [35, 0.0] - - - [42240, 2304, 1, 256] - - [47, 0.0] - - - [24576, 3072, 1, 256] - - [53, 0.0] - - - [27136, 1536, 1, 256] - - [53, 0.0] - - - [25344, 12801, 1, 256] - - [53, 0.0] - - - [32512, 20224, 1, 256] - - [47, 0.0] - - - [17664, 3072, 1, 256] - - [53, 0.0] - - - [28160, 15873, 1, 256] - - [53, 0.0] - - - [40960, 3072, 1, 256] - - [53, 0.0] - - - [14592, 9216, 1, 256] - - [53, 0.0] - - - [22784, 10497, 1, 256] - - [47, 0.0] - - - [22272, 3072, 1, 256] - - [53, 0.0] - - - [39680, 27137, 1, 256] - - [53, 0.0] - - - [20992, 8704, 1, 256] - - [53, 0.0] - - - [24320, 1536, 1, 256] - - [53, 0.0] - - - [7936, 4865, 1, 256] - - [47, 0.0] - - - [17664, 5376, 1, 256] - - [47, 0.0] - - - [37888, 25345, 1, 256] - - [47, 0.0] - - - [23296, 10753, 1, 256] - - [53, 0.0] - - - [28416, 15873, 1, 256] - - [53, 0.0] - - - [27648, 15361, 1, 256] - - [53, 0.0] - - - [39424, 1536, 1, 256] - - [53, 0.0] - - - [15104, 2817, 1, 256] - - [47, 0.0] - - - [19456, 9216, 1, 256] - - [53, 0.0] - - - [24064, 11777, 1, 256] - - [53, 0.0] - - - [40448, 1536, 1, 256] - - [53, 0.0] - - - [512, 3072, 1, 256] - - [59, 0.0] - - - [38912, 9216, 1, 256] - - [53, 0.0] - - - [19456, 6913, 1, 256] - - [47, 0.0] - - - [29440, 1792, 1, 256] - - [47, 0.0] - - - [41984, 9216, 1, 256] - - [53, 0.0] - - - [14080, 1793, 1, 256] - - [47, 0.0] - - - [20992, 8449, 1, 256] - - [47, 0.0] - - - [17920, 768, 1, 256] - - [47, 0.0] - - - [10496, 7169, 1, 256] - - [53, 0.0] - - - [40704, 27648, 1, 256] - - [53, 0.0] - - - [13568, 1025, 1, 256] - - [50, 0.0] - - - [38144, 9216, 1, 256] - - [53, 0.0] - - - [27392, 15104, 1, 256] - - [47, 0.0] - - - [2304, 3072, 1, 256] - - [47, 0.0] - - - [9472, 6401, 1, 256] - - [47, 0.0] - - - [39424, 1792, 1, 256] - - [47, 0.0] - - - [41728, 768, 1, 256] - - [47, 0.0] - - - [11264, 3072, 1, 256] - - [53, 0.0] - - - [25344, 3072, 1, 256] - - [53, 0.0] - - - [24576, 1792, 1, 256] - - [47, 0.0] - - - [27392, 14849, 1, 256] - - [53, 0.0] - - - [14848, 2561, 1, 256] - - [53, 0.0] - - - [28160, 3072, 1, 256] - - [53, 0.0] - - - [23552, 11009, 1, 256] - - [47, 0.0] - - - [11776, 8449, 1, 256] - - [47, 0.0] - - - [16640, 1792, 1, 256] - - [47, 0.0] - - - [24576, 12289, 1, 256] - - [53, 0.0] - - - [38656, 26369, 1, 256] - - [47, 0.0] - - - [13824, 9216, 1, 256] - - [53, 0.0] - - - [28928, 1792, 1, 256] - - [47, 0.0] - - - [27904, 15361, 1, 256] - - [53, 0.0] - - - [3840, 1792, 1, 256] - - [53, 0.0] - - - [14848, 3072, 1, 256] - - [53, 0.0] - - - [27904, 1536, 1, 256] - - [53, 0.0] - - - [34816, 1536, 1, 256] - - [53, 0.0] - - - [14592, 2305, 1, 256] - - [47, 0.0] - - - [22528, 9985, 1, 256] - - [47, 0.0] - - - [26368, 13825, 1, 256] - - [53, 0.0] - - - [4096, 1792, 1, 256] - - [47, 0.0] - - - [30720, 18177, 1, 256] - - [47, 0.0] - - - [37120, 24833, 1, 256] - - [47, 0.0] - - - [24320, 3072, 1, 256] - - [53, 0.0] - - - [2560, 1536, 1, 256] - - [57, 0.0] - - - [44032, 4097, 1, 256] - - [53, 0.0] - - - [44544, 27648, 1, 256] - - [53, 0.0] - - - [34048, 21761, 1, 256] - - [47, 0.0] - - - [24064, 1536, 1, 256] - - [53, 0.0] - - - [24832, 12545, 1, 256] - - [47, 0.0] - - - [44032, 3841, 1, 256] - - [47, 0.0] - - - [40448, 257, 1, 256] - - [49, 0.0] - - - [26624, 14337, 1, 256] - - [53, 0.0] - - - [8192, 5121, 1, 256] - - [53, 0.0] - - - [42240, 1536, 1, 256] - - [53, 0.0] - - - [5888, 2817, 1, 256] - - [47, 0.0] - - - [6144, 1792, 1, 256] - - [47, 0.0] - - - [16384, 1792, 1, 256] - - [47, 0.0] - - - [35584, 23297, 1, 256] - - [47, 0.0] - - - [36352, 24064, 1, 256] - - [53, 0.0] - - - [23040, 1536, 1, 256] - - [53, 0.0] - - - [8704, 1536, 1, 256] - - [53, 0.0] - - - [18432, 6145, 1, 256] - - [53, 0.0] - - - [12032, 3072, 1, 256] - - [53, 0.0] - - - [39168, 3072, 1, 256] - - [53, 0.0] - - - [28160, 1536, 1, 256] - - [53, 0.0] - - - [41728, 27648, 1, 256] - - [53, 0.0] - - - [28416, 1792, 1, 256] - - [47, 0.0] - - - [24320, 12032, 1, 256] - - [47, 0.0] - - - [28928, 16385, 1, 256] - - [53, 0.0] - - - [34816, 22528, 1, 256] - - [53, 0.0] - - - [26368, 1792, 1, 256] - - [47, 0.0] - - - [25856, 13569, 1, 256] - - [47, 0.0] - - - [25600, 13312, 1, 256] - - [53, 0.0] - - - [31232, 18689, 1, 256] - - [47, 0.0] - - - [20736, 9216, 1, 256] - - [53, 0.0] - - - [34304, 9216, 1, 256] - - [53, 0.0] - - - [43264, 3073, 1, 256] - - [53, 0.0] - - - [8704, 5633, 1, 256] - - [53, 0.0] - - - [4864, 1793, 1, 256] - - [49, 0.0] - - - [41984, 3072, 1, 256] - - [53, 0.0] - - - [20992, 3072, 1, 256] - - [53, 0.0] - - - [9728, 6401, 1, 256] - - [47, 0.0] - - - [16640, 4097, 1, 256] - - [53, 0.0] - - - [38400, 9216, 1, 256] - - [53, 0.0] - - - [38656, 1536, 1, 256] - - [53, 0.0] - - - [1536, 3072, 1, 256] - - [53, 0.0] - - - [12544, 1792, 1, 256] - - [47, 0.0] - - - [37632, 1792, 1, 256] - - [47, 0.0] - - - [17152, 4609, 1, 256] - - [53, 0.0] - - - [18944, 6656, 1, 256] - - [53, 0.0] - - - [34560, 22017, 1, 256] - - [53, 0.0] - - - [23296, 11008, 1, 256] - - [47, 0.0] - - - [14848, 768, 1, 256] - - [47, 0.0] - - - [38656, 1792, 1, 256] - - [47, 0.0] - - - [8448, 5377, 1, 256] - - [47, 0.0] - - - [29952, 17665, 1, 256] - - [47, 0.0] - - - [33792, 21504, 1, 256] - - [53, 0.0] - - - [24576, 1536, 1, 256] - - [53, 0.0] - - - [37376, 1792, 1, 256] - - [47, 0.0] - - - [42752, 768, 1, 256] - - [47, 0.0] - - - [4096, 1025, 1, 256] - - [69, 0.0] - - - [35840, 768, 1, 256] - - [53, 0.0] - - - [19200, 3072, 1, 256] - - [53, 0.0] - - - [33536, 1792, 1, 256] - - [47, 0.0] - - - [36864, 9216, 1, 256] - - [53, 0.0] - - - [38656, 26368, 1, 256] - - [47, 0.0] - - - [44288, 9216, 1, 256] - - [53, 0.0] - - - [44288, 4097, 1, 256] - - [53, 0.0] - - - [26112, 3072, 1, 256] - - [53, 0.0] - - - [512, 768, 1, 256] - - [55, 0.0] - - - [36096, 3072, 1, 256] - - [53, 0.0] - - - [4864, 1537, 1, 256] - - [53, 0.0] - - - [31232, 18944, 1, 256] - - [53, 0.0] - - - [20224, 7681, 1, 256] - - [53, 0.0] - - - [26112, 9216, 1, 256] - - [53, 0.0] - - - [21504, 3072, 1, 256] - - [53, 0.0] - - - [12544, 3072, 1, 256] - - [53, 0.0] - - - [32256, 19713, 1, 256] - - [47, 0.0] - - - [40704, 1792, 1, 256] - - [47, 0.0] - - - [18176, 5888, 1, 256] - - [47, 0.0] - - - [33792, 9216, 1, 256] - - [53, 0.0] - - - [26624, 14336, 1, 256] - - [53, 0.0] - - - [38912, 1792, 1, 256] - - [47, 0.0] - - - [7936, 1792, 1, 256] - - [53, 0.0] - - - [28672, 16385, 1, 256] - - [53, 0.0] - - - [18944, 3072, 1, 256] - - [53, 0.0] - - - [33280, 20993, 1, 256] - - [53, 0.0] - - - [37120, 24832, 1, 256] - - [47, 0.0] - - - [43520, 1792, 1, 256] - - [47, 0.0] - - - [16896, 4609, 1, 256] - - [53, 0.0] - - - [41472, 1536, 1, 256] - - [53, 0.0] - - - [39936, 768, 1, 256] - - [47, 0.0] - - - [23296, 11009, 1, 256] - - [47, 0.0] - - - [26624, 9216, 1, 256] - - [53, 0.0] - - - [29184, 9216, 1, 256] - - [53, 0.0] - - - [36352, 9216, 1, 256] - - [53, 0.0] - - - [37632, 25344, 1, 256] - - [47, 0.0] - - - [37888, 25600, 1, 256] - - [53, 0.0] - - - [16640, 9216, 1, 256] - - [53, 0.0] - - - [44544, 9216, 1, 256] - - [53, 0.0] - - - [14080, 1792, 1, 256] - - [47, 0.0] - - - [33536, 21249, 1, 256] - - [47, 0.0] - - - [34048, 21760, 1, 256] - - [47, 0.0] - - - [9984, 768, 1, 256] - - [47, 0.0] - - - [40192, 1536, 1, 256] - - [53, 0.0] - - - [41728, 3072, 1, 256] - - [53, 0.0] - - - [35328, 9216, 1, 256] - - [53, 0.0] - - - [32512, 768, 1, 256] - - [47, 0.0] - - - [14592, 2049, 1, 256] - - [53, 0.0] - - - [14848, 9216, 1, 256] - - [53, 0.0] - - - [23808, 3072, 1, 256] - - [53, 0.0] - - - [13568, 9216, 1, 256] - - [53, 0.0] - - - [42496, 2560, 1, 256] - - [53, 0.0] - - - [42752, 3072, 1, 256] - - [53, 0.0] - - - [39680, 27392, 1, 256] - - [47, 0.0] - - - [14592, 1792, 1, 256] - - [47, 0.0] - - - [25600, 13313, 1, 256] - - [53, 0.0] - - - [26624, 1792, 1, 256] - - [47, 0.0] - - - [20480, 8193, 1, 256] - - [53, 0.0] - - - [36096, 23808, 1, 256] - - [47, 0.0] - - - [15104, 2561, 1, 256] - - [53, 0.0] - - - [43520, 3072, 1, 256] - - [53, 0.0] - - - [1280, 2048, 1, 256] - - [57, 0.0] - - - [43008, 1792, 1, 256] - - [47, 0.0] - - - [18688, 3072, 1, 256] - - [53, 0.0] - - - [35328, 23040, 1, 256] - - [53, 0.0] - - - [18944, 6401, 1, 256] - - [47, 0.0] - - - [16128, 3585, 1, 256] - - [53, 0.0] - - - [29952, 1536, 1, 256] - - [53, 0.0] - - - [17408, 5121, 1, 256] - - [53, 0.0] - - - [36608, 1792, 1, 256] - - [47, 0.0] - - - [13056, 768, 1, 256] - - [53, 0.0] - - - [26112, 13824, 1, 256] - - [53, 0.0] - - - [43520, 3585, 1, 256] - - [53, 0.0] - - - [40704, 9216, 1, 256] - - [53, 0.0] - - - [27904, 15617, 1, 256] - - [47, 0.0] - - - [21248, 3072, 1, 256] - - [53, 0.0] - - - [38912, 1536, 1, 256] - - [53, 0.0] - - - [28672, 1792, 1, 256] - - [47, 0.0] - - - [18432, 1792, 1, 256] - - [47, 0.0] - - - [29952, 9216, 1, 256] - - [53, 0.0] - - - [4352, 1025, 1, 256] - - [49, 0.0] - - - [34304, 22017, 1, 256] - - [53, 0.0] - - - [28160, 15617, 1, 256] - - [47, 0.0] - - - [19968, 9216, 1, 256] - - [53, 0.0] - - - [7424, 4353, 1, 256] - - [47, 0.0] - - - [19200, 1792, 1, 256] - - [47, 0.0] - - - [27648, 15360, 1, 256] - - [53, 0.0] - - - [23040, 10497, 1, 256] - - [47, 0.0] - - - [21248, 8961, 1, 256] - - [47, 0.0] - - - [32256, 1792, 1, 256] - - [47, 0.0] - - - [26112, 13569, 1, 256] - - [47, 0.0] - - - [12288, 8961, 1, 256] - - [47, 0.0] - - - [6656, 3585, 1, 256] - - [53, 0.0] - - - [19968, 7425, 1, 256] - - [47, 0.0] - - - [9472, 768, 1, 256] - - [47, 0.0] - - - [33792, 3072, 1, 256] - - [53, 0.0] - - - [15616, 3072, 1, 256] - - [53, 0.0] - - - [8704, 5377, 1, 256] - - [47, 0.0] - - - [11520, 3072, 1, 256] - - [53, 0.0] - - - [25856, 1536, 1, 256] - - [53, 0.0] - - - [28416, 768, 1, 256] - - [47, 0.0] - - - [32256, 3072, 1, 256] - - [53, 0.0] - - - [20736, 1536, 1, 256] - - [53, 0.0] - - - [22784, 10241, 1, 256] - - [53, 0.0] - - - [36608, 24321, 1, 256] - - [47, 0.0] - - - [36096, 9216, 1, 256] - - [53, 0.0] - - - [10752, 768, 1, 256] - - [47, 0.0] - - - [38400, 26112, 1, 256] - - [53, 0.0] - - - [9216, 5889, 1, 256] - - [47, 0.0] - - - [41472, 27648, 1, 256] - - [53, 0.0] - - - [38144, 25856, 1, 256] - - [47, 0.0] - - - [15360, 3073, 1, 256] - - [53, 0.0] - - - [29184, 16896, 1, 256] - - [53, 0.0] - - - [16128, 1792, 1, 256] - - [47, 0.0] - - - [32768, 20225, 1, 256] - - [47, 0.0] - - - [23040, 10752, 1, 256] - - [53, 0.0] - - - [15872, 3585, 1, 256] - - [53, 0.0] - - - [11008, 7681, 1, 256] - - [53, 0.0] - - - [15360, 9216, 1, 256] - - [53, 0.0] - - - [28416, 16128, 1, 256] - - [47, 0.0] - - - [30208, 1792, 1, 256] - - [47, 0.0] - - - [41728, 1792, 1, 256] - - [47, 0.0] - - - [32256, 19968, 1, 256] - - [53, 0.0] - - - [18944, 1792, 1, 256] - - [47, 0.0] - - - [41728, 1793, 1, 256] - - [47, 0.0] - - - [31488, 19201, 1, 256] - - [47, 0.0] - - - [40192, 257, 1, 256] - - [50, 0.0] - - - [42752, 27648, 1, 256] - - [53, 0.0] - - - [40704, 768, 1, 256] - - [47, 0.0] - - - [25088, 12545, 1, 256] - - [47, 0.0] - - - [24576, 9216, 1, 256] - - [53, 0.0] - - - [33024, 20737, 1, 256] - - [47, 0.0] - - - [29696, 9216, 1, 256] - - [53, 0.0] - - - [31232, 1536, 1, 256] - - [53, 0.0] - - - [30208, 17920, 1, 256] - - [53, 0.0] - - - [44544, 4609, 1, 256] - - [53, 0.0] - - - [22016, 9728, 1, 256] - - [53, 0.0] - - - [30208, 17921, 1, 256] - - [53, 0.0] - - - [19200, 6657, 1, 256] - - [53, 0.0] - - - [22016, 9729, 1, 256] - - [53, 0.0] - - - [18176, 768, 1, 256] - - [47, 0.0] - - - [29184, 1792, 1, 256] - - [47, 0.0] - - - [12288, 1792, 1, 256] - - [47, 0.0] - - - [22528, 1536, 1, 256] - - [53, 0.0] - - - [14848, 2305, 1, 256] - - [47, 0.0] - - - [41216, 1025, 1, 256] - - [50, 0.0] - - - [8192, 3072, 1, 256] - - [53, 0.0] - - - [5888, 1792, 1, 256] - - [47, 0.0] - - - [21760, 3072, 1, 256] - - [53, 0.0] - - - [22272, 9985, 1, 256] - - [47, 0.0] - - - [29184, 1536, 1, 256] - - [53, 0.0] - - - [22016, 3072, 1, 256] - - [53, 0.0] - - - [30720, 9216, 1, 256] - - [53, 0.0] - - - [39680, 1792, 1, 256] - - [47, 0.0] - - - [9728, 1536, 1, 256] - - [53, 0.0] - - - [34560, 9216, 1, 256] - - [53, 0.0] - - - [12032, 8705, 1, 256] - - [53, 0.0] - - - [10752, 7425, 1, 256] - - [47, 0.0] - - - [18688, 1536, 1, 256] - - [53, 0.0] - - - [16128, 3840, 1, 256] - - [47, 0.0] - - - [38656, 768, 1, 256] - - [47, 0.0] - - - [21248, 1792, 1, 256] - - [47, 0.0] - - - [36352, 3072, 1, 256] - - [53, 0.0] - - - [19968, 7680, 1, 256] - - [53, 0.0] - - - [3840, 513, 1, 256] - - [49, 0.0] - - - [38400, 3072, 1, 256] - - [53, 0.0] - - - [5376, 768, 1, 256] - - [47, 0.0] - - - [20224, 9216, 1, 256] - - [53, 0.0] - - - [17408, 5120, 1, 256] - - [53, 0.0] - - - [28928, 9216, 1, 256] - - [53, 0.0] - - - [35072, 1792, 1, 256] - - [47, 0.0] - - - [31488, 19200, 1, 256] - - [47, 0.0] - - - [11008, 7937, 1, 256] - - [47, 0.0] - - - [21248, 8705, 1, 256] - - [53, 0.0] - - - [13568, 3072, 1, 256] - - [53, 0.0] - - - [34560, 22273, 1, 256] - - [47, 0.0] - - - [34048, 768, 1, 256] - - [47, 0.0] - - - [40448, 27648, 1, 256] - - [53, 0.0] - - - [28416, 16129, 1, 256] - - [47, 0.0] - - - [34816, 22529, 1, 256] - - [53, 0.0] - - - [22528, 3072, 1, 256] - - [53, 0.0] - - - [27136, 14593, 1, 256] - - [47, 0.0] - - - [35584, 3072, 1, 256] - - [53, 0.0] - - - [43008, 3072, 1, 256] - - [53, 0.0] - - - [30464, 1792, 1, 256] - - [47, 0.0] - - - [16384, 4097, 1, 256] - - [53, 0.0] - - - [20992, 9216, 1, 256] - - [53, 0.0] - - - [31488, 1792, 1, 256] - - [47, 0.0] - - - [31488, 9216, 1, 256] - - [53, 0.0] - - - [22272, 9984, 1, 256] - - [47, 0.0] - - - [41728, 1537, 1, 256] - - [53, 0.0] - - - [26880, 1792, 1, 256] - - [47, 0.0] - - - [30464, 768, 1, 256] - - [47, 0.0] - - - [2816, 1792, 1, 256] - - [47, 0.0] - - - [41472, 1537, 1, 256] - - [53, 0.0] - - - [43008, 27648, 1, 256] - - [53, 0.0] - - - [39424, 27137, 1, 256] - - [53, 0.0] - - - [24320, 1792, 1, 256] - - [47, 0.0] - - - [32000, 3072, 1, 256] - - [53, 0.0] - - - [12800, 1792, 1, 256] - - [47, 0.0] - - - [15872, 3072, 1, 256] - - [53, 0.0] - - - [15872, 1792, 1, 256] - - [47, 0.0] - - - [10496, 7425, 1, 256] - - [47, 0.0] - - - [16896, 4608, 1, 256] - - [53, 0.0] - - - [9984, 6913, 1, 256] - - [47, 0.0] - - - [21248, 8960, 1, 256] - - [47, 0.0] - - - [14336, 1792, 1, 256] - - [53, 0.0] - - - [24832, 12544, 1, 256] - - [47, 0.0] - - - [30464, 18176, 1, 256] - - [47, 0.0] - - - [31744, 19201, 1, 256] - - [47, 0.0] - - - [1792, 768, 1, 256] - - [67, 0.0] - - - [1536, 2048, 1, 256] - - [47, 0.0] - - - [40192, 3072, 1, 256] - - [53, 0.0] - - - [42240, 3072, 1, 256] - - [53, 0.0] - - - [32256, 9216, 1, 256] - - [53, 0.0] - - - [41984, 2049, 1, 256] - - [53, 0.0] - - - [6656, 1792, 1, 256] - - [47, 0.0] - - - [13824, 1537, 1, 256] - - [53, 0.0] - - - [20736, 3072, 1, 256] - - [53, 0.0] - - - [36096, 23809, 1, 256] - - [47, 0.0] - - - [41728, 9216, 1, 256] - - [53, 0.0] - - - [25600, 768, 1, 256] - - [47, 0.0] - - - [37632, 768, 1, 256] - - [47, 0.0] - - - [25600, 9216, 1, 256] - - [53, 0.0] - - - [19968, 3072, 1, 256] - - [53, 0.0] - - - [15616, 9216, 1, 256] - - [53, 0.0] - - - [29184, 16897, 1, 256] - - [53, 0.0] - - - [7168, 3841, 1, 256] - - [47, 0.0] - - - [40704, 769, 1, 256] - - [47, 0.0] - - - [6144, 3073, 1, 256] - - [53, 0.0] - - - [34304, 1792, 1, 256] - - [47, 0.0] - - - [18688, 6400, 1, 256] - - [47, 0.0] - - - [20992, 1536, 1, 256] - - [53, 0.0] - - - [21760, 768, 1, 256] - - [47, 0.0] - - - [43264, 3072, 1, 256] - - [53, 0.0] - - - [21760, 9216, 1, 256] - - [53, 0.0] - - - [11264, 768, 1, 256] - - [67, 0.0] - - - [42496, 3072, 1, 256] - - [53, 0.0] - - - [30208, 17665, 1, 256] - - [47, 0.0] - - - [27392, 15105, 1, 256] - - [47, 0.0] - - - [29952, 17409, 1, 256] - - [53, 0.0] - - - [44032, 3072, 1, 256] - - [53, 0.0] - - - [41216, 9216, 1, 256] - - [53, 0.0] - - - [8448, 1536, 1, 256] - - [53, 0.0] - - - [36352, 768, 1, 256] - - [47, 0.0] - - - [23552, 768, 1, 256] - - [47, 0.0] - - - [7168, 3072, 1, 256] - - [53, 0.0] - - - [44288, 4353, 1, 256] - - [47, 0.0] - - - [36608, 768, 1, 256] - - [53, 0.0] - - - [15616, 3073, 1, 256] - - [53, 0.0] - - - [37376, 24833, 1, 256] - - [47, 0.0] - - - [38144, 25857, 1, 256] - - [47, 0.0] - - - [26880, 14592, 1, 256] - - [47, 0.0] - - - [6144, 2817, 1, 256] - - [47, 0.0] - - - [23808, 768, 1, 256] - - [47, 0.0] - - - [39168, 26881, 1, 256] - - [47, 0.0] - - - [5120, 1793, 1, 256] - - [47, 0.0] - - - [32512, 19969, 1, 256] - - [53, 0.0] - - - [43008, 2817, 1, 256] - - [47, 0.0] - - - [26112, 13825, 1, 256] - - [53, 0.0] - - - [33536, 3072, 1, 256] - - [53, 0.0] - - - [9728, 6657, 1, 256] - - [53, 0.0] - - - [2048, 3072, 1, 256] - - [59, 0.0] - - - [24832, 9216, 1, 256] - - [53, 0.0] - - - [5632, 2561, 1, 256] - - [53, 0.0] - - - [33280, 20992, 1, 256] - - [53, 0.0] - - - [20224, 7936, 1, 256] - - [47, 0.0] - - - [28672, 16384, 1, 256] - - [53, 0.0] - - - [28416, 9216, 1, 256] - - [53, 0.0] - - - [7936, 768, 1, 256] - - [47, 0.0] - - - [23552, 11265, 1, 256] - - [53, 0.0] - - - [25088, 3072, 1, 256] - - [53, 0.0] - - - [32000, 19457, 1, 256] - - [53, 0.0] - - - [44800, 3072, 1, 256] - - [53, 0.0] - - - [37120, 1792, 1, 256] - - [47, 0.0] - - - [30464, 18177, 1, 256] - - [47, 0.0] - - - [44544, 4608, 1, 256] - - [53, 0.0] - - - [7168, 768, 1, 256] - - [69, 0.0] - - - [18944, 9216, 1, 256] - - [53, 0.0] - - - [33280, 20737, 1, 256] - - [47, 0.0] - - - [25856, 3072, 1, 256] - - [53, 0.0] - - - [27648, 9216, 1, 256] - - [53, 0.0] - - - [5120, 2049, 1, 256] - - [53, 0.0] - - - [28160, 9216, 1, 256] - - [53, 0.0] - - - [37632, 25089, 1, 256] - - [53, 0.0] - - - [22016, 1792, 1, 256] - - [47, 0.0] - - - [16384, 9216, 1, 256] - - [53, 0.0] - - - [21504, 9217, 1, 256] - - [53, 0.0] - - - [20480, 7937, 1, 256] - - [47, 0.0] - - - [33536, 21248, 1, 256] - - [47, 0.0] - - - [12800, 768, 1, 256] - - [53, 0.0] - - - [28672, 9216, 1, 256] - - [53, 0.0] - - - [32000, 9216, 1, 256] - - [53, 0.0] - - - [44544, 3072, 1, 256] - - [53, 0.0] - - - [5376, 3072, 1, 256] - - [53, 0.0] - - - [35840, 23297, 1, 256] - - [47, 0.0] - - - [23808, 11521, 1, 256] - - [47, 0.0] - - - [13312, 1025, 1, 256] - - [49, 0.0] - - - [18176, 9216, 1, 256] - - [53, 0.0] - - - [17920, 5633, 1, 256] - - [53, 0.0] - - - [27648, 3072, 1, 256] - - [53, 0.0] - - - [1024, 3072, 1, 256] - - [59, 0.0] - - - [22016, 9216, 1, 256] - - [53, 0.0] - - - [21760, 9473, 1, 256] - - [47, 0.0] - - - [6144, 1536, 1, 256] - - [53, 0.0] - - - [16896, 1536, 1, 256] - - [53, 0.0] - - - [19968, 768, 1, 256] - - [47, 0.0] - - - [23552, 11264, 1, 256] - - [53, 0.0] - - - [27904, 3072, 1, 256] - - [53, 0.0] - - - [19712, 7425, 1, 256] - - [47, 0.0] - - - [26624, 14081, 1, 256] - - [47, 0.0] - - - [3328, 257, 1, 256] - - [46, 0.0] - - - [24320, 9216, 1, 256] - - [53, 0.0] - - - [14080, 3072, 1, 256] - - [53, 0.0] - - - [17408, 3072, 1, 256] - - [53, 0.0] - - - [21504, 9216, 1, 256] - - [53, 0.0] - - - [14848, 2560, 1, 256] - - [53, 0.0] - - - [34304, 3072, 1, 256] - - [53, 0.0] - - - [15104, 9216, 1, 256] - - [53, 0.0] - - - [17152, 4865, 1, 256] - - [47, 0.0] - - - [38912, 26625, 1, 256] - - [53, 0.0] - - - [41216, 1792, 1, 256] - - [47, 0.0] - - - [39424, 3072, 1, 256] - - [53, 0.0] - - - [30720, 18433, 1, 256] - - [53, 0.0] - - - [18944, 6657, 1, 256] - - [53, 0.0] - - - [5632, 1792, 1, 256] - - [47, 0.0] - - - [18176, 1792, 1, 256] - - [47, 0.0] - - - [31232, 9216, 1, 256] - - [53, 0.0] - - - [42752, 2561, 1, 256] - - [53, 0.0] - - - [18688, 9216, 1, 256] - - [53, 0.0] - - - [43776, 1792, 1, 256] - - [47, 0.0] - - - [10240, 1792, 1, 256] - - [47, 0.0] - - - [33792, 21505, 1, 256] - - [53, 0.0] - - - [25856, 13313, 1, 256] - - [53, 0.0] - - - [29952, 3072, 1, 256] - - [53, 0.0] - - - [5888, 768, 1, 256] - - [47, 0.0] - - - [20480, 9216, 1, 256] - - [53, 0.0] - - - [17152, 1792, 1, 256] - - [47, 0.0] - - - [23040, 10753, 1, 256] - - [53, 0.0] - - - [8960, 5889, 1, 256] - - [47, 0.0] - - - [16640, 4352, 1, 256] - - [47, 0.0] - - - [30464, 3072, 1, 256] - - [53, 0.0] - - - [16128, 9216, 1, 256] - - [53, 0.0] - - - [25344, 13057, 1, 256] - - [47, 0.0] - - - [39424, 9216, 1, 256] - - [53, 0.0] - - - [25600, 3072, 1, 256] - - [53, 0.0] - - - [28416, 3072, 1, 256] - - [53, 0.0] - - - [12800, 257, 1, 256] - - [49, 0.0] - - - [43264, 1792, 1, 256] - - [47, 0.0] - - - [20736, 8193, 1, 256] - - [53, 0.0] - - - [30976, 9216, 1, 256] - - [53, 0.0] - - - [40192, 27648, 1, 256] - - [53, 0.0] - - - [31232, 1792, 1, 256] - - [47, 0.0] - - - [36352, 23809, 1, 256] - - [47, 0.0] - - - [9984, 3072, 1, 256] - - [53, 0.0] - - - [11776, 1792, 1, 256] - - [53, 0.0] - - - [37120, 1536, 1, 256] - - [53, 0.0] - - - [14592, 2304, 1, 256] - - [47, 0.0] - - - [7424, 768, 1, 256] - - [53, 0.0] - - - [10240, 1536, 1, 256] - - [53, 0.0] - - - [27392, 9216, 1, 256] - - [53, 0.0] - - - [15104, 3072, 1, 256] - - [53, 0.0] - - - [19200, 9216, 1, 256] - - [53, 0.0] - - - [36096, 23553, 1, 256] - - [53, 0.0] - - - [16128, 3841, 1, 256] - - [47, 0.0] - - - [18432, 5889, 1, 256] - - [47, 0.0] - - - [43776, 3841, 1, 256] - - [47, 0.0] - - - [22528, 10241, 1, 256] - - [53, 0.0] - - - [20224, 3072, 1, 256] - - [53, 0.0] - - - [39680, 3072, 1, 256] - - [53, 0.0] - - - [20736, 8449, 1, 256] - - [47, 0.0] - - - [30720, 1792, 1, 256] - - [47, 0.0] - - - [36864, 24321, 1, 256] - - [47, 0.0] - - - [22784, 1536, 1, 256] - - [53, 0.0] - - - [7424, 4097, 1, 256] - - [53, 0.0] - - - [7680, 4609, 1, 256] - - [53, 0.0] - - - [12032, 768, 1, 256] - - [47, 0.0] - - - [1792, 3072, 1, 256] - - [59, 0.0] - - - [6400, 3073, 1, 256] - - [53, 0.0] - - - [29440, 17153, 1, 256] - - [47, 0.0] - - - [8704, 1792, 1, 256] - - [53, 0.0] - - - [30720, 3072, 1, 256] - - [53, 0.0] - - - [16384, 3841, 1, 256] - - [47, 0.0] - - - [40192, 9216, 1, 256] - - [53, 0.0] - - - [23040, 1792, 1, 256] - - [47, 0.0] - - - [37888, 25601, 1, 256] - - [53, 0.0] - - - [26368, 14080, 1, 256] - - [47, 0.0] - - - [30208, 3072, 1, 256] - - [53, 0.0] - - - [33024, 20736, 1, 256] - - [47, 0.0] - - - [35072, 22784, 1, 256] - - [47, 0.0] - - - [9472, 6145, 1, 256] - - [53, 0.0] - - - [22784, 1792, 1, 256] - - [47, 0.0] - - - [768, 2048, 1, 256] - - [57, 0.0] - - - [1024, 1280, 1, 256] - - [58, 0.0] - - - [41984, 27648, 1, 256] - - [53, 0.0] - - - [33024, 20481, 1, 256] - - [53, 0.0] - - - [33280, 1536, 1, 256] - - [53, 0.0] - - - [9216, 3072, 1, 256] - - [53, 0.0] - - - [22528, 1792, 1, 256] - - [47, 0.0] - - - [25088, 768, 1, 256] - - [47, 0.0] - - - [13825, 128, 1, 128] - - [103, 0.0] - - - [20609, 128, 1, 256] - - [113, 0.0] - - - [6017, 128, 1, 256] - - [92, 0.0] - - - [2305, 128, 1, 128] - - [100, 0.0] - - - [15745, 128, 1, 256] - - [119, 0.0] - - - [8833, 128, 1, 128] - - [117, 0.0] - - - [641, 128, 1, 128] - - [101, 0.0] - - - [9217, 128, 1, 128] - - [110, 0.0] - - - [15361, 128, 1, 256] - - [119, 0.0] - - - [22913, 128, 1, 256] - - [113, 0.0] - - - [2177, 128, 1, 128] - - [100, 0.0] - - - [19073, 128, 1, 256] - - [113, 0.0] - - - [28289, 128, 1, 128] - - [111, 0.0] - - - [13057, 128, 1, 256] - - [120, 0.0] - - - [1793, 128, 1, 128] - - [102, 0.0] - - - [16769, 128, 1, 128] - - [112, 0.0] - - - [23681, 128, 1, 256] - - [117, 0.0] - - - [14593, 128, 1, 256] - - [119, 0.0] - - - [24449, 128, 1, 128] - - [118, 0.0] - - - [4609, 128, 1, 256] - - [93, 0.0] - - - [10625, 128, 1, 128] - - [111, 0.0] - - - [12545, 128, 1, 256] - - [111, 0.0] - - - [5633, 128, 1, 128] - - [105, 0.0] - - - [641, 128, 1, 256] - - [114, 0.0] - - - [18305, 128, 1, 256] - - [120, 0.0] - - - [23297, 128, 1, 256] - - [120, 0.0] - - - [21377, 128, 1, 256] - - [112, 0.0] - - - [9601, 128, 1, 128] - - [112, 0.0] - - - [13697, 128, 1, 256] - - [96, 0.0] - - - [23681, 128, 1, 128] - - [113, 0.0] - - - [24833, 128, 1, 256] - - [111, 0.0] - - - [25985, 128, 1, 128] - - [111, 0.0] - - - [9601, 128, 1, 256] - - [110, 0.0] - - - [17153, 128, 1, 128] - - [110, 0.0] - - - [9985, 128, 1, 128] - - [113, 0.0] - - - [23297, 128, 1, 128] - - [113, 0.0] - - - [19073, 128, 1, 128] - - [118, 0.0] - - - [2689, 128, 1, 256] - - [99, 0.0] - - - [4993, 128, 1, 128] - - [105, 0.0] - - - [6913, 128, 1, 256] - - [91, 0.0] - - - [6785, 128, 1, 128] - - [118, 0.0] - - - [27905, 128, 1, 128] - - [111, 0.0] - - - [7169, 128, 1, 256] - - [81, 0.0] - - - [11905, 128, 1, 256] - - [119, 0.0] - - - [1409, 128, 1, 128] - - [100, 0.0] - - - [12673, 128, 1, 128] - - [111, 0.0] - - - [27521, 128, 1, 256] - - [110, 0.0] - - - [1409, 128, 1, 256] - - [115, 0.0] - - - [25217, 128, 1, 128] - - [118, 0.0] - - - [7297, 128, 1, 128] - - [111, 0.0] - - - [14081, 128, 1, 128] - - [111, 0.0] - - - [22913, 128, 1, 128] - - [118, 0.0] - - - [10753, 128, 1, 256] - - [94, 0.0] - - - [7937, 128, 1, 128] - - [78, 0.0] - - - [11393, 128, 1, 128] - - [111, 0.0] - - - [26369, 128, 1, 128] - - [120, 0.0] - - - [12161, 128, 1, 256] - - [120, 0.0] - - - [8449, 128, 1, 128] - - [89, 0.0] - - - [22145, 128, 1, 256] - - [112, 0.0] - - - [20225, 128, 1, 256] - - [120, 0.0] - - - [10241, 128, 1, 256] - - [105, 0.0] - - - [6913, 128, 1, 128] - - [118, 0.0] - - - [4993, 128, 1, 256] - - [74, 0.0] - - - [6401, 128, 1, 256] - - [80, 0.0] - - - [13057, 128, 1, 128] - - [118, 0.0] - - - [2945, 128, 1, 128] - - [112, 0.0] - - - [3713, 128, 1, 256] - - [104, 0.0] - - - [10753, 128, 1, 128] - - [105, 0.0] - - - [14849, 128, 1, 256] - - [81, 0.0] - - - [3841, 128, 1, 128] - - [92, 0.0] - - - [28289, 128, 1, 256] - - [113, 0.0] - - - [12929, 128, 1, 128] - - [111, 0.0] - - - [14081, 128, 1, 256] - - [90, 0.0] - - - [14977, 128, 1, 256] - - [112, 0.0] - - - [12545, 128, 1, 128] - - [111, 0.0] - - - [16129, 128, 1, 256] - - [110, 0.0] - - - [11777, 128, 1, 256] - - [105, 0.0] - - - [11777, 128, 1, 128] - - [120, 0.0] - - - [17537, 128, 1, 256] - - [112, 0.0] - - - [5377, 128, 1, 128] - - [103, 0.0] - - - [8065, 128, 1, 256] - - [95, 0.0] - - - [6145, 128, 1, 128] - - [93, 0.0] - - - [20993, 128, 1, 128] - - [111, 0.0] - - - [15617, 128, 1, 128] - - [112, 0.0] - - - [5633, 128, 1, 256] - - [108, 0.0] - - - [4865, 128, 1, 128] - - [120, 0.0] - - - [385, 128, 1, 256] - - [101, 0.0] - - - [3841, 128, 1, 256] - - [71, 0.0] - - - [8833, 128, 1, 256] - - [112, 0.0] - - - [4225, 128, 1, 128] - - [85, 0.0] - - - [11009, 128, 1, 256] - - [113, 0.0] - - - [385, 128, 1, 128] - - [101, 0.0] - - - [9473, 128, 1, 256] - - [112, 0.0] - - - [5761, 128, 1, 128] - - [92, 0.0] - - - [11905, 128, 1, 128] - - [118, 0.0] - - - [4097, 128, 1, 256] - - [92, 0.0] - - - [25217, 128, 1, 256] - - [96, 0.0] - - - [9089, 128, 1, 256] - - [117, 0.0] - - - [10369, 128, 1, 256] - - [120, 0.0] - - - [14209, 128, 1, 256] - - [97, 0.0] - - - [6401, 128, 1, 128] - - [92, 0.0] - - - [27137, 128, 1, 256] - - [113, 0.0] - - - [16385, 128, 1, 256] - - [74, 0.0] - - - [24833, 128, 1, 128] - - [118, 0.0] - - - [18689, 128, 1, 128] - - [120, 0.0] - - - [7553, 128, 1, 256] - - [75, 0.0] - - - [8321, 128, 1, 128] - - [79, 0.0] - - - [15361, 128, 1, 128] - - [111, 0.0] - - - [1153, 128, 1, 128] - - [101, 0.0] - - - [1025, 128, 1, 128] - - [109, 0.0] - - - [19841, 128, 1, 256] - - [112, 0.0] - - - [15233, 128, 1, 128] - - [119, 0.0] - - - [21761, 128, 1, 256] - - [117, 0.0] - - - [17153, 128, 1, 256] - - [117, 0.0] - - - [15617, 128, 1, 256] - - [112, 0.0] - - - [4865, 128, 1, 256] - - [105, 0.0] - - - [14209, 128, 1, 128] - - [111, 0.0] - - - [19457, 128, 1, 256] - - [120, 0.0] - - - [9857, 128, 1, 256] - - [112, 0.0] - - - [11521, 128, 1, 128] - - [117, 0.0] - - - [8449, 128, 1, 256] - - [87, 0.0] - - - [4097, 128, 1, 128] - - [85, 0.0] - - - [28673, 128, 1, 256] - - [111, 0.0] - - - [12161, 128, 1, 128] - - [111, 0.0] - - - [1921, 128, 1, 256] - - [100, 0.0] - - - [9985, 128, 1, 256] - - [111, 0.0] - - - [7937, 128, 1, 256] - - [95, 0.0] - - - [9857, 128, 1, 128] - - [111, 0.0] - - - [13825, 128, 1, 256] - - [97, 0.0] - - - [9089, 128, 1, 128] - - [110, 0.0] - - - [6785, 128, 1, 256] - - [83, 0.0] - - - [5249, 128, 1, 256] - - [86, 0.0] - - - [7681, 128, 1, 256] - - [88, 0.0] - - - [3329, 128, 1, 128] - - [100, 0.0] - - - [14465, 128, 1, 128] - - [110, 0.0] - - - [11137, 128, 1, 256] - - [118, 0.0] - - - [1153, 128, 1, 256] - - [115, 0.0] - - - [16001, 128, 1, 128] - - [119, 0.0] - - - [26753, 128, 1, 128] - - [120, 0.0] - - - [13697, 128, 1, 128] - - [117, 0.0] - - - [3073, 128, 1, 128] - - [107, 0.0] - - - [22529, 128, 1, 256] - - [111, 0.0] - - - [18689, 128, 1, 256] - - [111, 0.0] - - - [257, 128, 1, 128] - - [99, 0.0] - - - [15233, 128, 1, 256] - - [112, 0.0] - - - [27521, 128, 1, 128] - - [118, 0.0] - - - [16385, 128, 1, 128] - - [74, 0.0] - - - [4481, 128, 1, 256] - - [93, 0.0] - - - [6017, 128, 1, 128] - - [92, 0.0] - - - [7297, 128, 1, 256] - - [82, 0.0] - - - [7553, 128, 1, 128] - - [111, 0.0] - - - [21761, 128, 1, 128] - - [118, 0.0] - - - [11393, 128, 1, 256] - - [120, 0.0] - - - [11521, 128, 1, 256] - - [112, 0.0] - - - [12929, 128, 1, 256] - - [118, 0.0] - - - [20225, 128, 1, 128] - - [118, 0.0] - - - [13313, 128, 1, 128] - - [105, 0.0] - - - [2561, 128, 1, 128] - - [115, 0.0] - - - [1537, 128, 1, 128] - - [115, 0.0] - - - [24449, 128, 1, 256] - - [120, 0.0] - - - [12289, 128, 1, 256] - - [103, 0.0] - - - [4225, 128, 1, 256] - - [73, 0.0] - - - [26369, 128, 1, 256] - - [118, 0.0] - - - [17921, 128, 1, 256] - - [113, 0.0] - - - [2945, 128, 1, 256] - - [102, 0.0] - - - [24065, 128, 1, 128] - - [113, 0.0] - - - [6529, 128, 1, 128] - - [84, 0.0] - - - [6145, 128, 1, 256] - - [81, 0.0] - - - [25985, 128, 1, 256] - - [113, 0.0] - - - [8705, 128, 1, 256] - - [104, 0.0] - - - [384, 128, 1, 256] - - [99, 0.0] - - - [25601, 128, 1, 256] - - [120, 0.0] - - - [28673, 128, 1, 128] - - [113, 0.0] - - - [20609, 128, 1, 128] - - [111, 0.0] - - - [19457, 128, 1, 128] - - [118, 0.0] - - - [16769, 128, 1, 256] - - [117, 0.0] - - - [12673, 128, 1, 256] - - [120, 0.0] - - - [8321, 128, 1, 256] - - [88, 0.0] - - - [5249, 128, 1, 128] - - [110, 0.0] - - - [16129, 128, 1, 128] - - [110, 0.0] - - - [13441, 128, 1, 256] - - [88, 0.0] - - - [5377, 128, 1, 256] - - [86, 0.0] - - - [21377, 128, 1, 128] - - [112, 0.0] - - - [14465, 128, 1, 256] - - [119, 0.0] - - - [11137, 128, 1, 128] - - [111, 0.0] - - - [7681, 128, 1, 128] - - [89, 0.0] - - - [7169, 128, 1, 128] - - [98, 0.0] - - - [22145, 128, 1, 128] - - [113, 0.0] - - - [11009, 128, 1, 128] - - [111, 0.0] - - - [20993, 128, 1, 256] - - [113, 0.0] - - - [13313, 128, 1, 256] - - [119, 0.0] - - - [25601, 128, 1, 128] - - [113, 0.0] - - - [4609, 128, 1, 128] - - [85, 0.0] - - - [5761, 128, 1, 256] - - [91, 0.0] - - - [17921, 128, 1, 128] - - [111, 0.0] - - - [2689, 128, 1, 128] - - [100, 0.0] - - - [8705, 128, 1, 128] - - [104, 0.0] - - - [10241, 128, 1, 128] - - [105, 0.0] - - - [14977, 128, 1, 128] - - [111, 0.0] - - - [18305, 128, 1, 128] - - [111, 0.0] - - - [3457, 128, 1, 128] - - [112, 0.0] - - - [24065, 128, 1, 256] - - [113, 0.0] - - - [12289, 128, 1, 128] - - [103, 0.0] - - - [14593, 128, 1, 128] - - [118, 0.0] - - - [2177, 128, 1, 256] - - [106, 0.0] - - - [4481, 128, 1, 128] - - [73, 0.0] - - - [8065, 128, 1, 128] - - [87, 0.0] - - - [3457, 128, 1, 256] - - [104, 0.0] - - - [6529, 128, 1, 256] - - [82, 0.0] - - - [26753, 128, 1, 256] - - [113, 0.0] - - - [17537, 128, 1, 128] - - [110, 0.0] - - - [22529, 128, 1, 128] - - [113, 0.0] - - - [10625, 128, 1, 256] - - [111, 0.0] - - - [14849, 128, 1, 128] - - [98, 0.0] - - - [9217, 128, 1, 256] - - [112, 0.0] - - - [19841, 128, 1, 128] - - [119, 0.0] - - - [15745, 128, 1, 128] - - [112, 0.0] - - - [13441, 128, 1, 128] - - [112, 0.0] - - - [3713, 128, 1, 128] - - [102, 0.0] - - - [27137, 128, 1, 128] - - [118, 0.0] - - - [16001, 128, 1, 256] - - [110, 0.0] - - - [10369, 128, 1, 128] - - [120, 0.0] - - - [1921, 128, 1, 128] - - [100, 0.0] - - - [9473, 128, 1, 128] - - [112, 0.0] - - - [27905, 128, 1, 256] - - [111, 0.0] - - - [30976, 1024, 1, 128] - - [120, 0.0] - - - [42240, 26369, 1, 128] - - [120, 0.0] - - - [33024, 17025, 1, 128] - - [120, 0.0] - - - [39168, 512, 1, 128] - - [120, 0.0] - - - [30848, 1024, 1, 128] - - [120, 0.0] - - - [41728, 8192, 1, 128] - - [120, 0.0] - - - [39552, 23553, 1, 128] - - [120, 0.0] - - - [35072, 512, 1, 128] - - [120, 0.0] - - - [29952, 14081, 1, 128] - - [120, 0.0] - - - [33280, 2048, 1, 128] - - [120, 0.0] - - - [40320, 128, 1, 128] - - [120, 0.0] - - - [35456, 1024, 1, 128] - - [120, 0.0] - - - [36096, 1024, 1, 128] - - [120, 0.0] - - - [36992, 20993, 1, 128] - - [120, 0.0] - - - [36096, 20097, 1, 128] - - [120, 0.0] - - - [31488, 15489, 1, 128] - - [120, 0.0] - - - [39552, 23681, 1, 128] - - [120, 0.0] - - - [36864, 128, 1, 128] - - [120, 0.0] - - - [40320, 4096, 1, 128] - - [120, 0.0] - - - [35200, 2048, 1, 128] - - [120, 0.0] - - - [29824, 2048, 1, 128] - - [120, 0.0] - - - [34688, 2048, 1, 128] - - [120, 0.0] - - - [42752, 26753, 1, 128] - - [120, 0.0] - - - [34304, 4096, 1, 128] - - [120, 0.0] - - - [36480, 20481, 1, 128] - - [120, 0.0] - - - [33408, 128, 1, 128] - - [120, 0.0] - - - [38784, 4096, 1, 128] - - [120, 0.0] - - - [43264, 27393, 1, 128] - - [120, 0.0] - - - [34560, 128, 1, 128] - - [120, 0.0] - - - [30336, 4096, 1, 128] - - [120, 0.0] - - - [29056, 2048, 1, 128] - - [120, 0.0] - - - [34816, 512, 1, 128] - - [120, 0.0] - - - [38272, 2048, 1, 128] - - [120, 0.0] - - - [39808, 23937, 1, 128] - - [120, 0.0] - - - [30848, 512, 1, 128] - - [120, 0.0] - - - [40448, 512, 1, 128] - - [120, 0.0] - - - [40448, 24577, 1, 128] - - [120, 0.0] - - - [44544, 28545, 1, 128] - - [120, 0.0] - - - [30208, 14209, 1, 128] - - [120, 0.0] - - - [34688, 18689, 1, 128] - - [120, 0.0] - - - [31360, 512, 1, 128] - - [120, 0.0] - - - [38912, 512, 1, 128] - - [120, 0.0] - - - [39680, 1024, 1, 128] - - [120, 0.0] - - - [34048, 1024, 1, 128] - - [120, 0.0] - - - [39552, 4096, 1, 128] - - [120, 0.0] - - - [40320, 24321, 1, 128] - - [120, 0.0] - - - [40832, 24833, 1, 128] - - [120, 0.0] - - - [36736, 1024, 1, 128] - - [120, 0.0] - - - [44672, 1024, 1, 128] - - [120, 0.0] - - - [32000, 128, 1, 128] - - [113, 0.0] - - - [40704, 4096, 1, 128] - - [120, 0.0] - - - [38144, 1024, 1, 128] - - [120, 0.0] - - - [30720, 14849, 1, 128] - - [120, 0.0] - - - [38144, 8192, 1, 128] - - [120, 0.0] - - - [30208, 1024, 1, 128] - - [120, 0.0] - - - [43136, 1024, 1, 128] - - [120, 0.0] - - - [38528, 1024, 1, 128] - - [120, 0.0] - - - [43264, 2048, 1, 128] - - [120, 0.0] - - - [38400, 22529, 1, 128] - - [120, 0.0] - - - [37120, 128, 1, 128] - - [119, 0.0] - - - [32256, 128, 1, 128] - - [111, 0.0] - - - [29952, 13953, 1, 128] - - [113, 0.0] - - - [34560, 8192, 1, 128] - - [120, 0.0] - - - [37504, 21505, 1, 128] - - [120, 0.0] - - - [33536, 128, 1, 128] - - [118, 0.0] - - - [41856, 2048, 1, 128] - - [120, 0.0] - - - [32896, 4096, 1, 128] - - [120, 0.0] - - - [41856, 8192, 1, 128] - - [120, 0.0] - - - [29440, 4096, 1, 128] - - [120, 0.0] - - - [33664, 8192, 1, 128] - - [120, 0.0] - - - [36992, 512, 1, 128] - - [120, 0.0] - - - [33280, 512, 1, 128] - - [120, 0.0] - - - [41728, 128, 1, 128] - - [113, 0.0] - - - [31744, 128, 1, 128] - - [111, 0.0] - - - [31360, 1024, 1, 128] - - [120, 0.0] - - - [29952, 8192, 1, 128] - - [120, 0.0] - - - [38016, 2048, 1, 128] - - [120, 0.0] - - - [34176, 8192, 1, 128] - - [120, 0.0] - - - [30464, 512, 1, 128] - - [120, 0.0] - - - [41984, 2048, 1, 128] - - [120, 0.0] - - - [40448, 4096, 1, 128] - - [120, 0.0] - - - [33920, 4096, 1, 128] - - [120, 0.0] - - - [41088, 8192, 1, 128] - - [120, 0.0] - - - [39808, 8192, 1, 128] - - [120, 0.0] - - - [40832, 4096, 1, 128] - - [120, 0.0] - - - [30592, 2048, 1, 128] - - [120, 0.0] - - - [36352, 1024, 1, 128] - - [120, 0.0] - - - [30336, 2048, 1, 128] - - [120, 0.0] - - - [30976, 512, 1, 128] - - [120, 0.0] - - - [42368, 1024, 1, 128] - - [120, 0.0] - - - [29056, 1024, 1, 128] - - [120, 0.0] - - - [38784, 22913, 1, 128] - - [120, 0.0] - - - [28928, 512, 1, 128] - - [120, 0.0] - - - [40576, 512, 1, 128] - - [120, 0.0] - - - [34816, 4096, 1, 128] - - [120, 0.0] - - - [41600, 2048, 1, 128] - - [120, 0.0] - - - [29696, 8192, 1, 128] - - [120, 0.0] - - - [41856, 4096, 1, 128] - - [120, 0.0] - - - [35584, 2048, 1, 128] - - [120, 0.0] - - - [30848, 14849, 1, 128] - - [120, 0.0] - - - [33280, 17281, 1, 128] - - [120, 0.0] - - - [43776, 2048, 1, 128] - - [120, 0.0] - - - [42112, 8192, 1, 128] - - [120, 0.0] - - - [37376, 128, 1, 128] - - [120, 0.0] - - - [41600, 4096, 1, 128] - - [120, 0.0] - - - [36224, 20353, 1, 128] - - [120, 0.0] - - - [29952, 1024, 1, 128] - - [120, 0.0] - - - [34176, 1024, 1, 128] - - [120, 0.0] - - - [31744, 512, 1, 128] - - [120, 0.0] - - - [42624, 8192, 1, 128] - - [120, 0.0] - - - [41216, 128, 1, 128] - - [113, 0.0] - - - [42624, 26753, 1, 128] - - [120, 0.0] - - - [32512, 2048, 1, 128] - - [120, 0.0] - - - [40064, 4096, 1, 128] - - [120, 0.0] - - - [32640, 4096, 1, 128] - - [120, 0.0] - - - [42112, 26241, 1, 128] - - [120, 0.0] - - - [32256, 512, 1, 128] - - [120, 0.0] - - - [40960, 1024, 1, 128] - - [120, 0.0] - - - [35968, 128, 1, 128] - - [120, 0.0] - - - [32384, 8192, 1, 128] - - [120, 0.0] - - - [42880, 512, 1, 128] - - [120, 0.0] - - - [33024, 8192, 1, 128] - - [120, 0.0] - - - [43904, 1024, 1, 128] - - [120, 0.0] - - - [33664, 17665, 1, 128] - - [120, 0.0] - - - [41856, 512, 1, 128] - - [120, 0.0] - - - [40704, 128, 1, 128] - - [118, 0.0] - - - [33408, 17537, 1, 128] - - [120, 0.0] - - - [37120, 512, 1, 128] - - [120, 0.0] - - - [41216, 25345, 1, 128] - - [120, 0.0] - - - [39680, 8192, 1, 128] - - [120, 0.0] - - - [40192, 24193, 1, 128] - - [120, 0.0] - - - [33024, 17153, 1, 128] - - [120, 0.0] - - - [38272, 1024, 1, 128] - - [120, 0.0] - - - [35328, 1024, 1, 128] - - [120, 0.0] - - - [31104, 8192, 1, 128] - - [120, 0.0] - - - [40320, 8192, 1, 128] - - [120, 0.0] - - - [29312, 2048, 1, 128] - - [120, 0.0] - - - [36608, 20737, 1, 128] - - [120, 0.0] - - - [42240, 4096, 1, 128] - - [120, 0.0] - - - [43520, 2048, 1, 128] - - [120, 0.0] - - - [29056, 512, 1, 128] - - [120, 0.0] - - - [35328, 19329, 1, 128] - - [120, 0.0] - - - [30464, 128, 1, 128] - - [113, 0.0] - - - [29696, 13697, 1, 128] - - [120, 0.0] - - - [43904, 28033, 1, 128] - - [120, 0.0] - - - [35584, 19713, 1, 128] - - [120, 0.0] - - - [41088, 4096, 1, 128] - - [120, 0.0] - - - [42368, 2048, 1, 128] - - [120, 0.0] - - - [36736, 128, 1, 128] - - [119, 0.0] - - - [30336, 8192, 1, 128] - - [120, 0.0] - - - [43008, 128, 1, 128] - - [118, 0.0] - - - [37120, 1024, 1, 128] - - [120, 0.0] - - - [31104, 2048, 1, 128] - - [120, 0.0] - - - [33152, 4096, 1, 128] - - [120, 0.0] - - - [43392, 27521, 1, 128] - - [120, 0.0] - - - [37248, 21249, 1, 128] - - [120, 0.0] - - - [33920, 17921, 1, 128] - - [120, 0.0] - - - [39680, 4096, 1, 128] - - [120, 0.0] - - - [43264, 512, 1, 128] - - [120, 0.0] - - - [35712, 8192, 1, 128] - - [120, 0.0] - - - [31616, 2048, 1, 128] - - [120, 0.0] - - - [35328, 512, 1, 128] - - [120, 0.0] - - - [43136, 27265, 1, 128] - - [120, 0.0] - - - [30208, 128, 1, 128] - - [113, 0.0] - - - [40320, 24449, 1, 128] - - [120, 0.0] - - - [44288, 2048, 1, 128] - - [120, 0.0] - - - [35072, 1024, 1, 128] - - [120, 0.0] - - - [30464, 14465, 1, 128] - - [120, 0.0] - - - [44160, 8192, 1, 128] - - [120, 0.0] - - - [33792, 17793, 1, 128] - - [120, 0.0] - - - [37632, 1024, 1, 128] - - [120, 0.0] - - - [35968, 2048, 1, 128] - - [120, 0.0] - - - [38400, 8192, 1, 128] - - [120, 0.0] - - - [32512, 4096, 1, 128] - - [120, 0.0] - - - [32512, 16641, 1, 128] - - [120, 0.0] - - - [39424, 128, 1, 128] - - [111, 0.0] - - - [30976, 8192, 1, 128] - - [120, 0.0] - - - [35968, 20097, 1, 128] - - [113, 0.0] - - - [38656, 512, 1, 128] - - [120, 0.0] - - - [34944, 18945, 1, 128] - - [120, 0.0] - - - [33664, 17793, 1, 128] - - [120, 0.0] - - - [38656, 22657, 1, 128] - - [120, 0.0] - - - [34944, 1024, 1, 128] - - [120, 0.0] - - - [31872, 16001, 1, 128] - - [113, 0.0] - - - [43392, 8192, 1, 128] - - [120, 0.0] - - - [38016, 512, 1, 128] - - [120, 0.0] - - - [29440, 8192, 1, 128] - - [120, 0.0] - - - [35200, 1024, 1, 128] - - [120, 0.0] - - - [34304, 18433, 1, 128] - - [120, 0.0] - - - [44672, 28801, 1, 128] - - [120, 0.0] - - - [29184, 4096, 1, 128] - - [120, 0.0] - - - [33408, 8192, 1, 128] - - [120, 0.0] - - - [39040, 128, 1, 128] - - [120, 0.0] - - - [39680, 23681, 1, 128] - - [120, 0.0] - - - [38144, 4096, 1, 128] - - [120, 0.0] - - - [42368, 26497, 1, 128] - - [120, 0.0] - - - [42368, 4096, 1, 128] - - [120, 0.0] - - - [31872, 128, 1, 128] - - [111, 0.0] - - - [41984, 512, 1, 128] - - [120, 0.0] - - - [39296, 2048, 1, 128] - - [120, 0.0] - - - [33920, 2048, 1, 128] - - [120, 0.0] - - - [36736, 20865, 1, 128] - - [120, 0.0] - - - [34432, 8192, 1, 128] - - [120, 0.0] - - - [30848, 14977, 1, 128] - - [120, 0.0] - - - [31744, 15873, 1, 128] - - [120, 0.0] - - - [42880, 27009, 1, 128] - - [120, 0.0] - - - [42240, 26241, 1, 128] - - [120, 0.0] - - - [38400, 4096, 1, 128] - - [120, 0.0] - - - [42624, 26625, 1, 128] - - [120, 0.0] - - - [35072, 4096, 1, 128] - - [120, 0.0] - - - [40576, 4096, 1, 128] - - [120, 0.0] - - - [39296, 8192, 1, 128] - - [120, 0.0] - - - [42624, 512, 1, 128] - - [120, 0.0] - - - [32768, 8192, 1, 128] - - [120, 0.0] - - - [36864, 1024, 1, 128] - - [120, 0.0] - - - [43392, 128, 1, 128] - - [120, 0.0] - - - [41344, 2048, 1, 128] - - [120, 0.0] - - - [35584, 4096, 1, 128] - - [120, 0.0] - - - [40064, 2048, 1, 128] - - [120, 0.0] - - - [40576, 24705, 1, 128] - - [120, 0.0] - - - [39808, 1024, 1, 128] - - [120, 0.0] - - - [36992, 1024, 1, 128] - - [120, 0.0] - - - [42496, 1024, 1, 128] - - [120, 0.0] - - - [43904, 128, 1, 128] - - [113, 0.0] - - - [31232, 512, 1, 128] - - [120, 0.0] - - - [42112, 128, 1, 128] - - [120, 0.0] - - - [37376, 2048, 1, 128] - - [120, 0.0] - - - [38016, 128, 1, 128] - - [118, 0.0] - - - [42368, 8192, 1, 128] - - [120, 0.0] - - - [43392, 512, 1, 128] - - [120, 0.0] - - - [41984, 1024, 1, 128] - - [120, 0.0] - - - [42240, 2048, 1, 128] - - [120, 0.0] - - - [29952, 128, 1, 128] - - [111, 0.0] - - - [36608, 8192, 1, 128] - - [120, 0.0] - - - [32512, 16513, 1, 128] - - [120, 0.0] - - - [29568, 512, 1, 128] - - [120, 0.0] - - - [34304, 1024, 1, 128] - - [120, 0.0] - - - [41984, 4096, 1, 128] - - [120, 0.0] - - - [30464, 4096, 1, 128] - - [120, 0.0] - - - [41216, 2048, 1, 128] - - [120, 0.0] - - - [36480, 20609, 1, 128] - - [120, 0.0] - - - [44800, 4096, 1, 128] - - [120, 0.0] - - - [36864, 512, 1, 128] - - [120, 0.0] - - - [39680, 2048, 1, 128] - - [120, 0.0] - - - [43648, 4096, 1, 128] - - [120, 0.0] - - - [33664, 128, 1, 128] - - [118, 0.0] - - - [41600, 512, 1, 128] - - [120, 0.0] - - - [43776, 1024, 1, 128] - - [120, 0.0] - - - [37632, 512, 1, 128] - - [118, 0.0] - - - [44160, 128, 1, 128] - - [111, 0.0] - - - [37248, 8192, 1, 128] - - [120, 0.0] - - - [34816, 18817, 1, 128] - - [120, 0.0] - - - [38528, 22529, 1, 128] - - [120, 0.0] - - - [40192, 24321, 1, 128] - - [120, 0.0] - - - [40832, 128, 1, 128] - - [111, 0.0] - - - [29312, 8192, 1, 128] - - [120, 0.0] - - - [43776, 27777, 1, 128] - - [120, 0.0] - - - [37632, 21633, 1, 128] - - [120, 0.0] - - - [33792, 4096, 1, 128] - - [120, 0.0] - - - [35968, 1024, 1, 128] - - [120, 0.0] - - - [37888, 512, 1, 128] - - [120, 0.0] - - - [35968, 512, 1, 128] - - [120, 0.0] - - - [30592, 1024, 1, 128] - - [120, 0.0] - - - [38400, 512, 1, 128] - - [120, 0.0] - - - [43264, 1024, 1, 128] - - [120, 0.0] - - - [38528, 4096, 1, 128] - - [120, 0.0] - - - [28928, 1024, 1, 128] - - [120, 0.0] - - - [33152, 1024, 1, 128] - - [120, 0.0] - - - [41344, 1024, 1, 128] - - [120, 0.0] - - - [30848, 8192, 1, 128] - - [120, 0.0] - - - [41344, 4096, 1, 128] - - [120, 0.0] - - - [38912, 2048, 1, 128] - - [120, 0.0] - - - [38272, 128, 1, 128] - - [118, 0.0] - - - [31488, 4096, 1, 128] - - [120, 0.0] - - - [44416, 4096, 1, 128] - - [120, 0.0] - - - [39552, 2048, 1, 128] - - [120, 0.0] - - - [37760, 1024, 1, 128] - - [120, 0.0] - - - [34304, 18305, 1, 128] - - [120, 0.0] - - - [44544, 28673, 1, 128] - - [120, 0.0] - - - [44416, 8192, 1, 128] - - [120, 0.0] - - - [38144, 512, 1, 128] - - [120, 0.0] - - - [30208, 14337, 1, 128] - - [120, 0.0] - - - [38144, 2048, 1, 128] - - [120, 0.0] - - - [40448, 128, 1, 128] - - [111, 0.0] - - - [42240, 8192, 1, 128] - - [120, 0.0] - - - [39424, 2048, 1, 128] - - [120, 0.0] - - - [41088, 512, 1, 128] - - [120, 0.0] - - - [36224, 2048, 1, 128] - - [120, 0.0] - - - [31744, 4096, 1, 128] - - [120, 0.0] - - - [44160, 512, 1, 128] - - [120, 0.0] - - - [32000, 1024, 1, 128] - - [120, 0.0] - - - [42752, 1024, 1, 128] - - [120, 0.0] - - - [42496, 2048, 1, 128] - - [120, 0.0] - - - [32640, 2048, 1, 128] - - [120, 0.0] - - - [42752, 26881, 1, 128] - - [120, 0.0] - - - [32256, 8192, 1, 128] - - [120, 0.0] - - - [44800, 512, 1, 128] - - [120, 0.0] - - - [34816, 128, 1, 128] - - [120, 0.0] - - - [38272, 8192, 1, 128] - - [120, 0.0] - - - [44800, 28929, 1, 128] - - [120, 0.0] - - - [37120, 8192, 1, 128] - - [120, 0.0] - - - [43776, 512, 1, 128] - - [120, 0.0] - - - [43008, 1024, 1, 128] - - [120, 0.0] - - - [34432, 18561, 1, 128] - - [113, 0.0] - - - [36736, 4096, 1, 128] - - [120, 0.0] - - - [36224, 512, 1, 128] - - [120, 0.0] - - - [32768, 512, 1, 128] - - [120, 0.0] - - - [30592, 128, 1, 128] - - [113, 0.0] - - - [43008, 27137, 1, 128] - - [120, 0.0] - - - [34048, 18177, 1, 128] - - [120, 0.0] - - - [43136, 2048, 1, 128] - - [120, 0.0] - - - [29184, 13313, 1, 128] - - [120, 0.0] - - - [40064, 24193, 1, 128] - - [113, 0.0] - - - [40960, 128, 1, 128] - - [113, 0.0] - - - [29184, 2048, 1, 128] - - [120, 0.0] - - - [37248, 128, 1, 128] - - [112, 0.0] - - - [35328, 128, 1, 128] - - [113, 0.0] - - - [43264, 128, 1, 128] - - [120, 0.0] - - - [29952, 4096, 1, 128] - - [120, 0.0] - - - [36736, 20737, 1, 128] - - [113, 0.0] - - - [34176, 4096, 1, 128] - - [120, 0.0] - - - [32768, 1024, 1, 128] - - [120, 0.0] - - - [44160, 4096, 1, 128] - - [120, 0.0] - - - [31104, 1024, 1, 128] - - [120, 0.0] - - - [33792, 512, 1, 128] - - [120, 0.0] - - - [41216, 25217, 1, 128] - - [120, 0.0] - - - [31872, 1024, 1, 128] - - [120, 0.0] - - - [38528, 8192, 1, 128] - - [120, 0.0] - - - [44672, 4096, 1, 128] - - [120, 0.0] - - - [32512, 1024, 1, 128] - - [120, 0.0] - - - [39168, 8192, 1, 128] - - [120, 0.0] - - - [31360, 15361, 1, 128] - - [120, 0.0] - - - [38016, 22145, 1, 128] - - [120, 0.0] - - - [35712, 128, 1, 128] - - [120, 0.0] - - - [30208, 4096, 1, 128] - - [120, 0.0] - - - [33920, 128, 1, 128] - - [118, 0.0] - - - [30336, 128, 1, 128] - - [120, 0.0] - - - [42368, 128, 1, 128] - - [111, 0.0] - - - [38912, 4096, 1, 128] - - [120, 0.0] - - - [34176, 512, 1, 128] - - [120, 0.0] - - - [42752, 8192, 1, 128] - - [120, 0.0] - - - [31488, 1024, 1, 128] - - [120, 0.0] - - - [36608, 1024, 1, 128] - - [120, 0.0] - - - [41856, 128, 1, 128] - - [120, 0.0] - - - [29312, 13441, 1, 128] - - [120, 0.0] - - - [43520, 128, 1, 128] - - [118, 0.0] - - - [31616, 8192, 1, 128] - - [120, 0.0] - - - [40448, 2048, 1, 128] - - [120, 0.0] - - - [35328, 2048, 1, 128] - - [120, 0.0] - - - [36864, 20865, 1, 128] - - [120, 0.0] - - - [32000, 2048, 1, 128] - - [120, 0.0] - - - [34176, 18177, 1, 128] - - [120, 0.0] - - - [37504, 128, 1, 128] - - [111, 0.0] - - - [33792, 1024, 1, 128] - - [120, 0.0] - - - [31872, 8192, 1, 128] - - [120, 0.0] - - - [40704, 512, 1, 128] - - [120, 0.0] - - - [37632, 128, 1, 128] - - [113, 0.0] - - - [32640, 1024, 1, 128] - - [120, 0.0] - - - [44544, 8192, 1, 128] - - [120, 0.0] - - - [39424, 8192, 1, 128] - - [120, 0.0] - - - [39296, 512, 1, 128] - - [120, 0.0] - - - [35840, 128, 1, 128] - - [120, 0.0] - - - [39168, 1024, 1, 128] - - [120, 0.0] - - - [35712, 19841, 1, 128] - - [120, 0.0] - - - [29568, 13569, 1, 128] - - [120, 0.0] - - - [34944, 4096, 1, 128] - - [120, 0.0] - - - [32768, 2048, 1, 128] - - [120, 0.0] - - - [39296, 128, 1, 128] - - [111, 0.0] - - - [29568, 4096, 1, 128] - - [120, 0.0] - - - [39040, 1024, 1, 128] - - [120, 0.0] - - - [37376, 1024, 1, 128] - - [120, 0.0] - - - [33536, 2048, 1, 128] - - [120, 0.0] - - - [31488, 8192, 1, 128] - - [120, 0.0] - - - [37888, 1024, 1, 128] - - [120, 0.0] - - - [41472, 4096, 1, 128] - - [120, 0.0] - - - [30592, 512, 1, 128] - - [120, 0.0] - - - [34560, 18561, 1, 128] - - [120, 0.0] - - - [29184, 512, 1, 128] - - [120, 0.0] - - - [32256, 16257, 1, 128] - - [120, 0.0] - - - [43392, 27393, 1, 128] - - [120, 0.0] - - - [29312, 4096, 1, 128] - - [120, 0.0] - - - [43648, 2048, 1, 128] - - [120, 0.0] - - - [44288, 1024, 1, 128] - - [120, 0.0] - - - [35456, 128, 1, 128] - - [118, 0.0] - - - [44160, 28289, 1, 128] - - [120, 0.0] - - - [40320, 1024, 1, 128] - - [120, 0.0] - - - [37888, 22017, 1, 128] - - [120, 0.0] - - - [29696, 512, 1, 128] - - [120, 0.0] - - - [35840, 2048, 1, 128] - - [120, 0.0] - - - [37504, 2048, 1, 128] - - [120, 0.0] - - - [41728, 4096, 1, 128] - - [120, 0.0] - - - [42752, 4096, 1, 128] - - [120, 0.0] - - - [29824, 4096, 1, 128] - - [120, 0.0] - - - [44800, 1024, 1, 128] - - [120, 0.0] - - - [30592, 4096, 1, 128] - - [120, 0.0] - - - [43904, 4096, 1, 128] - - [120, 0.0] - - - [39552, 8192, 1, 128] - - [120, 0.0] - - - [37632, 2048, 1, 128] - - [120, 0.0] - - - [29312, 128, 1, 128] - - [112, 0.0] - - - [30080, 512, 1, 128] - - [118, 0.0] - - - [33664, 2048, 1, 128] - - [120, 0.0] - - - [43520, 27521, 1, 128] - - [120, 0.0] - - - [36224, 128, 1, 128] - - [111, 0.0] - - - [28928, 12929, 1, 128] - - [113, 0.0] - - - [29440, 1024, 1, 128] - - [120, 0.0] - - - [35840, 19969, 1, 128] - - [120, 0.0] - - - [42880, 4096, 1, 128] - - [120, 0.0] - - - [42496, 8192, 1, 128] - - [120, 0.0] - - - [39936, 24065, 1, 128] - - [120, 0.0] - - - [33408, 1024, 1, 128] - - [120, 0.0] - - - [32256, 2048, 1, 128] - - [120, 0.0] - - - [35712, 19713, 1, 128] - - [113, 0.0] - - - [40192, 4096, 1, 128] - - [120, 0.0] - - - [32000, 16129, 1, 128] - - [120, 0.0] - - - [44032, 512, 1, 128] - - [120, 0.0] - - - [35584, 128, 1, 128] - - [111, 0.0] - - - [35584, 8192, 1, 128] - - [120, 0.0] - - - [37888, 21889, 1, 128] - - [120, 0.0] - - - [37504, 1024, 1, 128] - - [120, 0.0] - - - [33664, 512, 1, 128] - - [120, 0.0] - - - [32384, 1024, 1, 128] - - [120, 0.0] - - - [38400, 1024, 1, 128] - - [120, 0.0] - - - [35200, 128, 1, 128] - - [111, 0.0] - - - [43648, 1024, 1, 128] - - [120, 0.0] - - - [36608, 128, 1, 128] - - [117, 0.0] - - - [32768, 128, 1, 128] - - [118, 0.0] - - - [28928, 4096, 1, 128] - - [120, 0.0] - - - [35200, 19329, 1, 128] - - [120, 0.0] - - - [41216, 8192, 1, 128] - - [120, 0.0] - - - [36864, 8192, 1, 128] - - [120, 0.0] - - - [40064, 128, 1, 128] - - [113, 0.0] - - - [42624, 1024, 1, 128] - - [120, 0.0] - - - [34688, 128, 1, 128] - - [119, 0.0] - - - [43648, 27777, 1, 128] - - [120, 0.0] - - - [37888, 8192, 1, 128] - - [120, 0.0] - - - [41472, 25601, 1, 128] - - [120, 0.0] - - - [38272, 512, 1, 128] - - [120, 0.0] - - - [35456, 4096, 1, 128] - - [120, 0.0] - - - [42496, 26625, 1, 128] - - [120, 0.0] - - - [43136, 4096, 1, 128] - - [120, 0.0] - - - [44800, 8192, 1, 128] - - [120, 0.0] - - - [36480, 8192, 1, 128] - - [120, 0.0] - - - [37504, 4096, 1, 128] - - [120, 0.0] - - - [39040, 8192, 1, 128] - - [120, 0.0] - - - [31104, 512, 1, 128] - - [120, 0.0] - - - [34176, 2048, 1, 128] - - [120, 0.0] - - - [31616, 512, 1, 128] - - [120, 0.0] - - - [35456, 2048, 1, 128] - - [120, 0.0] - - - [43136, 8192, 1, 128] - - [120, 0.0] - - - [33024, 128, 1, 128] - - [119, 0.0] - - - [38656, 4096, 1, 128] - - [120, 0.0] - - - [33408, 17409, 1, 128] - - [120, 0.0] - - - [39424, 1024, 1, 128] - - [120, 0.0] - - - [29312, 13313, 1, 128] - - [120, 0.0] - - - [35840, 4096, 1, 128] - - [120, 0.0] - - - [42496, 512, 1, 128] - - [120, 0.0] - - - [37632, 8192, 1, 128] - - [120, 0.0] - - - [41088, 2048, 1, 128] - - [120, 0.0] - - - [38528, 512, 1, 128] - - [120, 0.0] - - - [35072, 2048, 1, 128] - - [120, 0.0] - - - [31104, 4096, 1, 128] - - [120, 0.0] - - - [33280, 4096, 1, 128] - - [120, 0.0] - - - [43904, 8192, 1, 128] - - [120, 0.0] - - - [34816, 8192, 1, 128] - - [120, 0.0] - - - [38016, 1024, 1, 128] - - [120, 0.0] - - - [33152, 128, 1, 128] - - [118, 0.0] - - - [42496, 128, 1, 128] - - [111, 0.0] - - - [40832, 24961, 1, 128] - - [120, 0.0] - - - [41728, 1024, 1, 128] - - [120, 0.0] - - - [41472, 25473, 1, 128] - - [120, 0.0] - - - [34560, 2048, 1, 128] - - [120, 0.0] - - - [31616, 15617, 1, 128] - - [120, 0.0] - - - [33664, 4096, 1, 128] - - [120, 0.0] - - - [35328, 8192, 1, 128] - - [120, 0.0] - - - [39808, 4096, 1, 128] - - [120, 0.0] - - - [37248, 512, 1, 128] - - [120, 0.0] - - - [31360, 4096, 1, 128] - - [120, 0.0] - - - [41344, 8192, 1, 128] - - [120, 0.0] - - - [32000, 512, 1, 128] - - [120, 0.0] - - - [35968, 19969, 1, 128] - - [120, 0.0] - - - [30080, 14081, 1, 128] - - [113, 0.0] - - - [35840, 8192, 1, 128] - - [120, 0.0] - - - [44672, 2048, 1, 128] - - [120, 0.0] - - - [31872, 2048, 1, 128] - - [120, 0.0] - - - [42496, 4096, 1, 128] - - [120, 0.0] - - - [43776, 128, 1, 128] - - [120, 0.0] - - - [40704, 2048, 1, 128] - - [120, 0.0] - - - [34432, 128, 1, 128] - - [111, 0.0] - - - [44544, 2048, 1, 128] - - [120, 0.0] - - - [32384, 16385, 1, 128] - - [120, 0.0] - - - [43776, 27905, 1, 128] - - [120, 0.0] - - - [44032, 4096, 1, 128] - - [120, 0.0] - - - [36480, 512, 1, 128] - - [120, 0.0] - - - [44160, 1024, 1, 128] - - [120, 0.0] - - - [41216, 4096, 1, 128] - - [120, 0.0] - - - [44032, 2048, 1, 128] - - [120, 0.0] - - - [33152, 2048, 1, 128] - - [120, 0.0] - - - [41984, 25985, 1, 128] - - [120, 0.0] - - - [39552, 512, 1, 128] - - [120, 0.0] - - - [41344, 25473, 1, 128] - - [120, 0.0] - - - [40960, 4096, 1, 128] - - [120, 0.0] - - - [32640, 128, 1, 128] - - [111, 0.0] - - - [35968, 4096, 1, 128] - - [120, 0.0] - - - [33536, 4096, 1, 128] - - [120, 0.0] - - - [30976, 15105, 1, 128] - - [120, 0.0] - - - [35072, 8192, 1, 128] - - [120, 0.0] - - - [39424, 23425, 1, 128] - - [120, 0.0] - - - [43520, 1024, 1, 128] - - [120, 0.0] - - - [44288, 28417, 1, 128] - - [120, 0.0] - - - [30848, 128, 1, 128] - - [119, 0.0] - - - [35712, 512, 1, 128] - - [120, 0.0] - - - [44160, 2048, 1, 128] - - [120, 0.0] - - - [34048, 8192, 1, 128] - - [120, 0.0] - - - [40448, 24449, 1, 128] - - [120, 0.0] - - - [39168, 23297, 1, 128] - - [120, 0.0] - - - [32128, 1024, 1, 128] - - [120, 0.0] - - - [36864, 20993, 1, 128] - - [120, 0.0] - - - [40064, 1024, 1, 128] - - [120, 0.0] - - - [38784, 8192, 1, 128] - - [120, 0.0] - - - [37248, 2048, 1, 128] - - [120, 0.0] - - - [34560, 4096, 1, 128] - - [120, 0.0] - - - [39040, 23041, 1, 128] - - [120, 0.0] - - - [36480, 1024, 1, 128] - - [120, 0.0] - - - [39040, 2048, 1, 128] - - [120, 0.0] - - - [39808, 23809, 1, 128] - - [113, 0.0] - - - [36992, 4096, 1, 128] - - [120, 0.0] - - - [32768, 16897, 1, 128] - - [120, 0.0] - - - [30976, 2048, 1, 128] - - [120, 0.0] - - - [32640, 16769, 1, 128] - - [113, 0.0] - - - [29824, 13953, 1, 128] - - [120, 0.0] - - - [29184, 128, 1, 128] - - [118, 0.0] - - - [30720, 8192, 1, 128] - - [120, 0.0] - - - [30848, 2048, 1, 128] - - [120, 0.0] - - - [38016, 4096, 1, 128] - - [120, 0.0] - - - [35456, 8192, 1, 128] - - [120, 0.0] - - - [36992, 21121, 1, 128] - - [120, 0.0] - - - [36736, 2048, 1, 128] - - [120, 0.0] - - - [37888, 128, 1, 128] - - [120, 0.0] - - - [39808, 2048, 1, 128] - - [120, 0.0] - - - [41856, 25985, 1, 128] - - [120, 0.0] - - - [34688, 4096, 1, 128] - - [120, 0.0] - - - [38784, 1024, 1, 128] - - [120, 0.0] - - - [40960, 25089, 1, 128] - - [120, 0.0] - - - [32000, 4096, 1, 128] - - [120, 0.0] - - - [41600, 25601, 1, 128] - - [120, 0.0] - - - [37504, 512, 1, 128] - - [120, 0.0] - - - [32128, 16129, 1, 128] - - [120, 0.0] - - - [37248, 21377, 1, 128] - - [120, 0.0] - - - [35840, 512, 1, 128] - - [120, 0.0] - - - [36096, 128, 1, 128] - - [111, 0.0] - - - [32512, 8192, 1, 128] - - [120, 0.0] - - - [36736, 8192, 1, 128] - - [120, 0.0] - - - [42880, 1024, 1, 128] - - [120, 0.0] - - - [44288, 8192, 1, 128] - - [120, 0.0] - - - [36224, 1024, 1, 128] - - [120, 0.0] - - - [41344, 25345, 1, 128] - - [120, 0.0] - - - [32384, 512, 1, 128] - - [120, 0.0] - - - [38272, 4096, 1, 128] - - [120, 0.0] - - - [37120, 2048, 1, 128] - - [120, 0.0] - - - [33152, 8192, 1, 128] - - [120, 0.0] - - - [36096, 4096, 1, 128] - - [120, 0.0] - - - [34560, 18689, 1, 128] - - [120, 0.0] - - - [36864, 4096, 1, 128] - - [120, 0.0] - - - [34944, 512, 1, 128] - - [120, 0.0] - - - [37760, 128, 1, 128] - - [118, 0.0] - - - [31616, 128, 1, 128] - - [111, 0.0] - - - [36224, 4096, 1, 128] - - [120, 0.0] - - - [40576, 24577, 1, 128] - - [120, 0.0] - - - [34688, 1024, 1, 128] - - [120, 0.0] - - - [40192, 1024, 1, 128] - - [120, 0.0] - - - [44672, 512, 1, 128] - - [120, 0.0] - - - [33664, 1024, 1, 128] - - [120, 0.0] - - - [39424, 512, 1, 128] - - [120, 0.0] - - - [44416, 1024, 1, 128] - - [120, 0.0] - - - [33408, 2048, 1, 128] - - [120, 0.0] - - - [43648, 8192, 1, 128] - - [120, 0.0] - - - [43520, 27649, 1, 128] - - [120, 0.0] - - - [40448, 1024, 1, 128] - - [120, 0.0] - - - [33152, 17153, 1, 128] - - [120, 0.0] - - - [33024, 512, 1, 128] - - [120, 0.0] - - - [39680, 128, 1, 128] - - [120, 0.0] - - - [29696, 4096, 1, 128] - - [120, 0.0] - - - [42112, 2048, 1, 128] - - [120, 0.0] - - - [38016, 8192, 1, 128] - - [120, 0.0] - - - [30464, 8192, 1, 128] - - [120, 0.0] - - - [43648, 128, 1, 128] - - [113, 0.0] - - - [32896, 16897, 1, 128] - - [120, 0.0] - - - [43008, 8192, 1, 128] - - [120, 0.0] - - - [34304, 512, 1, 128] - - [120, 0.0] - - - [38528, 128, 1, 128] - - [120, 0.0] - - - [41216, 1024, 1, 128] - - [120, 0.0] - - - [38272, 22401, 1, 128] - - [120, 0.0] - - - [34048, 4096, 1, 128] - - [120, 0.0] - - - [30720, 512, 1, 128] - - [120, 0.0] - - - [41728, 512, 1, 128] - - [120, 0.0] - - - [43136, 512, 1, 128] - - [120, 0.0] - - - [41088, 1024, 1, 128] - - [120, 0.0] - - - [33536, 1024, 1, 128] - - [120, 0.0] - - - [41088, 25089, 1, 128] - - [120, 0.0] - - - [36352, 20353, 1, 128] - - [120, 0.0] - - - [29184, 1024, 1, 128] - - [120, 0.0] - - - [44800, 128, 1, 128] - - [111, 0.0] - - - [41600, 8192, 1, 128] - - [120, 0.0] - - - [44416, 28545, 1, 128] - - [120, 0.0] - - - [34048, 512, 1, 128] - - [120, 0.0] - - - [32128, 16257, 1, 128] - - [120, 0.0] - - - [44288, 4096, 1, 128] - - [120, 0.0] - - - [34432, 18433, 1, 128] - - [120, 0.0] - - - [41856, 25857, 1, 128] - - [120, 0.0] - - - [32128, 2048, 1, 128] - - [120, 0.0] - - - [34688, 512, 1, 128] - - [120, 0.0] - - - [39936, 4096, 1, 128] - - [120, 0.0] - - - [38656, 1024, 1, 128] - - [120, 0.0] - - - [37760, 512, 1, 128] - - [120, 0.0] - - - [30336, 512, 1, 128] - - [120, 0.0] - - - [38016, 22017, 1, 128] - - [120, 0.0] - - - [44544, 4096, 1, 128] - - [120, 0.0] - - - [38912, 8192, 1, 128] - - [120, 0.0] - - - [39936, 128, 1, 128] - - [113, 0.0] - - - [36480, 2048, 1, 128] - - [120, 0.0] - - - [35200, 4096, 1, 128] - - [120, 0.0] - - - [30976, 14977, 1, 128] - - [120, 0.0] - - - [31104, 15105, 1, 128] - - [120, 0.0] - - - [40832, 1024, 1, 128] - - [120, 0.0] - - - [32384, 16513, 1, 128] - - [120, 0.0] - - - [43392, 4096, 1, 128] - - [120, 0.0] - - - [32768, 4096, 1, 128] - - [120, 0.0] - - - [38272, 22273, 1, 128] - - [120, 0.0] - - - [32128, 512, 1, 128] - - [120, 0.0] - - - [32896, 2048, 1, 128] - - [120, 0.0] - - - [37376, 21505, 1, 128] - - [120, 0.0] - - - [41856, 1024, 1, 128] - - [120, 0.0] - - - [33536, 8192, 1, 128] - - [120, 0.0] - - - [29568, 1024, 1, 128] - - [120, 0.0] - - - [44032, 28033, 1, 128] - - [120, 0.0] - - - [33280, 8192, 1, 128] - - [120, 0.0] - - - [39296, 4096, 1, 128] - - [120, 0.0] - - - [30592, 14593, 1, 128] - - [113, 0.0] - - - [37504, 8192, 1, 128] - - [120, 0.0] - - - [30336, 14465, 1, 128] - - [120, 0.0] - - - [29952, 2048, 1, 128] - - [120, 0.0] - - - [40832, 512, 1, 128] - - [120, 0.0] - - - [44672, 28673, 1, 128] - - [120, 0.0] - - - [30080, 4096, 1, 128] - - [120, 0.0] - - - [37888, 2048, 1, 128] - - [120, 0.0] - - - [37632, 21761, 1, 128] - - [120, 0.0] - - - [29824, 8192, 1, 128] - - [120, 0.0] - - - [35328, 19457, 1, 128] - - [120, 0.0] - - - [37376, 4096, 1, 128] - - [120, 0.0] - - - [33792, 17921, 1, 128] - - [120, 0.0] - - - [34304, 8192, 1, 128] - - [120, 0.0] - - - [42752, 512, 1, 128] - - [120, 0.0] - - - [36992, 2048, 1, 128] - - [120, 0.0] - - - [39168, 4096, 1, 128] - - [120, 0.0] - - - [31360, 15489, 1, 128] - - [120, 0.0] - - - [43520, 8192, 1, 128] - - [120, 0.0] - - - [30080, 2048, 1, 128] - - [120, 0.0] - - - [30720, 4096, 1, 128] - - [120, 0.0] - - - [34176, 128, 1, 128] - - [113, 0.0] - - - [32768, 16769, 1, 128] - - [120, 0.0] - - - [35072, 128, 1, 128] - - [112, 0.0] - - - [35712, 4096, 1, 128] - - [120, 0.0] - - - [36480, 4096, 1, 128] - - [120, 0.0] - - - [39424, 4096, 1, 128] - - [120, 0.0] - - - [38400, 128, 1, 128] - - [113, 0.0] - - - [34432, 2048, 1, 128] - - [120, 0.0] - - - [41344, 512, 1, 128] - - [120, 0.0] - - - [35200, 512, 1, 128] - - [120, 0.0] - - - [39936, 8192, 1, 128] - - [120, 0.0] - - - [31488, 128, 1, 128] - - [120, 0.0] - - - [43008, 512, 1, 128] - - [120, 0.0] - - - [33024, 4096, 1, 128] - - [120, 0.0] - - - [36608, 512, 1, 128] - - [120, 0.0] - - - [37376, 8192, 1, 128] - - [120, 0.0] - - - [29824, 13825, 1, 128] - - [120, 0.0] - - - [36352, 2048, 1, 128] - - [120, 0.0] - - - [30336, 1024, 1, 128] - - [120, 0.0] - - - [44416, 28417, 1, 128] - - [120, 0.0] - - - [38144, 22273, 1, 128] - - [120, 0.0] - - - [28928, 2048, 1, 128] - - [120, 0.0] - - - [29568, 13697, 1, 128] - - [120, 0.0] - - - [43136, 27137, 1, 128] - - [120, 0.0] - - - [42112, 4096, 1, 128] - - [120, 0.0] - - - [40960, 512, 1, 128] - - [120, 0.0] - - - [35584, 1024, 1, 128] - - [120, 0.0] - - - [31232, 15361, 1, 128] - - [120, 0.0] - - - [40960, 8192, 1, 128] - - [120, 0.0] - - - [31232, 1024, 1, 128] - - [120, 0.0] - - - [29312, 512, 1, 128] - - [120, 0.0] - - - [44416, 512, 1, 128] - - [120, 0.0] - - - [42240, 512, 1, 128] - - [120, 0.0] - - - [31232, 8192, 1, 128] - - [120, 0.0] - - - [35072, 19201, 1, 128] - - [120, 0.0] - - - [29568, 128, 1, 128] - - [112, 0.0] - - - [33792, 2048, 1, 128] - - [120, 0.0] - - - [35712, 2048, 1, 128] - - [120, 0.0] - - - [40576, 128, 1, 128] - - [119, 0.0] - - - [40704, 1024, 1, 128] - - [120, 0.0] - - - [29824, 1024, 1, 128] - - [120, 0.0] - - - [33536, 17665, 1, 128] - - [120, 0.0] - - - [43008, 27009, 1, 128] - - [120, 0.0] - - - [34304, 2048, 1, 128] - - [120, 0.0] - - - [37120, 21249, 1, 128] - - [120, 0.0] - - - [41600, 1024, 1, 128] - - [120, 0.0] - - - [33024, 1024, 1, 128] - - [120, 0.0] - - - [42368, 512, 1, 128] - - [120, 0.0] - - - [30592, 14721, 1, 128] - - [120, 0.0] - - - [29696, 2048, 1, 128] - - [120, 0.0] - - - [31232, 128, 1, 128] - - [120, 0.0] - - - [38784, 22785, 1, 128] - - [120, 0.0] - - - [32896, 1024, 1, 128] - - [120, 0.0] - - - [32128, 128, 1, 128] - - [113, 0.0] - - - [35968, 8192, 1, 128] - - [120, 0.0] - - - [38400, 2048, 1, 128] - - [120, 0.0] - - - [36864, 2048, 1, 128] - - [120, 0.0] - - - [31616, 4096, 1, 128] - - [120, 0.0] - - - [34688, 18817, 1, 128] - - [120, 0.0] - - - [42624, 4096, 1, 128] - - [120, 0.0] - - - [29312, 1024, 1, 128] - - [120, 0.0] - - - [37760, 2048, 1, 128] - - [120, 0.0] - - - [39808, 512, 1, 128] - - [120, 0.0] - - - [41472, 128, 1, 128] - - [111, 0.0] - - - [32128, 4096, 1, 128] - - [120, 0.0] - - - [43520, 4096, 1, 128] - - [120, 0.0] - - - [41472, 512, 1, 128] - - [120, 0.0] - - - [38912, 22913, 1, 128] - - [120, 0.0] - - - [30464, 1024, 1, 128] - - [120, 0.0] - - - [33280, 128, 1, 128] - - [113, 0.0] - - - [31872, 15873, 1, 128] - - [120, 0.0] - - - [36352, 4096, 1, 128] - - [120, 0.0] - - - [30720, 2048, 1, 128] - - [120, 0.0] - - - [33792, 128, 1, 128] - - [120, 0.0] - - - [36096, 8192, 1, 128] - - [120, 0.0] - - - [38784, 128, 1, 128] - - [120, 0.0] - - - [30208, 2048, 1, 128] - - [120, 0.0] - - - [34432, 4096, 1, 128] - - [120, 0.0] - - - [42880, 128, 1, 128] - - [120, 0.0] - - - [31616, 15745, 1, 128] - - [120, 0.0] - - - [40960, 2048, 1, 128] - - [120, 0.0] - - - [41344, 128, 1, 128] - - [113, 0.0] - - - [41728, 25857, 1, 128] - - [120, 0.0] - - - [32896, 512, 1, 128] - - [120, 0.0] - - - [41728, 2048, 1, 128] - - [120, 0.0] - - - [42368, 26369, 1, 128] - - [120, 0.0] - - - [30720, 14721, 1, 128] - - [120, 0.0] - - - [37376, 512, 1, 128] - - [120, 0.0] - - - [35456, 19457, 1, 128] - - [120, 0.0] - - - [29184, 13185, 1, 128] - - [120, 0.0] - - - [34944, 128, 1, 128] - - [119, 0.0] - - - [36608, 20609, 1, 128] - - [120, 0.0] - - - [35584, 19585, 1, 128] - - [120, 0.0] - - - [42880, 8192, 1, 128] - - [120, 0.0] - - - [39936, 1024, 1, 128] - - [120, 0.0] - - - [34944, 19073, 1, 128] - - [120, 0.0] - - - [32512, 128, 1, 128] - - [118, 0.0] - - - [40064, 512, 1, 128] - - [120, 0.0] - - - [30464, 2048, 1, 128] - - [120, 0.0] - - - [30592, 8192, 1, 128] - - [120, 0.0] - - - [39040, 512, 1, 128] - - [120, 0.0] - - - [41088, 128, 1, 128] - - [120, 0.0] - - - [29824, 128, 1, 128] - - [117, 0.0] - - - [32384, 128, 1, 128] - - [111, 0.0] - - - [41728, 25729, 1, 128] - - [120, 0.0] - - - [30976, 4096, 1, 128] - - [120, 0.0] - - - [42624, 128, 1, 128] - - [113, 0.0] - - - [42112, 512, 1, 128] - - [120, 0.0] - - - [38784, 2048, 1, 128] - - [120, 0.0] - - - [35200, 8192, 1, 128] - - [120, 0.0] - - - [30976, 128, 1, 128] - - [112, 0.0] - - - [32640, 16641, 1, 128] - - [120, 0.0] - - - [41984, 8192, 1, 128] - - [120, 0.0] - - - [30080, 128, 1, 128] - - [113, 0.0] - - - [35584, 512, 1, 128] - - [120, 0.0] - - - [44800, 2048, 1, 128] - - [120, 0.0] - - - [34048, 128, 1, 128] - - [118, 0.0] - - - [35712, 1024, 1, 128] - - [120, 0.0] - - - [43136, 128, 1, 128] - - [113, 0.0] - - - [33280, 1024, 1, 128] - - [120, 0.0] - - - [34816, 18945, 1, 128] - - [120, 0.0] - - - [40704, 8192, 1, 128] - - [120, 0.0] - - - [34304, 128, 1, 128] - - [111, 0.0] - - - [39936, 512, 1, 128] - - [120, 0.0] - - - [36096, 2048, 1, 128] - - [120, 0.0] - - - [40832, 8192, 1, 128] - - [120, 0.0] - - - [37760, 4096, 1, 128] - - [120, 0.0] - - - [36736, 512, 1, 128] - - [120, 0.0] - - - [31744, 8192, 1, 128] - - [120, 0.0] - - - [33920, 1024, 1, 128] - - [120, 0.0] - - - [39808, 128, 1, 128] - - [120, 0.0] - - - [36608, 2048, 1, 128] - - [120, 0.0] - - - [30464, 14593, 1, 128] - - [120, 0.0] - - - [35200, 19201, 1, 128] - - [113, 0.0] - - - [41472, 1024, 1, 128] - - [120, 0.0] - - - [30720, 128, 1, 128] - - [118, 0.0] - - - [41600, 128, 1, 128] - - [120, 0.0] - - - [38144, 22145, 1, 128] - - [120, 0.0] - - - [37120, 4096, 1, 128] - - [120, 0.0] - - - [40704, 24705, 1, 128] - - [120, 0.0] - - - [41088, 25217, 1, 128] - - [120, 0.0] - - - [43776, 8192, 1, 128] - - [120, 0.0] - - - [38912, 1024, 1, 128] - - [120, 0.0] - - - [43008, 2048, 1, 128] - - [120, 0.0] - - - [42496, 26497, 1, 128] - - [120, 0.0] - - - [33536, 512, 1, 128] - - [120, 0.0] - - - [43520, 512, 1, 128] - - [120, 0.0] - - - [39040, 23169, 1, 128] - - [120, 0.0] - - - [29568, 2048, 1, 128] - - [120, 0.0] - - - [44672, 8192, 1, 128] - - [120, 0.0] - - - [29824, 512, 1, 128] - - [120, 0.0] - - - [34944, 2048, 1, 128] - - [120, 0.0] - - - [33408, 4096, 1, 128] - - [120, 0.0] - - - [41600, 25729, 1, 128] - - [120, 0.0] - - - [40832, 2048, 1, 128] - - [120, 0.0] - - - [38912, 128, 1, 128] - - [111, 0.0] - - - [34048, 2048, 1, 128] - - [120, 0.0] - - - [43904, 2048, 1, 128] - - [120, 0.0] - - - [39296, 23297, 1, 128] - - [120, 0.0] - - - [31232, 4096, 1, 128] - - [120, 0.0] - - - [35840, 1024, 1, 128] - - [120, 0.0] - - - [28928, 128, 1, 128] - - [110, 0.0] - - - [42752, 2048, 1, 128] - - [120, 0.0] - - - [44032, 1024, 1, 128] - - [120, 0.0] - - - [29440, 13569, 1, 128] - - [120, 0.0] - - - [35456, 19585, 1, 128] - - [120, 0.0] - - - [35840, 19841, 1, 128] - - [120, 0.0] - - - [31360, 128, 1, 128] - - [112, 0.0] - - - [40192, 2048, 1, 128] - - [120, 0.0] - - - [33920, 8192, 1, 128] - - [120, 0.0] - - - [43648, 512, 1, 128] - - [120, 0.0] - - - [30080, 14209, 1, 128] - - [120, 0.0] - - - [39680, 23809, 1, 128] - - [120, 0.0] - - - [32512, 512, 1, 128] - - [120, 0.0] - - - [34816, 2048, 1, 128] - - [120, 0.0] - - - [43392, 1024, 1, 128] - - [120, 0.0] - - - [39040, 4096, 1, 128] - - [120, 0.0] - - - [43264, 4096, 1, 128] - - [120, 0.0] - - - [44416, 2048, 1, 128] - - [120, 0.0] - - - [31488, 512, 1, 128] - - [120, 0.0] - - - [31616, 1024, 1, 128] - - [120, 0.0] - - - [44032, 8192, 1, 128] - - [120, 0.0] - - - [39424, 23553, 1, 128] - - [120, 0.0] - - - [31360, 8192, 1, 128] - - [120, 0.0] - - - [42752, 128, 1, 128] - - [120, 0.0] - - - [40192, 512, 1, 128] - - [120, 0.0] - - - [36096, 20225, 1, 128] - - [120, 0.0] - - - [41984, 26113, 1, 128] - - [120, 0.0] - - - [39936, 2048, 1, 128] - - [120, 0.0] - - - [42880, 2048, 1, 128] - - [120, 0.0] - - - [29440, 128, 1, 128] - - [119, 0.0] - - - [40192, 128, 1, 128] - - [111, 0.0] - - - [36608, 4096, 1, 128] - - [120, 0.0] - - - [37760, 21761, 1, 128] - - [113, 0.0] - - - [44160, 28161, 1, 128] - - [120, 0.0] - - - [44288, 512, 1, 128] - - [120, 0.0] - - - [29056, 13185, 1, 128] - - [120, 0.0] - - - [43904, 512, 1, 128] - - [120, 0.0] - - - [29696, 128, 1, 128] - - [120, 0.0] - - - [36224, 8192, 1, 128] - - [120, 0.0] - - - [33024, 2048, 1, 128] - - [120, 0.0] - - - [44032, 28161, 1, 128] - - [120, 0.0] - - - [44032, 128, 1, 128] - - [113, 0.0] - - - [38784, 512, 1, 128] - - [120, 0.0] - - - [29056, 8192, 1, 128] - - [120, 0.0] - - - [33920, 18049, 1, 128] - - [120, 0.0] - - - [34816, 1024, 1, 128] - - [120, 0.0] - - - [29056, 128, 1, 128] - - [113, 0.0] - - - [39552, 1024, 1, 128] - - [120, 0.0] - - - [36992, 8192, 1, 128] - - [120, 0.0] - - - [44544, 1024, 1, 128] - - [120, 0.0] - - - [43904, 27905, 1, 128] - - [120, 0.0] - - - [29440, 512, 1, 128] - - [120, 0.0] - - - [29568, 8192, 1, 128] - - [120, 0.0] - - - [41472, 2048, 1, 128] - - [120, 0.0] - - - [29184, 8192, 1, 128] - - [120, 0.0] - - - [33408, 512, 1, 128] - - [120, 0.0] - - - [38656, 22785, 1, 128] - - [120, 0.0] - - - [31744, 15745, 1, 128] - - [120, 0.0] - - - [38656, 2048, 1, 128] - - [120, 0.0] - - - [30080, 8192, 1, 128] - - [120, 0.0] - - - [44672, 128, 1, 128] - - [120, 0.0] - - - [40704, 24833, 1, 128] - - [120, 0.0] - - - [33792, 8192, 1, 128] - - [120, 0.0] - - - [33920, 512, 1, 128] - - [120, 0.0] - - - [40576, 1024, 1, 128] - - [120, 0.0] - - - [36224, 20225, 1, 128] - - [120, 0.0] - - - [34432, 1024, 1, 128] - - [120, 0.0] - - - [31488, 15617, 1, 128] - - [120, 0.0] - - - [40576, 2048, 1, 128] - - [120, 0.0] - - - [30208, 512, 1, 128] - - [120, 0.0] - - - [36480, 128, 1, 128] - - [118, 0.0] - - - [37504, 21633, 1, 128] - - [120, 0.0] - - - [32896, 17025, 1, 128] - - [120, 0.0] - - - [39168, 2048, 1, 128] - - [120, 0.0] - - - [29440, 2048, 1, 128] - - [120, 0.0] - - - [29440, 13441, 1, 128] - - [120, 0.0] - - - [32640, 8192, 1, 128] - - [120, 0.0] - - - [35072, 19073, 1, 128] - - [120, 0.0] - - - [33152, 512, 1, 128] - - [120, 0.0] - - - [40576, 8192, 1, 128] - - [120, 0.0] - - - [34944, 8192, 1, 128] - - [120, 0.0] - - - [38656, 128, 1, 128] - - [119, 0.0] - - - [33536, 17537, 1, 128] - - [120, 0.0] - - - [29952, 512, 1, 128] - - [120, 0.0] - - - [31488, 2048, 1, 128] - - [120, 0.0] - - - [31872, 4096, 1, 128] - - [120, 0.0] - - - [31232, 15233, 1, 128] - - [120, 0.0] - - - [38912, 23041, 1, 128] - - [120, 0.0] - - - [31232, 2048, 1, 128] - - [120, 0.0] - - - [40448, 8192, 1, 128] - - [120, 0.0] - - - [36352, 128, 1, 128] - - [111, 0.0] - - - [43776, 4096, 1, 128] - - [120, 0.0] - - - [32000, 8192, 1, 128] - - [120, 0.0] - - - [37760, 8192, 1, 128] - - [120, 0.0] - - - [30080, 1024, 1, 128] - - [120, 0.0] - - - [44544, 128, 1, 128] - - [120, 0.0] - - - [29696, 1024, 1, 128] - - [120, 0.0] - - - [32640, 512, 1, 128] - - [120, 0.0] - - - [44416, 128, 1, 128] - - [120, 0.0] - - - [41216, 512, 1, 128] - - [120, 0.0] - - - [31872, 512, 1, 128] - - [120, 0.0] - - - [34432, 512, 1, 128] - - [120, 0.0] - - - [34560, 1024, 1, 128] - - [120, 0.0] - - - [42240, 128, 1, 128] - - [113, 0.0] - - - [44288, 28289, 1, 128] - - [120, 0.0] - - - [30336, 14337, 1, 128] - - [120, 0.0] - - - [32384, 2048, 1, 128] - - [120, 0.0] - - - [38400, 22401, 1, 128] - - [120, 0.0] - - - [39296, 1024, 1, 128] - - [120, 0.0] - - - [28928, 8192, 1, 128] - - [120, 0.0] - - - [40320, 2048, 1, 128] - - [120, 0.0] - - - [31104, 15233, 1, 128] - - [113, 0.0] - - - [39680, 512, 1, 128] - - [120, 0.0] - - - [34048, 18049, 1, 128] - - [120, 0.0] - - - [30720, 1024, 1, 128] - - [120, 0.0] - - - [42880, 26881, 1, 128] - - [120, 0.0] - - - [32896, 8192, 1, 128] - - [120, 0.0] - - - [43264, 8192, 1, 128] - - [120, 0.0] - - - [37632, 4096, 1, 128] - - [120, 0.0] - - - [32256, 4096, 1, 128] - - [120, 0.0] - - - [37248, 4096, 1, 128] - - [120, 0.0] - - - [33280, 17409, 1, 128] - - [120, 0.0] - - - [36096, 512, 1, 128] - - [120, 0.0] - - - [37120, 21121, 1, 128] - - [120, 0.0] - - - [32896, 128, 1, 128] - - [119, 0.0] - - - [36352, 20481, 1, 128] - - [120, 0.0] - - - [43392, 2048, 1, 128] - - [120, 0.0] - - - [36352, 512, 1, 128] - - [120, 0.0] - - - [29056, 13057, 1, 128] - - [120, 0.0] - - - [29056, 4096, 1, 128] - - [120, 0.0] - - - [37888, 4096, 1, 128] - - [120, 0.0] - - - [40320, 512, 1, 128] - - [120, 0.0] - - - [39168, 128, 1, 128] - - [119, 0.0] - - - [41472, 8192, 1, 128] - - [120, 0.0] - - - [34560, 512, 1, 128] - - [120, 0.0] - - - [34176, 18305, 1, 128] - - [120, 0.0] - - - [34688, 8192, 1, 128] - - [120, 0.0] - - - [29696, 13825, 1, 128] - - [120, 0.0] - - - [33152, 17281, 1, 128] - - [120, 0.0] - - - [30208, 8192, 1, 128] - - [120, 0.0] - - - [43648, 27649, 1, 128] - - [120, 0.0] - - - [31360, 2048, 1, 128] - - [120, 0.0] - - - [41984, 128, 1, 128] - - [118, 0.0] - - - [38528, 2048, 1, 128] - - [120, 0.0] - - - [32256, 16385, 1, 128] - - [120, 0.0] - - - [42240, 1024, 1, 128] - - [120, 0.0] - - - [32000, 16001, 1, 128] - - [113, 0.0] - - - [37248, 1024, 1, 128] - - [120, 0.0] - - - [32256, 1024, 1, 128] - - [120, 0.0] - - - [39296, 23425, 1, 128] - - [120, 0.0] - - - [43008, 4096, 1, 128] - - [120, 0.0] - - - [31104, 128, 1, 128] - - [119, 0.0] - - - [38656, 8192, 1, 128] - - [120, 0.0] - - - [44288, 128, 1, 128] - - [113, 0.0] - - - [38528, 22657, 1, 128] - - [120, 0.0] - - - [39552, 128, 1, 128] - - [120, 0.0] - - - [37376, 21377, 1, 128] - - [120, 0.0] - - - [28928, 13057, 1, 128] - - [120, 0.0] - - - [43264, 27265, 1, 128] - - [120, 0.0] - - - [35328, 4096, 1, 128] - - [120, 0.0] - - - [30848, 4096, 1, 128] - - [120, 0.0] - - - [44800, 28801, 1, 128] - - [120, 0.0] - - - [35456, 512, 1, 128] - - [120, 0.0] - - - [40960, 24961, 1, 128] - - [120, 0.0] - - - [39936, 23937, 1, 128] - - [120, 0.0] - - - [31744, 1024, 1, 128] - - [120, 0.0] - - - [32128, 8192, 1, 128] - - [120, 0.0] - - - [42112, 26113, 1, 128] - - [120, 0.0] - - - [31744, 2048, 1, 128] - - [120, 0.0] - - - [42112, 1024, 1, 128] - - [120, 0.0] - - - [40064, 8192, 1, 128] - - [120, 0.0] - - - [38144, 128, 1, 128] - - [120, 0.0] - - - [42624, 2048, 1, 128] - - [120, 0.0] - - - [36992, 128, 1, 128] - - [111, 0.0] - - - [40192, 8192, 1, 128] - - [120, 0.0] - - - [40064, 24065, 1, 128] - - [120, 0.0] - - - [37760, 21889, 1, 128] - - [120, 0.0] - - - [36352, 8192, 1, 128] - - [120, 0.0] - - - [44544, 512, 1, 128] - - [120, 0.0] - - - [32384, 4096, 1, 128] - - [120, 0.0] - - - [39168, 23169, 1, 128] - - [120, 0.0] - - - [1408, 897, 1, 128] - - [103, 0.0] - - - [16512, 512, 1, 128] - - [120, 0.0] - - - [20480, 12673, 1, 128] - - [120, 0.0] - - - [20992, 512, 1, 128] - - [120, 0.0] - - - [9344, 512, 1, 128] - - [120, 0.0] - - - [18048, 2048, 1, 128] - - [120, 0.0] - - - [20352, 12673, 1, 128] - - [120, 0.0] - - - [640, 128, 1, 128] - - [99, 0.0] - - - [28160, 512, 1, 128] - - [120, 0.0] - - - [20608, 4096, 1, 128] - - [120, 0.0] - - - [19328, 1024, 1, 128] - - [120, 0.0] - - - [26496, 4096, 1, 128] - - [120, 0.0] - - - [10624, 512, 1, 128] - - [120, 0.0] - - - [20352, 1024, 1, 128] - - [120, 0.0] - - - [10240, 6529, 1, 128] - - [120, 0.0] - - - [22144, 14465, 1, 128] - - [120, 0.0] - - - [13184, 2048, 1, 128] - - [120, 0.0] - - - [14720, 6913, 1, 128] - - [113, 0.0] - - - [21248, 512, 1, 128] - - [120, 0.0] - - - [10496, 128, 1, 128] - - [118, 0.0] - - - [13056, 5377, 1, 128] - - [120, 0.0] - - - [10880, 128, 1, 128] - - [113, 0.0] - - - [18688, 512, 1, 128] - - [120, 0.0] - - - [22656, 4096, 1, 128] - - [120, 0.0] - - - [15232, 1024, 1, 128] - - [120, 0.0] - - - [20224, 4096, 1, 128] - - [120, 0.0] - - - [6016, 2305, 1, 128] - - [113, 0.0] - - - [13184, 4096, 1, 128] - - [120, 0.0] - - - [256, 129, 1, 128] - - [116, 0.0] - - - [11264, 7553, 1, 128] - - [120, 0.0] - - - [18176, 128, 1, 128] - - [111, 0.0] - - - [15872, 8193, 1, 128] - - [120, 0.0] - - - [26112, 4096, 1, 128] - - [120, 0.0] - - - [22784, 2048, 1, 128] - - [120, 0.0] - - - [10880, 7297, 1, 128] - - [120, 0.0] - - - [14720, 2048, 1, 128] - - [120, 0.0] - - - [9216, 5633, 1, 128] - - [120, 0.0] - - - [23040, 15233, 1, 128] - - [120, 0.0] - - - [8832, 5121, 1, 128] - - [120, 0.0] - - - [18816, 1024, 1, 128] - - [120, 0.0] - - - [128, 129, 1, 128] - - [99, 0.0] - - - [15488, 512, 1, 128] - - [120, 0.0] - - - [18176, 1024, 1, 128] - - [120, 0.0] - - - [16128, 8449, 1, 128] - - [120, 0.0] - - - [16000, 2048, 1, 128] - - [120, 0.0] - - - [24960, 9089, 1, 128] - - [120, 0.0] - - - [14336, 1024, 1, 128] - - [120, 0.0] - - - [25472, 8192, 1, 128] - - [120, 0.0] - - - [23040, 128, 1, 128] - - [118, 0.0] - - - [9472, 512, 1, 128] - - [118, 0.0] - - - [19072, 128, 1, 128] - - [118, 0.0] - - - [10624, 6913, 1, 128] - - [120, 0.0] - - - [7808, 1024, 1, 128] - - [120, 0.0] - - - [27008, 11137, 1, 128] - - [120, 0.0] - - - [21504, 4096, 1, 128] - - [120, 0.0] - - - [7936, 1024, 1, 128] - - [118, 0.0] - - - [12928, 5121, 1, 128] - - [120, 0.0] - - - [26240, 8192, 1, 128] - - [120, 0.0] - - - [18304, 2048, 1, 128] - - [120, 0.0] - - - [24576, 1024, 1, 128] - - [120, 0.0] - - - [10624, 128, 1, 128] - - [111, 0.0] - - - [24576, 128, 1, 128] - - [120, 0.0] - - - [25600, 9601, 1, 128] - - [120, 0.0] - - - [5248, 128, 1, 128] - - [110, 0.0] - - - [24448, 4096, 1, 128] - - [120, 0.0] - - - [19328, 128, 1, 128] - - [119, 0.0] - - - [24064, 512, 1, 128] - - [118, 0.0] - - - [11136, 512, 1, 128] - - [118, 0.0] - - - [14592, 1024, 1, 128] - - [120, 0.0] - - - [12544, 4737, 1, 128] - - [120, 0.0] - - - [17280, 128, 1, 128] - - [112, 0.0] - - - [25344, 8192, 1, 128] - - [120, 0.0] - - - [4608, 512, 1, 128] - - [118, 0.0] - - - [4608, 128, 1, 128] - - [85, 0.0] - - - [21760, 512, 1, 128] - - [120, 0.0] - - - [7936, 128, 1, 128] - - [96, 0.0] - - - [11008, 7425, 1, 128] - - [113, 0.0] - - - [13824, 2048, 1, 128] - - [120, 0.0] - - - [18048, 512, 1, 128] - - [120, 0.0] - - - [19584, 11905, 1, 128] - - [120, 0.0] - - - [22656, 512, 1, 128] - - [120, 0.0] - - - [4608, 3073, 1, 128] - - [120, 0.0] - - - [5504, 128, 1, 128] - - [117, 0.0] - - - [4864, 1024, 1, 128] - - [120, 0.0] - - - [17664, 1024, 1, 128] - - [120, 0.0] - - - [18176, 2048, 1, 128] - - [120, 0.0] - - - [2048, 1537, 1, 128] - - [111, 0.0] - - - [22528, 128, 1, 128] - - [113, 0.0] - - - [21760, 13953, 1, 128] - - [120, 0.0] - - - [7040, 128, 1, 128] - - [111, 0.0] - - - [3328, 1665, 1, 128] - - [113, 0.0] - - - [768, 512, 1, 128] - - [104, 0.0] - - - [21504, 13697, 1, 128] - - [120, 0.0] - - - [18560, 10881, 1, 128] - - [113, 0.0] - - - [2560, 128, 1, 128] - - [100, 0.0] - - - [15616, 1024, 1, 128] - - [120, 0.0] - - - [19456, 4096, 1, 128] - - [120, 0.0] - - - [25600, 2048, 1, 128] - - [120, 0.0] - - - [2304, 128, 1, 128] - - [100, 0.0] - - - [1664, 1025, 1, 128] - - [124, 0.0] - - - [23168, 15361, 1, 128] - - [120, 0.0] - - - [9856, 128, 1, 128] - - [117, 0.0] - - - [13312, 2048, 1, 128] - - [120, 0.0] - - - [19200, 512, 1, 128] - - [120, 0.0] - - - [19200, 2048, 1, 128] - - [120, 0.0] - - - [23168, 2048, 1, 128] - - [120, 0.0] - - - [18688, 128, 1, 128] - - [111, 0.0] - - - [13568, 1024, 1, 128] - - [120, 0.0] - - - [17792, 9985, 1, 128] - - [113, 0.0] - - - [20608, 1024, 1, 128] - - [120, 0.0] - - - [11648, 8065, 1, 128] - - [120, 0.0] - - - [1280, 128, 1, 128] - - [110, 0.0] - - - [16256, 4096, 1, 128] - - [120, 0.0] - - - [17024, 1024, 1, 128] - - [120, 0.0] - - - [19456, 128, 1, 128] - - [119, 0.0] - - - [20736, 512, 1, 128] - - [120, 0.0] - - - [14464, 6785, 1, 128] - - [113, 0.0] - - - [20736, 13057, 1, 128] - - [120, 0.0] - - - [8704, 2048, 1, 128] - - [120, 0.0] - - - [640, 512, 1, 128] - - [100, 0.0] - - - [768, 129, 1, 128] - - [99, 0.0] - - - [27776, 1024, 1, 128] - - [120, 0.0] - - - [19200, 11521, 1, 128] - - [120, 0.0] - - - [6400, 2048, 1, 128] - - [120, 0.0] - - - [14976, 7297, 1, 128] - - [113, 0.0] - - - [7040, 2048, 1, 128] - - [120, 0.0] - - - [25984, 128, 1, 128] - - [119, 0.0] - - - [13696, 128, 1, 128] - - [112, 0.0] - - - [2688, 1153, 1, 128] - - [120, 0.0] - - - [15232, 2048, 1, 128] - - [120, 0.0] - - - [11776, 128, 1, 128] - - [118, 0.0] - - - [3328, 512, 1, 128] - - [111, 0.0] - - - [11648, 7937, 1, 128] - - [113, 0.0] - - - [19456, 2048, 1, 128] - - [120, 0.0] - - - [11008, 128, 1, 128] - - [111, 0.0] - - - [9984, 6401, 1, 128] - - [113, 0.0] - - - [25856, 9857, 1, 128] - - [120, 0.0] - - - [4224, 512, 1, 128] - - [119, 0.0] - - - [13568, 5761, 1, 128] - - [120, 0.0] - - - [5632, 2049, 1, 128] - - [118, 0.0] - - - [8832, 2048, 1, 128] - - [120, 0.0] - - - [5632, 3969, 1, 128] - - [120, 0.0] - - - [25856, 2048, 1, 128] - - [120, 0.0] - - - [25472, 2048, 1, 128] - - [120, 0.0] - - - [20736, 12929, 1, 128] - - [113, 0.0] - - - [14592, 128, 1, 128] - - [112, 0.0] - - - [1792, 512, 1, 128] - - [105, 0.0] - - - [14208, 2048, 1, 128] - - [120, 0.0] - - - [15360, 7681, 1, 128] - - [120, 0.0] - - - [5760, 2048, 1, 128] - - [120, 0.0] - - - [6400, 512, 1, 128] - - [118, 0.0] - - - [5248, 3713, 1, 128] - - [113, 0.0] - - - [16768, 1024, 1, 128] - - [120, 0.0] - - - [10752, 512, 1, 128] - - [118, 0.0] - - - [26624, 2048, 1, 128] - - [120, 0.0] - - - [384, 128, 1, 128] - - [101, 0.0] - - - [27392, 8192, 1, 128] - - [120, 0.0] - - - [24448, 512, 1, 128] - - [120, 0.0] - - - [11136, 7553, 1, 128] - - [113, 0.0] - - - [17024, 9345, 1, 128] - - [113, 0.0] - - - [16000, 8193, 1, 128] - - [120, 0.0] - - - [5888, 2048, 1, 128] - - [120, 0.0] - - - [18304, 10497, 1, 128] - - [113, 0.0] - - - [3968, 128, 1, 128] - - [92, 0.0] - - - [14336, 6529, 1, 128] - - [120, 0.0] - - - [19840, 128, 1, 128] - - [119, 0.0] - - - [25600, 8192, 1, 128] - - [120, 0.0] - - - [18688, 11009, 1, 128] - - [113, 0.0] - - - [7680, 1024, 1, 128] - - [120, 0.0] - - - [7168, 128, 1, 128] - - [118, 0.0] - - - [1664, 512, 1, 128] - - [84, 0.0] - - - [12544, 1024, 1, 128] - - [120, 0.0] - - - [6528, 2048, 1, 128] - - [120, 0.0] - - - [19072, 4096, 1, 128] - - [120, 0.0] - - - [2048, 512, 1, 128] - - [125, 0.0] - - - [13568, 5889, 1, 128] - - [120, 0.0] - - - [23680, 16001, 1, 128] - - [120, 0.0] - - - [26112, 10113, 1, 128] - - [120, 0.0] - - - [15872, 128, 1, 128] - - [112, 0.0] - - - [16384, 512, 1, 128] - - [120, 0.0] - - - [9856, 6273, 1, 128] - - [113, 0.0] - - - [26368, 1024, 1, 128] - - [120, 0.0] - - - [16256, 2048, 1, 128] - - [120, 0.0] - - - [3968, 2305, 1, 128] - - [113, 0.0] - - - [28672, 8192, 1, 128] - - [120, 0.0] - - - [10368, 1024, 1, 128] - - [120, 0.0] - - - [11008, 1024, 1, 128] - - [120, 0.0] - - - [11776, 4097, 1, 128] - - [120, 0.0] - - - [26496, 2048, 1, 128] - - [120, 0.0] - - - [17792, 4096, 1, 128] - - [120, 0.0] - - - [2304, 512, 1, 128] - - [119, 0.0] - - - [9216, 2048, 1, 128] - - [118, 0.0] - - - [12416, 512, 1, 128] - - [120, 0.0] - - - [18048, 128, 1, 128] - - [119, 0.0] - - - [21888, 14209, 1, 128] - - [120, 0.0] - - - [9344, 5761, 1, 128] - - [120, 0.0] - - - [19712, 2048, 1, 128] - - [120, 0.0] - - - [12288, 1024, 1, 128] - - [120, 0.0] - - - [3584, 1921, 1, 128] - - [120, 0.0] - - - [22784, 128, 1, 128] - - [118, 0.0] - - - [26880, 128, 1, 128] - - [120, 0.0] - - - [17408, 1024, 1, 128] - - [120, 0.0] - - - [15488, 4096, 1, 128] - - [120, 0.0] - - - [13312, 5633, 1, 128] - - [120, 0.0] - - - [22016, 14337, 1, 128] - - [120, 0.0] - - - [19328, 2048, 1, 128] - - [120, 0.0] - - - [25600, 128, 1, 128] - - [119, 0.0] - - - [22784, 15105, 1, 128] - - [120, 0.0] - - - [5376, 3713, 1, 128] - - [113, 0.0] - - - [14208, 512, 1, 128] - - [118, 0.0] - - - [12928, 4096, 1, 128] - - [120, 0.0] - - - [768, 257, 1, 128] - - [101, 0.0] - - - [27776, 11777, 1, 128] - - [120, 0.0] - - - [12032, 1024, 1, 128] - - [120, 0.0] - - - [14208, 4096, 1, 128] - - [120, 0.0] - - - [19840, 12161, 1, 128] - - [120, 0.0] - - - [17536, 512, 1, 128] - - [120, 0.0] - - - [19840, 4096, 1, 128] - - [120, 0.0] - - - [26624, 512, 1, 128] - - [120, 0.0] - - - [27136, 11137, 1, 128] - - [120, 0.0] - - - [11008, 512, 1, 128] - - [120, 0.0] - - - [1024, 513, 1, 128] - - [85, 0.0] - - - [15744, 512, 1, 128] - - [120, 0.0] - - - [22016, 128, 1, 128] - - [119, 0.0] - - - [9344, 1024, 1, 128] - - [120, 0.0] - - - [28544, 1024, 1, 128] - - [120, 0.0] - - - [13440, 5633, 1, 128] - - [120, 0.0] - - - [21632, 13825, 1, 128] - - [120, 0.0] - - - [24064, 4096, 1, 128] - - [120, 0.0] - - - [24192, 512, 1, 128] - - [120, 0.0] - - - [22912, 15233, 1, 128] - - [120, 0.0] - - - [20864, 13185, 1, 128] - - [113, 0.0] - - - [8064, 4353, 1, 128] - - [113, 0.0] - - - [8704, 5121, 1, 128] - - [120, 0.0] - - - [19840, 1024, 1, 128] - - [120, 0.0] - - - [15616, 128, 1, 128] - - [112, 0.0] - - - [21632, 512, 1, 128] - - [120, 0.0] - - - [13440, 512, 1, 128] - - [118, 0.0] - - - [23936, 128, 1, 128] - - [119, 0.0] - - - [8960, 5377, 1, 128] - - [120, 0.0] - - - [27008, 512, 1, 128] - - [120, 0.0] - - - [13440, 5761, 1, 128] - - [120, 0.0] - - - [3072, 512, 1, 128] - - [111, 0.0] - - - [4096, 1024, 1, 128] - - [118, 0.0] - - - [7296, 3585, 1, 128] - - [120, 0.0] - - - [12416, 4737, 1, 128] - - [120, 0.0] - - - [6912, 512, 1, 128] - - [120, 0.0] - - - [11136, 2048, 1, 128] - - [120, 0.0] - - - [18176, 10369, 1, 128] - - [120, 0.0] - - - [14976, 4096, 1, 128] - - [120, 0.0] - - - [19712, 4096, 1, 128] - - [120, 0.0] - - - [8064, 1024, 1, 128] - - [120, 0.0] - - - [9600, 128, 1, 128] - - [119, 0.0] - - - [26240, 1024, 1, 128] - - [120, 0.0] - - - [5248, 3585, 1, 128] - - [120, 0.0] - - - [16768, 2048, 1, 128] - - [120, 0.0] - - - [13184, 128, 1, 128] - - [113, 0.0] - - - [19328, 11521, 1, 128] - - [120, 0.0] - - - [4864, 512, 1, 128] - - [112, 0.0] - - - [3584, 2049, 1, 128] - - [118, 0.0] - - - [18560, 128, 1, 128] - - [111, 0.0] - - - [27392, 11393, 1, 128] - - [120, 0.0] - - - [27520, 512, 1, 128] - - [120, 0.0] - - - [18176, 4096, 1, 128] - - [120, 0.0] - - - [7808, 4225, 1, 128] - - [120, 0.0] - - - [15232, 128, 1, 128] - - [112, 0.0] - - - [25728, 1024, 1, 128] - - [120, 0.0] - - - [23936, 512, 1, 128] - - [120, 0.0] - - - [23424, 2048, 1, 128] - - [120, 0.0] - - - [28032, 12161, 1, 128] - - [120, 0.0] - - - [27136, 512, 1, 128] - - [120, 0.0] - - - [14336, 6657, 1, 128] - - [120, 0.0] - - - [15616, 4096, 1, 128] - - [120, 0.0] - - - [3328, 1793, 1, 128] - - [120, 0.0] - - - [28416, 512, 1, 128] - - [120, 0.0] - - - [16384, 8705, 1, 128] - - [120, 0.0] - - - [3200, 1537, 1, 128] - - [120, 0.0] - - - [26368, 128, 1, 128] - - [120, 0.0] - - - [16000, 512, 1, 128] - - [120, 0.0] - - - [25216, 9345, 1, 128] - - [120, 0.0] - - - [28288, 4096, 1, 128] - - [120, 0.0] - - - [24832, 512, 1, 128] - - [120, 0.0] - - - [18048, 10369, 1, 128] - - [120, 0.0] - - - [20480, 4096, 1, 128] - - [120, 0.0] - - - [17792, 10113, 1, 128] - - [120, 0.0] - - - [13312, 5505, 1, 128] - - [120, 0.0] - - - [17024, 2048, 1, 128] - - [120, 0.0] - - - [20608, 12929, 1, 128] - - [120, 0.0] - - - [16896, 4096, 1, 128] - - [120, 0.0] - - - [27776, 2048, 1, 128] - - [120, 0.0] - - - [6912, 3201, 1, 128] - - [113, 0.0] - - - [15744, 2048, 1, 128] - - [120, 0.0] - - - [24448, 128, 1, 128] - - [112, 0.0] - - - [2688, 128, 1, 128] - - [100, 0.0] - - - [7808, 2048, 1, 128] - - [120, 0.0] - - - [1408, 512, 1, 128] - - [110, 0.0] - - - [12032, 512, 1, 128] - - [120, 0.0] - - - [26752, 512, 1, 128] - - [120, 0.0] - - - [16128, 8321, 1, 128] - - [120, 0.0] - - - [25856, 128, 1, 128] - - [112, 0.0] - - - [24064, 8192, 1, 128] - - [120, 0.0] - - - [28160, 4096, 1, 128] - - [120, 0.0] - - - [13312, 128, 1, 128] - - [118, 0.0] - - - [10112, 6401, 1, 128] - - [113, 0.0] - - - [16384, 4096, 1, 128] - - [120, 0.0] - - - [16512, 2048, 1, 128] - - [120, 0.0] - - - [27520, 11521, 1, 128] - - [120, 0.0] - - - [8192, 4481, 1, 128] - - [120, 0.0] - - - [16768, 512, 1, 128] - - [120, 0.0] - - - [6144, 128, 1, 128] - - [92, 0.0] - - - [13568, 512, 1, 128] - - [120, 0.0] - - - [9344, 5633, 1, 128] - - [120, 0.0] - - - [13440, 4096, 1, 128] - - [120, 0.0] - - - [2176, 1665, 1, 128] - - [118, 0.0] - - - [28288, 128, 1, 128] - - [113, 0.0] - - - [11776, 4096, 1, 128] - - [120, 0.0] - - - [17280, 512, 1, 128] - - [120, 0.0] - - - [5504, 3841, 1, 128] - - [120, 0.0] - - - [14848, 7041, 1, 128] - - [120, 0.0] - - - [3584, 128, 1, 128] - - [102, 0.0] - - - [26880, 8192, 1, 128] - - [120, 0.0] - - - [2944, 1409, 1, 128] - - [111, 0.0] - - - [26368, 10369, 1, 128] - - [113, 0.0] - - - [21888, 512, 1, 128] - - [120, 0.0] - - - [15872, 2048, 1, 128] - - [120, 0.0] - - - [20224, 512, 1, 128] - - [120, 0.0] - - - [24320, 8449, 1, 128] - - [113, 0.0] - - - [5632, 1024, 1, 128] - - [120, 0.0] - - - [17152, 9473, 1, 128] - - [120, 0.0] - - - [4096, 128, 1, 128] - - [126, 0.0] - - - [8832, 128, 1, 128] - - [110, 0.0] - - - [2048, 1409, 1, 128] - - [112, 0.0] - - - [28160, 12289, 1, 128] - - [120, 0.0] - - - [9088, 5505, 1, 128] - - [113, 0.0] - - - [19200, 1024, 1, 128] - - [120, 0.0] - - - [18048, 4096, 1, 128] - - [120, 0.0] - - - [12928, 512, 1, 128] - - [118, 0.0] - - - [20864, 4096, 1, 128] - - [120, 0.0] - - - [27008, 2048, 1, 128] - - [120, 0.0] - - - [16640, 128, 1, 128] - - [117, 0.0] - - - [24960, 8192, 1, 128] - - [120, 0.0] - - - [24320, 1024, 1, 128] - - [120, 0.0] - - - [23552, 15873, 1, 128] - - [120, 0.0] - - - [26240, 4096, 1, 128] - - [120, 0.0] - - - [24320, 128, 1, 128] - - [120, 0.0] - - - [26240, 128, 1, 128] - - [118, 0.0] - - - [3200, 1665, 1, 128] - - [120, 0.0] - - - [11776, 2048, 1, 128] - - [120, 0.0] - - - [6144, 512, 1, 128] - - [118, 0.0] - - - [24960, 128, 1, 128] - - [120, 0.0] - - - [23424, 128, 1, 128] - - [110, 0.0] - - - [11776, 8065, 1, 128] - - [120, 0.0] - - - [19072, 11265, 1, 128] - - [120, 0.0] - - - [8192, 4609, 1, 128] - - [120, 0.0] - - - [21888, 4096, 1, 128] - - [120, 0.0] - - - [14976, 2048, 1, 128] - - [120, 0.0] - - - [23680, 4096, 1, 128] - - [120, 0.0] - - - [14080, 1024, 1, 128] - - [120, 0.0] - - - [19968, 4096, 1, 128] - - [120, 0.0] - - - [8704, 128, 1, 128] - - [110, 0.0] - - - [23424, 15745, 1, 128] - - [113, 0.0] - - - [8320, 2048, 1, 128] - - [120, 0.0] - - - [6144, 2433, 1, 128] - - [113, 0.0] - - - [19200, 11393, 1, 128] - - [120, 0.0] - - - [28416, 128, 1, 128] - - [111, 0.0] - - - [14080, 2048, 1, 128] - - [120, 0.0] - - - [12544, 4096, 1, 128] - - [120, 0.0] - - - [17024, 128, 1, 128] - - [112, 0.0] - - - [23936, 16257, 1, 128] - - [120, 0.0] - - - [12288, 128, 1, 128] - - [113, 0.0] - - - [28800, 1024, 1, 128] - - [120, 0.0] - - - [13824, 6017, 1, 128] - - [120, 0.0] - - - [23040, 2048, 1, 128] - - [120, 0.0] - - - [9984, 6273, 1, 128] - - [120, 0.0] - - - [23680, 512, 1, 128] - - [120, 0.0] - - - [7936, 4353, 1, 128] - - [113, 0.0] - - - [24192, 2048, 1, 128] - - [120, 0.0] - - - [8448, 512, 1, 128] - - [120, 0.0] - - - [5760, 2177, 1, 128] - - [113, 0.0] - - - [22656, 14977, 1, 128] - - [120, 0.0] - - - [17024, 4096, 1, 128] - - [120, 0.0] - - - [24960, 8961, 1, 128] - - [120, 0.0] - - - [5888, 1024, 1, 128] - - [120, 0.0] - - - [9344, 2048, 1, 128] - - [120, 0.0] - - - [11520, 1024, 1, 128] - - [120, 0.0] - - - [17024, 9217, 1, 128] - - [120, 0.0] - - - [10368, 6657, 1, 128] - - [120, 0.0] - - - [21632, 2048, 1, 128] - - [120, 0.0] - - - [26880, 2048, 1, 128] - - [120, 0.0] - - - [20736, 4096, 1, 128] - - [120, 0.0] - - - [26624, 8192, 1, 128] - - [120, 0.0] - - - [26752, 2048, 1, 128] - - [120, 0.0] - - - [24192, 8321, 1, 128] - - [120, 0.0] - - - [4736, 1024, 1, 128] - - [118, 0.0] - - - [27648, 8192, 1, 128] - - [120, 0.0] - - - [27392, 11521, 1, 128] - - [120, 0.0] - - - [27776, 4096, 1, 128] - - [120, 0.0] - - - [28672, 12801, 1, 128] - - [120, 0.0] - - - [13056, 512, 1, 128] - - [120, 0.0] - - - [25088, 2048, 1, 128] - - [120, 0.0] - - - [17408, 9601, 1, 128] - - [120, 0.0] - - - [5120, 3585, 1, 128] - - [120, 0.0] - - - [13824, 512, 1, 128] - - [118, 0.0] - - - [8576, 1024, 1, 128] - - [120, 0.0] - - - [16768, 4096, 1, 128] - - [120, 0.0] - - - [25728, 9729, 1, 128] - - [120, 0.0] - - - [27392, 512, 1, 128] - - [120, 0.0] - - - [13824, 128, 1, 128] - - [119, 0.0] - - - [27264, 1024, 1, 128] - - [120, 0.0] - - - [22272, 14465, 1, 128] - - [120, 0.0] - - - [19840, 2048, 1, 128] - - [120, 0.0] - - - [18176, 10497, 1, 128] - - [120, 0.0] - - - [4992, 3329, 1, 128] - - [113, 0.0] - - - [14976, 7169, 1, 128] - - [120, 0.0] - - - [10112, 512, 1, 128] - - [120, 0.0] - - - [24704, 128, 1, 128] - - [110, 0.0] - - - [16896, 128, 1, 128] - - [117, 0.0] - - - [10880, 7169, 1, 128] - - [120, 0.0] - - - [9600, 512, 1, 128] - - [118, 0.0] - - - [22528, 1024, 1, 128] - - [120, 0.0] - - - [27008, 128, 1, 128] - - [119, 0.0] - - - [4480, 2945, 1, 128] - - [113, 0.0] - - - [15872, 8065, 1, 128] - - [120, 0.0] - - - [28672, 128, 1, 128] - - [113, 0.0] - - - [9344, 128, 1, 128] - - [110, 0.0] - - - [15360, 2048, 1, 128] - - [120, 0.0] - - - [11392, 512, 1, 128] - - [118, 0.0] - - - [9216, 128, 1, 128] - - [112, 0.0] - - - [8192, 2048, 1, 128] - - [120, 0.0] - - - [14464, 1024, 1, 128] - - [120, 0.0] - - - [4096, 2433, 1, 128] - - [120, 0.0] - - - [6528, 2945, 1, 128] - - [120, 0.0] - - - [12672, 512, 1, 128] - - [120, 0.0] - - - [26624, 128, 1, 128] - - [111, 0.0] - - - [19712, 1024, 1, 128] - - [120, 0.0] - - - [4480, 2817, 1, 128] - - [113, 0.0] - - - [13440, 2048, 1, 128] - - [120, 0.0] - - - [256, 257, 1, 128] - - [114, 0.0] - - - [16000, 128, 1, 128] - - [112, 0.0] - - - [7552, 3969, 1, 128] - - [113, 0.0] - - - [12416, 2048, 1, 128] - - [120, 0.0] - - - [18432, 512, 1, 128] - - [120, 0.0] - - - [14464, 512, 1, 128] - - [118, 0.0] - - - [1280, 769, 1, 128] - - [121, 0.0] - - - [14976, 512, 1, 128] - - [120, 0.0] - - - [28032, 4096, 1, 128] - - [120, 0.0] - - - [27904, 128, 1, 128] - - [113, 0.0] - - - [20224, 12545, 1, 128] - - [120, 0.0] - - - [15872, 4096, 1, 128] - - [120, 0.0] - - - [3456, 1793, 1, 128] - - [113, 0.0] - - - [14336, 128, 1, 128] - - [112, 0.0] - - - [21248, 2048, 1, 128] - - [120, 0.0] - - - [23040, 1024, 1, 128] - - [120, 0.0] - - - [15232, 7425, 1, 128] - - [120, 0.0] - - - [14592, 512, 1, 128] - - [120, 0.0] - - - [22912, 15105, 1, 128] - - [120, 0.0] - - - [22528, 2048, 1, 128] - - [120, 0.0] - - - [3072, 1024, 1, 128] - - [120, 0.0] - - - [17536, 4096, 1, 128] - - [120, 0.0] - - - [384, 257, 1, 128] - - [101, 0.0] - - - [14464, 6657, 1, 128] - - [120, 0.0] - - - [20096, 1024, 1, 128] - - [120, 0.0] - - - [26880, 4096, 1, 128] - - [120, 0.0] - - - [18816, 2048, 1, 128] - - [120, 0.0] - - - [17152, 512, 1, 128] - - [120, 0.0] - - - [18432, 4096, 1, 128] - - [120, 0.0] - - - [10368, 2048, 1, 128] - - [120, 0.0] - - - [1408, 769, 1, 128] - - [102, 0.0] - - - [7168, 2048, 1, 128] - - [120, 0.0] - - - [17664, 128, 1, 128] - - [119, 0.0] - - - [1152, 513, 1, 128] - - [113, 0.0] - - - [7296, 3713, 1, 128] - - [120, 0.0] - - - [24064, 2048, 1, 128] - - [120, 0.0] - - - [8576, 2048, 1, 128] - - [120, 0.0] - - - [23168, 15489, 1, 128] - - [113, 0.0] - - - [14848, 7169, 1, 128] - - [120, 0.0] - - - [2432, 512, 1, 128] - - [110, 0.0] - - - [19712, 12033, 1, 128] - - [120, 0.0] - - - [25856, 4096, 1, 128] - - [120, 0.0] - - - [17152, 9345, 1, 128] - - [120, 0.0] - - - [3712, 128, 1, 128] - - [104, 0.0] - - - [22272, 128, 1, 128] - - [112, 0.0] - - - [25600, 9729, 1, 128] - - [120, 0.0] - - - [6016, 2433, 1, 128] - - [113, 0.0] - - - [12928, 128, 1, 128] - - [118, 0.0] - - - [25088, 8192, 1, 128] - - [120, 0.0] - - - [7040, 1024, 1, 128] - - [120, 0.0] - - - [4736, 3201, 1, 128] - - [120, 0.0] - - - [16000, 1024, 1, 128] - - [120, 0.0] - - - [1920, 512, 1, 128] - - [111, 0.0] - - - [8192, 1024, 1, 128] - - [120, 0.0] - - - [8448, 4865, 1, 128] - - [113, 0.0] - - - [11136, 7425, 1, 128] - - [113, 0.0] - - - [23296, 4096, 1, 128] - - [120, 0.0] - - - [27904, 2048, 1, 128] - - [120, 0.0] - - - [23552, 4096, 1, 128] - - [120, 0.0] - - - [24960, 2048, 1, 128] - - [120, 0.0] - - - [2816, 128, 1, 128] - - [100, 0.0] - - - [7424, 3841, 1, 128] - - [113, 0.0] - - - [20480, 128, 1, 128] - - [111, 0.0] - - - [18816, 11137, 1, 128] - - [120, 0.0] - - - [26496, 128, 1, 128] - - [111, 0.0] - - - [16896, 9217, 1, 128] - - [120, 0.0] - - - [23296, 512, 1, 128] - - [120, 0.0] - - - [8064, 2048, 1, 128] - - [120, 0.0] - - - [19968, 128, 1, 128] - - [119, 0.0] - - - [8320, 4737, 1, 128] - - [113, 0.0] - - - [27648, 1024, 1, 128] - - [120, 0.0] - - - [3712, 512, 1, 128] - - [119, 0.0] - - - [256, 128, 1, 128] - - [99, 0.0] - - - [3072, 1537, 1, 128] - - [120, 0.0] - - - [5504, 1024, 1, 128] - - [120, 0.0] - - - [20992, 2048, 1, 128] - - [120, 0.0] - - - [20480, 1024, 1, 128] - - [120, 0.0] - - - [20864, 128, 1, 128] - - [113, 0.0] - - - [28544, 12545, 1, 128] - - [120, 0.0] - - - [1152, 512, 1, 128] - - [70, 0.0] - - - [24320, 8321, 1, 128] - - [120, 0.0] - - - [2688, 512, 1, 128] - - [118, 0.0] - - - [27904, 8192, 1, 128] - - [120, 0.0] - - - [3840, 2177, 1, 128] - - [113, 0.0] - - - [25344, 128, 1, 128] - - [117, 0.0] - - - [13184, 512, 1, 128] - - [120, 0.0] - - - [7680, 512, 1, 128] - - [118, 0.0] - - - [11904, 2048, 1, 128] - - [120, 0.0] - - - [12544, 512, 1, 128] - - [118, 0.0] - - - [8448, 4737, 1, 128] - - [120, 0.0] - - - [28544, 128, 1, 128] - - [111, 0.0] - - - [21760, 14081, 1, 128] - - [120, 0.0] - - - [12800, 128, 1, 128] - - [111, 0.0] - - - [17664, 4096, 1, 128] - - [120, 0.0] - - - [2432, 1793, 1, 128] - - [120, 0.0] - - - [16384, 8577, 1, 128] - - [120, 0.0] - - - [28544, 512, 1, 128] - - [120, 0.0] - - - [28032, 12033, 1, 128] - - [120, 0.0] - - - [4864, 3329, 1, 128] - - [113, 0.0] - - - [12928, 5249, 1, 128] - - [113, 0.0] - - - [4736, 512, 1, 128] - - [118, 0.0] - - - [27264, 2048, 1, 128] - - [120, 0.0] - - - [19840, 12033, 1, 128] - - [113, 0.0] - - - [19584, 4096, 1, 128] - - [120, 0.0] - - - [21376, 4096, 1, 128] - - [120, 0.0] - - - [20352, 4096, 1, 128] - - [120, 0.0] - - - [6400, 2689, 1, 128] - - [120, 0.0] - - - [24704, 8192, 1, 128] - - [120, 0.0] - - - [22528, 14849, 1, 128] - - [120, 0.0] - - - [18304, 512, 1, 128] - - [120, 0.0] - - - [6656, 1024, 1, 128] - - [118, 0.0] - - - [13568, 4096, 1, 128] - - [120, 0.0] - - - [6016, 512, 1, 128] - - [118, 0.0] - - - [17664, 2048, 1, 128] - - [120, 0.0] - - - [17408, 512, 1, 128] - - [120, 0.0] - - - [24960, 4096, 1, 128] - - [120, 0.0] - - - [20608, 12801, 1, 128] - - [120, 0.0] - - - [27648, 11649, 1, 128] - - [120, 0.0] - - - [5760, 128, 1, 128] - - [102, 0.0] - - - [17792, 512, 1, 128] - - [120, 0.0] - - - [17664, 512, 1, 128] - - [120, 0.0] - - - [19968, 12161, 1, 128] - - [120, 0.0] - - - [19840, 512, 1, 128] - - [120, 0.0] - - - [12032, 4353, 1, 128] - - [120, 0.0] - - - [25984, 512, 1, 128] - - [120, 0.0] - - - [27648, 4096, 1, 128] - - [120, 0.0] - - - [10752, 7041, 1, 128] - - [120, 0.0] - - - [28544, 2048, 1, 128] - - [120, 0.0] - - - [7680, 2048, 1, 128] - - [120, 0.0] - - - [13184, 5377, 1, 128] - - [113, 0.0] - - - [6784, 3201, 1, 128] - - [120, 0.0] - - - [16384, 2048, 1, 128] - - [120, 0.0] - - - [22656, 1024, 1, 128] - - [120, 0.0] - - - [12800, 512, 1, 128] - - [120, 0.0] - - - [23936, 1024, 1, 128] - - [120, 0.0] - - - [15360, 1024, 1, 128] - - [120, 0.0] - - - [15488, 2048, 1, 128] - - [120, 0.0] - - - [11392, 1024, 1, 128] - - [120, 0.0] - - - [15744, 1024, 1, 128] - - [120, 0.0] - - - [9856, 2048, 1, 128] - - [120, 0.0] - - - [5888, 2305, 1, 128] - - [113, 0.0] - - - [10496, 512, 1, 128] - - [120, 0.0] - - - [1664, 1153, 1, 128] - - [117, 0.0] - - - [3456, 1024, 1, 128] - - [120, 0.0] - - - [20992, 13313, 1, 128] - - [120, 0.0] - - - [11904, 4096, 1, 128] - - [120, 0.0] - - - [13056, 1024, 1, 128] - - [120, 0.0] - - - [12800, 2048, 1, 128] - - [120, 0.0] - - - [12160, 512, 1, 128] - - [120, 0.0] - - - [5760, 2049, 1, 128] - - [120, 0.0] - - - [11392, 128, 1, 128] - - [111, 0.0] - - - [5632, 128, 1, 128] - - [110, 0.0] - - - [11520, 2048, 1, 128] - - [120, 0.0] - - - [11648, 2048, 1, 128] - - [120, 0.0] - - - [28544, 8192, 1, 128] - - [120, 0.0] - - - [22912, 1024, 1, 128] - - [120, 0.0] - - - [10752, 7169, 1, 128] - - [120, 0.0] - - - [8320, 128, 1, 128] - - [96, 0.0] - - - [23808, 1024, 1, 128] - - [120, 0.0] - - - [25984, 8192, 1, 128] - - [120, 0.0] - - - [22656, 2048, 1, 128] - - [120, 0.0] - - - [7296, 1024, 1, 128] - - [120, 0.0] - - - [28032, 512, 1, 128] - - [120, 0.0] - - - [22400, 2048, 1, 128] - - [120, 0.0] - - - [22144, 512, 1, 128] - - [120, 0.0] - - - [13312, 4096, 1, 128] - - [120, 0.0] - - - [10240, 2048, 1, 128] - - [120, 0.0] - - - [12672, 128, 1, 128] - - [111, 0.0] - - - [10752, 2048, 1, 128] - - [120, 0.0] - - - [1152, 128, 1, 128] - - [101, 0.0] - - - [13696, 5889, 1, 128] - - [113, 0.0] - - - [9216, 1024, 1, 128] - - [120, 0.0] - - - [17152, 128, 1, 128] - - [119, 0.0] - - - [24320, 2048, 1, 128] - - [120, 0.0] - - - [16512, 8705, 1, 128] - - [120, 0.0] - - - [3072, 1409, 1, 128] - - [113, 0.0] - - - [1024, 128, 1, 128] - - [110, 0.0] - - - [22400, 14593, 1, 128] - - [120, 0.0] - - - [4096, 512, 1, 128] - - [119, 0.0] - - - [4992, 128, 1, 128] - - [113, 0.0] - - - [9472, 5889, 1, 128] - - [113, 0.0] - - - [9472, 5761, 1, 128] - - [120, 0.0] - - - [27136, 1024, 1, 128] - - [120, 0.0] - - - [6528, 1024, 1, 128] - - [120, 0.0] - - - [25472, 1024, 1, 128] - - [120, 0.0] - - - [5120, 512, 1, 128] - - [118, 0.0] - - - [5504, 512, 1, 128] - - [118, 0.0] - - - [21120, 13441, 1, 128] - - [120, 0.0] - - - [4352, 128, 1, 128] - - [85, 0.0] - - - [8832, 5249, 1, 128] - - [113, 0.0] - - - [1536, 1025, 1, 128] - - [120, 0.0] - - - [11520, 512, 1, 128] - - [120, 0.0] - - - [5632, 2048, 1, 128] - - [120, 0.0] - - - [7424, 128, 1, 128] - - [120, 0.0] - - - [18432, 128, 1, 128] - - [111, 0.0] - - - [12672, 2048, 1, 128] - - [120, 0.0] - - - [14208, 128, 1, 128] - - [112, 0.0] - - - [15360, 7553, 1, 128] - - [120, 0.0] - - - [26496, 1024, 1, 128] - - [120, 0.0] - - - [27136, 128, 1, 128] - - [118, 0.0] - - - [12032, 2048, 1, 128] - - [120, 0.0] - - - [11648, 1024, 1, 128] - - [120, 0.0] - - - [11776, 512, 1, 128] - - [118, 0.0] - - - [1024, 512, 1, 128] - - [122, 0.0] - - - [11264, 7681, 1, 128] - - [120, 0.0] - - - [19456, 11777, 1, 128] - - [120, 0.0] - - - [14080, 4096, 1, 128] - - [120, 0.0] - - - [7040, 3329, 1, 128] - - [113, 0.0] - - - [27392, 4096, 1, 128] - - [120, 0.0] - - - [14720, 7041, 1, 128] - - [113, 0.0] - - - [19584, 1024, 1, 128] - - [120, 0.0] - - - [21376, 13569, 1, 128] - - [120, 0.0] - - - [20480, 12801, 1, 128] - - [120, 0.0] - - - [21248, 128, 1, 128] - - [112, 0.0] - - - [9728, 1024, 1, 128] - - [120, 0.0] - - - [18688, 10881, 1, 128] - - [120, 0.0] - - - [21120, 13313, 1, 128] - - [120, 0.0] - - - [20096, 2048, 1, 128] - - [120, 0.0] - - - [16640, 4096, 1, 128] - - [120, 0.0] - - - [28160, 12161, 1, 128] - - [120, 0.0] - - - [640, 129, 1, 128] - - [106, 0.0] - - - [28672, 512, 1, 128] - - [120, 0.0] - - - [12416, 4096, 1, 128] - - [120, 0.0] - - - [25344, 9473, 1, 128] - - [120, 0.0] - - - [18304, 1024, 1, 128] - - [120, 0.0] - - - [25600, 4096, 1, 128] - - [120, 0.0] - - - [22272, 512, 1, 128] - - [120, 0.0] - - - [21504, 13825, 1, 128] - - [120, 0.0] - - - [4736, 128, 1, 128] - - [73, 0.0] - - - [26496, 10625, 1, 128] - - [120, 0.0] - - - [7040, 512, 1, 128] - - [120, 0.0] - - - [14336, 4096, 1, 128] - - [120, 0.0] - - - [9216, 512, 1, 128] - - [120, 0.0] - - - [1280, 641, 1, 128] - - [73, 0.0] - - - [16768, 8961, 1, 128] - - [113, 0.0] - - - [18944, 11137, 1, 128] - - [120, 0.0] - - - [21504, 2048, 1, 128] - - [120, 0.0] - - - [21888, 1024, 1, 128] - - [120, 0.0] - - - [11264, 512, 1, 128] - - [120, 0.0] - - - [27776, 8192, 1, 128] - - [120, 0.0] - - - [10368, 6785, 1, 128] - - [113, 0.0] - - - [18432, 10753, 1, 128] - - [120, 0.0] - - - [19968, 2048, 1, 128] - - [120, 0.0] - - - [16640, 512, 1, 128] - - [120, 0.0] - - - [24576, 8577, 1, 128] - - [120, 0.0] - - - [28672, 2048, 1, 128] - - [120, 0.0] - - - [11136, 128, 1, 128] - - [120, 0.0] - - - [12288, 4609, 1, 128] - - [120, 0.0] - - - [14848, 1024, 1, 128] - - [120, 0.0] - - - [14848, 128, 1, 128] - - [113, 0.0] - - - [7424, 1024, 1, 128] - - [120, 0.0] - - - [2560, 1024, 1, 128] - - [111, 0.0] - - - [6400, 128, 1, 128] - - [84, 0.0] - - - [15488, 7809, 1, 128] - - [120, 0.0] - - - [17920, 2048, 1, 128] - - [120, 0.0] - - - [5760, 512, 1, 128] - - [118, 0.0] - - - [16640, 1024, 1, 128] - - [120, 0.0] - - - [28160, 2048, 1, 128] - - [120, 0.0] - - - [5504, 3969, 1, 128] - - [120, 0.0] - - - [11776, 1024, 1, 128] - - [120, 0.0] - - - [18816, 128, 1, 128] - - [111, 0.0] - - - [27904, 12033, 1, 128] - - [113, 0.0] - - - [11520, 7937, 1, 128] - - [120, 0.0] - - - [18944, 11265, 1, 128] - - [120, 0.0] - - - [5376, 1024, 1, 128] - - [120, 0.0] - - - [12032, 4225, 1, 128] - - [120, 0.0] - - - [5376, 128, 1, 128] - - [102, 0.0] - - - [9856, 1024, 1, 128] - - [120, 0.0] - - - [26752, 10881, 1, 128] - - [120, 0.0] - - - [20352, 128, 1, 128] - - [118, 0.0] - - - [14464, 128, 1, 128] - - [112, 0.0] - - - [1024, 385, 1, 128] - - [100, 0.0] - - - [3840, 128, 1, 128] - - [110, 0.0] - - - [24192, 128, 1, 128] - - [119, 0.0] - - - [28544, 12673, 1, 128] - - [120, 0.0] - - - [1664, 128, 1, 128] - - [102, 0.0] - - - [26752, 8192, 1, 128] - - [120, 0.0] - - - [16896, 1024, 1, 128] - - [120, 0.0] - - - [9728, 128, 1, 128] - - [112, 0.0] - - - [11264, 2048, 1, 128] - - [120, 0.0] - - - [11392, 2048, 1, 128] - - [120, 0.0] - - - [20224, 2048, 1, 128] - - [120, 0.0] - - - [26880, 1024, 1, 128] - - [120, 0.0] - - - [15104, 512, 1, 128] - - [120, 0.0] - - - [26368, 2048, 1, 128] - - [120, 0.0] - - - [6784, 3073, 1, 128] - - [120, 0.0] - - - [23168, 128, 1, 128] - - [119, 0.0] - - - [8448, 1024, 1, 128] - - [118, 0.0] - - - [16896, 9089, 1, 128] - - [120, 0.0] - - - [17536, 128, 1, 128] - - [112, 0.0] - - - [22912, 512, 1, 128] - - [120, 0.0] - - - [28032, 128, 1, 128] - - [120, 0.0] - - - [19584, 512, 1, 128] - - [120, 0.0] - - - [27136, 11265, 1, 128] - - [120, 0.0] - - - [4992, 512, 1, 128] - - [118, 0.0] - - - [8448, 128, 1, 128] - - [87, 0.0] - - - [27648, 128, 1, 128] - - [120, 0.0] - - - [16640, 2048, 1, 128] - - [120, 0.0] - - - [26752, 10753, 1, 128] - - [120, 0.0] - - - [2944, 1281, 1, 128] - - [112, 0.0] - - - [5376, 3841, 1, 128] - - [113, 0.0] - - - [10496, 6913, 1, 128] - - [113, 0.0] - - - [17024, 512, 1, 128] - - [120, 0.0] - - - [11008, 7297, 1, 128] - - [120, 0.0] - - - [14080, 128, 1, 128] - - [119, 0.0] - - - [5888, 512, 1, 128] - - [119, 0.0] - - - [19200, 128, 1, 128] - - [111, 0.0] - - - [14208, 6529, 1, 128] - - [120, 0.0] - - - [22912, 4096, 1, 128] - - [120, 0.0] - - - [14336, 2048, 1, 128] - - [120, 0.0] - - - [17792, 128, 1, 128] - - [112, 0.0] - - - [22656, 14849, 1, 128] - - [120, 0.0] - - - [19712, 512, 1, 128] - - [120, 0.0] - - - [5248, 1024, 1, 128] - - [120, 0.0] - - - [3712, 2049, 1, 128] - - [120, 0.0] - - - [24448, 8449, 1, 128] - - [113, 0.0] - - - [8192, 512, 1, 128] - - [120, 0.0] - - - [25472, 4096, 1, 128] - - [120, 0.0] - - - [25088, 512, 1, 128] - - [120, 0.0] - - - [23168, 1024, 1, 128] - - [120, 0.0] - - - [24320, 8192, 1, 128] - - [120, 0.0] - - - [24192, 8192, 1, 128] - - [120, 0.0] - - - [2176, 512, 1, 128] - - [110, 0.0] - - - [4992, 3457, 1, 128] - - [120, 0.0] - - - [896, 257, 1, 128] - - [72, 0.0] - - - [28288, 1024, 1, 128] - - [120, 0.0] - - - [20864, 1024, 1, 128] - - [120, 0.0] - - - [18432, 2048, 1, 128] - - [120, 0.0] - - - [17280, 9601, 1, 128] - - [120, 0.0] - - - [18944, 4096, 1, 128] - - [120, 0.0] - - - [13440, 128, 1, 128] - - [118, 0.0] - - - [7424, 2048, 1, 128] - - [120, 0.0] - - - [768, 128, 1, 128] - - [99, 0.0] - - - [16128, 512, 1, 128] - - [120, 0.0] - - - [28288, 12289, 1, 128] - - [120, 0.0] - - - [23552, 128, 1, 128] - - [112, 0.0] - - - [24832, 8192, 1, 128] - - [120, 0.0] - - - [10240, 1024, 1, 128] - - [120, 0.0] - - - [8960, 2048, 1, 128] - - [120, 0.0] - - - [17664, 9985, 1, 128] - - [120, 0.0] - - - [25088, 4096, 1, 128] - - [120, 0.0] - - - [7552, 2048, 1, 128] - - [120, 0.0] - - - [15104, 7297, 1, 128] - - [120, 0.0] - - - [7168, 1024, 1, 128] - - [120, 0.0] - - - [26112, 8192, 1, 128] - - [120, 0.0] - - - [24192, 1024, 1, 128] - - [120, 0.0] - - - [22912, 2048, 1, 128] - - [120, 0.0] - - - [10368, 512, 1, 128] - - [120, 0.0] - - - [22528, 4096, 1, 128] - - [120, 0.0] - - - [6528, 128, 1, 128] - - [84, 0.0] - - - [26752, 4096, 1, 128] - - [120, 0.0] - - - [2816, 512, 1, 128] - - [118, 0.0] - - - [22016, 14209, 1, 128] - - [120, 0.0] - - - [8832, 1024, 1, 128] - - [120, 0.0] - - - [16384, 128, 1, 128] - - [112, 0.0] - - - [5120, 1024, 1, 128] - - [118, 0.0] - - - [24832, 8833, 1, 128] - - [120, 0.0] - - - [11520, 128, 1, 128] - - [119, 0.0] - - - [24960, 512, 1, 128] - - [120, 0.0] - - - [27520, 2048, 1, 128] - - [120, 0.0] - - - [22272, 14593, 1, 128] - - [120, 0.0] - - - [2048, 128, 1, 128] - - [115, 0.0] - - - [2176, 1537, 1, 128] - - [118, 0.0] - - - [10496, 1024, 1, 128] - - [120, 0.0] - - - [12160, 4353, 1, 128] - - [113, 0.0] - - - [6144, 1024, 1, 128] - - [120, 0.0] - - - [26752, 1024, 1, 128] - - [120, 0.0] - - - [17280, 4096, 1, 128] - - [120, 0.0] - - - [16896, 512, 1, 128] - - [120, 0.0] - - - [4480, 128, 1, 128] - - [71, 0.0] - - - [18944, 128, 1, 128] - - [111, 0.0] - - - [9600, 2048, 1, 128] - - [120, 0.0] - - - [19456, 1024, 1, 128] - - [120, 0.0] - - - [9984, 2048, 1, 128] - - [120, 0.0] - - - [25216, 9217, 1, 128] - - [120, 0.0] - - - [19968, 1024, 1, 128] - - [120, 0.0] - - - [13952, 2048, 1, 128] - - [120, 0.0] - - - [10496, 2048, 1, 128] - - [120, 0.0] - - - [12672, 1024, 1, 128] - - [120, 0.0] - - - [19072, 11393, 1, 128] - - [120, 0.0] - - - [11008, 2048, 1, 128] - - [120, 0.0] - - - [27520, 11649, 1, 128] - - [120, 0.0] - - - [10880, 512, 1, 128] - - [118, 0.0] - - - [14592, 6785, 1, 128] - - [120, 0.0] - - - [7424, 512, 1, 128] - - [120, 0.0] - - - [13056, 5249, 1, 128] - - [120, 0.0] - - - [23296, 15489, 1, 128] - - [120, 0.0] - - - [28416, 8192, 1, 128] - - [120, 0.0] - - - [11392, 7681, 1, 128] - - [120, 0.0] - - - [18048, 1024, 1, 128] - - [120, 0.0] - - - [15616, 7809, 1, 128] - - [120, 0.0] - - - [128, 128, 1, 128] - - [99, 0.0] - - - [24704, 512, 1, 128] - - [120, 0.0] - - - [7680, 4097, 1, 128] - - [120, 0.0] - - - [16640, 8961, 1, 128] - - [120, 0.0] - - - [18944, 1024, 1, 128] - - [120, 0.0] - - - [12928, 2048, 1, 128] - - [120, 0.0] - - - [22272, 2048, 1, 128] - - [120, 0.0] - - - [27904, 11905, 1, 128] - - [120, 0.0] - - - [26240, 2048, 1, 128] - - [120, 0.0] - - - [9728, 6017, 1, 128] - - [120, 0.0] - - - [20736, 1024, 1, 128] - - [120, 0.0] - - - [3456, 1921, 1, 128] - - [113, 0.0] - - - [8064, 512, 1, 128] - - [118, 0.0] - - - [4224, 1024, 1, 128] - - [118, 0.0] - - - [25984, 10113, 1, 128] - - [120, 0.0] - - - [13696, 6017, 1, 128] - - [120, 0.0] - - - [27520, 8192, 1, 128] - - [120, 0.0] - - - [18944, 512, 1, 128] - - [120, 0.0] - - - [6272, 128, 1, 128] - - [92, 0.0] - - - [27264, 4096, 1, 128] - - [120, 0.0] - - - [1792, 1153, 1, 128] - - [119, 0.0] - - - [17536, 9729, 1, 128] - - [120, 0.0] - - - [13184, 5505, 1, 128] - - [120, 0.0] - - - [2944, 128, 1, 128] - - [112, 0.0] - - - [25344, 512, 1, 128] - - [120, 0.0] - - - [23040, 15361, 1, 128] - - [120, 0.0] - - - [8704, 512, 1, 128] - - [118, 0.0] - - - [20864, 13057, 1, 128] - - [120, 0.0] - - - [19328, 4096, 1, 128] - - [120, 0.0] - - - [28288, 8192, 1, 128] - - [120, 0.0] - - - [10112, 1024, 1, 128] - - [120, 0.0] - - - [17536, 2048, 1, 128] - - [120, 0.0] - - - [7552, 128, 1, 128] - - [118, 0.0] - - - [15616, 7937, 1, 128] - - [120, 0.0] - - - [23040, 512, 1, 128] - - [120, 0.0] - - - [25984, 2048, 1, 128] - - [120, 0.0] - - - [14720, 128, 1, 128] - - [112, 0.0] - - - [23424, 1024, 1, 128] - - [120, 0.0] - - - [1920, 1281, 1, 128] - - [112, 0.0] - - - [27136, 2048, 1, 128] - - [120, 0.0] - - - [28800, 8192, 1, 128] - - [120, 0.0] - - - [15488, 128, 1, 128] - - [112, 0.0] - - - [28800, 12929, 1, 128] - - [113, 0.0] - - - [21888, 14081, 1, 128] - - [120, 0.0] - - - [25600, 1024, 1, 128] - - [120, 0.0] - - - [21632, 1024, 1, 128] - - [120, 0.0] - - - [24448, 1024, 1, 128] - - [120, 0.0] - - - [4352, 2689, 1, 128] - - [120, 0.0] - - - [20480, 512, 1, 128] - - [118, 0.0] - - - [7296, 128, 1, 128] - - [118, 0.0] - - - [4992, 1024, 1, 128] - - [120, 0.0] - - - [27264, 11393, 1, 128] - - [120, 0.0] - - - [26752, 128, 1, 128] - - [111, 0.0] - - - [24960, 1024, 1, 128] - - [120, 0.0] - - - [21504, 512, 1, 128] - - [120, 0.0] - - - [6272, 2561, 1, 128] - - [120, 0.0] - - - [25088, 9089, 1, 128] - - [120, 0.0] - - - [20864, 512, 1, 128] - - [120, 0.0] - - - [4224, 2561, 1, 128] - - [120, 0.0] - - - [15744, 8065, 1, 128] - - [120, 0.0] - - - [21632, 128, 1, 128] - - [119, 0.0] - - - [15104, 4096, 1, 128] - - [120, 0.0] - - - [20352, 512, 1, 128] - - [120, 0.0] - - - [25472, 9601, 1, 128] - - [120, 0.0] - - - [27904, 512, 1, 128] - - [120, 0.0] - - - [19968, 512, 1, 128] - - [120, 0.0] - - - [5760, 1024, 1, 128] - - [120, 0.0] - - - [28416, 12545, 1, 128] - - [120, 0.0] - - - [16512, 8833, 1, 128] - - [113, 0.0] - - - [6016, 128, 1, 128] - - [85, 0.0] - - - [13056, 4096, 1, 128] - - [120, 0.0] - - - [19968, 12289, 1, 128] - - [120, 0.0] - - - [7424, 3713, 1, 128] - - [113, 0.0] - - - [28800, 128, 1, 128] - - [120, 0.0] - - - [512, 512, 1, 128] - - [100, 0.0] - - - [24832, 2048, 1, 128] - - [120, 0.0] - - - [20736, 128, 1, 128] - - [111, 0.0] - - - [26368, 512, 1, 128] - - [120, 0.0] - - - [26496, 8192, 1, 128] - - [120, 0.0] - - - [13824, 4096, 1, 128] - - [120, 0.0] - - - [27264, 128, 1, 128] - - [112, 0.0] - - - [21760, 1024, 1, 128] - - [120, 0.0] - - - [2432, 1921, 1, 128] - - [113, 0.0] - - - [27136, 8192, 1, 128] - - [120, 0.0] - - - [6784, 2048, 1, 128] - - [120, 0.0] - - - [11264, 128, 1, 128] - - [118, 0.0] - - - [7552, 512, 1, 128] - - [118, 0.0] - - - [19328, 11649, 1, 128] - - [120, 0.0] - - - [17152, 2048, 1, 128] - - [120, 0.0] - - - [23808, 16129, 1, 128] - - [120, 0.0] - - - [20224, 12417, 1, 128] - - [120, 0.0] - - - [27904, 1024, 1, 128] - - [120, 0.0] - - - [3456, 512, 1, 128] - - [119, 0.0] - - - [13312, 512, 1, 128] - - [120, 0.0] - - - [26368, 4096, 1, 128] - - [120, 0.0] - - - [23296, 15617, 1, 128] - - [120, 0.0] - - - [26112, 10241, 1, 128] - - [120, 0.0] - - - [26240, 512, 1, 128] - - [120, 0.0] - - - [4352, 1024, 1, 128] - - [118, 0.0] - - - [10624, 2048, 1, 128] - - [120, 0.0] - - - [23808, 16001, 1, 128] - - [113, 0.0] - - - [17536, 9857, 1, 128] - - [113, 0.0] - - - [23936, 4096, 1, 128] - - [120, 0.0] - - - [1408, 128, 1, 128] - - [100, 0.0] - - - [14848, 512, 1, 128] - - [120, 0.0] - - - [8704, 4993, 1, 128] - - [113, 0.0] - - - [15104, 2048, 1, 128] - - [120, 0.0] - - - [2560, 512, 1, 128] - - [118, 0.0] - - - [27264, 8192, 1, 128] - - [120, 0.0] - - - [23808, 4096, 1, 128] - - [120, 0.0] - - - [14080, 6273, 1, 128] - - [120, 0.0] - - - [10112, 6529, 1, 128] - - [120, 0.0] - - - [27648, 512, 1, 128] - - [120, 0.0] - - - [20992, 128, 1, 128] - - [113, 0.0] - - - [15104, 128, 1, 128] - - [119, 0.0] - - - [7808, 128, 1, 128] - - [76, 0.0] - - - [3584, 1024, 1, 128] - - [120, 0.0] - - - [15232, 512, 1, 128] - - [120, 0.0] - - - [21376, 13697, 1, 128] - - [120, 0.0] - - - [11392, 7809, 1, 128] - - [120, 0.0] - - - [11904, 1024, 1, 128] - - [120, 0.0] - - - [28800, 2048, 1, 128] - - [120, 0.0] - - - [8960, 512, 1, 128] - - [120, 0.0] - - - [19456, 11649, 1, 128] - - [120, 0.0] - - - [11904, 128, 1, 128] - - [112, 0.0] - - - [18560, 512, 1, 128] - - [120, 0.0] - - - [6656, 128, 1, 128] - - [92, 0.0] - - - [17792, 2048, 1, 128] - - [120, 0.0] - - - [21632, 4096, 1, 128] - - [120, 0.0] - - - [25728, 4096, 1, 128] - - [120, 0.0] - - - [18048, 10241, 1, 128] - - [120, 0.0] - - - [1792, 1281, 1, 128] - - [103, 0.0] - - - [512, 385, 1, 128] - - [115, 0.0] - - - [26112, 512, 1, 128] - - [120, 0.0] - - - [16128, 1024, 1, 128] - - [120, 0.0] - - - [4480, 1024, 1, 128] - - [120, 0.0] - - - [14720, 4096, 1, 128] - - [120, 0.0] - - - [23552, 2048, 1, 128] - - [120, 0.0] - - - [22528, 512, 1, 128] - - [120, 0.0] - - - [22912, 128, 1, 128] - - [120, 0.0] - - - [25344, 1024, 1, 128] - - [120, 0.0] - - - [24064, 16257, 1, 128] - - [120, 0.0] - - - [9088, 5377, 1, 128] - - [113, 0.0] - - - [27776, 128, 1, 128] - - [112, 0.0] - - - [15616, 512, 1, 128] - - [120, 0.0] - - - [13568, 128, 1, 128] - - [110, 0.0] - - - [15488, 7681, 1, 128] - - [120, 0.0] - - - [20096, 512, 1, 128] - - [120, 0.0] - - - [24832, 4096, 1, 128] - - [120, 0.0] - - - [28800, 4096, 1, 128] - - [120, 0.0] - - - [11904, 4225, 1, 128] - - [120, 0.0] - - - [3968, 1024, 1, 128] - - [120, 0.0] - - - [6400, 2817, 1, 128] - - [113, 0.0] - - - [24576, 4096, 1, 128] - - [120, 0.0] - - - [9088, 128, 1, 128] - - [119, 0.0] - - - [17152, 4096, 1, 128] - - [120, 0.0] - - - [22528, 14721, 1, 128] - - [120, 0.0] - - - [27392, 2048, 1, 128] - - [120, 0.0] - - - [8832, 512, 1, 128] - - [118, 0.0] - - - [8960, 5249, 1, 128] - - [120, 0.0] - - - [3200, 1024, 1, 128] - - [118, 0.0] - - - [4736, 3073, 1, 128] - - [120, 0.0] - - - [28032, 2048, 1, 128] - - [120, 0.0] - - - [14592, 2048, 1, 128] - - [120, 0.0] - - - [13440, 1024, 1, 128] - - [120, 0.0] - - - [14464, 2048, 1, 128] - - [120, 0.0] - - - [6912, 2048, 1, 128] - - [120, 0.0] - - - [19584, 2048, 1, 128] - - [120, 0.0] - - - [17920, 128, 1, 128] - - [112, 0.0] - - - [19584, 11777, 1, 128] - - [120, 0.0] - - - [23936, 16129, 1, 128] - - [113, 0.0] - - - [10496, 6785, 1, 128] - - [120, 0.0] - - - [27648, 2048, 1, 128] - - [120, 0.0] - - - [23808, 128, 1, 128] - - [119, 0.0] - - - [20864, 2048, 1, 128] - - [120, 0.0] - - - [9088, 512, 1, 128] - - [120, 0.0] - - - [3584, 512, 1, 128] - - [118, 0.0] - - - [8576, 4993, 1, 128] - - [113, 0.0] - - - [3328, 1024, 1, 128] - - [120, 0.0] - - - [20608, 2048, 1, 128] - - [120, 0.0] - - - [23552, 15745, 1, 128] - - [120, 0.0] - - - [23424, 15617, 1, 128] - - [113, 0.0] - - - [21120, 512, 1, 128] - - [120, 0.0] - - - [6656, 512, 1, 128] - - [118, 0.0] - - - [12544, 128, 1, 128] - - [118, 0.0] - - - [24448, 8577, 1, 128] - - [113, 0.0] - - - [9984, 512, 1, 128] - - [120, 0.0] - - - [18304, 4096, 1, 128] - - [120, 0.0] - - - [17920, 512, 1, 128] - - [120, 0.0] - - - [12160, 4096, 1, 128] - - [120, 0.0] - - - [3968, 2433, 1, 128] - - [113, 0.0] - - - [27008, 4096, 1, 128] - - [120, 0.0] - - - [22272, 1024, 1, 128] - - [120, 0.0] - - - [14336, 512, 1, 128] - - [120, 0.0] - - - [18560, 10753, 1, 128] - - [120, 0.0] - - - [6272, 2048, 1, 128] - - [120, 0.0] - - - [12800, 1024, 1, 128] - - [120, 0.0] - - - [9600, 5889, 1, 128] - - [120, 0.0] - - - [13056, 128, 1, 128] - - [111, 0.0] - - - [7296, 2048, 1, 128] - - [120, 0.0] - - - [21376, 512, 1, 128] - - [120, 0.0] - - - [11904, 512, 1, 128] - - [120, 0.0] - - - [6400, 1024, 1, 128] - - [118, 0.0] - - - [27008, 1024, 1, 128] - - [120, 0.0] - - - [22400, 14721, 1, 128] - - [120, 0.0] - - - [6272, 1024, 1, 128] - - [120, 0.0] - - - [17408, 128, 1, 128] - - [112, 0.0] - - - [26624, 10625, 1, 128] - - [120, 0.0] - - - [22400, 1024, 1, 128] - - [120, 0.0] - - - [18304, 10625, 1, 128] - - [120, 0.0] - - - [15872, 1024, 1, 128] - - [120, 0.0] - - - [21120, 128, 1, 128] - - [111, 0.0] - - - [22784, 4096, 1, 128] - - [120, 0.0] - - - [25728, 9857, 1, 128] - - [120, 0.0] - - - [16256, 1024, 1, 128] - - [120, 0.0] - - - [18560, 4096, 1, 128] - - [120, 0.0] - - - [7936, 4225, 1, 128] - - [113, 0.0] - - - [7680, 3969, 1, 128] - - [120, 0.0] - - - [9472, 2048, 1, 128] - - [120, 0.0] - - - [28160, 128, 1, 128] - - [120, 0.0] - - - [18816, 512, 1, 128] - - [120, 0.0] - - - [9856, 512, 1, 128] - - [118, 0.0] - - - [17664, 9857, 1, 128] - - [113, 0.0] - - - [27392, 128, 1, 128] - - [119, 0.0] - - - [24448, 2048, 1, 128] - - [120, 0.0] - - - [7808, 512, 1, 128] - - [119, 0.0] - - - [13952, 512, 1, 128] - - [120, 0.0] - - - [24576, 512, 1, 128] - - [120, 0.0] - - - [27520, 128, 1, 128] - - [112, 0.0] - - - [26496, 512, 1, 128] - - [120, 0.0] - - - [8576, 512, 1, 128] - - [120, 0.0] - - - [11648, 512, 1, 128] - - [120, 0.0] - - - [17408, 2048, 1, 128] - - [120, 0.0] - - - [17920, 10241, 1, 128] - - [120, 0.0] - - - [16384, 1024, 1, 128] - - [120, 0.0] - - - [6016, 2048, 1, 128] - - [120, 0.0] - - - [9728, 512, 1, 128] - - [120, 0.0] - - - [19712, 128, 1, 128] - - [112, 0.0] - - - [26112, 1024, 1, 128] - - [120, 0.0] - - - [16768, 128, 1, 128] - - [112, 0.0] - - - [8960, 1024, 1, 128] - - [120, 0.0] - - - [6784, 128, 1, 128] - - [111, 0.0] - - - [12800, 4993, 1, 128] - - [120, 0.0] - - - [6144, 2561, 1, 128] - - [120, 0.0] - - - [26880, 10881, 1, 128] - - [120, 0.0] - - - [12928, 1024, 1, 128] - - [120, 0.0] - - - [7040, 3457, 1, 128] - - [120, 0.0] - - - [15744, 4096, 1, 128] - - [120, 0.0] - - - [20096, 4096, 1, 128] - - [120, 0.0] - - - [21760, 128, 1, 128] - - [119, 0.0] - - - [7936, 2048, 1, 128] - - [120, 0.0] - - - [24448, 8192, 1, 128] - - [120, 0.0] - - - [21120, 2048, 1, 128] - - [120, 0.0] - - - [12160, 1024, 1, 128] - - [120, 0.0] - - - [7168, 3457, 1, 128] - - [120, 0.0] - - - [15232, 7553, 1, 128] - - [120, 0.0] - - - [26624, 1024, 1, 128] - - [120, 0.0] - - - [25344, 2048, 1, 128] - - [120, 0.0] - - - [12544, 4865, 1, 128] - - [120, 0.0] - - - [21120, 4096, 1, 128] - - [120, 0.0] - - - [20224, 128, 1, 128] - - [112, 0.0] - - - [14592, 4096, 1, 128] - - [120, 0.0] - - - [16256, 8577, 1, 128] - - [120, 0.0] - - - [24192, 4096, 1, 128] - - [120, 0.0] - - - [21248, 1024, 1, 128] - - [120, 0.0] - - - [25216, 1024, 1, 128] - - [120, 0.0] - - - [5888, 2177, 1, 128] - - [113, 0.0] - - - [21504, 1024, 1, 128] - - [120, 0.0] - - - [17536, 1024, 1, 128] - - [120, 0.0] - - - [9728, 2048, 1, 128] - - [120, 0.0] - - - [13952, 6273, 1, 128] - - [120, 0.0] - - - [28800, 512, 1, 128] - - [120, 0.0] - - - [2304, 1793, 1, 128] - - [113, 0.0] - - - [12416, 128, 1, 128] - - [118, 0.0] - - - [20224, 1024, 1, 128] - - [120, 0.0] - - - [22144, 128, 1, 128] - - [119, 0.0] - - - [22784, 1024, 1, 128] - - [120, 0.0] - - - [27136, 4096, 1, 128] - - [120, 0.0] - - - [27264, 512, 1, 128] - - [120, 0.0] - - - [26240, 10241, 1, 128] - - [120, 0.0] - - - [27904, 4096, 1, 128] - - [120, 0.0] - - - [21504, 128, 1, 128] - - [119, 0.0] - - - [3712, 2177, 1, 128] - - [120, 0.0] - - - [18432, 1024, 1, 128] - - [120, 0.0] - - - [28672, 4096, 1, 128] - - [120, 0.0] - - - [25344, 4096, 1, 128] - - [120, 0.0] - - - [26880, 512, 1, 128] - - [120, 0.0] - - - [21888, 2048, 1, 128] - - [120, 0.0] - - - [1792, 128, 1, 128] - - [117, 0.0] - - - [6016, 1024, 1, 128] - - [118, 0.0] - - - [15104, 7425, 1, 128] - - [113, 0.0] - - - [22016, 2048, 1, 128] - - [120, 0.0] - - - [13952, 4096, 1, 128] - - [120, 0.0] - - - [20992, 4096, 1, 128] - - [120, 0.0] - - - [8064, 4481, 1, 128] - - [113, 0.0] - - - [12672, 4096, 1, 128] - - [120, 0.0] - - - [20096, 12289, 1, 128] - - [120, 0.0] - - - [14848, 2048, 1, 128] - - [120, 0.0] - - - [23168, 512, 1, 128] - - [120, 0.0] - - - [7680, 128, 1, 128] - - [118, 0.0] - - - [13312, 1024, 1, 128] - - [120, 0.0] - - - [10624, 1024, 1, 128] - - [120, 0.0] - - - [3840, 512, 1, 128] - - [118, 0.0] - - - [22144, 14337, 1, 128] - - [120, 0.0] - - - [3200, 128, 1, 128] - - [104, 0.0] - - - [25472, 9473, 1, 128] - - [113, 0.0] - - - [16768, 9089, 1, 128] - - [120, 0.0] - - - [12288, 2048, 1, 128] - - [120, 0.0] - - - [20608, 512, 1, 128] - - [120, 0.0] - - - [2816, 1024, 1, 128] - - [120, 0.0] - - - [7552, 1024, 1, 128] - - [120, 0.0] - - - [5120, 3457, 1, 128] - - [113, 0.0] - - - [25216, 2048, 1, 128] - - [120, 0.0] - - - [12672, 4865, 1, 128] - - [113, 0.0] - - - [10880, 2048, 1, 128] - - [120, 0.0] - - - [18176, 512, 1, 128] - - [120, 0.0] - - - [8320, 4609, 1, 128] - - [120, 0.0] - - - [16000, 4096, 1, 128] - - [120, 0.0] - - - [22144, 2048, 1, 128] - - [120, 0.0] - - - [22784, 512, 1, 128] - - [120, 0.0] - - - [4096, 2561, 1, 128] - - [120, 0.0] - - - [24576, 2048, 1, 128] - - [120, 0.0] - - - [26624, 4096, 1, 128] - - [120, 0.0] - - - [18560, 2048, 1, 128] - - [120, 0.0] - - - [19584, 128, 1, 128] - - [112, 0.0] - - - [23936, 2048, 1, 128] - - [120, 0.0] - - - [23552, 512, 1, 128] - - [120, 0.0] - - - [12032, 4096, 1, 128] - - [120, 0.0] - - - [3840, 2305, 1, 128] - - [113, 0.0] - - - [25088, 128, 1, 128] - - [120, 0.0] - - - [16640, 8833, 1, 128] - - [120, 0.0] - - - [896, 128, 1, 128] - - [99, 0.0] - - - [17280, 2048, 1, 128] - - [120, 0.0] - - - [16896, 2048, 1, 128] - - [120, 0.0] - - - [22656, 128, 1, 128] - - [111, 0.0] - - - [25728, 8192, 1, 128] - - [120, 0.0] - - - [16128, 128, 1, 128] - - [112, 0.0] - - - [3840, 1024, 1, 128] - - [118, 0.0] - - - [2944, 512, 1, 128] - - [110, 0.0] - - - [24064, 1024, 1, 128] - - [120, 0.0] - - - [896, 385, 1, 128] - - [100, 0.0] - - - [8064, 128, 1, 128] - - [89, 0.0] - - - [12416, 1024, 1, 128] - - [120, 0.0] - - - [20608, 128, 1, 128] - - [111, 0.0] - - - [2944, 1024, 1, 128] - - [112, 0.0] - - - [6656, 2048, 1, 128] - - [120, 0.0] - - - [24064, 128, 1, 128] - - [111, 0.0] - - - [15744, 7937, 1, 128] - - [113, 0.0] - - - [2688, 1024, 1, 128] - - [112, 0.0] - - - [24192, 8193, 1, 128] - - [120, 0.0] - - - [24320, 4096, 1, 128] - - [120, 0.0] - - - [24576, 8705, 1, 128] - - [120, 0.0] - - - [13824, 1024, 1, 128] - - [120, 0.0] - - - [27776, 512, 1, 128] - - [120, 0.0] - - - [10240, 128, 1, 128] - - [117, 0.0] - - - [26240, 10369, 1, 128] - - [120, 0.0] - - - [16512, 4096, 1, 128] - - [120, 0.0] - - - [9856, 6145, 1, 128] - - [120, 0.0] - - - [27392, 1024, 1, 128] - - [120, 0.0] - - - [14976, 1024, 1, 128] - - [120, 0.0] - - - [1280, 512, 1, 128] - - [113, 0.0] - - - [6528, 2817, 1, 128] - - [120, 0.0] - - - [12288, 512, 1, 128] - - [120, 0.0] - - - [5248, 512, 1, 128] - - [120, 0.0] - - - [28544, 4096, 1, 128] - - [120, 0.0] - - - [21248, 13569, 1, 128] - - [120, 0.0] - - - [26112, 2048, 1, 128] - - [120, 0.0] - - - [14208, 6401, 1, 128] - - [113, 0.0] - - - [13952, 128, 1, 128] - - [112, 0.0] - - - [2304, 1665, 1, 128] - - [118, 0.0] - - - [6912, 1024, 1, 128] - - [120, 0.0] - - - [28672, 1024, 1, 128] - - [120, 0.0] - - - [14592, 6913, 1, 128] - - [120, 0.0] - - - [24704, 1024, 1, 128] - - [120, 0.0] - - - [22400, 512, 1, 128] - - [120, 0.0] - - - [23424, 4096, 1, 128] - - [120, 0.0] - - - [24832, 128, 1, 128] - - [112, 0.0] - - - [23680, 2048, 1, 128] - - [120, 0.0] - - - [25984, 9985, 1, 128] - - [113, 0.0] - - - [15360, 512, 1, 128] - - [118, 0.0] - - - [21376, 2048, 1, 128] - - [120, 0.0] - - - [16128, 2048, 1, 128] - - [120, 0.0] - - - [15872, 512, 1, 128] - - [120, 0.0] - - - [3072, 128, 1, 128] - - [119, 0.0] - - - [27520, 4096, 1, 128] - - [120, 0.0] - - - [25216, 4096, 1, 128] - - [120, 0.0] - - - [28672, 12673, 1, 128] - - [120, 0.0] - - - [28288, 512, 1, 128] - - [120, 0.0] - - - [22400, 4096, 1, 128] - - [120, 0.0] - - - [25344, 9345, 1, 128] - - [120, 0.0] - - - [9984, 128, 1, 128] - - [119, 0.0] - - - [28416, 1024, 1, 128] - - [120, 0.0] - - - [27008, 8192, 1, 128] - - [120, 0.0] - - - [13184, 1024, 1, 128] - - [120, 0.0] - - - [10240, 512, 1, 128] - - [120, 0.0] - - - [3456, 128, 1, 128] - - [112, 0.0] - - - [16000, 8321, 1, 128] - - [113, 0.0] - - - [27520, 1024, 1, 128] - - [120, 0.0] - - - [25088, 1024, 1, 128] - - [120, 0.0] - - - [6784, 512, 1, 128] - - [113, 0.0] - - - [18432, 10625, 1, 128] - - [120, 0.0] - - - [16128, 4096, 1, 128] - - [120, 0.0] - - - [26880, 11009, 1, 128] - - [120, 0.0] - - - [28800, 12801, 1, 128] - - [120, 0.0] - - - [12288, 4096, 1, 128] - - [120, 0.0] - - - [20096, 12417, 1, 128] - - [120, 0.0] - - - [1920, 128, 1, 128] - - [116, 0.0] - - - [13056, 2048, 1, 128] - - [120, 0.0] - - - [384, 385, 1, 128] - - [101, 0.0] - - - [9088, 1024, 1, 128] - - [120, 0.0] - - - [6784, 1024, 1, 128] - - [118, 0.0] - - - [21760, 4096, 1, 128] - - [120, 0.0] - - - [27008, 11009, 1, 128] - - [120, 0.0] - - - [14208, 1024, 1, 128] - - [120, 0.0] - - - [25600, 512, 1, 128] - - [120, 0.0] - - - [23680, 1024, 1, 128] - - [120, 0.0] - - - [28160, 8192, 1, 128] - - [120, 0.0] - - - [22016, 4096, 1, 128] - - [120, 0.0] - - - [18688, 4096, 1, 128] - - [120, 0.0] - - - [10752, 1024, 1, 128] - - [120, 0.0] - - - [2432, 128, 1, 128] - - [100, 0.0] - - - [7296, 512, 1, 128] - - [118, 0.0] - - - [19200, 4096, 1, 128] - - [120, 0.0] - - - [4608, 2945, 1, 128] - - [113, 0.0] - - - [18816, 11009, 1, 128] - - [113, 0.0] - - - [9600, 1024, 1, 128] - - [120, 0.0] - - - [7168, 512, 1, 128] - - [118, 0.0] - - - [11904, 4097, 1, 128] - - [120, 0.0] - - - [17920, 1024, 1, 128] - - [120, 0.0] - - - [11520, 7809, 1, 128] - - [113, 0.0] - - - [22784, 14977, 1, 128] - - [120, 0.0] - - - [13696, 1024, 1, 128] - - [120, 0.0] - - - [15104, 1024, 1, 128] - - [120, 0.0] - - - [25216, 512, 1, 128] - - [120, 0.0] - - - [5376, 512, 1, 128] - - [117, 0.0] - - - [17408, 4096, 1, 128] - - [120, 0.0] - - - [25728, 512, 1, 128] - - [120, 0.0] - - - [896, 512, 1, 128] - - [104, 0.0] - - - [6912, 3329, 1, 128] - - [113, 0.0] - - - [22016, 512, 1, 128] - - [118, 0.0] - - - [22144, 4096, 1, 128] - - [120, 0.0] - - - [10368, 128, 1, 128] - - [118, 0.0] - - - [23296, 2048, 1, 128] - - [120, 0.0] - - - [17920, 10113, 1, 128] - - [120, 0.0] - - - [14848, 4096, 1, 128] - - [120, 0.0] - - - [26112, 128, 1, 128] - - [113, 0.0] - - - [28032, 8192, 1, 128] - - [120, 0.0] - - - [20096, 128, 1, 128] - - [112, 0.0] - - - [15360, 4096, 1, 128] - - [120, 0.0] - - - [3328, 128, 1, 128] - - [119, 0.0] - - - [25472, 512, 1, 128] - - [120, 0.0] - - - [18304, 128, 1, 128] - - [119, 0.0] - - - [20352, 12545, 1, 128] - - [113, 0.0] - - - [26624, 10753, 1, 128] - - [120, 0.0] - - - [20480, 2048, 1, 128] - - [120, 0.0] - - - [26496, 10497, 1, 128] - - [120, 0.0] - - - [22400, 128, 1, 128] - - [119, 0.0] - - - [9216, 5505, 1, 128] - - [120, 0.0] - - - [24064, 8193, 1, 128] - - [120, 0.0] - - - [4224, 128, 1, 128] - - [123, 0.0] - - - [6656, 3073, 1, 128] - - [118, 0.0] - - - [10880, 1024, 1, 128] - - [120, 0.0] - - - [23808, 512, 1, 128] - - [120, 0.0] - - - [15488, 1024, 1, 128] - - [120, 0.0] - - - [24704, 8705, 1, 128] - - [120, 0.0] - - - [12416, 4609, 1, 128] - - [120, 0.0] - - - [3712, 1024, 1, 128] - - [120, 0.0] - - - [25856, 8192, 1, 128] - - [120, 0.0] - - - [8320, 1024, 1, 128] - - [120, 0.0] - - - [16256, 512, 1, 128] - - [120, 0.0] - - - [18944, 2048, 1, 128] - - [120, 0.0] - - - [23168, 4096, 1, 128] - - [120, 0.0] - - - [15616, 2048, 1, 128] - - [120, 0.0] - - - [24320, 512, 1, 128] - - [120, 0.0] - - - [2688, 1025, 1, 128] - - [118, 0.0] - - - [12800, 5121, 1, 128] - - [120, 0.0] - - - [5120, 128, 1, 128] - - [103, 0.0] - - - [4352, 512, 1, 128] - - [119, 0.0] - - - [24576, 8192, 1, 128] - - [120, 0.0] - - - [8320, 512, 1, 128] - - [118, 0.0] - - - [12160, 4481, 1, 128] - - [120, 0.0] - - - [2560, 1025, 1, 128] - - [111, 0.0] - - - [19072, 1024, 1, 128] - - [120, 0.0] - - - [2816, 1153, 1, 128] - - [120, 0.0] - - - [6912, 128, 1, 128] - - [103, 0.0] - - - [9088, 2048, 1, 128] - - [120, 0.0] - - - [26368, 8192, 1, 128] - - [120, 0.0] - - - [17408, 9729, 1, 128] - - [120, 0.0] - - - [18816, 4096, 1, 128] - - [120, 0.0] - - - [4480, 512, 1, 128] - - [118, 0.0] - - - [11648, 128, 1, 128] - - [110, 0.0] - - - [1536, 897, 1, 128] - - [103, 0.0] - - - [11136, 1024, 1, 128] - - [120, 0.0] - - - [8704, 1024, 1, 128] - - [120, 0.0] - - - [19072, 2048, 1, 128] - - [120, 0.0] - - - [25856, 1024, 1, 128] - - [120, 0.0] - - - [7552, 3841, 1, 128] - - [113, 0.0] - - - [23296, 128, 1, 128] - - [112, 0.0] - - - [23424, 512, 1, 128] - - [120, 0.0] - - - [26368, 10497, 1, 128] - - [120, 0.0] - - - [18560, 1024, 1, 128] - - [120, 0.0] - - - [8192, 128, 1, 128] - - [76, 0.0] - - - [27776, 11905, 1, 128] - - [120, 0.0] - - - [18688, 1024, 1, 128] - - [120, 0.0] - - - [21248, 4096, 1, 128] - - [120, 0.0] - - - [16256, 8449, 1, 128] - - [113, 0.0] - - - [1920, 1409, 1, 128] - - [104, 0.0] - - - [24704, 4096, 1, 128] - - [120, 0.0] - - - [13824, 6145, 1, 128] - - [120, 0.0] - - - [6528, 512, 1, 128] - - [120, 0.0] - - - [21376, 128, 1, 128] - - [119, 0.0] - - - [11264, 1024, 1, 128] - - [120, 0.0] - - - [4352, 2817, 1, 128] - - [113, 0.0] - - - [22272, 4096, 1, 128] - - [120, 0.0] - - - [27264, 11265, 1, 128] - - [120, 0.0] - - - [28160, 1024, 1, 128] - - [120, 0.0] - - - [16256, 128, 1, 128] - - [112, 0.0] - - - [18688, 2048, 1, 128] - - [120, 0.0] - - - [9600, 6017, 1, 128] - - [120, 0.0] - - - [23552, 1024, 1, 128] - - [120, 0.0] - - - [8576, 128, 1, 128] - - [77, 0.0] - - - [20992, 13185, 1, 128] - - [120, 0.0] - - - [20992, 1024, 1, 128] - - [120, 0.0] - - - [14720, 512, 1, 128] - - [120, 0.0] - - - [28032, 1024, 1, 128] - - [120, 0.0] - - - [20352, 2048, 1, 128] - - [120, 0.0] - - - [15360, 128, 1, 128] - - [118, 0.0] - - - [8448, 2048, 1, 128] - - [120, 0.0] - - - [6272, 2689, 1, 128] - - [120, 0.0] - - - [7808, 4097, 1, 128] - - [120, 0.0] - - - [25472, 128, 1, 128] - - [112, 0.0] - - - [12288, 4481, 1, 128] - - [120, 0.0] - - - [28416, 4096, 1, 128] - - [120, 0.0] - - - [2176, 128, 1, 128] - - [115, 0.0] - - - [21760, 2048, 1, 128] - - [120, 0.0] - - - [21376, 1024, 1, 128] - - [120, 0.0] - - - [13696, 2048, 1, 128] - - [120, 0.0] - - - [28288, 12417, 1, 128] - - [120, 0.0] - - - [5632, 512, 1, 128] - - [120, 0.0] - - - [22016, 1024, 1, 128] - - [120, 0.0] - - - [25216, 128, 1, 128] - - [112, 0.0] - - - [25216, 8192, 1, 128] - - [120, 0.0] - - - [12032, 128, 1, 128] - - [113, 0.0] - - - [6144, 2048, 1, 128] - - [120, 0.0] - - - [23680, 128, 1, 128] - - [119, 0.0] - - - [15744, 128, 1, 128] - - [112, 0.0] - - - [3968, 512, 1, 128] - - [112, 0.0] - - - [16512, 1024, 1, 128] - - [120, 0.0] - - - [1536, 128, 1, 128] - - [102, 0.0] - - - [25984, 4096, 1, 128] - - [120, 0.0] - - - [19456, 512, 1, 128] - - [120, 0.0] - - - [9984, 1024, 1, 128] - - [120, 0.0] - - - [14080, 6401, 1, 128] - - [120, 0.0] - - - [20736, 2048, 1, 128] - - [120, 0.0] - - - [4224, 2689, 1, 128] - - [113, 0.0] - - - [13696, 512, 1, 128] - - [120, 0.0] - - - [17280, 1024, 1, 128] - - [120, 0.0] - - - [10752, 128, 1, 128] - - [111, 0.0] - - - [1536, 512, 1, 128] - - [84, 0.0] - - - [25728, 2048, 1, 128] - - [120, 0.0] - - - [9472, 128, 1, 128] - - [110, 0.0] - - - [7168, 3585, 1, 128] - - [120, 0.0] - - - [14720, 1024, 1, 128] - - [120, 0.0] - - - [25728, 128, 1, 128] - - [119, 0.0] - - - [14976, 128, 1, 128] - - [119, 0.0] - - - [24832, 1024, 1, 128] - - [120, 0.0] - - - [14080, 512, 1, 128] - - [120, 0.0] - - - [17152, 1024, 1, 128] - - [120, 0.0] - - - [19072, 512, 1, 128] - - [120, 0.0] - - - [21120, 1024, 1, 128] - - [120, 0.0] - - - [4864, 128, 1, 128] - - [113, 0.0] - - - [7936, 512, 1, 128] - - [120, 0.0] - - - [21248, 13441, 1, 128] - - [113, 0.0] - - - [12160, 2048, 1, 128] - - [120, 0.0] - - - [19712, 11905, 1, 128] - - [113, 0.0] - - - [23296, 1024, 1, 128] - - [120, 0.0] - - - [24832, 8961, 1, 128] - - [113, 0.0] - - - [13568, 2048, 1, 128] - - [120, 0.0] - - - [13696, 4096, 1, 128] - - [120, 0.0] - - - [5888, 128, 1, 128] - - [92, 0.0] - - - [10112, 2048, 1, 128] - - [120, 0.0] - - - [21632, 13953, 1, 128] - - [120, 0.0] - - - [19328, 512, 1, 128] - - [120, 0.0] - - - [6272, 512, 1, 128] - - [119, 0.0] - - - [4864, 3201, 1, 128] - - [113, 0.0] - - - [15232, 4096, 1, 128] - - [120, 0.0] - - - [23040, 4096, 1, 128] - - [120, 0.0] - - - [2816, 1281, 1, 128] - - [113, 0.0] - - - [8960, 128, 1, 128] - - [119, 0.0] - - - [9472, 1024, 1, 128] - - [118, 0.0] - - - [27648, 11777, 1, 128] - - [120, 0.0] - - - [28416, 2048, 1, 128] - - [120, 0.0] - - - [13952, 6145, 1, 128] - - [120, 0.0] - - - [13952, 1024, 1, 128] - - [120, 0.0] - - - [12544, 2048, 1, 128] - - [120, 0.0] - - - [10624, 7041, 1, 128] - - [120, 0.0] - - - [24704, 2048, 1, 128] - - [120, 0.0] - - - [17280, 9473, 1, 128] - - [120, 0.0] - - - [25088, 9217, 1, 128] - - [120, 0.0] - - - [10240, 6657, 1, 128] - - [120, 0.0] - - - [12800, 4096, 1, 128] - - [120, 0.0] - - - [17792, 1024, 1, 128] - - [120, 0.0] - - - [12160, 128, 1, 128] - - [113, 0.0] - - - [16512, 128, 1, 128] - - [117, 0.0] - - - [25856, 512, 1, 128] - - [120, 0.0] - - - [8576, 4865, 1, 128] - - [113, 0.0] - - - [25984, 1024, 1, 128] - - [120, 0.0] - - - [512, 128, 1, 128] - - [99, 0.0] - - - [10112, 128, 1, 128] - - [110, 0.0] - - - [28288, 2048, 1, 128] - - [120, 0.0] - - - [1152, 641, 1, 128] - - [93, 0.0] - - - [17920, 4096, 1, 128] - - [120, 0.0] - - - [2560, 1921, 1, 128] - - [118, 0.0] - - - [24704, 8833, 1, 128] - - [120, 0.0] - - - [3200, 512, 1, 128] - - [118, 0.0] - - - [6656, 2945, 1, 128] - - [120, 0.0] - - - [12672, 4993, 1, 128] - - [120, 0.0] - - - [4608, 1024, 1, 128] - - [118, 0.0] - - - [25856, 9985, 1, 128] - - [120, 0.0] - - - [23808, 2048, 1, 128] - - [120, 0.0] - - - [9728, 6145, 1, 128] - - [120, 0.0] - - - [28416, 12417, 1, 128] - - [120, 0.0] - - - [14464, 4096, 1, 128] - - [120, 0.0] - - - [21888, 128, 1, 128] - - [112, 0.0] - - - [23680, 15873, 1, 128] - - [120, 0.0] - - - [22144, 1024, 1, 128] - - [120, 0.0] - - - [17664, 512, 1, 256] - - [64, 0.0] - - - [25600, 1024, 1, 256] - - [64, 0.0] - - - [28928, 512, 1, 256] - - [64, 0.0] - - - [15104, 512, 1, 256] - - [64, 0.0] - - - [38912, 1024, 1, 256] - - [69, 0.0] - - - [34304, 8192, 1, 256] - - [69, 0.0] - - - [23552, 1024, 1, 256] - - [64, 0.0] - - - [39424, 23552, 1, 256] - - [69, 0.0] - - - [9472, 1024, 1, 256] - - [64, 0.0] - - - [28928, 13056, 1, 256] - - [45, 0.0] - - - [42496, 1024, 1, 256] - - [64, 0.0] - - - [18432, 1024, 1, 256] - - [64, 0.0] - - - [40192, 24320, 1, 256] - - [45, 0.0] - - - [33280, 17152, 1, 256] - - [45, 0.0] - - - [27904, 512, 1, 256] - - [64, 0.0] - - - [39680, 8192, 1, 256] - - [69, 0.0] - - - [28160, 8192, 1, 256] - - [69, 0.0] - - - [25088, 8192, 1, 256] - - [69, 0.0] - - - [23040, 15360, 1, 256] - - [69, 0.0] - - - [19712, 11776, 1, 256] - - [64, 0.0] - - - [43520, 27648, 1, 256] - - [69, 0.0] - - - [44544, 4096, 1, 256] - - [69, 0.0] - - - [20224, 4096, 1, 256] - - [69, 0.0] - - - [31744, 4096, 1, 256] - - [69, 0.0] - - - [33024, 16896, 1, 256] - - [64, 0.0] - - - [32768, 8192, 1, 256] - - [69, 0.0] - - - [42752, 4096, 1, 256] - - [69, 0.0] - - - [19968, 512, 1, 256] - - [64, 0.0] - - - [10496, 512, 1, 256] - - [63, 0.0] - - - [36864, 4096, 1, 256] - - [69, 0.0] - - - [12288, 1024, 1, 256] - - [64, 0.0] - - - [22784, 14848, 1, 256] - - [64, 0.0] - - - [17152, 9472, 1, 256] - - [45, 0.0] - - - [31488, 1024, 1, 256] - - [64, 0.0] - - - [25344, 1024, 1, 256] - - [64, 0.0] - - - [33536, 512, 1, 256] - - [64, 0.0] - - - [28672, 8192, 1, 256] - - [69, 0.0] - - - [15104, 7168, 1, 256] - - [64, 0.0] - - - [38144, 22272, 1, 256] - - [45, 0.0] - - - [25344, 4096, 1, 256] - - [69, 0.0] - - - [6400, 2560, 1, 256] - - [64, 0.0] - - - [21248, 13568, 1, 256] - - [45, 0.0] - - - [2304, 1536, 1, 256] - - [57, 0.0] - - - [20992, 512, 1, 256] - - [64, 0.0] - - - [3072, 1024, 1, 256] - - [64, 0.0] - - - [36864, 20736, 1, 256] - - [45, 0.0] - - - [39936, 24064, 1, 256] - - [69, 0.0] - - - [2816, 512, 1, 256] - - [64, 0.0] - - - [37888, 512, 1, 256] - - [64, 0.0] - - - [39680, 1024, 1, 256] - - [64, 0.0] - - - [35584, 19712, 1, 256] - - [45, 0.0] - - - [25600, 9728, 1, 256] - - [69, 0.0] - - - [2816, 1024, 1, 256] - - [57, 0.0] - - - [13056, 1024, 1, 256] - - [64, 0.0] - - - [39680, 4096, 1, 256] - - [69, 0.0] - - - [4864, 3072, 1, 256] - - [64, 0.0] - - - [27648, 11776, 1, 256] - - [69, 0.0] - - - [13056, 4096, 1, 256] - - [64, 0.0] - - - [4096, 2304, 1, 256] - - [59, 0.0] - - - [34048, 1024, 1, 256] - - [64, 0.0] - - - [6400, 512, 1, 256] - - [63, 0.0] - - - [15872, 4096, 1, 256] - - [69, 0.0] - - - [29440, 1024, 1, 256] - - [64, 0.0] - - - [7424, 512, 1, 256] - - [63, 0.0] - - - [19200, 4096, 1, 256] - - [69, 0.0] - - - [37376, 21504, 1, 256] - - [69, 0.0] - - - [37888, 1024, 1, 256] - - [69, 0.0] - - - [40704, 24832, 1, 256] - - [45, 0.0] - - - [26112, 1024, 1, 256] - - [64, 0.0] - - - [25088, 8960, 1, 256] - - [45, 0.0] - - - [27136, 512, 1, 256] - - [63, 0.0] - - - [4608, 512, 1, 256] - - [63, 0.0] - - - [31232, 8192, 1, 256] - - [69, 0.0] - - - [33024, 512, 1, 256] - - [64, 0.0] - - - [27648, 512, 1, 256] - - [64, 0.0] - - - [28928, 4096, 1, 256] - - [64, 0.0] - - - [44544, 2048, 1, 256] - - [69, 0.0] - - - [43776, 27648, 1, 256] - - [69, 0.0] - - - [19456, 4096, 1, 256] - - [69, 0.0] - - - [33536, 17664, 1, 256] - - [45, 0.0] - - - [35328, 4096, 1, 256] - - [69, 0.0] - - - [13312, 5376, 1, 256] - - [45, 0.0] - - - [32768, 1024, 1, 256] - - [69, 0.0] - - - [39168, 4096, 1, 256] - - [69, 0.0] - - - [15616, 7936, 1, 256] - - [45, 0.0] - - - [41472, 25600, 1, 256] - - [69, 0.0] - - - [14592, 4096, 1, 256] - - [69, 0.0] - - - [37632, 21760, 1, 256] - - [45, 0.0] - - - [37376, 21248, 1, 256] - - [45, 0.0] - - - [14336, 6656, 1, 256] - - [69, 0.0] - - - [36608, 20480, 1, 256] - - [69, 0.0] - - - [32256, 16384, 1, 256] - - [69, 0.0] - - - [44544, 28416, 1, 256] - - [45, 0.0] - - - [26112, 512, 1, 256] - - [64, 0.0] - - - [41216, 25344, 1, 256] - - [45, 0.0] - - - [16640, 512, 1, 256] - - [64, 0.0] - - - [30464, 14336, 1, 256] - - [64, 0.0] - - - [13312, 4096, 1, 256] - - [69, 0.0] - - - [22528, 1024, 1, 256] - - [64, 0.0] - - - [5632, 1024, 1, 256] - - [64, 0.0] - - - [27392, 1024, 1, 256] - - [64, 0.0] - - - [27648, 8192, 1, 256] - - [69, 0.0] - - - [26368, 1024, 1, 256] - - [64, 0.0] - - - [43776, 4096, 1, 256] - - [69, 0.0] - - - [23552, 15872, 1, 256] - - [69, 0.0] - - - [26624, 10496, 1, 256] - - [45, 0.0] - - - [27392, 8192, 1, 256] - - [69, 0.0] - - - [17408, 9728, 1, 256] - - [69, 0.0] - - - [16896, 9216, 1, 256] - - [64, 0.0] - - - [26880, 11008, 1, 256] - - [45, 0.0] - - - [31488, 512, 1, 256] - - [64, 0.0] - - - [14336, 6400, 1, 256] - - [44, 0.0] - - - [17152, 512, 1, 256] - - [64, 0.0] - - - [7168, 512, 1, 256] - - [64, 0.0] - - - [41984, 26112, 1, 256] - - [69, 0.0] - - - [11776, 512, 1, 256] - - [64, 0.0] - - - [16128, 8448, 1, 256] - - [45, 0.0] - - - [11520, 1024, 1, 256] - - [63, 0.0] - - - [27904, 1024, 1, 256] - - [64, 0.0] - - - [37888, 8192, 1, 256] - - [69, 0.0] - - - [20480, 12544, 1, 256] - - [45, 0.0] - - - [23552, 15616, 1, 256] - - [45, 0.0] - - - [21504, 13824, 1, 256] - - [69, 0.0] - - - [27136, 11008, 1, 256] - - [45, 0.0] - - - [32000, 512, 1, 256] - - [64, 0.0] - - - [26624, 1024, 1, 256] - - [64, 0.0] - - - [34816, 8192, 1, 256] - - [69, 0.0] - - - [23040, 512, 1, 256] - - [64, 0.0] - - - [36608, 1024, 1, 256] - - [69, 0.0] - - - [43264, 8192, 1, 256] - - [69, 0.0] - - - [30208, 14336, 1, 256] - - [69, 0.0] - - - [43520, 512, 1, 256] - - [64, 0.0] - - - [32256, 4096, 1, 256] - - [69, 0.0] - - - [33792, 17664, 1, 256] - - [45, 0.0] - - - [10752, 6912, 1, 256] - - [45, 0.0] - - - [29696, 8192, 1, 256] - - [69, 0.0] - - - [41472, 512, 1, 256] - - [64, 0.0] - - - [44544, 8192, 1, 256] - - [69, 0.0] - - - [41472, 8192, 1, 256] - - [69, 0.0] - - - [38656, 4096, 1, 256] - - [69, 0.0] - - - [44800, 512, 1, 256] - - [63, 0.0] - - - [37376, 4096, 1, 256] - - [69, 0.0] - - - [19200, 1024, 1, 256] - - [64, 0.0] - - - [39680, 23552, 1, 256] - - [69, 0.0] - - - [30976, 8192, 1, 256] - - [69, 0.0] - - - [25856, 1024, 1, 256] - - [64, 0.0] - - - [22016, 14336, 1, 256] - - [69, 0.0] - - - [17152, 9216, 1, 256] - - [64, 0.0] - - - [18432, 10752, 1, 256] - - [69, 0.0] - - - [5376, 1024, 1, 256] - - [64, 0.0] - - - [21760, 13824, 1, 256] - - [64, 0.0] - - - [15360, 512, 1, 256] - - [64, 0.0] - - - [2560, 512, 1, 256] - - [35, 0.0] - - - [36096, 8192, 1, 256] - - [69, 0.0] - - - [42752, 26624, 1, 256] - - [69, 0.0] - - - [35584, 19456, 1, 256] - - [69, 0.0] - - - [6144, 2304, 1, 256] - - [64, 0.0] - - - [42240, 1024, 1, 256] - - [64, 0.0] - - - [26880, 4096, 1, 256] - - [69, 0.0] - - - [28160, 12032, 1, 256] - - [45, 0.0] - - - [18688, 10752, 1, 256] - - [64, 0.0] - - - [43520, 8192, 1, 256] - - [69, 0.0] - - - [8192, 4352, 1, 256] - - [64, 0.0] - - - [6912, 3072, 1, 256] - - [64, 0.0] - - - [31744, 15616, 1, 256] - - [45, 0.0] - - - [36352, 20224, 1, 256] - - [45, 0.0] - - - [41216, 25088, 1, 256] - - [64, 0.0] - - - [37632, 1024, 1, 256] - - [69, 0.0] - - - [18944, 512, 1, 256] - - [64, 0.0] - - - [15616, 1024, 1, 256] - - [64, 0.0] - - - [44288, 512, 1, 256] - - [64, 0.0] - - - [24832, 8704, 1, 256] - - [64, 0.0] - - - [21504, 13568, 1, 256] - - [45, 0.0] - - - [18176, 10496, 1, 256] - - [45, 0.0] - - - [21248, 1024, 1, 256] - - [64, 0.0] - - - [16384, 1024, 1, 256] - - [64, 0.0] - - - [25600, 8192, 1, 256] - - [69, 0.0] - - - [28672, 12544, 1, 256] - - [45, 0.0] - - - [16128, 1024, 1, 256] - - [64, 0.0] - - - [22272, 14592, 1, 256] - - [45, 0.0] - - - [1280, 512, 1, 256] - - [57, 0.0] - - - [36864, 20992, 1, 256] - - [69, 0.0] - - - [3584, 1792, 1, 256] - - [57, 0.0] - - - [35072, 19200, 1, 256] - - [45, 0.0] - - - [32000, 4096, 1, 256] - - [69, 0.0] - - - [28416, 1024, 1, 256] - - [69, 0.0] - - - [20480, 12800, 1, 256] - - [69, 0.0] - - - [21760, 4096, 1, 256] - - [64, 0.0] - - - [44288, 8192, 1, 256] - - [69, 0.0] - - - [33280, 4096, 1, 256] - - [69, 0.0] - - - [32512, 1024, 1, 256] - - [64, 0.0] - - - [38400, 22528, 1, 256] - - [69, 0.0] - - - [40448, 1024, 1, 256] - - [69, 0.0] - - - [5120, 512, 1, 256] - - [63, 0.0] - - - [29952, 8192, 1, 256] - - [69, 0.0] - - - [40448, 24576, 1, 256] - - [69, 0.0] - - - [29696, 4096, 1, 256] - - [69, 0.0] - - - [21504, 1024, 1, 256] - - [63, 0.0] - - - [19968, 1024, 1, 256] - - [64, 0.0] - - - [16896, 512, 1, 256] - - [63, 0.0] - - - [33536, 17408, 1, 256] - - [69, 0.0] - - - [19712, 512, 1, 256] - - [64, 0.0] - - - [16384, 8704, 1, 256] - - [69, 0.0] - - - [29952, 13824, 1, 256] - - [64, 0.0] - - - [14592, 6656, 1, 256] - - [64, 0.0] - - - [36864, 1024, 1, 256] - - [69, 0.0] - - - [31744, 15872, 1, 256] - - [69, 0.0] - - - [24832, 8960, 1, 256] - - [45, 0.0] - - - [23808, 1024, 1, 256] - - [64, 0.0] - - - [19200, 11264, 1, 256] - - [64, 0.0] - - - [23296, 15360, 1, 256] - - [69, 0.0] - - - [34304, 18432, 1, 256] - - [69, 0.0] - - - [22016, 1024, 1, 256] - - [64, 0.0] - - - [40704, 4096, 1, 256] - - [69, 0.0] - - - [25600, 4096, 1, 256] - - [69, 0.0] - - - [3328, 1024, 1, 256] - - [63, 0.0] - - - [30464, 8192, 1, 256] - - [69, 0.0] - - - [39424, 8192, 1, 256] - - [69, 0.0] - - - [23808, 15872, 1, 256] - - [69, 0.0] - - - [8960, 1024, 1, 256] - - [64, 0.0] - - - [44032, 4096, 1, 256] - - [69, 0.0] - - - [35584, 8192, 1, 256] - - [69, 0.0] - - - [29184, 8192, 1, 256] - - [69, 0.0] - - - [13824, 1024, 1, 256] - - [64, 0.0] - - - [36608, 8192, 1, 256] - - [69, 0.0] - - - [30976, 512, 1, 256] - - [64, 0.0] - - - [33024, 4096, 1, 256] - - [69, 0.0] - - - [11776, 7936, 1, 256] - - [64, 0.0] - - - [23808, 16128, 1, 256] - - [45, 0.0] - - - [22272, 14336, 1, 256] - - [69, 0.0] - - - [27392, 11520, 1, 256] - - [45, 0.0] - - - [30464, 4096, 1, 256] - - [69, 0.0] - - - [20992, 13312, 1, 256] - - [69, 0.0] - - - [44800, 1024, 1, 256] - - [69, 0.0] - - - [32512, 4096, 1, 256] - - [69, 0.0] - - - [23296, 15616, 1, 256] - - [45, 0.0] - - - [9216, 1024, 1, 256] - - [64, 0.0] - - - [20224, 12544, 1, 256] - - [45, 0.0] - - - [32256, 1024, 1, 256] - - [64, 0.0] - - - [38400, 512, 1, 256] - - [64, 0.0] - - - [29952, 1024, 1, 256] - - [64, 0.0] - - - [36352, 512, 1, 256] - - [64, 0.0] - - - [41728, 25600, 1, 256] - - [69, 0.0] - - - [32000, 1024, 1, 256] - - [64, 0.0] - - - [38144, 22016, 1, 256] - - [69, 0.0] - - - [27136, 11264, 1, 256] - - [69, 0.0] - - - [34048, 18176, 1, 256] - - [45, 0.0] - - - [22016, 14080, 1, 256] - - [45, 0.0] - - - [19712, 12032, 1, 256] - - [45, 0.0] - - - [23552, 4096, 1, 256] - - [69, 0.0] - - - [15872, 1024, 1, 256] - - [64, 0.0] - - - [37120, 512, 1, 256] - - [63, 0.0] - - - [9984, 1024, 1, 256] - - [64, 0.0] - - - [32512, 8192, 1, 256] - - [69, 0.0] - - - [15360, 4096, 1, 256] - - [69, 0.0] - - - [13056, 512, 1, 256] - - [64, 0.0] - - - [44032, 8192, 1, 256] - - [69, 0.0] - - - [24576, 8192, 1, 256] - - [69, 0.0] - - - [36352, 8192, 1, 256] - - [69, 0.0] - - - [26368, 8192, 1, 256] - - [69, 0.0] - - - [20480, 1024, 1, 256] - - [64, 0.0] - - - [35072, 8192, 1, 256] - - [69, 0.0] - - - [32000, 15872, 1, 256] - - [69, 0.0] - - - [40704, 24576, 1, 256] - - [69, 0.0] - - - [15104, 7424, 1, 256] - - [45, 0.0] - - - [25856, 4096, 1, 256] - - [69, 0.0] - - - [14848, 512, 1, 256] - - [64, 0.0] - - - [39424, 4096, 1, 256] - - [69, 0.0] - - - [24832, 512, 1, 256] - - [64, 0.0] - - - [44288, 28416, 1, 256] - - [45, 0.0] - - - [12544, 4608, 1, 256] - - [64, 0.0] - - - [12800, 4864, 1, 256] - - [64, 0.0] - - - [29440, 512, 1, 256] - - [63, 0.0] - - - [40192, 24064, 1, 256] - - [69, 0.0] - - - [18176, 4096, 1, 256] - - [69, 0.0] - - - [40960, 8192, 1, 256] - - [69, 0.0] - - - [42240, 512, 1, 256] - - [64, 0.0] - - - [9728, 512, 1, 256] - - [63, 0.0] - - - [14848, 7168, 1, 256] - - [64, 0.0] - - - [44800, 28672, 1, 256] - - [69, 0.0] - - - [15616, 7680, 1, 256] - - [64, 0.0] - - - [33280, 17408, 1, 256] - - [69, 0.0] - - - [42752, 1024, 1, 256] - - [69, 0.0] - - - [35328, 8192, 1, 256] - - [69, 0.0] - - - [36352, 1024, 1, 256] - - [64, 0.0] - - - [35840, 1024, 1, 256] - - [64, 0.0] - - - [41472, 4096, 1, 256] - - [69, 0.0] - - - [3584, 1024, 1, 256] - - [63, 0.0] - - - [22528, 14592, 1, 256] - - [45, 0.0] - - - [44032, 512, 1, 256] - - [63, 0.0] - - - [30720, 1024, 1, 256] - - [64, 0.0] - - - [39680, 512, 1, 256] - - [64, 0.0] - - - [22272, 1024, 1, 256] - - [64, 0.0] - - - [42240, 26368, 1, 256] - - [45, 0.0] - - - [10240, 6400, 1, 256] - - [45, 0.0] - - - [30976, 14848, 1, 256] - - [69, 0.0] - - - [41728, 25856, 1, 256] - - [45, 0.0] - - - [28928, 12800, 1, 256] - - [64, 0.0] - - - [21760, 14080, 1, 256] - - [45, 0.0] - - - [5888, 1024, 1, 256] - - [64, 0.0] - - - [24576, 8704, 1, 256] - - [69, 0.0] - - - [38912, 4096, 1, 256] - - [69, 0.0] - - - [15360, 1024, 1, 256] - - [64, 0.0] - - - [18688, 512, 1, 256] - - [64, 0.0] - - - [27392, 512, 1, 256] - - [64, 0.0] - - - [22784, 512, 1, 256] - - [64, 0.0] - - - [40448, 4096, 1, 256] - - [69, 0.0] - - - [19200, 512, 1, 256] - - [64, 0.0] - - - [26368, 10496, 1, 256] - - [45, 0.0] - - - [25088, 9216, 1, 256] - - [69, 0.0] - - - [33536, 1024, 1, 256] - - [64, 0.0] - - - [25600, 9472, 1, 256] - - [45, 0.0] - - - [13824, 4096, 1, 256] - - [64, 0.0] - - - [5632, 3840, 1, 256] - - [64, 0.0] - - - [9216, 5376, 1, 256] - - [45, 0.0] - - - [8960, 5120, 1, 256] - - [64, 0.0] - - - [19456, 512, 1, 256] - - [64, 0.0] - - - [24576, 4096, 1, 256] - - [64, 0.0] - - - [27392, 11264, 1, 256] - - [69, 0.0] - - - [35072, 4096, 1, 256] - - [69, 0.0] - - - [44288, 4096, 1, 256] - - [69, 0.0] - - - [40448, 8192, 1, 256] - - [69, 0.0] - - - [33280, 512, 1, 256] - - [64, 0.0] - - - [22272, 4096, 1, 256] - - [69, 0.0] - - - [35584, 512, 1, 256] - - [64, 0.0] - - - [10752, 512, 1, 256] - - [64, 0.0] - - - [19968, 4096, 1, 256] - - [69, 0.0] - - - [34304, 1024, 1, 256] - - [64, 0.0] - - - [41216, 8192, 1, 256] - - [69, 0.0] - - - [35840, 19712, 1, 256] - - [45, 0.0] - - - [43520, 27392, 1, 256] - - [45, 0.0] - - - [30720, 14848, 1, 256] - - [69, 0.0] - - - [38400, 22272, 1, 256] - - [45, 0.0] - - - [1536, 1024, 1, 256] - - [57, 0.0] - - - [40192, 1024, 1, 256] - - [69, 0.0] - - - [44800, 256, 1, 256] - - [67, 0.0] - - - [1536, 512, 1, 256] - - [34, 0.0] - - - [34560, 18432, 1, 256] - - [69, 0.0] - - - [1792, 1024, 1, 256] - - [57, 0.0] - - - [5376, 3584, 1, 256] - - [64, 0.0] - - - [30208, 1024, 1, 256] - - [64, 0.0] - - - [31232, 512, 1, 256] - - [63, 0.0] - - - [23040, 4096, 1, 256] - - [69, 0.0] - - - [35840, 4096, 1, 256] - - [69, 0.0] - - - [38144, 512, 1, 256] - - [64, 0.0] - - - [31744, 512, 1, 256] - - [64, 0.0] - - - [14592, 6912, 1, 256] - - [45, 0.0] - - - [19456, 11520, 1, 256] - - [45, 0.0] - - - [7168, 1024, 1, 256] - - [63, 0.0] - - - [18944, 11264, 1, 256] - - [69, 0.0] - - - [19712, 1024, 1, 256] - - [64, 0.0] - - - [26112, 9984, 1, 256] - - [45, 0.0] - - - [38656, 22784, 1, 256] - - [45, 0.0] - - - [24320, 8192, 1, 256] - - [69, 0.0] - - - [4864, 1024, 1, 256] - - [63, 0.0] - - - [20480, 4096, 1, 256] - - [69, 0.0] - - - [10240, 1024, 1, 256] - - [64, 0.0] - - - [31232, 15360, 1, 256] - - [69, 0.0] - - - [24320, 4096, 1, 256] - - [67, 0.0] - - - [33792, 1024, 1, 256] - - [64, 0.0] - - - [12032, 1024, 1, 256] - - [64, 0.0] - - - [39168, 512, 1, 256] - - [64, 0.0] - - - [16896, 4096, 1, 256] - - [69, 0.0] - - - [36096, 1024, 1, 256] - - [69, 0.0] - - - [28416, 12544, 1, 256] - - [45, 0.0] - - - [30720, 4096, 1, 256] - - [69, 0.0] - - - [19712, 4096, 1, 256] - - [69, 0.0] - - - [37120, 21248, 1, 256] - - [45, 0.0] - - - [16384, 4096, 1, 256] - - [69, 0.0] - - - [18688, 11008, 1, 256] - - [45, 0.0] - - - [38400, 8192, 1, 256] - - [69, 0.0] - - - [11264, 7424, 1, 256] - - [45, 0.0] - - - [23296, 512, 1, 256] - - [64, 0.0] - - - [25344, 512, 1, 256] - - [64, 0.0] - - - [44544, 256, 1, 256] - - [64, 0.0] - - - [43264, 4096, 1, 256] - - [69, 0.0] - - - [32512, 16640, 1, 256] - - [45, 0.0] - - - [39936, 8192, 1, 256] - - [69, 0.0] - - - [43264, 512, 1, 256] - - [63, 0.0] - - - [16640, 8704, 1, 256] - - [64, 0.0] - - - [26624, 8192, 1, 256] - - [69, 0.0] - - - [35328, 19456, 1, 256] - - [69, 0.0] - - - [42752, 26880, 1, 256] - - [45, 0.0] - - - [25344, 9216, 1, 256] - - [69, 0.0] - - - [34048, 8192, 1, 256] - - [69, 0.0] - - - [18688, 4096, 1, 256] - - [69, 0.0] - - - [37632, 8192, 1, 256] - - [69, 0.0] - - - [19968, 12032, 1, 256] - - [45, 0.0] - - - [8448, 4608, 1, 256] - - [64, 0.0] - - - [2048, 1536, 1, 256] - - [57, 0.0] - - - [31488, 15616, 1, 256] - - [45, 0.0] - - - [35328, 512, 1, 256] - - [64, 0.0] - - - [37376, 8192, 1, 256] - - [69, 0.0] - - - [33792, 8192, 1, 256] - - [69, 0.0] - - - [36608, 4096, 1, 256] - - [69, 0.0] - - - [28416, 8192, 1, 256] - - [69, 0.0] - - - [5632, 512, 1, 256] - - [40, 0.0] - - - [13568, 4096, 1, 256] - - [64, 0.0] - - - [17664, 9728, 1, 256] - - [64, 0.0] - - - [13568, 1024, 1, 256] - - [64, 0.0] - - - [8448, 512, 1, 256] - - [64, 0.0] - - - [22528, 4096, 1, 256] - - [69, 0.0] - - - [33536, 8192, 1, 256] - - [69, 0.0] - - - [23296, 1024, 1, 256] - - [64, 0.0] - - - [43520, 4096, 1, 256] - - [69, 0.0] - - - [39936, 23808, 1, 256] - - [45, 0.0] - - - [12544, 4096, 1, 256] - - [64, 0.0] - - - [22016, 4096, 1, 256] - - [69, 0.0] - - - [14592, 512, 1, 256] - - [63, 0.0] - - - [39936, 4096, 1, 256] - - [69, 0.0] - - - [18176, 1024, 1, 256] - - [64, 0.0] - - - [44800, 2048, 1, 256] - - [67, 0.0] - - - [14848, 4096, 1, 256] - - [69, 0.0] - - - [20224, 12288, 1, 256] - - [69, 0.0] - - - [16896, 8960, 1, 256] - - [45, 0.0] - - - [43264, 27392, 1, 256] - - [45, 0.0] - - - [24064, 16128, 1, 256] - - [45, 0.0] - - - [1024, 512, 1, 256] - - [127, 0.0] - - - [24576, 8448, 1, 256] - - [45, 0.0] - - - [25344, 9472, 1, 256] - - [45, 0.0] - - - [3328, 1536, 1, 256] - - [64, 0.0] - - - [31488, 4096, 1, 256] - - [69, 0.0] - - - [43008, 8192, 1, 256] - - [69, 0.0] - - - [28672, 12800, 1, 256] - - [69, 0.0] - - - [20736, 13056, 1, 256] - - [45, 0.0] - - - [17664, 9984, 1, 256] - - [45, 0.0] - - - [17920, 1024, 1, 256] - - [64, 0.0] - - - [11008, 1024, 1, 256] - - [63, 0.0] - - - [44800, 4096, 1, 256] - - [69, 0.0] - - - [29952, 14080, 1, 256] - - [45, 0.0] - - - [39168, 23296, 1, 256] - - [45, 0.0] - - - [9472, 512, 1, 256] - - [64, 0.0] - - - [27904, 8192, 1, 256] - - [69, 0.0] - - - [5120, 1024, 1, 256] - - [64, 0.0] - - - [15872, 7936, 1, 256] - - [64, 0.0] - - - [13568, 5632, 1, 256] - - [64, 0.0] - - - [17920, 9984, 1, 256] - - [45, 0.0] - - - [16640, 8960, 1, 256] - - [45, 0.0] - - - [41984, 4096, 1, 256] - - [69, 0.0] - - - [6912, 512, 1, 256] - - [64, 0.0] - - - [28416, 4096, 1, 256] - - [69, 0.0] - - - [27648, 11520, 1, 256] - - [45, 0.0] - - - [7680, 3840, 1, 256] - - [64, 0.0] - - - [34048, 4096, 1, 256] - - [69, 0.0] - - - [11264, 512, 1, 256] - - [63, 0.0] - - - [26368, 4096, 1, 256] - - [69, 0.0] - - - [21248, 13312, 1, 256] - - [69, 0.0] - - - [15104, 1024, 1, 256] - - [64, 0.0] - - - [35072, 18944, 1, 256] - - [69, 0.0] - - - [6144, 1024, 1, 256] - - [64, 0.0] - - - [44800, 8192, 1, 256] - - [69, 0.0] - - - [25088, 512, 1, 256] - - [64, 0.0] - - - [27904, 12032, 1, 256] - - [45, 0.0] - - - [27648, 1024, 1, 256] - - [64, 0.0] - - - [28928, 8192, 1, 256] - - [69, 0.0] - - - [29440, 13312, 1, 256] - - [69, 0.0] - - - [43264, 27136, 1, 256] - - [69, 0.0] - - - [23552, 512, 1, 256] - - [64, 0.0] - - - [26880, 10752, 1, 256] - - [64, 0.0] - - - [44032, 28160, 1, 256] - - [69, 0.0] - - - [36096, 512, 1, 256] - - [64, 0.0] - - - [4352, 2560, 1, 256] - - [64, 0.0] - - - [38912, 8192, 1, 256] - - [69, 0.0] - - - [12032, 4096, 1, 256] - - [64, 0.0] - - - [37632, 512, 1, 256] - - [64, 0.0] - - - [30208, 512, 1, 256] - - [64, 0.0] - - - [2304, 512, 1, 256] - - [58, 0.0] - - - [24320, 8448, 1, 256] - - [45, 0.0] - - - [39424, 512, 1, 256] - - [64, 0.0] - - - [37632, 21504, 1, 256] - - [69, 0.0] - - - [17152, 1024, 1, 256] - - [64, 0.0] - - - [22784, 15104, 1, 256] - - [45, 0.0] - - - [27904, 11776, 1, 256] - - [64, 0.0] - - - [43008, 26880, 1, 256] - - [45, 0.0] - - - [41728, 4096, 1, 256] - - [69, 0.0] - - - [25344, 8192, 1, 256] - - [69, 0.0] - - - [44800, 28928, 1, 256] - - [45, 0.0] - - - [38912, 22784, 1, 256] - - [45, 0.0] - - - [44032, 1024, 1, 256] - - [69, 0.0] - - - [30976, 4096, 1, 256] - - [69, 0.0] - - - [15872, 8192, 1, 256] - - [64, 0.0] - - - [40960, 4096, 1, 256] - - [69, 0.0] - - - [35584, 1024, 1, 256] - - [69, 0.0] - - - [18944, 4096, 1, 256] - - [64, 0.0] - - - [36096, 20224, 1, 256] - - [45, 0.0] - - - [11008, 7168, 1, 256] - - [64, 0.0] - - - [7936, 1024, 1, 256] - - [64, 0.0] - - - [44288, 1024, 1, 256] - - [64, 0.0] - - - [38656, 8192, 1, 256] - - [69, 0.0] - - - [38144, 1024, 1, 256] - - [64, 0.0] - - - [41984, 1024, 1, 256] - - [69, 0.0] - - - [20736, 512, 1, 256] - - [63, 0.0] - - - [32768, 16640, 1, 256] - - [45, 0.0] - - - [40960, 1024, 1, 256] - - [69, 0.0] - - - [25856, 9984, 1, 256] - - [45, 0.0] - - - [29696, 13824, 1, 256] - - [69, 0.0] - - - [37120, 4096, 1, 256] - - [69, 0.0] - - - [37120, 20992, 1, 256] - - [69, 0.0] - - - [35072, 512, 1, 256] - - [64, 0.0] - - - [38656, 1024, 1, 256] - - [64, 0.0] - - - [37376, 512, 1, 256] - - [64, 0.0] - - - [32000, 16128, 1, 256] - - [45, 0.0] - - - [41984, 25856, 1, 256] - - [45, 0.0] - - - [23040, 15104, 1, 256] - - [45, 0.0] - - - [31232, 15104, 1, 256] - - [45, 0.0] - - - [25088, 4096, 1, 256] - - [64, 0.0] - - - [15360, 7424, 1, 256] - - [45, 0.0] - - - [16384, 8448, 1, 256] - - [45, 0.0] - - - [26624, 4096, 1, 256] - - [69, 0.0] - - - [14080, 6400, 1, 256] - - [64, 0.0] - - - [16128, 4096, 1, 256] - - [69, 0.0] - - - [43776, 27904, 1, 256] - - [45, 0.0] - - - [15872, 512, 1, 256] - - [64, 0.0] - - - [43776, 8192, 1, 256] - - [69, 0.0] - - - [10496, 6656, 1, 256] - - [64, 0.0] - - - [13312, 512, 1, 256] - - [63, 0.0] - - - [29184, 512, 1, 256] - - [64, 0.0] - - - [15360, 7680, 1, 256] - - [69, 0.0] - - - [40192, 8192, 1, 256] - - [69, 0.0] - - - [34560, 8192, 1, 256] - - [69, 0.0] - - - [25856, 8192, 1, 256] - - [69, 0.0] - - - [32512, 16384, 1, 256] - - [69, 0.0] - - - [12288, 4352, 1, 256] - - [45, 0.0] - - - [29440, 13568, 1, 256] - - [45, 0.0] - - - [28160, 1024, 1, 256] - - [64, 0.0] - - - [32768, 4096, 1, 256] - - [69, 0.0] - - - [24832, 4096, 1, 256] - - [69, 0.0] - - - [39680, 23808, 1, 256] - - [45, 0.0] - - - [22784, 4096, 1, 256] - - [69, 0.0] - - - [7936, 4096, 1, 256] - - [64, 0.0] - - - [8704, 4864, 1, 256] - - [64, 0.0] - - - [29696, 512, 1, 256] - - [64, 0.0] - - - [39424, 23296, 1, 256] - - [45, 0.0] - - - [17408, 9472, 1, 256] - - [45, 0.0] - - - [33792, 4096, 1, 256] - - [69, 0.0] - - - [17920, 512, 1, 256] - - [64, 0.0] - - - [25856, 512, 1, 256] - - [64, 0.0] - - - [44288, 28160, 1, 256] - - [69, 0.0] - - - [40192, 4096, 1, 256] - - [69, 0.0] - - - [21248, 512, 1, 256] - - [64, 0.0] - - - [3072, 512, 1, 256] - - [64, 0.0] - - - [29184, 13312, 1, 256] - - [69, 0.0] - - - [44544, 1024, 1, 256] - - [69, 0.0] - - - [37888, 21760, 1, 256] - - [44, 0.0] - - - [33792, 17920, 1, 256] - - [69, 0.0] - - - [6912, 1024, 1, 256] - - [63, 0.0] - - - [41216, 512, 1, 256] - - [64, 0.0] - - - [42240, 26112, 1, 256] - - [69, 0.0] - - - [30720, 8192, 1, 256] - - [69, 0.0] - - - [11776, 1024, 1, 256] - - [64, 0.0] - - - [43008, 4096, 1, 256] - - [69, 0.0] - - - [34560, 18688, 1, 256] - - [45, 0.0] - - - [41984, 512, 1, 256] - - [64, 0.0] - - - [41728, 512, 1, 256] - - [64, 0.0] - - - [2560, 1792, 1, 256] - - [59, 0.0] - - - [36864, 8192, 1, 256] - - [69, 0.0] - - - [40704, 8192, 1, 256] - - [69, 0.0] - - - [30720, 14592, 1, 256] - - [44, 0.0] - - - [32256, 512, 1, 256] - - [64, 0.0] - - - [40192, 512, 1, 256] - - [64, 0.0] - - - [8960, 512, 1, 256] - - [63, 0.0] - - - [16640, 4096, 1, 256] - - [64, 0.0] - - - [30976, 15104, 1, 256] - - [45, 0.0] - - - [27136, 8192, 1, 256] - - [69, 0.0] - - - [30208, 8192, 1, 256] - - [69, 0.0] - - - [21504, 512, 1, 256] - - [64, 0.0] - - - [9728, 5888, 1, 256] - - [64, 0.0] - - - [38912, 23040, 1, 256] - - [69, 0.0] - - - [7424, 1024, 1, 256] - - [64, 0.0] - - - [38656, 22528, 1, 256] - - [69, 0.0] - - - [26880, 512, 1, 256] - - [64, 0.0] - - - [29184, 13056, 1, 256] - - [45, 0.0] - - - [44032, 27904, 1, 256] - - [45, 0.0] - - - [38144, 8192, 1, 256] - - [69, 0.0] - - - [29952, 512, 1, 256] - - [64, 0.0] - - - [18432, 4096, 1, 256] - - [69, 0.0] - - - [28160, 12288, 1, 256] - - [69, 0.0] - - - [29696, 1024, 1, 256] - - [64, 0.0] - - - [39936, 1024, 1, 256] - - [69, 0.0] - - - [25600, 512, 1, 256] - - [63, 0.0] - - - [40448, 24320, 1, 256] - - [45, 0.0] - - - [40448, 512, 1, 256] - - [63, 0.0] - - - [7424, 3584, 1, 256] - - [64, 0.0] - - - [5376, 512, 1, 256] - - [63, 0.0] - - - [27136, 4096, 1, 256] - - [69, 0.0] - - - [35840, 19968, 1, 256] - - [69, 0.0] - - - [18944, 11008, 1, 256] - - [45, 0.0] - - - [34816, 18688, 1, 256] - - [45, 0.0] - - - [38400, 1024, 1, 256] - - [69, 0.0] - - - [36352, 20480, 1, 256] - - [69, 0.0] - - - [36608, 20736, 1, 256] - - [45, 0.0] - - - [28672, 1024, 1, 256] - - [64, 0.0] - - - [42496, 26624, 1, 256] - - [69, 0.0] - - - [31488, 15360, 1, 256] - - [64, 0.0] - - - [20992, 4096, 1, 256] - - [69, 0.0] - - - [12544, 512, 1, 256] - - [64, 0.0] - - - [24064, 8192, 1, 256] - - [69, 0.0] - - - [26880, 8192, 1, 256] - - [69, 0.0] - - - [4352, 512, 1, 256] - - [64, 0.0] - - - [7680, 1024, 1, 256] - - [64, 0.0] - - - [16128, 8192, 1, 256] - - [64, 0.0] - - - [39168, 8192, 1, 256] - - [69, 0.0] - - - [29440, 4096, 1, 256] - - [69, 0.0] - - - [33536, 4096, 1, 256] - - [69, 0.0] - - - [33024, 17152, 1, 256] - - [45, 0.0] - - - [34816, 18944, 1, 256] - - [69, 0.0] - - - [22016, 512, 1, 256] - - [64, 0.0] - - - [14848, 6912, 1, 256] - - [45, 0.0] - - - [20736, 12800, 1, 256] - - [64, 0.0] - - - [32256, 16128, 1, 256] - - [45, 0.0] - - - [7680, 512, 1, 256] - - [63, 0.0] - - - [19968, 12288, 1, 256] - - [69, 0.0] - - - [29184, 4096, 1, 256] - - [69, 0.0] - - - [15616, 4096, 1, 256] - - [64, 0.0] - - - [44544, 28672, 1, 256] - - [69, 0.0] - - - [26112, 4096, 1, 256] - - [69, 0.0] - - - [26624, 10752, 1, 256] - - [69, 0.0] - - - [15104, 4096, 1, 256] - - [64, 0.0] - - - [23296, 4096, 1, 256] - - [69, 0.0] - - - [37888, 22016, 1, 256] - - [69, 0.0] - - - [11520, 7680, 1, 256] - - [64, 0.0] - - - [41728, 1024, 1, 256] - - [69, 0.0] - - - [2304, 1792, 1, 256] - - [57, 0.0] - - - [34048, 17920, 1, 256] - - [69, 0.0] - - - [1536, 768, 1, 256] - - [56, 0.0] - - - [33280, 8192, 1, 256] - - [69, 0.0] - - - [11264, 1024, 1, 256] - - [64, 0.0] - - - [21760, 1024, 1, 256] - - [64, 0.0] - - - [18432, 10496, 1, 256] - - [45, 0.0] - - - [41216, 4096, 1, 256] - - [69, 0.0] - - - [41472, 25344, 1, 256] - - [45, 0.0] - - - [17408, 1024, 1, 256] - - [64, 0.0] - - - [19456, 1024, 1, 256] - - [64, 0.0] - - - [36096, 19968, 1, 256] - - [69, 0.0] - - - [8704, 512, 1, 256] - - [63, 0.0] - - - [30464, 1024, 1, 256] - - [64, 0.0] - - - [8192, 1024, 1, 256] - - [64, 0.0] - - - [11520, 512, 1, 256] - - [64, 0.0] - - - [44544, 512, 1, 256] - - [63, 0.0] - - - [20736, 4096, 1, 256] - - [64, 0.0] - - - [42752, 8192, 1, 256] - - [69, 0.0] - - - [39936, 512, 1, 256] - - [64, 0.0] - - - [42496, 26368, 1, 256] - - [45, 0.0] - - - [28672, 4096, 1, 256] - - [69, 0.0] - - - [35840, 8192, 1, 256] - - [69, 0.0] - - - [17664, 1024, 1, 256] - - [64, 0.0] - - - [21248, 4096, 1, 256] - - [69, 0.0] - - - [1280, 768, 1, 256] - - [58, 0.0] - - - [28160, 512, 1, 256] - - [64, 0.0] - - - [34304, 18176, 1, 256] - - [45, 0.0] - - - [19200, 11520, 1, 256] - - [45, 0.0] - - - [25856, 9728, 1, 256] - - [64, 0.0] - - - [35328, 19200, 1, 256] - - [45, 0.0] - - - [29440, 8192, 1, 256] - - [69, 0.0] - - - [20992, 13056, 1, 256] - - [45, 0.0] - - - [21760, 512, 1, 256] - - [64, 0.0] - - - [12800, 512, 1, 256] - - [63, 0.0] - - - [28416, 12288, 1, 256] - - [64, 0.0] - - - [29696, 13568, 1, 256] - - [45, 0.0] - - - [21504, 4096, 1, 256] - - [69, 0.0] - - - [30464, 14592, 1, 256] - - [45, 0.0] - - - [13056, 5120, 1, 256] - - [64, 0.0] - - - [34560, 4096, 1, 256] - - [69, 0.0] - - - [32768, 16896, 1, 256] - - [69, 0.0] - - - [13824, 5888, 1, 256] - - [45, 0.0] - - - [33024, 8192, 1, 256] - - [69, 0.0] - - - [14080, 4096, 1, 256] - - [64, 0.0] - - - [43008, 1024, 1, 256] - - [69, 0.0] - - - [31744, 1024, 1, 256] - - [64, 0.0] - - - [11008, 512, 1, 256] - - [64, 0.0] - - - [24832, 8192, 1, 256] - - [69, 0.0] - - - [43776, 512, 1, 256] - - [64, 0.0] - - - [24064, 1024, 1, 256] - - [63, 0.0] - - - [12800, 4096, 1, 256] - - [64, 0.0] - - - [19456, 11776, 1, 256] - - [69, 0.0] - - - [22528, 14848, 1, 256] - - [69, 0.0] - - - [30208, 14080, 1, 256] - - [45, 0.0] - - - [40704, 1024, 1, 256] - - [69, 0.0] - - - [35584, 4096, 1, 256] - - [69, 0.0] - - - [26112, 8192, 1, 256] - - [69, 0.0] - - - [9472, 5632, 1, 256] - - [64, 0.0] - - - [15616, 512, 1, 256] - - [64, 0.0] - - - [34816, 4096, 1, 256] - - [69, 0.0] - - - [31232, 4096, 1, 256] - - [69, 0.0] - - - [9728, 1024, 1, 256] - - [63, 0.0] - - - [13312, 1024, 1, 256] - - [64, 0.0] - - - [20224, 1024, 1, 256] - - [64, 0.0] - - - [4864, 512, 1, 256] - - [64, 0.0] - - - [34304, 4096, 1, 256] - - [69, 0.0] - - - [43776, 1024, 1, 256] - - [69, 0.0] - - - [37120, 8192, 1, 256] - - [69, 0.0] - - - [33792, 512, 1, 256] - - [64, 0.0] - - - [42496, 512, 1, 256] - - [64, 0.0] - - - [9216, 512, 1, 256] - - [63, 0.0] - - - [14336, 4096, 1, 256] - - [64, 0.0] - - - [43008, 27136, 1, 256] - - [69, 0.0] - - - [35840, 512, 1, 256] - - [64, 0.0] - - - [40960, 25088, 1, 256] - - [69, 0.0] - - - [17408, 512, 1, 256] - - [64, 0.0] - - - [12288, 4096, 1, 256] - - [64, 0.0] - - - [6656, 512, 1, 256] - - [64, 0.0] - - - [40960, 24832, 1, 256] - - [45, 0.0] - - - [39168, 23040, 1, 256] - - [69, 0.0] - - - [512, 1, 1, 128] - - [134, 0.0] - - - [384, 1, 1, 384] - - [134, 0.0] - - - [256, 1, 1, 256] - - [134, 0.0] - - - [128, 1, 1, 128] - - [134, 0.0] - - - [640, 1, 1, 128] - - [134, 0.0] - - - [1, 128, 1, 256] - - [134, 0.0] - - - [512, 128, 1, 256] - - [145, 0.0] - - - [2049, 128, 1, 256] - - [56, 0.0] - - - [49, 128, 1, 256] - - [147, 0.0] - - - [1537, 128, 1, 256] - - [56, 0.0] - - - [257, 128, 1, 256] - - [54, 0.0] - - - [9728, 128, 1, 256] - - [68, 0.0] - - - [3840, 128, 1, 256] - - [58, 0.0] - - - [1280, 128, 1, 256] - - [56, 0.0] - - - [7168, 128, 1, 256] - - [132, 0.0] - - - [6656, 128, 1, 256] - - [132, 0.0] - - - [2561, 128, 1, 256] - - [62, 0.0] - - - [6912, 128, 1, 256] - - [132, 0.0] - - - [2048, 128, 1, 256] - - [62, 0.0] - - - [2304, 128, 1, 256] - - [58, 0.0] - - - [1536, 128, 1, 256] - - [56, 0.0] - - - [4864, 128, 1, 256] - - [34, 0.0] - - - [8448, 128, 1, 256] - - [131, 0.0] - - - [3072, 128, 1, 256] - - [65, 0.0] - - - [3329, 128, 1, 256] - - [58, 0.0] - - - [3328, 128, 1, 256] - - [58, 0.0] - - - [8960, 128, 1, 256] - - [130, 0.0] - - - [9216, 128, 1, 256] - - [53, 0.0] - - - [2817, 128, 1, 256] - - [56, 0.0] - - - [6400, 128, 1, 256] - - [138, 0.0] - - - [561, 128, 1, 256] - - [143, 0.0] - - - [2816, 128, 1, 256] - - [58, 0.0] - - - [3073, 128, 1, 256] - - [56, 0.0] - - - [2097, 128, 1, 256] - - [58, 0.0] - - - [768, 128, 1, 256] - - [54, 0.0] - - - [9984, 128, 1, 256] - - [62, 0.0] - - - [3584, 128, 1, 256] - - [58, 0.0] - - - [817, 128, 1, 256] - - [60, 0.0] - - - [5632, 128, 1, 256] - - [140, 0.0] - - - [9472, 128, 1, 256] - - [58, 0.0] - - - [2305, 128, 1, 256] - - [58, 0.0] - - - [1329, 128, 1, 256] - - [62, 0.0] - - - [5888, 128, 1, 256] - - [136, 0.0] - - - [7680, 128, 1, 256] - - [132, 0.0] - - - [4608, 128, 1, 256] - - [139, 0.0] - - - [2353, 128, 1, 256] - - [58, 0.0] - - - [5120, 128, 1, 256] - - [137, 0.0] - - - [769, 128, 1, 256] - - [54, 0.0] - - - [1792, 128, 1, 256] - - [56, 0.0] - - - [1073, 128, 1, 256] - - [62, 0.0] - - - [513, 128, 1, 256] - - [60, 0.0] - - - [4096, 128, 1, 256] - - [58, 0.0] - - - [7424, 128, 1, 256] - - [132, 0.0] - - - [4352, 128, 1, 256] - - [34, 0.0] - - - [1793, 128, 1, 256] - - [58, 0.0] - - - [8192, 128, 1, 256] - - [40, 0.0] - - - [1281, 128, 1, 256] - - [56, 0.0] - - - [305, 128, 1, 256] - - [142, 0.0] - - - [2560, 128, 1, 256] - - [58, 0.0] - - - [2609, 128, 1, 256] - - [58, 0.0] - - - [1585, 128, 1, 256] - - [62, 0.0] - - - [8704, 128, 1, 256] - - [135, 0.0] - - - [10240, 128, 1, 256] - - [144, 0.0] - - - [256, 128, 1, 256] - - [146, 0.0] - - - [1025, 128, 1, 256] - - [56, 0.0] - - - [2865, 128, 1, 256] - - [58, 0.0] - - - [5376, 128, 1, 256] - - [59, 0.0] - - - [1841, 128, 1, 256] - - [56, 0.0] - - - [7936, 128, 1, 256] - - [42, 0.0] - - - [6144, 128, 1, 256] - - [141, 0.0] - - - [1024, 128, 1, 256] - - [145, 0.0] - - - [36096, 1281, 1, 256] - - [45, 0.0] - - - [38656, 2816, 1, 256] - - [47, 0.0] - - - [35072, 2048, 1, 256] - - [53, 0.0] - - - [39424, 2865, 1, 256] - - [53, 0.0] - - - [39168, 3328, 1, 256] - - [47, 0.0] - - - [36096, 2865, 1, 256] - - [47, 0.0] - - - [39216, 5632, 1, 256] - - [53, 0.0] - - - [38144, 6144, 1, 256] - - [53, 0.0] - - - [35328, 3072, 1, 256] - - [53, 0.0] - - - [39936, 256, 1, 256] - - [47, 0.0] - - - [36864, 3328, 1, 256] - - [47, 0.0] - - - [39168, 6144, 1, 256] - - [53, 0.0] - - - [36352, 4352, 1, 256] - - [47, 0.0] - - - [37680, 10240, 1, 256] - - [53, 0.0] - - - [38144, 256, 1, 256] - - [47, 0.0] - - - [37632, 1281, 1, 256] - - [45, 0.0] - - - [35632, 1792, 1, 256] - - [53, 0.0] - - - [36096, 4096, 1, 256] - - [53, 0.0] - - - [36144, 2816, 1, 256] - - [47, 0.0] - - - [36352, 256, 1, 256] - - [53, 0.0] - - - [35888, 2865, 1, 256] - - [47, 0.0] - - - [38912, 1280, 1, 256] - - [47, 0.0] - - - [37120, 3072, 1, 256] - - [53, 0.0] - - - [38448, 10240, 1, 256] - - [53, 0.0] - - - [39936, 3328, 1, 256] - - [47, 0.0] - - - [39168, 10240, 1, 256] - - [53, 0.0] - - - [39680, 3329, 1, 256] - - [47, 0.0] - - - [37168, 2865, 1, 256] - - [53, 0.0] - - - [38144, 5888, 1, 256] - - [47, 0.0] - - - [37120, 1281, 1, 256] - - [45, 0.0] - - - [37376, 10240, 1, 256] - - [53, 0.0] - - - [38704, 5120, 1, 256] - - [53, 0.0] - - - [39168, 5376, 1, 256] - - [47, 0.0] - - - [38656, 2865, 1, 256] - - [47, 0.0] - - - [37376, 3584, 1, 256] - - [53, 0.0] - - - [35072, 6144, 1, 256] - - [53, 0.0] - - - [39936, 6144, 1, 256] - - [53, 0.0] - - - [37632, 5376, 1, 256] - - [47, 0.0] - - - [36352, 2304, 1, 256] - - [47, 0.0] - - - [35840, 2048, 1, 256] - - [53, 0.0] - - - [36608, 1280, 1, 256] - - [47, 0.0] - - - [39936, 1792, 1, 256] - - [47, 0.0] - - - [36608, 3329, 1, 256] - - [47, 0.0] - - - [35072, 3329, 1, 256] - - [47, 0.0] - - - [37168, 3584, 1, 256] - - [53, 0.0] - - - [36096, 1792, 1, 256] - - [47, 0.0] - - - [39424, 3329, 1, 256] - - [47, 0.0] - - - [39424, 2048, 1, 256] - - [53, 0.0] - - - [39984, 2865, 1, 256] - - [47, 0.0] - - - [38448, 256, 1, 256] - - [53, 0.0] - - - [35584, 256, 1, 256] - - [53, 0.0] - - - [36608, 10240, 1, 256] - - [53, 0.0] - - - [38960, 5376, 1, 256] - - [53, 0.0] - - - [36352, 2048, 1, 256] - - [53, 0.0] - - - [39680, 1281, 1, 256] - - [45, 0.0] - - - [36608, 2304, 1, 256] - - [47, 0.0] - - - [39936, 1280, 1, 256] - - [47, 0.0] - - - [39680, 5376, 1, 256] - - [47, 0.0] - - - [35584, 10240, 1, 256] - - [53, 0.0] - - - [36864, 512, 1, 256] - - [53, 0.0] - - - [39424, 2816, 1, 256] - - [47, 0.0] - - - [35840, 2816, 1, 256] - - [47, 0.0] - - - [38192, 2816, 1, 256] - - [53, 0.0] - - - [35584, 2048, 1, 256] - - [53, 0.0] - - - [37936, 2865, 1, 256] - - [53, 0.0] - - - [39936, 2865, 1, 256] - - [47, 0.0] - - - [38656, 10240, 1, 256] - - [53, 0.0] - - - [36608, 2048, 1, 256] - - [53, 0.0] - - - [35120, 2816, 1, 256] - - [47, 0.0] - - - [39424, 5888, 1, 256] - - [47, 0.0] - - - [37680, 2816, 1, 256] - - [47, 0.0] - - - [36096, 6144, 1, 256] - - [53, 0.0] - - - [38144, 1281, 1, 256] - - [45, 0.0] - - - [37632, 2048, 1, 256] - - [53, 0.0] - - - [39680, 256, 1, 256] - - [47, 0.0] - - - [37680, 3840, 1, 256] - - [53, 0.0] - - - [39168, 2816, 1, 256] - - [47, 0.0] - - - [38192, 2865, 1, 256] - - [53, 0.0] - - - [38912, 4608, 1, 256] - - [53, 0.0] - - - [37120, 2048, 1, 256] - - [53, 0.0] - - - [35376, 1536, 1, 256] - - [53, 0.0] - - - [38448, 4864, 1, 256] - - [47, 0.0] - - - [38192, 10240, 1, 256] - - [53, 0.0] - - - [37632, 2816, 1, 256] - - [53, 0.0] - - - [39424, 1024, 1, 256] - - [53, 0.0] - - - [39168, 256, 1, 256] - - [47, 0.0] - - - [39984, 6144, 1, 256] - - [53, 0.0] - - - [38144, 4608, 1, 256] - - [53, 0.0] - - - [35840, 2865, 1, 256] - - [47, 0.0] - - - [36352, 6144, 1, 256] - - [53, 0.0] - - - [36864, 768, 1, 256] - - [47, 0.0] - - - [37888, 3328, 1, 256] - - [47, 0.0] - - - [36912, 3328, 1, 256] - - [47, 0.0] - - - [37120, 3584, 1, 256] - - [53, 0.0] - - - [38912, 1281, 1, 256] - - [45, 0.0] - - - [39472, 256, 1, 256] - - [53, 0.0] - - - [39936, 1281, 1, 256] - - [45, 0.0] - - - [37376, 5120, 1, 256] - - [53, 0.0] - - - [37888, 2048, 1, 256] - - [53, 0.0] - - - [37632, 1280, 1, 256] - - [47, 0.0] - - - [35376, 2816, 1, 256] - - [47, 0.0] - - - [38656, 3329, 1, 256] - - [47, 0.0] - - - [36912, 256, 1, 256] - - [53, 0.0] - - - [39168, 768, 1, 256] - - [47, 0.0] - - - [37424, 256, 1, 256] - - [47, 0.0] - - - [38448, 2816, 1, 256] - - [53, 0.0] - - - [35840, 3840, 1, 256] - - [47, 0.0] - - - [38912, 2865, 1, 256] - - [47, 0.0] - - - [36096, 1280, 1, 256] - - [47, 0.0] - - - [35328, 1024, 1, 256] - - [53, 0.0] - - - [39680, 3328, 1, 256] - - [47, 0.0] - - - [36352, 2816, 1, 256] - - [47, 0.0] - - - [38912, 256, 1, 256] - - [53, 0.0] - - - [39424, 3328, 1, 256] - - [47, 0.0] - - - [35888, 2816, 1, 256] - - [53, 0.0] - - - [36096, 2816, 1, 256] - - [47, 0.0] - - - [38960, 10240, 1, 256] - - [53, 0.0] - - - [35840, 3584, 1, 256] - - [53, 0.0] - - - [39424, 5120, 1, 256] - - [53, 0.0] - - - [37376, 1024, 1, 256] - - [53, 0.0] - - - [37632, 4096, 1, 256] - - [53, 0.0] - - - [36400, 2865, 1, 256] - - [47, 0.0] - - - [36144, 2560, 1, 256] - - [53, 0.0] - - - [36864, 1281, 1, 256] - - [45, 0.0] - - - [39424, 5376, 1, 256] - - [47, 0.0] - - - [36400, 2816, 1, 256] - - [47, 0.0] - - - [38656, 6144, 1, 256] - - [53, 0.0] - - - [37888, 5632, 1, 256] - - [53, 0.0] - - - [36912, 2865, 1, 256] - - [47, 0.0] - - - [38656, 4352, 1, 256] - - [47, 0.0] - - - [37632, 1536, 1, 256] - - [53, 0.0] - - - [35072, 2865, 1, 256] - - [47, 0.0] - - - [35888, 2304, 1, 256] - - [47, 0.0] - - - [38912, 3329, 1, 256] - - [47, 0.0] - - - [37680, 4096, 1, 256] - - [53, 0.0] - - - [38400, 6144, 1, 256] - - [53, 0.0] - - - [37888, 3840, 1, 256] - - [47, 0.0] - - - [36608, 3328, 1, 256] - - [47, 0.0] - - - [35328, 256, 1, 256] - - [47, 0.0] - - - [36096, 3329, 1, 256] - - [47, 0.0] - - - [37888, 5888, 1, 256] - - [47, 0.0] - - - [36864, 3329, 1, 256] - - [47, 0.0] - - - [35632, 256, 1, 256] - - [47, 0.0] - - - [38656, 4864, 1, 256] - - [47, 0.0] - - - [37888, 2816, 1, 256] - - [47, 0.0] - - - [37120, 3328, 1, 256] - - [47, 0.0] - - - [35328, 1536, 1, 256] - - [53, 0.0] - - - [35328, 1280, 1, 256] - - [47, 0.0] - - - [35888, 10240, 1, 256] - - [53, 0.0] - - - [36400, 10240, 1, 256] - - [53, 0.0] - - - [35072, 10240, 1, 256] - - [53, 0.0] - - - [39680, 2816, 1, 256] - - [47, 0.0] - - - [35584, 3329, 1, 256] - - [47, 0.0] - - - [36656, 256, 1, 256] - - [47, 0.0] - - - [38144, 4096, 1, 256] - - [53, 0.0] - - - [39936, 2816, 1, 256] - - [47, 0.0] - - - [36864, 3072, 1, 256] - - [53, 0.0] - - - [37936, 2816, 1, 256] - - [47, 0.0] - - - [37632, 3584, 1, 256] - - [53, 0.0] - - - [39984, 10240, 1, 256] - - [53, 0.0] - - - [38656, 512, 1, 256] - - [53, 0.0] - - - [35328, 10240, 1, 256] - - [53, 0.0] - - - [36096, 2048, 1, 256] - - [53, 0.0] - - - [37120, 4864, 1, 256] - - [47, 0.0] - - - [35840, 10240, 1, 256] - - [53, 0.0] - - - [39680, 5632, 1, 256] - - [53, 0.0] - - - [38144, 4352, 1, 256] - - [47, 0.0] - - - [36400, 2560, 1, 256] - - [53, 0.0] - - - [35840, 3329, 1, 256] - - [47, 0.0] - - - [37424, 10240, 1, 256] - - [53, 0.0] - - - [38912, 10240, 1, 256] - - [53, 0.0] - - - [35072, 768, 1, 256] - - [47, 0.0] - - - [36096, 3840, 1, 256] - - [47, 0.0] - - - [36656, 3072, 1, 256] - - [53, 0.0] - - - [39680, 1536, 1, 256] - - [53, 0.0] - - - [36656, 2865, 1, 256] - - [53, 0.0] - - - [38912, 512, 1, 256] - - [53, 0.0] - - - [38400, 256, 1, 256] - - [47, 0.0] - - - [38704, 10240, 1, 256] - - [53, 0.0] - - - [38912, 5376, 1, 256] - - [47, 0.0] - - - [35120, 256, 1, 256] - - [47, 0.0] - - - [38656, 3328, 1, 256] - - [47, 0.0] - - - [37888, 1536, 1, 256] - - [53, 0.0] - - - [39216, 5376, 1, 256] - - [53, 0.0] - - - [37376, 3329, 1, 256] - - [47, 0.0] - - - [37680, 256, 1, 256] - - [53, 0.0] - - - [39680, 6144, 1, 256] - - [53, 0.0] - - - [38400, 2865, 1, 256] - - [47, 0.0] - - - [36608, 2865, 1, 256] - - [47, 0.0] - - - [38912, 768, 1, 256] - - [47, 0.0] - - - [35584, 1792, 1, 256] - - [47, 0.0] - - - [39424, 256, 1, 256] - - [47, 0.0] - - - [36352, 1281, 1, 256] - - [45, 0.0] - - - [38400, 2048, 1, 256] - - [53, 0.0] - - - [38144, 3329, 1, 256] - - [47, 0.0] - - - [39680, 2048, 1, 256] - - [53, 0.0] - - - [38656, 256, 1, 256] - - [47, 0.0] - - - [39728, 2816, 1, 256] - - [53, 0.0] - - - [36352, 3329, 1, 256] - - [47, 0.0] - - - [38400, 10240, 1, 256] - - [53, 0.0] - - - [39984, 6400, 1, 256] - - [53, 0.0] - - - [37888, 4352, 1, 256] - - [47, 0.0] - - - [37888, 4096, 1, 256] - - [53, 0.0] - - - [35584, 1536, 1, 256] - - [53, 0.0] - - - [36096, 256, 1, 256] - - [53, 0.0] - - - [36864, 2048, 1, 256] - - [53, 0.0] - - - [36144, 2865, 1, 256] - - [53, 0.0] - - - [35584, 3584, 1, 256] - - [53, 0.0] - - - [35072, 1024, 1, 256] - - [53, 0.0] - - - [36352, 3328, 1, 256] - - [47, 0.0] - - - [39424, 1281, 1, 256] - - [50, 0.0] - - - [39728, 10240, 1, 256] - - [53, 0.0] - - - [37632, 2865, 1, 256] - - [47, 0.0] - - - [37168, 3328, 1, 256] - - [47, 0.0] - - - [37376, 5376, 1, 256] - - [47, 0.0] - - - [35328, 2865, 1, 256] - - [47, 0.0] - - - [35584, 6144, 1, 256] - - [53, 0.0] - - - [38704, 2816, 1, 256] - - [53, 0.0] - - - [36608, 3072, 1, 256] - - [53, 0.0] - - - [39680, 1280, 1, 256] - - [47, 0.0] - - - [35328, 1281, 1, 256] - - [45, 0.0] - - - [36608, 512, 1, 256] - - [53, 0.0] - - - [39936, 1536, 1, 256] - - [53, 0.0] - - - [39728, 5888, 1, 256] - - [47, 0.0] - - - [39168, 1281, 1, 256] - - [45, 0.0] - - - [37120, 256, 1, 256] - - [53, 0.0] - - - [38960, 2865, 1, 256] - - [47, 0.0] - - - [39168, 5120, 1, 256] - - [53, 0.0] - - - [36864, 256, 1, 256] - - [53, 0.0] - - - [36912, 2816, 1, 256] - - [53, 0.0] - - - [36096, 2304, 1, 256] - - [47, 0.0] - - - [35840, 3328, 1, 256] - - [47, 0.0] - - - [38704, 2865, 1, 256] - - [47, 0.0] - - - [38144, 1792, 1, 256] - - [47, 0.0] - - - [36608, 2560, 1, 256] - - [53, 0.0] - - - [35376, 10240, 1, 256] - - [53, 0.0] - - - [35840, 2304, 1, 256] - - [47, 0.0] - - - [35840, 1280, 1, 256] - - [53, 0.0] - - - [37376, 1280, 1, 256] - - [47, 0.0] - - - [35584, 3328, 1, 256] - - [47, 0.0] - - - [35584, 2865, 1, 256] - - [47, 0.0] - - - [39936, 10240, 1, 256] - - [53, 0.0] - - - [38912, 5120, 1, 256] - - [53, 0.0] - - - [37632, 3329, 1, 256] - - [47, 0.0] - - - [37888, 1792, 1, 256] - - [47, 0.0] - - - [36608, 1281, 1, 256] - - [45, 0.0] - - - [38192, 4352, 1, 256] - - [53, 0.0] - - - [39936, 2048, 1, 256] - - [53, 0.0] - - - [35072, 1281, 1, 256] - - [45, 0.0] - - - [39472, 2816, 1, 256] - - [53, 0.0] - - - [39728, 2865, 1, 256] - - [53, 0.0] - - - [38400, 2816, 1, 256] - - [47, 0.0] - - - [38400, 4608, 1, 256] - - [53, 0.0] - - - [39216, 10240, 1, 256] - - [53, 0.0] - - - [35072, 3072, 1, 256] - - [53, 0.0] - - - [38400, 4352, 1, 256] - - [47, 0.0] - - - [39216, 2816, 1, 256] - - [53, 0.0] - - - [35840, 1792, 1, 256] - - [47, 0.0] - - - [35632, 2048, 1, 256] - - [53, 0.0] - - - [38704, 256, 1, 256] - - [47, 0.0] - - - [37888, 3329, 1, 256] - - [47, 0.0] - - - [37888, 6144, 1, 256] - - [53, 0.0] - - - [37376, 6144, 1, 256] - - [53, 0.0] - - - [37376, 256, 1, 256] - - [47, 0.0] - - - [36400, 256, 1, 256] - - [53, 0.0] - - - [37936, 4096, 1, 256] - - [53, 0.0] - - - [38144, 10240, 1, 256] - - [53, 0.0] - - - [35376, 1792, 1, 256] - - [53, 0.0] - - - [37168, 10240, 1, 256] - - [53, 0.0] - - - [39984, 2816, 1, 256] - - [53, 0.0] - - - [37168, 2816, 1, 256] - - [53, 0.0] - - - [39424, 5632, 1, 256] - - [53, 0.0] - - - [36352, 1280, 1, 256] - - [47, 0.0] - - - [39680, 10240, 1, 256] - - [53, 0.0] - - - [38144, 3328, 1, 256] - - [47, 0.0] - - - [39168, 2048, 1, 256] - - [53, 0.0] - - - [35328, 6144, 1, 256] - - [53, 0.0] - - - [35632, 2865, 1, 256] - - [53, 0.0] - - - [36656, 10240, 1, 256] - - [53, 0.0] - - - [36608, 4352, 1, 256] - - [47, 0.0] - - - [35120, 2865, 1, 256] - - [47, 0.0] - - - [36608, 6144, 1, 256] - - [53, 0.0] - - - [37888, 2865, 1, 256] - - [47, 0.0] - - - [39168, 1024, 1, 256] - - [53, 0.0] - - - [38704, 4864, 1, 256] - - [47, 0.0] - - - [39168, 2865, 1, 256] - - [47, 0.0] - - - [38960, 5120, 1, 256] - - [53, 0.0] - - - [36864, 2816, 1, 256] - - [47, 0.0] - - - [38656, 1280, 1, 256] - - [47, 0.0] - - - [35584, 1281, 1, 256] - - [44, 0.0] - - - [39216, 2865, 1, 256] - - [53, 0.0] - - - [35120, 1280, 1, 256] - - [47, 0.0] - - - [36096, 3328, 1, 256] - - [47, 0.0] - - - [38912, 6144, 1, 256] - - [53, 0.0] - - - [37376, 3840, 1, 256] - - [47, 0.0] - - - [37424, 2816, 1, 256] - - [53, 0.0] - - - [36864, 10240, 1, 256] - - [53, 0.0] - - - [35328, 3328, 1, 256] - - [47, 0.0] - - - [37632, 5632, 1, 256] - - [53, 0.0] - - - [35072, 1536, 1, 256] - - [53, 0.0] - - - [36864, 2865, 1, 256] - - [47, 0.0] - - - [36864, 4608, 1, 256] - - [53, 0.0] - - - [37888, 1280, 1, 256] - - [47, 0.0] - - - [36864, 4864, 1, 256] - - [47, 0.0] - - - [37632, 256, 1, 256] - - [53, 0.0] - - - [38912, 2816, 1, 256] - - [47, 0.0] - - - [38656, 5120, 1, 256] - - [53, 0.0] - - - [35072, 1280, 1, 256] - - [47, 0.0] - - - [38400, 3329, 1, 256] - - [47, 0.0] - - - [35840, 1281, 1, 256] - - [45, 0.0] - - - [39680, 2865, 1, 256] - - [47, 0.0] - - - [38192, 256, 1, 256] - - [47, 0.0] - - - [37632, 10240, 1, 256] - - [53, 0.0] - - - [39984, 256, 1, 256] - - [47, 0.0] - - - [37424, 2865, 1, 256] - - [47, 0.0] - - - [37888, 256, 1, 256] - - [53, 0.0] - - - [36864, 6144, 1, 256] - - [53, 0.0] - - - [38656, 1281, 1, 256] - - [50, 0.0] - - - [37936, 256, 1, 256] - - [47, 0.0] - - - [39168, 4864, 1, 256] - - [47, 0.0] - - - [35840, 256, 1, 256] - - [53, 0.0] - - - [37888, 10240, 1, 256] - - [53, 0.0] - - - [39728, 6144, 1, 256] - - [53, 0.0] - - - [39680, 5888, 1, 256] - - [47, 0.0] - - - [38144, 2816, 1, 256] - - [47, 0.0] - - - [39728, 256, 1, 256] - - [47, 0.0] - - - [37376, 2816, 1, 256] - - [47, 0.0] - - - [36352, 2865, 1, 256] - - [47, 0.0] - - - [39216, 256, 1, 256] - - [47, 0.0] - - - [37888, 1281, 1, 256] - - [45, 0.0] - - - [39472, 10240, 1, 256] - - [53, 0.0] - - - [37376, 2048, 1, 256] - - [53, 0.0] - - - [36096, 10240, 1, 256] - - [53, 0.0] - - - [35584, 1280, 1, 256] - - [47, 0.0] - - - [39168, 5632, 1, 256] - - [53, 0.0] - - - [39936, 5632, 1, 256] - - [53, 0.0] - - - [35072, 256, 1, 256] - - [47, 0.0] - - - [35376, 2865, 1, 256] - - [53, 0.0] - - - [38400, 4864, 1, 256] - - [47, 0.0] - - - [35888, 256, 1, 256] - - [53, 0.0] - - - [35072, 3328, 1, 256] - - [47, 0.0] - - - [37936, 10240, 1, 256] - - [53, 0.0] - - - [36352, 10240, 1, 256] - - [53, 0.0] - - - [38656, 2048, 1, 256] - - [53, 0.0] - - - [35632, 2816, 1, 256] - - [47, 0.0] - - - [36912, 10240, 1, 256] - - [53, 0.0] - - - [39936, 5888, 1, 256] - - [47, 0.0] - - - [38448, 2865, 1, 256] - - [47, 0.0] - - - [38144, 3840, 1, 256] - - [47, 0.0] - - - [37632, 6144, 1, 256] - - [53, 0.0] - - - [37376, 3328, 1, 256] - - [53, 0.0] - - - [36608, 2816, 1, 256] - - [53, 0.0] - - - [36912, 3072, 1, 256] - - [53, 0.0] - - - [37120, 2816, 1, 256] - - [47, 0.0] - - - [38144, 2865, 1, 256] - - [47, 0.0] - - - [38912, 2048, 1, 256] - - [53, 0.0] - - - [38192, 4608, 1, 256] - - [53, 0.0] - - - [37120, 5120, 1, 256] - - [53, 0.0] - - - [38400, 3328, 1, 256] - - [53, 0.0] - - - [35632, 10240, 1, 256] - - [53, 0.0] - - - [38912, 4864, 1, 256] - - [47, 0.0] - - - [37120, 10240, 1, 256] - - [53, 0.0] - - - [37120, 3329, 1, 256] - - [47, 0.0] - - - [35840, 6144, 1, 256] - - [53, 0.0] - - - [38400, 1281, 1, 256] - - [44, 0.0] - - - [36144, 10240, 1, 256] - - [53, 0.0] - - - [38144, 1280, 1, 256] - - [47, 0.0] - - - [39424, 10240, 1, 256] - - [53, 0.0] - - - [39424, 6144, 1, 256] - - [53, 0.0] - - - [39424, 1280, 1, 256] - - [47, 0.0] - - - [35328, 3329, 1, 256] - - [47, 0.0] - - - [39472, 5888, 1, 256] - - [53, 0.0] - - - [36352, 4096, 1, 256] - - [53, 0.0] - - - [38656, 4608, 1, 256] - - [53, 0.0] - - - [37168, 256, 1, 256] - - [53, 0.0] - - - [38144, 2048, 1, 256] - - [53, 0.0] - - - [35840, 1536, 1, 256] - - [53, 0.0] - - - [37120, 1280, 1, 256] - - [47, 0.0] - - - [37424, 3840, 1, 256] - - [53, 0.0] - - - [37424, 3584, 1, 256] - - [53, 0.0] - - - [36864, 2560, 1, 256] - - [53, 0.0] - - - [39936, 6400, 1, 256] - - [47, 0.0] - - - [36096, 2560, 1, 256] - - [53, 0.0] - - - [37120, 768, 1, 256] - - [47, 0.0] - - - [35328, 2048, 1, 256] - - [53, 0.0] - - - [36608, 4608, 1, 256] - - [53, 0.0] - - - [38400, 4096, 1, 256] - - [53, 0.0] - - - [35328, 2816, 1, 256] - - [47, 0.0] - - - [36144, 256, 1, 256] - - [47, 0.0] - - - [36608, 256, 1, 256] - - [53, 0.0] - - - [39168, 3329, 1, 256] - - [47, 0.0] - - - [38448, 4608, 1, 256] - - [53, 0.0] - - - [37632, 3328, 1, 256] - - [47, 0.0] - - - [37680, 2865, 1, 256] - - [47, 0.0] - - - [35120, 10240, 1, 256] - - [53, 0.0] - - - [37120, 6144, 1, 256] - - [53, 0.0] - - - [36656, 2816, 1, 256] - - [53, 0.0] - - - [39936, 3329, 1, 256] - - [47, 0.0] - - - [35328, 1792, 1, 256] - - [47, 0.0] - - - [35120, 1536, 1, 256] - - [53, 0.0] - - - [39472, 2865, 1, 256] - - [53, 0.0] - - - [37936, 4352, 1, 256] - - [53, 0.0] - - - [35888, 2048, 1, 256] - - [53, 0.0] - - - [37888, 3584, 1, 256] - - [53, 0.0] - - - [37376, 2865, 1, 256] - - [47, 0.0] - - - [36864, 1280, 1, 256] - - [47, 0.0] - - - [39472, 5632, 1, 256] - - [53, 0.0] - - - [37120, 1024, 1, 256] - - [53, 0.0] - - - [37120, 2865, 1, 256] - - [47, 0.0] - - - [38400, 1280, 1, 256] - - [47, 0.0] - - - [35584, 2816, 1, 256] - - [47, 0.0] - - - [37376, 1281, 1, 256] - - [50, 0.0] - - - [36352, 2560, 1, 256] - - [53, 0.0] - - - [36144, 2304, 1, 256] - - [47, 0.0] - - - [37632, 3840, 1, 256] - - [47, 0.0] - - - [38960, 2816, 1, 256] - - [53, 0.0] - - - [37376, 3072, 1, 256] - - [53, 0.0] - - - [35072, 2816, 1, 256] - - [47, 0.0] - - - [38912, 3328, 1, 256] - - [47, 0.0] - - - [38960, 256, 1, 256] - - [53, 0.0] - - - [35376, 256, 1, 256] - - [47, 0.0] - - - [39168, 1280, 1, 256] - - [53, 0.0] - - - [44032, 5888, 1, 256] - - [47, 0.0] - - - [40192, 2865, 1, 256] - - [47, 0.0] - - - [43312, 256, 1, 256] - - [47, 0.0] - - - [43520, 1280, 1, 256] - - [47, 0.0] - - - [41216, 2816, 1, 256] - - [47, 0.0] - - - [41520, 7936, 1, 256] - - [47, 0.0] - - - [43008, 2048, 1, 256] - - [53, 0.0] - - - [42496, 2048, 1, 256] - - [53, 0.0] - - - [40704, 3328, 1, 256] - - [47, 0.0] - - - [41776, 7936, 1, 256] - - [53, 0.0] - - - [40192, 1792, 1, 256] - - [47, 0.0] - - - [43520, 6144, 1, 256] - - [53, 0.0] - - - [42032, 2865, 1, 256] - - [53, 0.0] - - - [41472, 3329, 1, 256] - - [47, 0.0] - - - [41008, 7424, 1, 256] - - [47, 0.0] - - - [40448, 2865, 1, 256] - - [47, 0.0] - - - [41264, 2865, 1, 256] - - [53, 0.0] - - - [43312, 9728, 1, 256] - - [53, 0.0] - - - [40704, 2816, 1, 256] - - [47, 0.0] - - - [42544, 8704, 1, 256] - - [53, 0.0] - - - [40960, 7168, 1, 256] - - [53, 0.0] - - - [41216, 3329, 1, 256] - - [47, 0.0] - - - [41984, 6144, 1, 256] - - [53, 0.0] - - - [42240, 10240, 1, 256] - - [53, 0.0] - - - [42752, 2865, 1, 256] - - [47, 0.0] - - - [41216, 1280, 1, 256] - - [47, 0.0] - - - [40704, 7168, 1, 256] - - [53, 0.0] - - - [41216, 10240, 1, 256] - - [53, 0.0] - - - [40960, 256, 1, 256] - - [47, 0.0] - - - [40704, 2560, 1, 256] - - [53, 0.0] - - - [42752, 3329, 1, 256] - - [47, 0.0] - - - [43264, 3329, 1, 256] - - [47, 0.0] - - - [40192, 6144, 1, 256] - - [53, 0.0] - - - [43008, 10240, 1, 256] - - [53, 0.0] - - - [43520, 1281, 1, 256] - - [45, 0.0] - - - [42496, 8960, 1, 256] - - [47, 0.0] - - - [43312, 10240, 1, 256] - - [53, 0.0] - - - [44032, 6144, 1, 256] - - [53, 0.0] - - - [40192, 256, 1, 256] - - [47, 0.0] - - - [41984, 1536, 1, 256] - - [53, 0.0] - - - [41216, 768, 1, 256] - - [47, 0.0] - - - [40752, 256, 1, 256] - - [53, 0.0] - - - [44288, 1280, 1, 256] - - [47, 0.0] - - - [43520, 9216, 1, 256] - - [53, 0.0] - - - [42032, 8192, 1, 256] - - [53, 0.0] - - - [41728, 3584, 1, 256] - - [53, 0.0] - - - [40448, 1280, 1, 256] - - [47, 0.0] - - - [41216, 7168, 1, 256] - - [53, 0.0] - - - [42496, 1280, 1, 256] - - [47, 0.0] - - - [40448, 6656, 1, 256] - - [53, 0.0] - - - [40240, 256, 1, 256] - - [47, 0.0] - - - [41264, 2816, 1, 256] - - [53, 0.0] - - - [43264, 3328, 1, 256] - - [47, 0.0] - - - [43008, 9216, 1, 256] - - [53, 0.0] - - - [42240, 1281, 1, 256] - - [45, 0.0] - - - [42288, 2865, 1, 256] - - [53, 0.0] - - - [43008, 3328, 1, 256] - - [47, 0.0] - - - [40496, 256, 1, 256] - - [47, 0.0] - - - [43264, 8960, 1, 256] - - [47, 0.0] - - - [43056, 9472, 1, 256] - - [47, 0.0] - - - [40448, 3328, 1, 256] - - [47, 0.0] - - - [41776, 8192, 1, 256] - - [53, 0.0] - - - [40704, 6400, 1, 256] - - [47, 0.0] - - - [41984, 7680, 1, 256] - - [53, 0.0] - - - [43312, 9472, 1, 256] - - [47, 0.0] - - - [40192, 1280, 1, 256] - - [47, 0.0] - - - [43776, 5632, 1, 256] - - [53, 0.0] - - - [41984, 2865, 1, 256] - - [47, 0.0] - - - [40448, 2816, 1, 256] - - [47, 0.0] - - - [42240, 3328, 1, 256] - - [47, 0.0] - - - [42752, 2048, 1, 256] - - [53, 0.0] - - - [42240, 256, 1, 256] - - [47, 0.0] - - - [43008, 3329, 1, 256] - - [47, 0.0] - - - [44032, 5632, 1, 256] - - [53, 0.0] - - - [40192, 2048, 1, 256] - - [53, 0.0] - - - [41216, 256, 1, 256] - - [47, 0.0] - - - [44288, 9984, 1, 256] - - [47, 0.0] - - - [43008, 1280, 1, 256] - - [47, 0.0] - - - [41984, 2816, 1, 256] - - [47, 0.0] - - - [42752, 6144, 1, 256] - - [53, 0.0] - - - [43776, 3329, 1, 256] - - [47, 0.0] - - - [43008, 2865, 1, 256] - - [47, 0.0] - - - [43776, 9728, 1, 256] - - [53, 0.0] - - - [42240, 7936, 1, 256] - - [47, 0.0] - - - [41472, 7424, 1, 256] - - [47, 0.0] - - - [43776, 5376, 1, 256] - - [47, 0.0] - - - [43008, 6144, 1, 256] - - [53, 0.0] - - - [41216, 3072, 1, 256] - - [53, 0.0] - - - [42496, 8192, 1, 256] - - [53, 0.0] - - - [40704, 6144, 1, 256] - - [53, 0.0] - - - [44032, 3329, 1, 256] - - [47, 0.0] - - - [43520, 2048, 1, 256] - - [53, 0.0] - - - [43264, 2048, 1, 256] - - [53, 0.0] - - - [40448, 1281, 1, 256] - - [44, 0.0] - - - [40496, 2865, 1, 256] - - [53, 0.0] - - - [40448, 6144, 1, 256] - - [53, 0.0] - - - [41008, 10240, 1, 256] - - [53, 0.0] - - - [43056, 2865, 1, 256] - - [53, 0.0] - - - [43264, 1280, 1, 256] - - [47, 0.0] - - - [40192, 10240, 1, 256] - - [53, 0.0] - - - [41216, 7680, 1, 256] - - [53, 0.0] - - - [41008, 7168, 1, 256] - - [53, 0.0] - - - [44288, 2048, 1, 256] - - [53, 0.0] - - - [41472, 6144, 1, 256] - - [53, 0.0] - - - [43264, 2865, 1, 256] - - [47, 0.0] - - - [40448, 6912, 1, 256] - - [53, 0.0] - - - [41216, 6912, 1, 256] - - [47, 0.0] - - - [41984, 1792, 1, 256] - - [47, 0.0] - - - [40192, 1281, 1, 256] - - [44, 0.0] - - - [40960, 3329, 1, 256] - - [47, 0.0] - - - [41520, 10240, 1, 256] - - [53, 0.0] - - - [44032, 10240, 1, 256] - - [53, 0.0] - - - [43264, 2816, 1, 256] - - [47, 0.0] - - - [43008, 4608, 1, 256] - - [53, 0.0] - - - [43776, 1281, 1, 256] - - [45, 0.0] - - - [40240, 6656, 1, 256] - - [53, 0.0] - - - [43264, 9216, 1, 256] - - [53, 0.0] - - - [40704, 3329, 1, 256] - - [47, 0.0] - - - [42752, 3328, 1, 256] - - [47, 0.0] - - - [41984, 2048, 1, 256] - - [53, 0.0] - - - [44288, 3329, 1, 256] - - [47, 0.0] - - - [40192, 3328, 1, 256] - - [47, 0.0] - - - [40960, 10240, 1, 256] - - [53, 0.0] - - - [42496, 256, 1, 256] - - [53, 0.0] - - - [40496, 10240, 1, 256] - - [53, 0.0] - - - [40496, 6912, 1, 256] - - [53, 0.0] - - - [43776, 6144, 1, 256] - - [53, 0.0] - - - [40960, 1280, 1, 256] - - [47, 0.0] - - - [42288, 8704, 1, 256] - - [53, 0.0] - - - [42496, 3328, 1, 256] - - [47, 0.0] - - - [41216, 2865, 1, 256] - - [47, 0.0] - - - [42496, 3329, 1, 256] - - [47, 0.0] - - - [41984, 7936, 1, 256] - - [47, 0.0] - - - [41472, 1281, 1, 256] - - [50, 0.0] - - - [41776, 256, 1, 256] - - [47, 0.0] - - - [42752, 8960, 1, 256] - - [47, 0.0] - - - [41472, 7168, 1, 256] - - [53, 0.0] - - - [40240, 10240, 1, 256] - - [53, 0.0] - - - [41728, 1280, 1, 256] - - [47, 0.0] - - - [40752, 2865, 1, 256] - - [53, 0.0] - - - [40960, 2048, 1, 256] - - [53, 0.0] - - - [41472, 7680, 1, 256] - - [53, 0.0] - - - [41472, 10240, 1, 256] - - [53, 0.0] - - - [41264, 7680, 1, 256] - - [53, 0.0] - - - [42800, 8960, 1, 256] - - [47, 0.0] - - - [41728, 10240, 1, 256] - - [53, 0.0] - - - [44032, 3328, 1, 256] - - [53, 0.0] - - - [40704, 6912, 1, 256] - - [47, 0.0] - - - [41472, 2048, 1, 256] - - [53, 0.0] - - - [40960, 6144, 1, 256] - - [53, 0.0] - - - [43776, 3328, 1, 256] - - [47, 0.0] - - - [42496, 2865, 1, 256] - - [47, 0.0] - - - [40960, 3328, 1, 256] - - [47, 0.0] - - - [41728, 7936, 1, 256] - - [47, 0.0] - - - [41984, 3329, 1, 256] - - [47, 0.0] - - - [43008, 256, 1, 256] - - [53, 0.0] - - - [42240, 1280, 1, 256] - - [47, 0.0] - - - [43776, 10240, 1, 256] - - [53, 0.0] - - - [42752, 8448, 1, 256] - - [47, 0.0] - - - [42496, 1281, 1, 256] - - [45, 0.0] - - - [44032, 1536, 1, 256] - - [53, 0.0] - - - [40960, 2816, 1, 256] - - [47, 0.0] - - - [44288, 1792, 1, 256] - - [47, 0.0] - - - [43264, 1281, 1, 256] - - [45, 0.0] - - - [43008, 8704, 1, 256] - - [53, 0.0] - - - [41728, 1536, 1, 256] - - [53, 0.0] - - - [41728, 2048, 1, 256] - - [53, 0.0] - - - [43520, 9728, 1, 256] - - [53, 0.0] - - - [42032, 256, 1, 256] - - [47, 0.0] - - - [43776, 256, 1, 256] - - [53, 0.0] - - - [43008, 9472, 1, 256] - - [47, 0.0] - - - [44032, 1792, 1, 256] - - [47, 0.0] - - - [40704, 2865, 1, 256] - - [47, 0.0] - - - [42240, 1792, 1, 256] - - [47, 0.0] - - - [40704, 2304, 1, 256] - - [47, 0.0] - - - [42800, 9216, 1, 256] - - [53, 0.0] - - - [42240, 8704, 1, 256] - - [53, 0.0] - - - [42496, 6144, 1, 256] - - [53, 0.0] - - - [43568, 9728, 1, 256] - - [53, 0.0] - - - [40704, 2048, 1, 256] - - [53, 0.0] - - - [41472, 7936, 1, 256] - - [47, 0.0] - - - [42752, 2816, 1, 256] - - [47, 0.0] - - - [41008, 2865, 1, 256] - - [53, 0.0] - - - [40960, 6912, 1, 256] - - [47, 0.0] - - - [44032, 256, 1, 256] - - [47, 0.0] - - - [42496, 4352, 1, 256] - - [47, 0.0] - - - [42032, 8448, 1, 256] - - [47, 0.0] - - - [42752, 4608, 1, 256] - - [53, 0.0] - - - [44032, 1280, 1, 256] - - [53, 0.0] - - - [44288, 6144, 1, 256] - - [53, 0.0] - - - [42800, 2865, 1, 256] - - [47, 0.0] - - - [41008, 2816, 1, 256] - - [53, 0.0] - - - [41984, 8192, 1, 256] - - [53, 0.0] - - - [43264, 256, 1, 256] - - [53, 0.0] - - - [41728, 2865, 1, 256] - - [47, 0.0] - - - [43520, 5120, 1, 256] - - [53, 0.0] - - - [41984, 3584, 1, 256] - - [53, 0.0] - - - [41216, 3328, 1, 256] - - [47, 0.0] - - - [43520, 9472, 1, 256] - - [47, 0.0] - - - [43264, 9728, 1, 256] - - [53, 0.0] - - - [41728, 1281, 1, 256] - - [45, 0.0] - - - [40704, 1281, 1, 256] - - [45, 0.0] - - - [42288, 256, 1, 256] - - [53, 0.0] - - - [40960, 512, 1, 256] - - [53, 0.0] - - - [42752, 4352, 1, 256] - - [47, 0.0] - - - [40752, 10240, 1, 256] - - [53, 0.0] - - - [41728, 3328, 1, 256] - - [47, 0.0] - - - [43568, 2816, 1, 256] - - [47, 0.0] - - - [43008, 512, 1, 256] - - [53, 0.0] - - - [41216, 2048, 1, 256] - - [53, 0.0] - - - [42800, 256, 1, 256] - - [47, 0.0] - - - [43312, 2816, 1, 256] - - [53, 0.0] - - - [40192, 6400, 1, 256] - - [53, 0.0] - - - [41264, 7424, 1, 256] - - [53, 0.0] - - - [42544, 8960, 1, 256] - - [53, 0.0] - - - [41472, 256, 1, 256] - - [47, 0.0] - - - [42288, 10240, 1, 256] - - [53, 0.0] - - - [43520, 1024, 1, 256] - - [53, 0.0] - - - [42288, 8448, 1, 256] - - [53, 0.0] - - - [43776, 9472, 1, 256] - - [47, 0.0] - - - [43008, 1281, 1, 256] - - [50, 0.0] - - - [43008, 8960, 1, 256] - - [47, 0.0] - - - [41728, 256, 1, 256] - - [47, 0.0] - - - [41520, 7680, 1, 256] - - [53, 0.0] - - - [42240, 3329, 1, 256] - - [47, 0.0] - - - [41472, 2816, 1, 256] - - [47, 0.0] - - - [41216, 6144, 1, 256] - - [53, 0.0] - - - [40752, 2816, 1, 256] - - [47, 0.0] - - - [42496, 8704, 1, 256] - - [53, 0.0] - - - [40448, 6400, 1, 256] - - [47, 0.0] - - - [44032, 1281, 1, 256] - - [45, 0.0] - - - [41472, 1024, 1, 256] - - [53, 0.0] - - - [41216, 7424, 1, 256] - - [47, 0.0] - - - [43312, 2865, 1, 256] - - [53, 0.0] - - - [40960, 768, 1, 256] - - [47, 0.0] - - - [40240, 2865, 1, 256] - - [53, 0.0] - - - [43264, 768, 1, 256] - - [47, 0.0] - - - [40192, 3329, 1, 256] - - [47, 0.0] - - - [42800, 10240, 1, 256] - - [53, 0.0] - - - [42752, 512, 1, 256] - - [53, 0.0] - - - [40752, 6912, 1, 256] - - [53, 0.0] - - - [42240, 8192, 1, 256] - - [53, 0.0] - - - [42288, 2816, 1, 256] - - [53, 0.0] - - - [40960, 2865, 1, 256] - - [47, 0.0] - - - [42800, 2816, 1, 256] - - [47, 0.0] - - - [42496, 2816, 1, 256] - - [47, 0.0] - - - [41728, 7680, 1, 256] - - [53, 0.0] - - - [42240, 8448, 1, 256] - - [47, 0.0] - - - [41984, 1281, 1, 256] - - [45, 0.0] - - - [41984, 3328, 1, 256] - - [47, 0.0] - - - [40240, 6400, 1, 256] - - [53, 0.0] - - - [44288, 256, 1, 256] - - [53, 0.0] - - - [42496, 4096, 1, 256] - - [53, 0.0] - - - [43520, 3329, 1, 256] - - [47, 0.0] - - - [44288, 5888, 1, 256] - - [47, 0.0] - - - [42752, 1281, 1, 256] - - [45, 0.0] - - - [43776, 9984, 1, 256] - - [47, 0.0] - - - [41008, 256, 1, 256] - - [47, 0.0] - - - [40960, 1281, 1, 256] - - [45, 0.0] - - - [40704, 6656, 1, 256] - - [53, 0.0] - - - [40192, 2816, 1, 256] - - [47, 0.0] - - - [43264, 10240, 1, 256] - - [53, 0.0] - - - [44032, 9984, 1, 256] - - [47, 0.0] - - - [43520, 2865, 1, 256] - - [47, 0.0] - - - [42240, 3840, 1, 256] - - [47, 0.0] - - - [43056, 9216, 1, 256] - - [53, 0.0] - - - [43520, 10240, 1, 256] - - [53, 0.0] - - - [42544, 10240, 1, 256] - - [53, 0.0] - - - [40448, 2304, 1, 256] - - [47, 0.0] - - - [40704, 1280, 1, 256] - - [47, 0.0] - - - [43520, 2816, 1, 256] - - [47, 0.0] - - - [43520, 5376, 1, 256] - - [47, 0.0] - - - [41984, 256, 1, 256] - - [47, 0.0] - - - [43776, 1280, 1, 256] - - [47, 0.0] - - - [43568, 2865, 1, 256] - - [53, 0.0] - - - [41520, 256, 1, 256] - - [47, 0.0] - - - [41472, 3328, 1, 256] - - [47, 0.0] - - - [40192, 6656, 1, 256] - - [53, 0.0] - - - [40448, 2048, 1, 256] - - [53, 0.0] - - - [41520, 2816, 1, 256] - - [53, 0.0] - - - [43520, 9984, 1, 256] - - [47, 0.0] - - - [42544, 2865, 1, 256] - - [47, 0.0] - - - [42240, 2048, 1, 256] - - [53, 0.0] - - - [41472, 1280, 1, 256] - - [53, 0.0] - - - [40192, 5888, 1, 256] - - [47, 0.0] - - - [42240, 2865, 1, 256] - - [47, 0.0] - - - [41984, 10240, 1, 256] - - [53, 0.0] - - - [41264, 10240, 1, 256] - - [53, 0.0] - - - [42752, 10240, 1, 256] - - [53, 0.0] - - - [41216, 1024, 1, 256] - - [53, 0.0] - - - [41776, 10240, 1, 256] - - [53, 0.0] - - - [40960, 7424, 1, 256] - - [47, 0.0] - - - [40960, 2560, 1, 256] - - [53, 0.0] - - - [41216, 1281, 1, 256] - - [45, 0.0] - - - [41984, 1280, 1, 256] - - [47, 0.0] - - - [40448, 3329, 1, 256] - - [47, 0.0] - - - [41776, 2816, 1, 256] - - [47, 0.0] - - - [40704, 256, 1, 256] - - [47, 0.0] - - - [43264, 4864, 1, 256] - - [47, 0.0] - - - [42240, 6144, 1, 256] - - [53, 0.0] - - - [43520, 3328, 1, 256] - - [47, 0.0] - - - [42752, 256, 1, 256] - - [47, 0.0] - - - [40752, 7168, 1, 256] - - [53, 0.0] - - - [43776, 1536, 1, 256] - - [53, 0.0] - - - [42032, 10240, 1, 256] - - [53, 0.0] - - - [43008, 4864, 1, 256] - - [47, 0.0] - - - [40704, 10240, 1, 256] - - [53, 0.0] - - - [44288, 1281, 1, 256] - - [45, 0.0] - - - [41520, 2865, 1, 256] - - [53, 0.0] - - - [41264, 256, 1, 256] - - [53, 0.0] - - - [40496, 6656, 1, 256] - - [53, 0.0] - - - [42240, 4096, 1, 256] - - [53, 0.0] - - - [43568, 9984, 1, 256] - - [47, 0.0] - - - [43264, 9472, 1, 256] - - [47, 0.0] - - - [43008, 768, 1, 256] - - [47, 0.0] - - - [43776, 2816, 1, 256] - - [47, 0.0] - - - [43008, 2816, 1, 256] - - [47, 0.0] - - - [41984, 8448, 1, 256] - - [47, 0.0] - - - [43520, 256, 1, 256] - - [53, 0.0] - - - [43776, 2865, 1, 256] - - [47, 0.0] - - - [41984, 3840, 1, 256] - - [47, 0.0] - - - [42544, 256, 1, 256] - - [47, 0.0] - - - [43056, 2816, 1, 256] - - [53, 0.0] - - - [41472, 3072, 1, 256] - - [53, 0.0] - - - [41776, 2865, 1, 256] - - [47, 0.0] - - - [43056, 256, 1, 256] - - [47, 0.0] - - - [41728, 6144, 1, 256] - - [53, 0.0] - - - [42496, 8448, 1, 256] - - [47, 0.0] - - - [43568, 256, 1, 256] - - [53, 0.0] - - - [42752, 8704, 1, 256] - - [53, 0.0] - - - [42544, 2816, 1, 256] - - [47, 0.0] - - - [40448, 10240, 1, 256] - - [53, 0.0] - - - [41728, 2816, 1, 256] - - [47, 0.0] - - - [43568, 10240, 1, 256] - - [53, 0.0] - - - [44032, 2048, 1, 256] - - [53, 0.0] - - - [41472, 2865, 1, 256] - - [47, 0.0] - - - [40448, 256, 1, 256] - - [47, 0.0] - - - [41728, 3329, 1, 256] - - [47, 0.0] - - - [43264, 6144, 1, 256] - - [53, 0.0] - - - [40960, 6656, 1, 256] - - [53, 0.0] - - - [42752, 9216, 1, 256] - - [53, 0.0] - - - [40496, 2816, 1, 256] - - [53, 0.0] - - - [40704, 512, 1, 256] - - [53, 0.0] - - - [43056, 10240, 1, 256] - - [53, 0.0] - - - [44032, 9728, 1, 256] - - [53, 0.0] - - - [41728, 8192, 1, 256] - - [53, 0.0] - - - [43264, 1024, 1, 256] - - [53, 0.0] - - - [43776, 2048, 1, 256] - - [53, 0.0] - - - [40240, 2816, 1, 256] - - [53, 0.0] - - - [42752, 1280, 1, 256] - - [47, 0.0] - - - [44288, 10240, 1, 256] - - [53, 0.0] - - - [42240, 2816, 1, 256] - - [47, 0.0] - - - [41728, 7424, 1, 256] - - [47, 0.0] - - - [44288, 3328, 1, 256] - - [47, 0.0] - - - [43264, 5120, 1, 256] - - [53, 0.0] - - - [42032, 2816, 1, 256] - - [53, 0.0] - - - [11776, 6144, 1, 256] - - [53, 0.0] - - - [11264, 1792, 1, 256] - - [47, 0.0] - - - [4352, 2865, 1, 256] - - [53, 0.0] - - - [14640, 1536, 1, 256] - - [53, 0.0] - - - [4096, 2865, 1, 256] - - [47, 0.0] - - - [5168, 256, 1, 256] - - [62, 0.0] - - - [19968, 3328, 1, 256] - - [47, 0.0] - - - [12544, 3328, 1, 256] - - [47, 0.0] - - - [15408, 2816, 1, 256] - - [53, 0.0] - - - [16640, 3329, 1, 256] - - [47, 0.0] - - - [768, 768, 1, 256] - - [34, 0.0] - - - [3840, 512, 1, 256] - - [67, 0.0] - - - [7424, 5888, 1, 256] - - [53, 0.0] - - - [48, 49, 1, 256] - - [145, 0.0] - - - [16384, 768, 1, 256] - - [47, 0.0] - - - [15664, 2865, 1, 256] - - [53, 0.0] - - - [12544, 2048, 1, 256] - - [53, 0.0] - - - [7680, 4096, 1, 256] - - [53, 0.0] - - - [8240, 5376, 1, 256] - - [47, 0.0] - - - [11520, 256, 1, 256] - - [53, 0.0] - - - [12800, 256, 1, 256] - - [47, 0.0] - - - [10544, 2865, 1, 256] - - [53, 0.0] - - - [10032, 6912, 1, 256] - - [53, 0.0] - - - [3072, 3072, 1, 256] - - [53, 0.0] - - - [5888, 2865, 1, 256] - - [47, 0.0] - - - [8448, 3328, 1, 256] - - [47, 0.0] - - - [17920, 4096, 1, 256] - - [53, 0.0] - - - [19200, 5376, 1, 256] - - [47, 0.0] - - - [16432, 2865, 1, 256] - - [47, 0.0] - - - [12032, 3329, 1, 256] - - [47, 0.0] - - - [11776, 8704, 1, 256] - - [53, 0.0] - - - [11520, 1281, 1, 256] - - [44, 0.0] - - - [19760, 10240, 1, 256] - - [53, 0.0] - - - [15360, 1281, 1, 256] - - [50, 0.0] - - - [19712, 2865, 1, 256] - - [47, 0.0] - - - [9216, 6400, 1, 256] - - [53, 0.0] - - - [18944, 3329, 1, 256] - - [47, 0.0] - - - [5632, 2816, 1, 256] - - [47, 0.0] - - - [13872, 256, 1, 256] - - [53, 0.0] - - - [9984, 1280, 1, 256] - - [47, 0.0] - - - [19248, 10240, 1, 256] - - [53, 0.0] - - - [14128, 256, 1, 256] - - [68, 0.0] - - - [12080, 9216, 1, 256] - - [53, 0.0] - - - [18224, 5120, 1, 256] - - [53, 0.0] - - - [2352, 256, 1, 256] - - [34, 0.0] - - - [17712, 4608, 1, 256] - - [53, 0.0] - - - [8192, 5376, 1, 256] - - [47, 0.0] - - - [8752, 5888, 1, 256] - - [53, 0.0] - - - [11264, 3584, 1, 256] - - [53, 0.0] - - - [816, 256, 1, 256] - - [143, 0.0] - - - [5376, 3328, 1, 256] - - [47, 0.0] - - - [6144, 2560, 1, 256] - - [53, 0.0] - - - [9264, 256, 1, 256] - - [68, 0.0] - - - [8960, 5376, 1, 256] - - [47, 0.0] - - - [2608, 2353, 1, 256] - - [53, 0.0] - - - [2096, 256, 1, 256] - - [51, 0.0] - - - [9984, 7168, 1, 256] - - [53, 0.0] - - - [7424, 3329, 1, 256] - - [47, 0.0] - - - [2352, 2304, 1, 256] - - [53, 0.0] - - - [9984, 512, 1, 256] - - [53, 0.0] - - - [6656, 3840, 1, 256] - - [47, 0.0] - - - [17408, 3329, 1, 256] - - [47, 0.0] - - - [8496, 5376, 1, 256] - - [53, 0.0] - - - [11264, 3840, 1, 256] - - [47, 0.0] - - - [13312, 2865, 1, 256] - - [47, 0.0] - - - [3584, 768, 1, 256] - - [47, 0.0] - - - [11520, 6144, 1, 256] - - [53, 0.0] - - - [15360, 2048, 1, 256] - - [53, 0.0] - - - [7936, 3328, 1, 256] - - [47, 0.0] - - - [6144, 1281, 1, 256] - - [50, 0.0] - - - [19968, 6656, 1, 256] - - [53, 0.0] - - - [15152, 256, 1, 256] - - [144, 0.0] - - - [18432, 4608, 1, 256] - - [53, 0.0] - - - [1072, 256, 1, 256] - - [55, 0.0] - - - [6400, 4864, 1, 256] - - [47, 0.0] - - - [19712, 1281, 1, 256] - - [45, 0.0] - - - [1792, 1280, 1, 256] - - [58, 0.0] - - - [8192, 2865, 1, 256] - - [53, 0.0] - - - [3376, 256, 1, 256] - - [40, 0.0] - - - [10544, 2816, 1, 256] - - [47, 0.0] - - - [14336, 2816, 1, 256] - - [47, 0.0] - - - [16384, 1280, 1, 256] - - [47, 0.0] - - - [1280, 256, 1, 256] - - [55, 0.0] - - - [12544, 8960, 1, 256] - - [47, 0.0] - - - [13824, 1281, 1, 256] - - [50, 0.0] - - - [3072, 256, 1, 256] - - [42, 0.0] - - - [19760, 2816, 1, 256] - - [53, 0.0] - - - [8448, 5376, 1, 256] - - [47, 0.0] - - - [11824, 2865, 1, 256] - - [53, 0.0] - - - [6656, 3584, 1, 256] - - [53, 0.0] - - - [12288, 8704, 1, 256] - - [53, 0.0] - - - [11312, 256, 1, 256] - - [47, 0.0] - - - [15920, 2816, 1, 256] - - [53, 0.0] - - - [12032, 8448, 1, 256] - - [47, 0.0] - - - [14080, 2048, 1, 256] - - [53, 0.0] - - - [6400, 5120, 1, 256] - - [53, 0.0] - - - [7216, 2865, 1, 256] - - [47, 0.0] - - - [4400, 1280, 1, 256] - - [47, 0.0] - - - [5376, 3840, 1, 256] - - [47, 0.0] - - - [7168, 2816, 1, 256] - - [47, 0.0] - - - [19200, 5632, 1, 256] - - [53, 0.0] - - - [4144, 1024, 1, 256] - - [53, 0.0] - - - [12800, 3329, 1, 256] - - [47, 0.0] - - - [6400, 2865, 1, 256] - - [47, 0.0] - - - [12800, 5376, 1, 256] - - [47, 0.0] - - - [7168, 1536, 1, 256] - - [53, 0.0] - - - [19968, 1281, 1, 256] - - [45, 0.0] - - - [17664, 1281, 1, 256] - - [44, 0.0] - - - [11264, 3329, 1, 256] - - [47, 0.0] - - - [17712, 256, 1, 256] - - [53, 0.0] - - - [6656, 5376, 1, 256] - - [53, 0.0] - - - [13056, 5376, 1, 256] - - [47, 0.0] - - - [11568, 2865, 1, 256] - - [47, 0.0] - - - [3328, 1281, 1, 256] - - [50, 0.0] - - - [19968, 2048, 1, 256] - - [53, 0.0] - - - [2304, 2048, 1, 256] - - [53, 0.0] - - - [7728, 256, 1, 256] - - [47, 0.0] - - - [7424, 4352, 1, 256] - - [53, 0.0] - - - [5376, 2048, 1, 256] - - [53, 0.0] - - - [19456, 2816, 1, 256] - - [47, 0.0] - - - [7216, 2816, 1, 256] - - [53, 0.0] - - - [18688, 5376, 1, 256] - - [47, 0.0] - - - [4656, 1792, 1, 256] - - [53, 0.0] - - - [10240, 768, 1, 256] - - [47, 0.0] - - - [19456, 1280, 1, 256] - - [47, 0.0] - - - [18432, 3329, 1, 256] - - [47, 0.0] - - - [17920, 6144, 1, 256] - - [53, 0.0] - - - [1536, 1280, 1, 256] - - [58, 0.0] - - - [19456, 6400, 1, 256] - - [47, 0.0] - - - [15360, 6144, 1, 256] - - [53, 0.0] - - - [15664, 10240, 1, 256] - - [53, 0.0] - - - [3840, 256, 1, 256] - - [39, 0.0] - - - [4864, 3328, 1, 256] - - [47, 0.0] - - - [18224, 2865, 1, 256] - - [47, 0.0] - - - [13056, 9984, 1, 256] - - [47, 0.0] - - - [12288, 256, 1, 256] - - [69, 0.0] - - - [7168, 3840, 1, 256] - - [47, 0.0] - - - [17712, 4352, 1, 256] - - [53, 0.0] - - - [14592, 10240, 1, 256] - - [53, 0.0] - - - [8704, 5376, 1, 256] - - [53, 0.0] - - - [16128, 2816, 1, 256] - - [47, 0.0] - - - [4352, 3329, 1, 256] - - [47, 0.0] - - - [13568, 512, 1, 256] - - [53, 0.0] - - - [15872, 2865, 1, 256] - - [47, 0.0] - - - [12032, 1281, 1, 256] - - [45, 0.0] - - - [11520, 2048, 1, 256] - - [53, 0.0] - - - [12032, 2048, 1, 256] - - [53, 0.0] - - - [5632, 1281, 1, 256] - - [50, 0.0] - - - [13312, 9984, 1, 256] - - [47, 0.0] - - - [4912, 2865, 1, 256] - - [47, 0.0] - - - [15408, 2304, 1, 256] - - [47, 0.0] - - - [7472, 2816, 1, 256] - - [53, 0.0] - - - [18688, 10240, 1, 256] - - [53, 0.0] - - - [10752, 7936, 1, 256] - - [47, 0.0] - - - [2048, 1793, 1, 256] - - [69, 0.0] - - - [11776, 1280, 1, 256] - - [47, 0.0] - - - [10032, 256, 1, 256] - - [144, 0.0] - - - [17408, 1536, 1, 256] - - [53, 0.0] - - - [14080, 2865, 1, 256] - - [47, 0.0] - - - [16688, 3328, 1, 256] - - [47, 0.0] - - - [18944, 1024, 1, 256] - - [53, 0.0] - - - [2352, 2097, 1, 256] - - [53, 0.0] - - - [11008, 2048, 1, 256] - - [53, 0.0] - - - [10240, 6912, 1, 256] - - [47, 0.0] - - - [8448, 768, 1, 256] - - [47, 0.0] - - - [16640, 1024, 1, 256] - - [53, 0.0] - - - [11824, 8960, 1, 256] - - [47, 0.0] - - - [7936, 1280, 1, 256] - - [53, 0.0] - - - [6960, 3840, 1, 256] - - [53, 0.0] - - - [3328, 2048, 1, 256] - - [53, 0.0] - - - [16944, 2865, 1, 256] - - [53, 0.0] - - - [1024, 256, 1, 256] - - [54, 0.0] - - - [16944, 3840, 1, 256] - - [53, 0.0] - - - [3376, 2816, 1, 256] - - [47, 0.0] - - - [12288, 768, 1, 256] - - [53, 0.0] - - - [17152, 3329, 1, 256] - - [47, 0.0] - - - [6192, 2865, 1, 256] - - [53, 0.0] - - - [5888, 1281, 1, 256] - - [45, 0.0] - - - [11824, 256, 1, 256] - - [144, 0.0] - - - [18688, 1280, 1, 256] - - [47, 0.0] - - - [11520, 7936, 1, 256] - - [47, 0.0] - - - [15616, 1281, 1, 256] - - [50, 0.0] - - - [16944, 10240, 1, 256] - - [53, 0.0] - - - [12032, 4352, 1, 256] - - [47, 0.0] - - - [9984, 6656, 1, 256] - - [53, 0.0] - - - [17408, 1281, 1, 256] - - [45, 0.0] - - - [6912, 3329, 1, 256] - - [47, 0.0] - - - [16176, 2865, 1, 256] - - [53, 0.0] - - - [7936, 4864, 1, 256] - - [47, 0.0] - - - [7168, 256, 1, 256] - - [47, 0.0] - - - [9728, 6144, 1, 256] - - [53, 0.0] - - - [10752, 7680, 1, 256] - - [53, 0.0] - - - [13056, 5632, 1, 256] - - [53, 0.0] - - - [17152, 2865, 1, 256] - - [47, 0.0] - - - [4096, 512, 1, 256] - - [47, 0.0] - - - [3584, 2304, 1, 256] - - [47, 0.0] - - - [11264, 2048, 1, 256] - - [53, 0.0] - - - [18944, 5376, 1, 256] - - [47, 0.0] - - - [8960, 3329, 1, 256] - - [47, 0.0] - - - [7936, 1281, 1, 256] - - [50, 0.0] - - - [12848, 2816, 1, 256] - - [53, 0.0] - - - [9472, 3328, 1, 256] - - [47, 0.0] - - - [2816, 2816, 1, 256] - - [53, 0.0] - - - [15616, 10240, 1, 256] - - [53, 0.0] - - - [2816, 256, 1, 256] - - [148, 0.0] - - - [48, 256, 1, 256] - - [142, 0.0] - - - [17408, 1792, 1, 256] - - [47, 0.0] - - - [10032, 2865, 1, 256] - - [47, 0.0] - - - [3584, 2865, 1, 256] - - [47, 0.0] - - - [9472, 2816, 1, 256] - - [47, 0.0] - - - [2096, 2048, 1, 256] - - [53, 0.0] - - - [9216, 1536, 1, 256] - - [53, 0.0] - - - [5936, 256, 1, 256] - - [144, 0.0] - - - [11520, 1280, 1, 256] - - [47, 0.0] - - - [16896, 3328, 1, 256] - - [47, 0.0] - - - [7984, 4864, 1, 256] - - [47, 0.0] - - - [11008, 1280, 1, 256] - - [53, 0.0] - - - [18432, 6144, 1, 256] - - [53, 0.0] - - - [2096, 1841, 1, 256] - - [53, 0.0] - - - [8448, 1024, 1, 256] - - [53, 0.0] - - - [17968, 10240, 1, 256] - - [53, 0.0] - - - [1536, 1536, 1, 256] - - [57, 0.0] - - - [7728, 4864, 1, 256] - - [53, 0.0] - - - [18944, 3328, 1, 256] - - [47, 0.0] - - - [4608, 1792, 1, 256] - - [53, 0.0] - - - [8960, 6144, 1, 256] - - [53, 0.0] - - - [18736, 2816, 1, 256] - - [53, 0.0] - - - [8704, 5120, 1, 256] - - [53, 0.0] - - - [19456, 6144, 1, 256] - - [53, 0.0] - - - [19456, 1281, 1, 256] - - [50, 0.0] - - - [17200, 3840, 1, 256] - - [47, 0.0] - - - [2352, 2353, 1, 256] - - [53, 0.0] - - - [17408, 2816, 1, 256] - - [47, 0.0] - - - [13312, 2816, 1, 256] - - [47, 0.0] - - - [8960, 2816, 1, 256] - - [47, 0.0] - - - [2048, 1792, 1, 256] - - [57, 0.0] - - - [17152, 10240, 1, 256] - - [53, 0.0] - - - [16176, 10240, 1, 256] - - [53, 0.0] - - - [10288, 2865, 1, 256] - - [47, 0.0] - - - [8704, 2816, 1, 256] - - [47, 0.0] - - - [7424, 4096, 1, 256] - - [53, 0.0] - - - [6656, 1024, 1, 256] - - [53, 0.0] - - - [2304, 256, 1, 256] - - [48, 0.0] - - - [16384, 2865, 1, 256] - - [47, 0.0] - - - [7680, 2816, 1, 256] - - [47, 0.0] - - - [11520, 3329, 1, 256] - - [47, 0.0] - - - [10752, 1280, 1, 256] - - [47, 0.0] - - - [3120, 2816, 1, 256] - - [53, 0.0] - - - [15872, 1281, 1, 256] - - [45, 0.0] - - - [13824, 6144, 1, 256] - - [53, 0.0] - - - [6912, 3584, 1, 256] - - [53, 0.0] - - - [12032, 3328, 1, 256] - - [47, 0.0] - - - [11264, 1281, 1, 256] - - [45, 0.0] - - - [19456, 5632, 1, 256] - - [53, 0.0] - - - [17200, 2816, 1, 256] - - [53, 0.0] - - - [11520, 3840, 1, 256] - - [47, 0.0] - - - [11520, 2865, 1, 256] - - [47, 0.0] - - - [14848, 1280, 1, 256] - - [53, 0.0] - - - [16176, 256, 1, 256] - - [68, 0.0] - - - [16384, 256, 1, 256] - - [47, 0.0] - - - [4096, 768, 1, 256] - - [69, 0.0] - - - [4864, 2816, 1, 256] - - [53, 0.0] - - - [13568, 256, 1, 256] - - [53, 0.0] - - - [4608, 2048, 1, 256] - - [53, 0.0] - - - [9984, 6144, 1, 256] - - [53, 0.0] - - - [3632, 768, 1, 256] - - [53, 0.0] - - - [19200, 5888, 1, 256] - - [47, 0.0] - - - [5632, 2865, 1, 256] - - [47, 0.0] - - - [15360, 1280, 1, 256] - - [47, 0.0] - - - [12800, 1280, 1, 256] - - [47, 0.0] - - - [7168, 3328, 1, 256] - - [47, 0.0] - - - [11264, 8448, 1, 256] - - [47, 0.0] - - - [18176, 3328, 1, 256] - - [47, 0.0] - - - [4096, 2560, 1, 256] - - [53, 0.0] - - - [12544, 768, 1, 256] - - [53, 0.0] - - - [11568, 8448, 1, 256] - - [47, 0.0] - - - [8704, 1280, 1, 256] - - [47, 0.0] - - - [13056, 1536, 1, 256] - - [53, 0.0] - - - [2304, 1024, 1, 256] - - [69, 0.0] - - - [3072, 1281, 1, 256] - - [129, 0.0] - - - [6912, 1280, 1, 256] - - [47, 0.0] - - - [9216, 2816, 1, 256] - - [47, 0.0] - - - [17152, 6144, 1, 256] - - [53, 0.0] - - - [18992, 2865, 1, 256] - - [53, 0.0] - - - [10240, 2560, 1, 256] - - [53, 0.0] - - - [560, 256, 1, 256] - - [158, 0.0] - - - [2304, 1280, 1, 256] - - [53, 0.0] - - - [7680, 6144, 1, 256] - - [53, 0.0] - - - [15920, 2560, 1, 256] - - [53, 0.0] - - - [17456, 10240, 1, 256] - - [53, 0.0] - - - [14080, 3328, 1, 256] - - [47, 0.0] - - - [13360, 10240, 1, 256] - - [53, 0.0] - - - [8448, 5632, 1, 256] - - [53, 0.0] - - - [17408, 3584, 1, 256] - - [53, 0.0] - - - [6704, 2865, 1, 256] - - [47, 0.0] - - - [12592, 9472, 1, 256] - - [53, 0.0] - - - [18992, 10240, 1, 256] - - [53, 0.0] - - - [5376, 2865, 1, 256] - - [47, 0.0] - - - [18480, 5120, 1, 256] - - [53, 0.0] - - - [14336, 2048, 1, 256] - - [53, 0.0] - - - [7424, 3328, 1, 256] - - [47, 0.0] - - - [256, 49, 1, 256] - - [145, 0.0] - - - [12288, 1280, 1, 256] - - [47, 0.0] - - - [13568, 3329, 1, 256] - - [47, 0.0] - - - [15360, 1792, 1, 256] - - [53, 0.0] - - - [7168, 3584, 1, 256] - - [53, 0.0] - - - [10240, 3328, 1, 256] - - [47, 0.0] - - - [6400, 1281, 1, 256] - - [50, 0.0] - - - [11008, 6144, 1, 256] - - [53, 0.0] - - - [512, 513, 1, 256] - - [145, 0.0] - - - [19248, 256, 1, 256] - - [47, 0.0] - - - [2608, 256, 1, 256] - - [154, 0.0] - - - [16688, 3584, 1, 256] - - [53, 0.0] - - - [17920, 4864, 1, 256] - - [53, 0.0] - - - [18688, 1281, 1, 256] - - [45, 0.0] - - - [18224, 4864, 1, 256] - - [47, 0.0] - - - [10496, 2816, 1, 256] - - [47, 0.0] - - - [12288, 4864, 1, 256] - - [53, 0.0] - - - [9216, 2865, 1, 256] - - [47, 0.0] - - - [17664, 3329, 1, 256] - - [47, 0.0] - - - [3632, 512, 1, 256] - - [68, 0.0] - - - [11776, 3329, 1, 256] - - [47, 0.0] - - - [19456, 1792, 1, 256] - - [47, 0.0] - - - [12592, 256, 1, 256] - - [53, 0.0] - - - [10752, 3072, 1, 256] - - [53, 0.0] - - - [10800, 2816, 1, 256] - - [53, 0.0] - - - [6192, 3072, 1, 256] - - [53, 0.0] - - - [17152, 1536, 1, 256] - - [53, 0.0] - - - [2096, 2097, 1, 256] - - [47, 0.0] - - - [8192, 4608, 1, 256] - - [53, 0.0] - - - [13056, 6144, 1, 256] - - [53, 0.0] - - - [16640, 10240, 1, 256] - - [53, 0.0] - - - [12592, 9728, 1, 256] - - [53, 0.0] - - - [18176, 2816, 1, 256] - - [47, 0.0] - - - [18176, 4864, 1, 256] - - [47, 0.0] - - - [18944, 6144, 1, 256] - - [53, 0.0] - - - [12544, 256, 1, 256] - - [53, 0.0] - - - [13056, 1281, 1, 256] - - [44, 0.0] - - - [304, 49, 1, 256] - - [142, 0.0] - - - [17920, 2816, 1, 256] - - [47, 0.0] - - - [4656, 256, 1, 256] - - [144, 0.0] - - - [7728, 2865, 1, 256] - - [47, 0.0] - - - [15872, 1280, 1, 256] - - [47, 0.0] - - - [17456, 256, 1, 256] - - [53, 0.0] - - - [18176, 4608, 1, 256] - - [53, 0.0] - - - [7168, 5632, 1, 256] - - [53, 0.0] - - - [13616, 256, 1, 256] - - [47, 0.0] - - - [15104, 3329, 1, 256] - - [47, 0.0] - - - [19712, 3328, 1, 256] - - [47, 0.0] - - - [10032, 7168, 1, 256] - - [53, 0.0] - - - [11008, 3328, 1, 256] - - [53, 0.0] - - - [10496, 6144, 1, 256] - - [53, 0.0] - - - [6656, 2865, 1, 256] - - [47, 0.0] - - - [17664, 3840, 1, 256] - - [47, 0.0] - - - [6960, 4096, 1, 256] - - [53, 0.0] - - - [4608, 256, 1, 256] - - [68, 0.0] - - - [10496, 6912, 1, 256] - - [47, 0.0] - - - [16128, 2560, 1, 256] - - [53, 0.0] - - - [15872, 256, 1, 256] - - [64, 0.0] - - - [6656, 1281, 1, 256] - - [45, 0.0] - - - [3584, 512, 1, 256] - - [53, 0.0] - - - [11520, 1792, 1, 256] - - [47, 0.0] - - - [11264, 8192, 1, 256] - - [53, 0.0] - - - [10752, 2048, 1, 256] - - [53, 0.0] - - - [18688, 3329, 1, 256] - - [47, 0.0] - - - [4352, 768, 1, 256] - - [64, 0.0] - - - [18432, 512, 1, 256] - - [53, 0.0] - - - [18992, 256, 1, 256] - - [69, 0.0] - - - [13568, 2865, 1, 256] - - [47, 0.0] - - - [14640, 256, 1, 256] - - [47, 0.0] - - - [11264, 256, 1, 256] - - [53, 0.0] - - - [16896, 3329, 1, 256] - - [47, 0.0] - - - [18944, 5120, 1, 256] - - [53, 0.0] - - - [768, 513, 1, 256] - - [58, 0.0] - - - [14080, 10240, 1, 256] - - [53, 0.0] - - - [15872, 2560, 1, 256] - - [53, 0.0] - - - [6912, 5632, 1, 256] - - [53, 0.0] - - - [13360, 2865, 1, 256] - - [53, 0.0] - - - [6400, 3072, 1, 256] - - [53, 0.0] - - - [15616, 3329, 1, 256] - - [47, 0.0] - - - [9264, 2816, 1, 256] - - [47, 0.0] - - - [18176, 512, 1, 256] - - [53, 0.0] - - - [11264, 1280, 1, 256] - - [47, 0.0] - - - [1328, 1329, 1, 256] - - [53, 0.0] - - - [18736, 5376, 1, 256] - - [53, 0.0] - - - [5376, 1792, 1, 256] - - [47, 0.0] - - - [6144, 4608, 1, 256] - - [53, 0.0] - - - [6400, 3328, 1, 256] - - [47, 0.0] - - - [12032, 2865, 1, 256] - - [47, 0.0] - - - [12288, 4608, 1, 256] - - [53, 0.0] - - - [16128, 3072, 1, 256] - - [53, 0.0] - - - [2048, 256, 1, 256] - - [149, 0.0] - - - [4096, 256, 1, 256] - - [153, 0.0] - - - [5888, 2304, 1, 256] - - [47, 0.0] - - - [2816, 2561, 1, 256] - - [53, 0.0] - - - [3072, 1536, 1, 256] - - [47, 0.0] - - - [2304, 1281, 1, 256] - - [47, 0.0] - - - [15616, 2048, 1, 256] - - [53, 0.0] - - - [12800, 1024, 1, 256] - - [53, 0.0] - - - [8960, 3328, 1, 256] - - [47, 0.0] - - - [18432, 1280, 1, 256] - - [47, 0.0] - - - [8448, 2048, 1, 256] - - [53, 0.0] - - - [19712, 6400, 1, 256] - - [47, 0.0] - - - [14384, 1280, 1, 256] - - [47, 0.0] - - - [6448, 2816, 1, 256] - - [47, 0.0] - - - [18176, 2048, 1, 256] - - [53, 0.0] - - - [3072, 1792, 1, 256] - - [47, 0.0] - - - [12080, 8960, 1, 256] - - [47, 0.0] - - - [13312, 1281, 1, 256] - - [45, 0.0] - - - [16688, 2816, 1, 256] - - [53, 0.0] - - - [6400, 256, 1, 256] - - [64, 0.0] - - - [2048, 2048, 1, 256] - - [53, 0.0] - - - [14336, 256, 1, 256] - - [47, 0.0] - - - [11520, 2816, 1, 256] - - [47, 0.0] - - - [14384, 10240, 1, 256] - - [53, 0.0] - - - [7472, 256, 1, 256] - - [144, 0.0] - - - [1280, 1280, 1, 256] - - [59, 0.0] - - - [8704, 1024, 1, 256] - - [53, 0.0] - - - [9520, 2865, 1, 256] - - [47, 0.0] - - - [18480, 256, 1, 256] - - [47, 0.0] - - - [18176, 256, 1, 256] - - [53, 0.0] - - - [15872, 6144, 1, 256] - - [53, 0.0] - - - [304, 256, 1, 256] - - [146, 0.0] - - - [13568, 5888, 1, 256] - - [47, 0.0] - - - [3328, 3328, 1, 256] - - [53, 0.0] - - - [6656, 5120, 1, 256] - - [53, 0.0] - - - [9520, 2816, 1, 256] - - [47, 0.0] - - - [1536, 1537, 1, 256] - - [69, 0.0] - - - [3072, 2865, 1, 256] - - [47, 0.0] - - - [10032, 2816, 1, 256] - - [53, 0.0] - - - [12032, 9216, 1, 256] - - [53, 0.0] - - - [13872, 10240, 1, 256] - - [53, 0.0] - - - [13824, 2048, 1, 256] - - [53, 0.0] - - - [12544, 9728, 1, 256] - - [53, 0.0] - - - [17664, 4352, 1, 256] - - [47, 0.0] - - - [4096, 1281, 1, 256] - - [50, 0.0] - - - [17408, 1280, 1, 256] - - [47, 0.0] - - - [18432, 2816, 1, 256] - - [47, 0.0] - - - [5120, 256, 1, 256] - - [67, 0.0] - - - [18736, 2865, 1, 256] - - [53, 0.0] - - - [19200, 256, 1, 256] - - [69, 0.0] - - - [2048, 512, 1, 256] - - [36, 0.0] - - - [11008, 7680, 1, 256] - - [53, 0.0] - - - [5888, 3072, 1, 256] - - [53, 0.0] - - - [11776, 8192, 1, 256] - - [53, 0.0] - - - [5888, 512, 1, 256] - - [68, 0.0] - - - [7936, 2816, 1, 256] - - [47, 0.0] - - - [5120, 2865, 1, 256] - - [47, 0.0] - - - [12032, 2816, 1, 256] - - [47, 0.0] - - - [256, 257, 1, 256] - - [143, 0.0] - - - [13104, 2865, 1, 256] - - [47, 0.0] - - - [5680, 2865, 1, 256] - - [53, 0.0] - - - [15408, 10240, 1, 256] - - [53, 0.0] - - - [18432, 4864, 1, 256] - - [47, 0.0] - - - [17712, 2865, 1, 256] - - [47, 0.0] - - - [768, 256, 1, 256] - - [142, 0.0] - - - [9728, 3328, 1, 256] - - [47, 0.0] - - - [12848, 9728, 1, 256] - - [53, 0.0] - - - [2304, 2305, 1, 256] - - [47, 0.0] - - - [10240, 6144, 1, 256] - - [53, 0.0] - - - [13312, 1280, 1, 256] - - [53, 0.0] - - - [9008, 5888, 1, 256] - - [53, 0.0] - - - [7424, 3840, 1, 256] - - [47, 0.0] - - - [12032, 1280, 1, 256] - - [47, 0.0] - - - [18480, 2816, 1, 256] - - [53, 0.0] - - - [18432, 5120, 1, 256] - - [53, 0.0] - - - [7424, 4608, 1, 256] - - [47, 0.0] - - - [9776, 2865, 1, 256] - - [47, 0.0] - - - [5632, 2560, 1, 256] - - [53, 0.0] - - - [7680, 2048, 1, 256] - - [53, 0.0] - - - [6704, 2816, 1, 256] - - [53, 0.0] - - - [13872, 2816, 1, 256] - - [47, 0.0] - - - [17968, 2816, 1, 256] - - [53, 0.0] - - - [4144, 2865, 1, 256] - - [53, 0.0] - - - [14640, 1280, 1, 256] - - [47, 0.0] - - - [16432, 2816, 1, 256] - - [47, 0.0] - - - [16128, 1280, 1, 256] - - [47, 0.0] - - - [8240, 5120, 1, 256] - - [53, 0.0] - - - [4352, 2816, 1, 256] - - [47, 0.0] - - - [12544, 2865, 1, 256] - - [47, 0.0] - - - [6144, 2048, 1, 256] - - [53, 0.0] - - - [13616, 512, 1, 256] - - [53, 0.0] - - - [5632, 2048, 1, 256] - - [53, 0.0] - - - [13312, 2048, 1, 256] - - [53, 0.0] - - - [9728, 1281, 1, 256] - - [49, 0.0] - - - [7424, 1281, 1, 256] - - [49, 0.0] - - - [10800, 256, 1, 256] - - [53, 0.0] - - - [2048, 1281, 1, 256] - - [47, 0.0] - - - [5376, 1280, 1, 256] - - [47, 0.0] - - - [15664, 2816, 1, 256] - - [53, 0.0] - - - [256, 256, 1, 256] - - [142, 0.0] - - - [2048, 1280, 1, 256] - - [59, 0.0] - - - [9776, 256, 1, 256] - - [47, 0.0] - - - [4096, 3329, 1, 256] - - [53, 0.0] - - - [9728, 2304, 1, 256] - - [47, 0.0] - - - [19968, 2865, 1, 256] - - [47, 0.0] - - - [13568, 6144, 1, 256] - - [53, 0.0] - - - [15360, 2304, 1, 256] - - [47, 0.0] - - - [9264, 6400, 1, 256] - - [47, 0.0] - - - [19200, 2048, 1, 256] - - [53, 0.0] - - - [11520, 4096, 1, 256] - - [53, 0.0] - - - [18688, 5632, 1, 256] - - [53, 0.0] - - - [11776, 256, 1, 256] - - [144, 0.0] - - - [17152, 256, 1, 256] - - [47, 0.0] - - - [5120, 1280, 1, 256] - - [47, 0.0] - - - [14896, 1792, 1, 256] - - [47, 0.0] - - - [10288, 2816, 1, 256] - - [53, 0.0] - - - [7984, 2865, 1, 256] - - [53, 0.0] - - - [4864, 1281, 1, 256] - - [44, 0.0] - - - [7216, 256, 1, 256] - - [66, 0.0] - - - [5888, 3328, 1, 256] - - [47, 0.0] - - - [7424, 2816, 1, 256] - - [47, 0.0] - - - [15360, 3328, 1, 256] - - [47, 0.0] - - - [10544, 256, 1, 256] - - [68, 0.0] - - - [9776, 2816, 1, 256] - - [53, 0.0] - - - [8240, 2816, 1, 256] - - [47, 0.0] - - - [6656, 3072, 1, 256] - - [53, 0.0] - - - [18224, 10240, 1, 256] - - [53, 0.0] - - - [13824, 2865, 1, 256] - - [47, 0.0] - - - [5376, 1281, 1, 256] - - [45, 0.0] - - - [13568, 9984, 1, 256] - - [47, 0.0] - - - [18176, 4352, 1, 256] - - [47, 0.0] - - - [11776, 1281, 1, 256] - - [49, 0.0] - - - [15616, 6144, 1, 256] - - [53, 0.0] - - - [4400, 256, 1, 256] - - [66, 0.0] - - - [18992, 2816, 1, 256] - - [47, 0.0] - - - [14640, 10240, 1, 256] - - [53, 0.0] - - - [5120, 2048, 1, 256] - - [53, 0.0] - - - [19968, 10240, 1, 256] - - [53, 0.0] - - - [19200, 2865, 1, 256] - - [47, 0.0] - - - [15152, 2816, 1, 256] - - [53, 0.0] - - - [2560, 2560, 1, 256] - - [53, 0.0] - - - [8448, 2816, 1, 256] - - [53, 0.0] - - - [8704, 5632, 1, 256] - - [53, 0.0] - - - [1024, 769, 1, 256] - - [34, 0.0] - - - [17200, 4096, 1, 256] - - [53, 0.0] - - - [5376, 256, 1, 256] - - [67, 0.0] - - - [6656, 256, 1, 256] - - [64, 0.0] - - - [18688, 3328, 1, 256] - - [47, 0.0] - - - [13056, 256, 1, 256] - - [63, 0.0] - - - [13104, 2816, 1, 256] - - [53, 0.0] - - - [7424, 1792, 1, 256] - - [47, 0.0] - - - [14592, 2816, 1, 256] - - [47, 0.0] - - - [12336, 2865, 1, 256] - - [53, 0.0] - - - [17920, 256, 1, 256] - - [69, 0.0] - - - [12800, 2048, 1, 256] - - [53, 0.0] - - - [3632, 256, 1, 256] - - [155, 0.0] - - - [18688, 768, 1, 256] - - [53, 0.0] - - - [16384, 2816, 1, 256] - - [47, 0.0] - - - [14896, 10240, 1, 256] - - [53, 0.0] - - - [816, 817, 1, 256] - - [136, 0.0] - - - [9008, 2865, 1, 256] - - [47, 0.0] - - - [14848, 1024, 1, 256] - - [53, 0.0] - - - [16640, 256, 1, 256] - - [47, 0.0] - - - [7424, 256, 1, 256] - - [68, 0.0] - - - [10240, 3329, 1, 256] - - [47, 0.0] - - - [18176, 5120, 1, 256] - - [53, 0.0] - - - [6912, 2865, 1, 256] - - [47, 0.0] - - - [1024, 1025, 1, 256] - - [129, 0.0] - - - [5632, 4096, 1, 256] - - [53, 0.0] - - - [12544, 1024, 1, 256] - - [53, 0.0] - - - [2864, 2609, 1, 256] - - [53, 0.0] - - - [16896, 2048, 1, 256] - - [53, 0.0] - - - [3840, 1280, 1, 256] - - [59, 0.0] - - - [11008, 1281, 1, 256] - - [45, 0.0] - - - [15104, 1281, 1, 256] - - [45, 0.0] - - - [7168, 3329, 1, 256] - - [47, 0.0] - - - [12800, 5120, 1, 256] - - [53, 0.0] - - - [512, 257, 1, 256] - - [56, 0.0] - - - [12288, 8960, 1, 256] - - [47, 0.0] - - - [9728, 6912, 1, 256] - - [47, 0.0] - - - [9728, 6656, 1, 256] - - [53, 0.0] - - - [2560, 2304, 1, 256] - - [69, 0.0] - - - [10544, 7424, 1, 256] - - [47, 0.0] - - - [5888, 3329, 1, 256] - - [47, 0.0] - - - [3888, 2816, 1, 256] - - [53, 0.0] - - - [18944, 10240, 1, 256] - - [53, 0.0] - - - [17200, 10240, 1, 256] - - [53, 0.0] - - - [4144, 1280, 1, 256] - - [47, 0.0] - - - [9728, 1280, 1, 256] - - [47, 0.0] - - - [14896, 1536, 1, 256] - - [53, 0.0] - - - [5888, 4352, 1, 256] - - [53, 0.0] - - - [1024, 1024, 1, 256] - - [36, 0.0] - - - [4912, 2816, 1, 256] - - [53, 0.0] - - - [19456, 3329, 1, 256] - - [47, 0.0] - - - [7680, 4608, 1, 256] - - [53, 0.0] - - - [8496, 2865, 1, 256] - - [53, 0.0] - - - [3584, 2048, 1, 256] - - [53, 0.0] - - - [9984, 3329, 1, 256] - - [47, 0.0] - - - [10800, 7680, 1, 256] - - [53, 0.0] - - - [13616, 2816, 1, 256] - - [47, 0.0] - - - [15104, 10240, 1, 256] - - [53, 0.0] - - - [10240, 6656, 1, 256] - - [47, 0.0] - - - [16128, 1281, 1, 256] - - [45, 0.0] - - - [16896, 1280, 1, 256] - - [47, 0.0] - - - [12544, 9472, 1, 256] - - [47, 0.0] - - - [11008, 7424, 1, 256] - - [47, 0.0] - - - [9472, 3329, 1, 256] - - [47, 0.0] - - - [6912, 2816, 1, 256] - - [53, 0.0] - - - [2048, 1841, 1, 256] - - [58, 0.0] - - - [17152, 4096, 1, 256] - - [53, 0.0] - - - [12544, 5120, 1, 256] - - [53, 0.0] - - - [13824, 3328, 1, 256] - - [47, 0.0] - - - [6912, 2048, 1, 256] - - [53, 0.0] - - - [9472, 256, 1, 256] - - [63, 0.0] - - - [9216, 1281, 1, 256] - - [45, 0.0] - - - [7168, 1281, 1, 256] - - [50, 0.0] - - - [10752, 7424, 1, 256] - - [47, 0.0] - - - [16176, 3072, 1, 256] - - [53, 0.0] - - - [12288, 9216, 1, 256] - - [53, 0.0] - - - [14336, 512, 1, 256] - - [53, 0.0] - - - [14336, 3328, 1, 256] - - [53, 0.0] - - - [4864, 1280, 1, 256] - - [47, 0.0] - - - [19760, 2865, 1, 256] - - [47, 0.0] - - - [8240, 256, 1, 256] - - [47, 0.0] - - - [18688, 1024, 1, 256] - - [53, 0.0] - - - [16128, 10240, 1, 256] - - [53, 0.0] - - - [5632, 256, 1, 256] - - [64, 0.0] - - - [5680, 2560, 1, 256] - - [53, 0.0] - - - [7680, 1281, 1, 256] - - [49, 0.0] - - - [17408, 2048, 1, 256] - - [53, 0.0] - - - [10752, 2865, 1, 256] - - [47, 0.0] - - - [14848, 1281, 1, 256] - - [44, 0.0] - - - [560, 512, 1, 256] - - [65, 0.0] - - - [19968, 1280, 1, 256] - - [47, 0.0] - - - [16384, 10240, 1, 256] - - [53, 0.0] - - - [512, 305, 1, 256] - - [56, 0.0] - - - [19200, 6144, 1, 256] - - [53, 0.0] - - - [8448, 5120, 1, 256] - - [53, 0.0] - - - [13824, 3329, 1, 256] - - [47, 0.0] - - - [7984, 2816, 1, 256] - - [53, 0.0] - - - [17920, 3329, 1, 256] - - [47, 0.0] - - - [16688, 2865, 1, 256] - - [53, 0.0] - - - [12032, 256, 1, 256] - - [62, 0.0] - - - [7424, 2865, 1, 256] - - [47, 0.0] - - - [14336, 10240, 1, 256] - - [53, 0.0] - - - [17152, 2048, 1, 256] - - [53, 0.0] - - - [14896, 2816, 1, 256] - - [53, 0.0] - - - [16384, 2048, 1, 256] - - [53, 0.0] - - - [8192, 2816, 1, 256] - - [47, 0.0] - - - [6192, 256, 1, 256] - - [69, 0.0] - - - [2304, 768, 1, 256] - - [47, 0.0] - - - [18688, 256, 1, 256] - - [63, 0.0] - - - [8960, 1281, 1, 256] - - [50, 0.0] - - - [19968, 6400, 1, 256] - - [47, 0.0] - - - [8752, 2816, 1, 256] - - [53, 0.0] - - - [19456, 3328, 1, 256] - - [53, 0.0] - - - [2560, 2561, 1, 256] - - [50, 0.0] - - - [15920, 2865, 1, 256] - - [53, 0.0] - - - [12544, 6144, 1, 256] - - [53, 0.0] - - - [19200, 3328, 1, 256] - - [47, 0.0] - - - [3328, 2865, 1, 256] - - [59, 0.0] - - - [7936, 3329, 1, 256] - - [47, 0.0] - - - [11264, 2865, 1, 256] - - [53, 0.0] - - - [6144, 3329, 1, 256] - - [47, 0.0] - - - [16128, 3329, 1, 256] - - [47, 0.0] - - - [12800, 9728, 1, 256] - - [53, 0.0] - - - [512, 256, 1, 256] - - [56, 0.0] - - - [11264, 2816, 1, 256] - - [47, 0.0] - - - [12544, 3329, 1, 256] - - [47, 0.0] - - - [14848, 3329, 1, 256] - - [47, 0.0] - - - [1328, 256, 1, 256] - - [147, 0.0] - - - [3120, 256, 1, 256] - - [132, 0.0] - - - [1024, 768, 1, 256] - - [132, 0.0] - - - [7728, 2816, 1, 256] - - [47, 0.0] - - - [1024, 817, 1, 256] - - [138, 0.0] - - - [10288, 7424, 1, 256] - - [53, 0.0] - - - [19968, 6144, 1, 256] - - [53, 0.0] - - - [13616, 10240, 1, 256] - - [53, 0.0] - - - [1536, 1329, 1, 256] - - [49, 0.0] - - - [9984, 3328, 1, 256] - - [47, 0.0] - - - [9472, 5888, 1, 256] - - [47, 0.0] - - - [11264, 7936, 1, 256] - - [47, 0.0] - - - [8496, 256, 1, 256] - - [47, 0.0] - - - [17664, 1792, 1, 256] - - [47, 0.0] - - - [11824, 2816, 1, 256] - - [47, 0.0] - - - [16944, 2816, 1, 256] - - [53, 0.0] - - - [19968, 6912, 1, 256] - - [47, 0.0] - - - [3376, 2865, 1, 256] - - [47, 0.0] - - - [3840, 2560, 1, 256] - - [53, 0.0] - - - [11776, 8448, 1, 256] - - [47, 0.0] - - - [19248, 6144, 1, 256] - - [53, 0.0] - - - [14080, 512, 1, 256] - - [53, 0.0] - - - [16128, 3328, 1, 256] - - [47, 0.0] - - - [6656, 2048, 1, 256] - - [53, 0.0] - - - [15664, 256, 1, 256] - - [53, 0.0] - - - [17664, 1280, 1, 256] - - [47, 0.0] - - - [16384, 6144, 1, 256] - - [53, 0.0] - - - [9984, 256, 1, 256] - - [53, 0.0] - - - [14592, 1281, 1, 256] - - [45, 0.0] - - - [4608, 3329, 1, 256] - - [53, 0.0] - - - [8960, 2048, 1, 256] - - [53, 0.0] - - - [2864, 2865, 1, 256] - - [47, 0.0] - - - [2816, 2609, 1, 256] - - [47, 0.0] - - - [14080, 1281, 1, 256] - - [50, 0.0] - - - [1792, 1536, 1, 256] - - [47, 0.0] - - - [10240, 7424, 1, 256] - - [47, 0.0] - - - [5936, 2816, 1, 256] - - [47, 0.0] - - - [19712, 256, 1, 256] - - [47, 0.0] - - - [18944, 5888, 1, 256] - - [47, 0.0] - - - [9728, 3329, 1, 256] - - [47, 0.0] - - - [19248, 2816, 1, 256] - - [53, 0.0] - - - [13568, 1792, 1, 256] - - [47, 0.0] - - - [1584, 1585, 1, 256] - - [53, 0.0] - - - [8704, 2048, 1, 256] - - [53, 0.0] - - - [13056, 9728, 1, 256] - - [53, 0.0] - - - [12800, 2865, 1, 256] - - [47, 0.0] - - - [14336, 6144, 1, 256] - - [53, 0.0] - - - [5120, 1536, 1, 256] - - [53, 0.0] - - - [18432, 1281, 1, 256] - - [50, 0.0] - - - [10240, 256, 1, 256] - - [69, 0.0] - - - [12544, 9216, 1, 256] - - [53, 0.0] - - - [12800, 1281, 1, 256] - - [45, 0.0] - - - [8704, 5888, 1, 256] - - [47, 0.0] - - - [15360, 3329, 1, 256] - - [47, 0.0] - - - [11312, 8448, 1, 256] - - [47, 0.0] - - - [17152, 3328, 1, 256] - - [47, 0.0] - - - [16384, 3328, 1, 256] - - [47, 0.0] - - - [13824, 2816, 1, 256] - - [47, 0.0] - - - [560, 305, 1, 256] - - [65, 0.0] - - - [16432, 256, 1, 256] - - [47, 0.0] - - - [3632, 2865, 1, 256] - - [47, 0.0] - - - [3584, 3328, 1, 256] - - [53, 0.0] - - - [3840, 768, 1, 256] - - [53, 0.0] - - - [19504, 256, 1, 256] - - [53, 0.0] - - - [1280, 1073, 1, 256] - - [59, 0.0] - - - [17712, 10240, 1, 256] - - [53, 0.0] - - - [2816, 1536, 1, 256] - - [53, 0.0] - - - [12800, 6144, 1, 256] - - [53, 0.0] - - - [4656, 2816, 1, 256] - - [47, 0.0] - - - [17920, 10240, 1, 256] - - [53, 0.0] - - - [9984, 2816, 1, 256] - - [47, 0.0] - - - [4352, 256, 1, 256] - - [66, 0.0] - - - [11312, 2865, 1, 256] - - [47, 0.0] - - - [18432, 3328, 1, 256] - - [47, 0.0] - - - [4096, 1280, 1, 256] - - [47, 0.0] - - - [4864, 3329, 1, 256] - - [45, 0.0] - - - [14640, 2865, 1, 256] - - [47, 0.0] - - - [17152, 2816, 1, 256] - - [47, 0.0] - - - [7680, 1280, 1, 256] - - [53, 0.0] - - - [1584, 1536, 1, 256] - - [69, 0.0] - - - [14080, 1280, 1, 256] - - [47, 0.0] - - - [13824, 512, 1, 256] - - [53, 0.0] - - - [7936, 256, 1, 256] - - [53, 0.0] - - - [12592, 2865, 1, 256] - - [53, 0.0] - - - [2816, 2560, 1, 256] - - [53, 0.0] - - - [6912, 1536, 1, 256] - - [53, 0.0] - - - [12800, 9984, 1, 256] - - [47, 0.0] - - - [10496, 256, 1, 256] - - [64, 0.0] - - - [18176, 2865, 1, 256] - - [47, 0.0] - - - [4608, 1536, 1, 256] - - [53, 0.0] - - - [3328, 2816, 1, 256] - - [47, 0.0] - - - [3840, 1024, 1, 256] - - [69, 0.0] - - - [13824, 1280, 1, 256] - - [47, 0.0] - - - [3840, 1281, 1, 256] - - [49, 0.0] - - - [17152, 1281, 1, 256] - - [45, 0.0] - - - [13568, 1281, 1, 256] - - [45, 0.0] - - - [14848, 1792, 1, 256] - - [47, 0.0] - - - [13056, 9472, 1, 256] - - [47, 0.0] - - - [18176, 1281, 1, 256] - - [50, 0.0] - - - [5680, 256, 1, 256] - - [69, 0.0] - - - [13056, 2816, 1, 256] - - [47, 0.0] - - - [11824, 8704, 1, 256] - - [53, 0.0] - - - [7936, 4352, 1, 256] - - [47, 0.0] - - - [8704, 256, 1, 256] - - [68, 0.0] - - - [5424, 2304, 1, 256] - - [53, 0.0] - - - [14128, 768, 1, 256] - - [53, 0.0] - - - [10752, 1024, 1, 256] - - [53, 0.0] - - - [9264, 6144, 1, 256] - - [53, 0.0] - - - [4352, 3328, 1, 256] - - [47, 0.0] - - - [18944, 5632, 1, 256] - - [53, 0.0] - - - [12032, 8704, 1, 256] - - [53, 0.0] - - - [2048, 2049, 1, 256] - - [50, 0.0] - - - [6400, 3329, 1, 256] - - [47, 0.0] - - - [15616, 2560, 1, 256] - - [53, 0.0] - - - [7472, 2865, 1, 256] - - [47, 0.0] - - - [14848, 1536, 1, 256] - - [53, 0.0] - - - [18736, 10240, 1, 256] - - [53, 0.0] - - - [6400, 1024, 1, 256] - - [53, 0.0] - - - [7936, 5120, 1, 256] - - [53, 0.0] - - - [4656, 1536, 1, 256] - - [53, 0.0] - - - [3328, 256, 1, 256] - - [40, 0.0] - - - [3072, 1280, 1, 256] - - [57, 0.0] - - - [2864, 2816, 1, 256] - - [47, 0.0] - - - [9472, 6144, 1, 256] - - [53, 0.0] - - - [3840, 2304, 1, 256] - - [47, 0.0] - - - [17408, 2865, 1, 256] - - [47, 0.0] - - - [16384, 2560, 1, 256] - - [53, 0.0] - - - [16384, 3329, 1, 256] - - [47, 0.0] - - - [16688, 10240, 1, 256] - - [53, 0.0] - - - [18688, 2048, 1, 256] - - [53, 0.0] - - - [7936, 4608, 1, 256] - - [53, 0.0] - - - [9472, 6400, 1, 256] - - [47, 0.0] - - - [14336, 3329, 1, 256] - - [47, 0.0] - - - [4608, 1024, 1, 256] - - [53, 0.0] - - - [16896, 6144, 1, 256] - - [53, 0.0] - - - [10752, 3329, 1, 256] - - [47, 0.0] - - - [6704, 256, 1, 256] - - [68, 0.0] - - - [17408, 6144, 1, 256] - - [53, 0.0] - - - [9984, 2048, 1, 256] - - [53, 0.0] - - - [17968, 4864, 1, 256] - - [53, 0.0] - - - [5120, 3584, 1, 256] - - [53, 0.0] - - - [14336, 2865, 1, 256] - - [47, 0.0] - - - [18736, 256, 1, 256] - - [68, 0.0] - - - [13568, 2048, 1, 256] - - [53, 0.0] - - - [17456, 4352, 1, 256] - - [53, 0.0] - - - [5424, 2816, 1, 256] - - [53, 0.0] - - - [17664, 4608, 1, 256] - - [53, 0.0] - - - [7984, 256, 1, 256] - - [47, 0.0] - - - [6400, 3584, 1, 256] - - [53, 0.0] - - - [19712, 3329, 1, 256] - - [47, 0.0] - - - [10752, 2816, 1, 256] - - [47, 0.0] - - - [17152, 1280, 1, 256] - - [47, 0.0] - - - [560, 561, 1, 256] - - [133, 0.0] - - - [8192, 1281, 1, 256] - - [44, 0.0] - - - [4864, 1792, 1, 256] - - [53, 0.0] - - - [3632, 2816, 1, 256] - - [47, 0.0] - - - [11520, 8448, 1, 256] - - [47, 0.0] - - - [5168, 2865, 1, 256] - - [53, 0.0] - - - [13568, 10240, 1, 256] - - [53, 0.0] - - - [12544, 2816, 1, 256] - - [47, 0.0] - - - [19504, 6144, 1, 256] - - [53, 0.0] - - - [11776, 2048, 1, 256] - - [53, 0.0] - - - [18688, 2865, 1, 256] - - [47, 0.0] - - - [14336, 768, 1, 256] - - [53, 0.0] - - - [18688, 6144, 1, 256] - - [53, 0.0] - - - [11776, 2816, 1, 256] - - [47, 0.0] - - - [12288, 9472, 1, 256] - - [47, 0.0] - - - [5120, 1792, 1, 256] - - [53, 0.0] - - - [16128, 512, 1, 256] - - [53, 0.0] - - - [5376, 3329, 1, 256] - - [47, 0.0] - - - [9216, 3329, 1, 256] - - [47, 0.0] - - - [9008, 2816, 1, 256] - - [47, 0.0] - - - [6448, 3328, 1, 256] - - [53, 0.0] - - - [19968, 3329, 1, 256] - - [47, 0.0] - - - [11520, 8704, 1, 256] - - [53, 0.0] - - - [13824, 256, 1, 256] - - [47, 0.0] - - - [1584, 256, 1, 256] - - [58, 0.0] - - - [10496, 7168, 1, 256] - - [53, 0.0] - - - [5376, 2304, 1, 256] - - [53, 0.0] - - - [10752, 7168, 1, 256] - - [53, 0.0] - - - [18432, 2048, 1, 256] - - [53, 0.0] - - - [12080, 256, 1, 256] - - [144, 0.0] - - - [8704, 3328, 1, 256] - - [47, 0.0] - - - [4608, 1280, 1, 256] - - [53, 0.0] - - - [6192, 3328, 1, 256] - - [53, 0.0] - - - [8704, 3329, 1, 256] - - [47, 0.0] - - - [5424, 2560, 1, 256] - - [53, 0.0] - - - [11008, 2816, 1, 256] - - [47, 0.0] - - - [11776, 4352, 1, 256] - - [47, 0.0] - - - [11008, 1536, 1, 256] - - [53, 0.0] - - - [13312, 3328, 1, 256] - - [47, 0.0] - - - [7168, 4096, 1, 256] - - [53, 0.0] - - - [9216, 256, 1, 256] - - [64, 0.0] - - - [19504, 2865, 1, 256] - - [53, 0.0] - - - [5936, 2865, 1, 256] - - [53, 0.0] - - - [1840, 1792, 1, 256] - - [144, 0.0] - - - [19968, 2816, 1, 256] - - [47, 0.0] - - - [9008, 256, 1, 256] - - [68, 0.0] - - - [9728, 256, 1, 256] - - [47, 0.0] - - - [11056, 7936, 1, 256] - - [53, 0.0] - - - [7680, 3329, 1, 256] - - [47, 0.0] - - - [1792, 256, 1, 256] - - [58, 0.0] - - - [17664, 10240, 1, 256] - - [53, 0.0] - - - [11776, 2865, 1, 256] - - [47, 0.0] - - - [512, 512, 1, 256] - - [156, 0.0] - - - [16640, 768, 1, 256] - - [47, 0.0] - - - [4352, 2048, 1, 256] - - [53, 0.0] - - - [19504, 2816, 1, 256] - - [47, 0.0] - - - [12080, 2865, 1, 256] - - [53, 0.0] - - - [14080, 768, 1, 256] - - [47, 0.0] - - - [7936, 512, 1, 256] - - [69, 0.0] - - - [5376, 2560, 1, 256] - - [53, 0.0] - - - [5632, 3329, 1, 256] - - [47, 0.0] - - - [5120, 3840, 1, 256] - - [47, 0.0] - - - [6192, 2816, 1, 256] - - [53, 0.0] - - - [4608, 3072, 1, 256] - - [53, 0.0] - - - [19712, 6656, 1, 256] - - [53, 0.0] - - - [14896, 256, 1, 256] - - [53, 0.0] - - - [6400, 1280, 1, 256] - - [47, 0.0] - - - [12800, 9216, 1, 256] - - [53, 0.0] - - - [15616, 256, 1, 256] - - [53, 0.0] - - - [17920, 4608, 1, 256] - - [53, 0.0] - - - [7936, 2865, 1, 256] - - [47, 0.0] - - - [13312, 3329, 1, 256] - - [47, 0.0] - - - [5168, 2304, 1, 256] - - [53, 0.0] - - - [14128, 10240, 1, 256] - - [53, 0.0] - - - [3840, 2816, 1, 256] - - [47, 0.0] - - - [8960, 1536, 1, 256] - - [53, 0.0] - - - [3328, 3073, 1, 256] - - [53, 0.0] - - - [4096, 3328, 1, 256] - - [47, 0.0] - - - [14592, 2048, 1, 256] - - [53, 0.0] - - - [9728, 2048, 1, 256] - - [53, 0.0] - - - [13312, 5888, 1, 256] - - [47, 0.0] - - - [15616, 2304, 1, 256] - - [47, 0.0] - - - [19712, 2816, 1, 256] - - [47, 0.0] - - - [9216, 5888, 1, 256] - - [47, 0.0] - - - [7168, 4352, 1, 256] - - [47, 0.0] - - - [9520, 6400, 1, 256] - - [47, 0.0] - - - [13568, 3328, 1, 256] - - [47, 0.0] - - - [17408, 4352, 1, 256] - - [47, 0.0] - - - [11056, 2865, 1, 256] - - [47, 0.0] - - - [18480, 2865, 1, 256] - - [47, 0.0] - - - [13824, 768, 1, 256] - - [53, 0.0] - - - [17664, 6144, 1, 256] - - [53, 0.0] - - - [7216, 4352, 1, 256] - - [47, 0.0] - - - [14128, 2865, 1, 256] - - [53, 0.0] - - - [11520, 3328, 1, 256] - - [47, 0.0] - - - [18992, 5888, 1, 256] - - [47, 0.0] - - - [17408, 10240, 1, 256] - - [53, 0.0] - - - [15104, 6144, 1, 256] - - [53, 0.0] - - - [16640, 1280, 1, 256] - - [47, 0.0] - - - [13056, 2865, 1, 256] - - [47, 0.0] - - - [11776, 8960, 1, 256] - - [47, 0.0] - - - [11312, 2816, 1, 256] - - [47, 0.0] - - - [11264, 3328, 1, 256] - - [47, 0.0] - - - [8192, 512, 1, 256] - - [53, 0.0] - - - [14848, 6144, 1, 256] - - [53, 0.0] - - - [10496, 7680, 1, 256] - - [53, 0.0] - - - [2816, 2817, 1, 256] - - [45, 0.0] - - - [15104, 2865, 1, 256] - - [47, 0.0] - - - [18176, 3329, 1, 256] - - [47, 0.0] - - - [3328, 1792, 1, 256] - - [47, 0.0] - - - [6144, 3328, 1, 256] - - [47, 0.0] - - - [12288, 6144, 1, 256] - - [53, 0.0] - - - [8960, 5888, 1, 256] - - [47, 0.0] - - - [3584, 1280, 1, 256] - - [69, 0.0] - - - [7728, 4608, 1, 256] - - [53, 0.0] - - - [18176, 6144, 1, 256] - - [53, 0.0] - - - [16944, 256, 1, 256] - - [47, 0.0] - - - [3888, 768, 1, 256] - - [62, 0.0] - - - [8448, 3329, 1, 256] - - [47, 0.0] - - - [3072, 3073, 1, 256] - - [53, 0.0] - - - [4912, 256, 1, 256] - - [68, 0.0] - - - [5936, 3072, 1, 256] - - [53, 0.0] - - - [7168, 2865, 1, 256] - - [47, 0.0] - - - [19456, 10240, 1, 256] - - [53, 0.0] - - - [1840, 1585, 1, 256] - - [58, 0.0] - - - [18992, 5632, 1, 256] - - [53, 0.0] - - - [4912, 1792, 1, 256] - - [53, 0.0] - - - [8704, 6144, 1, 256] - - [53, 0.0] - - - [816, 768, 1, 256] - - [157, 0.0] - - - [18432, 2865, 1, 256] - - [47, 0.0] - - - [3120, 2865, 1, 256] - - [47, 0.0] - - - [6448, 2865, 1, 256] - - [47, 0.0] - - - [12080, 2816, 1, 256] - - [53, 0.0] - - - [10496, 3328, 1, 256] - - [53, 0.0] - - - [15920, 10240, 1, 256] - - [53, 0.0] - - - [15872, 2048, 1, 256] - - [53, 0.0] - - - [11568, 2816, 1, 256] - - [53, 0.0] - - - [19200, 10240, 1, 256] - - [53, 0.0] - - - [13312, 5632, 1, 256] - - [53, 0.0] - - - [15360, 2816, 1, 256] - - [47, 0.0] - - - [12288, 2865, 1, 256] - - [47, 0.0] - - - [19760, 6400, 1, 256] - - [53, 0.0] - - - [19968, 256, 1, 256] - - [53, 0.0] - - - [7680, 4352, 1, 256] - - [47, 0.0] - - - [11008, 3584, 1, 256] - - [53, 0.0] - - - [3072, 2817, 1, 256] - - [53, 0.0] - - - [11264, 6144, 1, 256] - - [53, 0.0] - - - [5424, 256, 1, 256] - - [144, 0.0] - - - [13568, 1280, 1, 256] - - [47, 0.0] - - - [3840, 2048, 1, 256] - - [53, 0.0] - - - [6144, 3072, 1, 256] - - [53, 0.0] - - - [19200, 1536, 1, 256] - - [53, 0.0] - - - [10240, 1280, 1, 256] - - [47, 0.0] - - - [3376, 512, 1, 256] - - [53, 0.0] - - - [12544, 1281, 1, 256] - - [50, 0.0] - - - [9776, 6656, 1, 256] - - [53, 0.0] - - - [7680, 2865, 1, 256] - - [47, 0.0] - - - [10544, 7680, 1, 256] - - [53, 0.0] - - - [15616, 3328, 1, 256] - - [47, 0.0] - - - [3328, 1280, 1, 256] - - [47, 0.0] - - - [2560, 1024, 1, 256] - - [69, 0.0] - - - [17456, 4096, 1, 256] - - [53, 0.0] - - - [6912, 3328, 1, 256] - - [47, 0.0] - - - [3584, 2816, 1, 256] - - [47, 0.0] - - - [17408, 3328, 1, 256] - - [47, 0.0] - - - [19200, 2816, 1, 256] - - [47, 0.0] - - - [15104, 1792, 1, 256] - - [47, 0.0] - - - [6144, 256, 1, 256] - - [68, 0.0] - - - [8192, 6144, 1, 256] - - [53, 0.0] - - - [12032, 4608, 1, 256] - - [53, 0.0] - - - [1840, 256, 1, 256] - - [58, 0.0] - - - [13312, 256, 1, 256] - - [64, 0.0] - - - [9216, 1792, 1, 256] - - [47, 0.0] - - - [14592, 3329, 1, 256] - - [47, 0.0] - - - [8448, 1280, 1, 256] - - [47, 0.0] - - - [11520, 8192, 1, 256] - - [53, 0.0] - - - [2608, 2560, 1, 256] - - [47, 0.0] - - - [5120, 2304, 1, 256] - - [47, 0.0] - - - [13056, 3328, 1, 256] - - [47, 0.0] - - - [11008, 8192, 1, 256] - - [47, 0.0] - - - [14896, 2865, 1, 256] - - [47, 0.0] - - - [6704, 3840, 1, 256] - - [47, 0.0] - - - [15872, 3329, 1, 256] - - [47, 0.0] - - - [7168, 1792, 1, 256] - - [47, 0.0] - - - [4656, 2865, 1, 256] - - [47, 0.0] - - - [18736, 5632, 1, 256] - - [53, 0.0] - - - [768, 512, 1, 256] - - [58, 0.0] - - - [16432, 3072, 1, 256] - - [53, 0.0] - - - [14848, 2865, 1, 256] - - [47, 0.0] - - - [4864, 1536, 1, 256] - - [53, 0.0] - - - [9472, 2865, 1, 256] - - [47, 0.0] - - - [10496, 2048, 1, 256] - - [53, 0.0] - - - [14336, 1024, 1, 256] - - [53, 0.0] - - - [18432, 256, 1, 256] - - [47, 0.0] - - - [16896, 3840, 1, 256] - - [47, 0.0] - - - [10240, 512, 1, 256] - - [53, 0.0] - - - [15664, 2304, 1, 256] - - [53, 0.0] - - - [10496, 3329, 1, 256] - - [47, 0.0] - - - [19456, 1536, 1, 256] - - [53, 0.0] - - - [17920, 1281, 1, 256] - - [50, 0.0] - - - [8960, 256, 1, 256] - - [67, 0.0] - - - [10496, 768, 1, 256] - - [47, 0.0] - - - [5120, 2816, 1, 256] - - [47, 0.0] - - - [12288, 2048, 1, 256] - - [53, 0.0] - - - [11568, 8704, 1, 256] - - [47, 0.0] - - - [10496, 1024, 1, 256] - - [53, 0.0] - - - [10288, 256, 1, 256] - - [144, 0.0] - - - [5168, 2048, 1, 256] - - [53, 0.0] - - - [11776, 3328, 1, 256] - - [47, 0.0] - - - [15152, 10240, 1, 256] - - [53, 0.0] - - - [14384, 2865, 1, 256] - - [47, 0.0] - - - [12288, 512, 1, 256] - - [47, 0.0] - - - [16688, 256, 1, 256] - - [47, 0.0] - - - [6912, 4096, 1, 256] - - [53, 0.0] - - - [4864, 2048, 1, 256] - - [53, 0.0] - - - [4096, 1024, 1, 256] - - [53, 0.0] - - - [12848, 9984, 1, 256] - - [53, 0.0] - - - [16896, 1281, 1, 256] - - [45, 0.0] - - - [768, 561, 1, 256] - - [58, 0.0] - - - [16896, 3584, 1, 256] - - [53, 0.0] - - - [14592, 6144, 1, 256] - - [53, 0.0] - - - [17664, 4096, 1, 256] - - [53, 0.0] - - - [8448, 2865, 1, 256] - - [47, 0.0] - - - [18432, 768, 1, 256] - - [47, 0.0] - - - [12032, 512, 1, 256] - - [53, 0.0] - - - [11008, 256, 1, 256] - - [53, 0.0] - - - [15360, 1536, 1, 256] - - [53, 0.0] - - - [5888, 2048, 1, 256] - - [53, 0.0] - - - [13104, 256, 1, 256] - - [144, 0.0] - - - [11264, 7680, 1, 256] - - [53, 0.0] - - - [19248, 2865, 1, 256] - - [53, 0.0] - - - [17200, 2865, 1, 256] - - [47, 0.0] - - - [8192, 2048, 1, 256] - - [53, 0.0] - - - [7472, 4608, 1, 256] - - [53, 0.0] - - - [7168, 2048, 1, 256] - - [53, 0.0] - - - [13360, 2816, 1, 256] - - [53, 0.0] - - - [17920, 4352, 1, 256] - - [47, 0.0] - - - [15408, 256, 1, 256] - - [47, 0.0] - - - [19200, 1281, 1, 256] - - [50, 0.0] - - - [15360, 256, 1, 256] - - [47, 0.0] - - - [9984, 6400, 1, 256] - - [47, 0.0] - - - [18944, 2865, 1, 256] - - [47, 0.0] - - - [3840, 2865, 1, 256] - - [47, 0.0] - - - [8192, 3328, 1, 256] - - [47, 0.0] - - - [5888, 256, 1, 256] - - [62, 0.0] - - - [15616, 2816, 1, 256] - - [47, 0.0] - - - [17664, 2865, 1, 256] - - [47, 0.0] - - - [14592, 768, 1, 256] - - [47, 0.0] - - - [18944, 1281, 1, 256] - - [50, 0.0] - - - [11264, 1536, 1, 256] - - [53, 0.0] - - - [8496, 5632, 1, 256] - - [53, 0.0] - - - [17664, 3328, 1, 256] - - [47, 0.0] - - - [14848, 2048, 1, 256] - - [53, 0.0] - - - [15408, 2865, 1, 256] - - [47, 0.0] - - - [4096, 2048, 1, 256] - - [53, 0.0] - - - [14128, 1024, 1, 256] - - [53, 0.0] - - - [1072, 817, 1, 256] - - [151, 0.0] - - - [17152, 3840, 1, 256] - - [47, 0.0] - - - [17664, 2048, 1, 256] - - [53, 0.0] - - - [16896, 256, 1, 256] - - [47, 0.0] - - - [2304, 2097, 1, 256] - - [59, 0.0] - - - [5888, 2560, 1, 256] - - [53, 0.0] - - - [9472, 1792, 1, 256] - - [47, 0.0] - - - [1328, 1280, 1, 256] - - [68, 0.0] - - - [19200, 1280, 1, 256] - - [47, 0.0] - - - [12544, 1280, 1, 256] - - [47, 0.0] - - - [16432, 3328, 1, 256] - - [47, 0.0] - - - [17920, 1280, 1, 256] - - [47, 0.0] - - - [8752, 5632, 1, 256] - - [53, 0.0] - - - [7936, 2048, 1, 256] - - [53, 0.0] - - - [9472, 1280, 1, 256] - - [53, 0.0] - - - [16896, 1024, 1, 256] - - [53, 0.0] - - - [6656, 3329, 1, 256] - - [47, 0.0] - - - [17456, 2865, 1, 256] - - [47, 0.0] - - - [5632, 2304, 1, 256] - - [47, 0.0] - - - [14080, 1024, 1, 256] - - [53, 0.0] - - - [15872, 3328, 1, 256] - - [47, 0.0] - - - [5168, 2816, 1, 256] - - [53, 0.0] - - - [13312, 9728, 1, 256] - - [53, 0.0] - - - [1584, 1329, 1, 256] - - [47, 0.0] - - - [15664, 2560, 1, 256] - - [53, 0.0] - - - [2048, 768, 1, 256] - - [59, 0.0] - - - [17712, 2816, 1, 256] - - [53, 0.0] - - - [16128, 2865, 1, 256] - - [47, 0.0] - - - [15872, 2816, 1, 256] - - [47, 0.0] - - - [18224, 2816, 1, 256] - - [53, 0.0] - - - [5632, 4352, 1, 256] - - [47, 0.0] - - - [1792, 1281, 1, 256] - - [69, 0.0] - - - [6656, 2816, 1, 256] - - [47, 0.0] - - - [16640, 1281, 1, 256] - - [50, 0.0] - - - [13056, 10240, 1, 256] - - [53, 0.0] - - - [17968, 256, 1, 256] - - [47, 0.0] - - - [5376, 4096, 1, 256] - - [53, 0.0] - - - [15152, 2048, 1, 256] - - [53, 0.0] - - - [13568, 2816, 1, 256] - - [53, 0.0] - - - [12800, 2816, 1, 256] - - [47, 0.0] - - - [6960, 2816, 1, 256] - - [53, 0.0] - - - [17968, 4608, 1, 256] - - [53, 0.0] - - - [15104, 3328, 1, 256] - - [47, 0.0] - - - [7472, 4352, 1, 256] - - [47, 0.0] - - - [15872, 2304, 1, 256] - - [47, 0.0] - - - [4400, 2816, 1, 256] - - [53, 0.0] - - - [16128, 6144, 1, 256] - - [53, 0.0] - - - [18944, 2816, 1, 256] - - [47, 0.0] - - - [5424, 2865, 1, 256] - - [47, 0.0] - - - [8192, 768, 1, 256] - - [47, 0.0] - - - [12848, 256, 1, 256] - - [47, 0.0] - - - [12288, 1281, 1, 256] - - [45, 0.0] - - - [13872, 512, 1, 256] - - [53, 0.0] - - - [5888, 1280, 1, 256] - - [47, 0.0] - - - [2816, 1281, 1, 256] - - [150, 0.0] - - - [19200, 3329, 1, 256] - - [47, 0.0] - - - [12800, 3328, 1, 256] - - [47, 0.0] - - - [15360, 2865, 1, 256] - - [47, 0.0] - - - [17152, 3584, 1, 256] - - [53, 0.0] - - - [17456, 2816, 1, 256] - - [47, 0.0] - - - [18176, 10240, 1, 256] - - [53, 0.0] - - - [6144, 2816, 1, 256] - - [47, 0.0] - - - [18176, 1280, 1, 256] - - [53, 0.0] - - - [16384, 1281, 1, 256] - - [50, 0.0] - - - [9216, 3328, 1, 256] - - [47, 0.0] - - - [14080, 2816, 1, 256] - - [47, 0.0] - - - [18688, 2816, 1, 256] - - [47, 0.0] - - - [15872, 10240, 1, 256] - - [53, 0.0] - - - [10800, 7936, 1, 256] - - [53, 0.0] - - - [9984, 1281, 1, 256] - - [45, 0.0] - - - [4144, 256, 1, 256] - - [153, 0.0] - - - [16640, 6144, 1, 256] - - [53, 0.0] - - - [11776, 4096, 1, 256] - - [53, 0.0] - - - [11056, 8192, 1, 256] - - [53, 0.0] - - - [5376, 2816, 1, 256] - - [47, 0.0] - - - [19712, 1280, 1, 256] - - [47, 0.0] - - - [4608, 2816, 1, 256] - - [53, 0.0] - - - [19456, 2865, 1, 256] - - [47, 0.0] - - - [14080, 256, 1, 256] - - [47, 0.0] - - - [7216, 4096, 1, 256] - - [53, 0.0] - - - [2816, 1280, 1, 256] - - [47, 0.0] - - - [10496, 3072, 1, 256] - - [53, 0.0] - - - [12544, 4864, 1, 256] - - [47, 0.0] - - - [9984, 6912, 1, 256] - - [47, 0.0] - - - [4912, 2048, 1, 256] - - [53, 0.0] - - - [9984, 2304, 1, 256] - - [47, 0.0] - - - [19248, 5888, 1, 256] - - [47, 0.0] - - - [19712, 2048, 1, 256] - - [53, 0.0] - - - [9728, 2816, 1, 256] - - [47, 0.0] - - - [19504, 6400, 1, 256] - - [53, 0.0] - - - [16896, 3072, 1, 256] - - [53, 0.0] - - - [15104, 1536, 1, 256] - - [53, 0.0] - - - [2608, 2609, 1, 256] - - [53, 0.0] - - - [14384, 256, 1, 256] - - [144, 0.0] - - - [17664, 2816, 1, 256] - - [47, 0.0] - - - [9776, 6912, 1, 256] - - [47, 0.0] - - - [1792, 512, 1, 256] - - [42, 0.0] - - - [13312, 1536, 1, 256] - - [53, 0.0] - - - [1072, 1073, 1, 256] - - [56, 0.0] - - - [9472, 2048, 1, 256] - - [53, 0.0] - - - [4608, 3328, 1, 256] - - [47, 0.0] - - - [9984, 2560, 1, 256] - - [53, 0.0] - - - [6912, 5376, 1, 256] - - [47, 0.0] - - - [16640, 2865, 1, 256] - - [47, 0.0] - - - [4352, 3072, 1, 256] - - [53, 0.0] - - - [5632, 3328, 1, 256] - - [47, 0.0] - - - [9216, 5632, 1, 256] - - [53, 0.0] - - - [3328, 3329, 1, 256] - - [47, 0.0] - - - [13824, 10240, 1, 256] - - [53, 0.0] - - - [12288, 3329, 1, 256] - - [47, 0.0] - - - [2864, 256, 1, 256] - - [154, 0.0] - - - [19712, 1792, 1, 256] - - [47, 0.0] - - - [6656, 1280, 1, 256] - - [47, 0.0] - - - [13056, 2048, 1, 256] - - [53, 0.0] - - - [6912, 1281, 1, 256] - - [44, 0.0] - - - [16176, 2816, 1, 256] - - [53, 0.0] - - - [14592, 3328, 1, 256] - - [47, 0.0] - - - [10496, 2865, 1, 256] - - [47, 0.0] - - - [9728, 6400, 1, 256] - - [47, 0.0] - - - [5888, 4608, 1, 256] - - [53, 0.0] - - - [16432, 10240, 1, 256] - - [53, 0.0] - - - [19456, 5888, 1, 256] - - [47, 0.0] - - - [3888, 256, 1, 256] - - [53, 0.0] - - - [12336, 2816, 1, 256] - - [47, 0.0] - - - [19456, 256, 1, 256] - - [47, 0.0] - - - [14384, 2816, 1, 256] - - [47, 0.0] - - - [14384, 1024, 1, 256] - - [53, 0.0] - - - [16640, 2816, 1, 256] - - [53, 0.0] - - - [3840, 3329, 1, 256] - - [47, 0.0] - - - [2304, 2304, 1, 256] - - [47, 0.0] - - - [10240, 2816, 1, 256] - - [47, 0.0] - - - [13104, 10240, 1, 256] - - [53, 0.0] - - - [1536, 256, 1, 256] - - [58, 0.0] - - - [11008, 2865, 1, 256] - - [47, 0.0] - - - [13104, 9984, 1, 256] - - [47, 0.0] - - - [10240, 7168, 1, 256] - - [53, 0.0] - - - [3888, 2865, 1, 256] - - [53, 0.0] - - - [8192, 4864, 1, 256] - - [47, 0.0] - - - [15920, 256, 1, 256] - - [62, 0.0] - - - [6448, 3584, 1, 256] - - [53, 0.0] - - - [16128, 2304, 1, 256] - - [47, 0.0] - - - [9728, 2865, 1, 256] - - [53, 0.0] - - - [6144, 4864, 1, 256] - - [47, 0.0] - - - [14848, 256, 1, 256] - - [53, 0.0] - - - [4352, 1024, 1, 256] - - [53, 0.0] - - - [15360, 10240, 1, 256] - - [53, 0.0] - - - [19504, 10240, 1, 256] - - [53, 0.0] - - - [3328, 3072, 1, 256] - - [53, 0.0] - - - [1536, 1281, 1, 256] - - [49, 0.0] - - - [19760, 6656, 1, 256] - - [53, 0.0] - - - [3584, 3329, 1, 256] - - [47, 0.0] - - - [14848, 2816, 1, 256] - - [47, 0.0] - - - [4400, 2865, 1, 256] - - [53, 0.0] - - - [3888, 1024, 1, 256] - - [53, 0.0] - - - [16640, 2048, 1, 256] - - [53, 0.0] - - - [4096, 2816, 1, 256] - - [47, 0.0] - - - [14640, 2816, 1, 256] - - [47, 0.0] - - - [9472, 1281, 1, 256] - - [45, 0.0] - - - [8192, 1280, 1, 256] - - [47, 0.0] - - - [8960, 2865, 1, 256] - - [47, 0.0] - - - [4144, 2816, 1, 256] - - [53, 0.0] - - - [10288, 7168, 1, 256] - - [53, 0.0] - - - [14592, 256, 1, 256] - - [53, 0.0] - - - [10240, 2048, 1, 256] - - [53, 0.0] - - - [17920, 2865, 1, 256] - - [47, 0.0] - - - [12592, 2816, 1, 256] - - [53, 0.0] - - - [14592, 1536, 1, 256] - - [53, 0.0] - - - [11568, 256, 1, 256] - - [144, 0.0] - - - [6704, 3584, 1, 256] - - [53, 0.0] - - - [5120, 3328, 1, 256] - - [47, 0.0] - - - [4400, 1536, 1, 256] - - [53, 0.0] - - - [18944, 256, 1, 256] - - [69, 0.0] - - - [19712, 5888, 1, 256] - - [47, 0.0] - - - [7984, 5120, 1, 256] - - [53, 0.0] - - - [8240, 2865, 1, 256] - - [53, 0.0] - - - [6144, 1280, 1, 256] - - [47, 0.0] - - - [8496, 2816, 1, 256] - - [47, 0.0] - - - [14592, 1024, 1, 256] - - [53, 0.0] - - - [14592, 2865, 1, 256] - - [47, 0.0] - - - [13360, 256, 1, 256] - - [144, 0.0] - - - [8448, 256, 1, 256] - - [47, 0.0] - - - [16896, 2816, 1, 256] - - [47, 0.0] - - - [15152, 2865, 1, 256] - - [53, 0.0] - - - [11056, 2816, 1, 256] - - [47, 0.0] - - - [15616, 1280, 1, 256] - - [47, 0.0] - - - [8192, 5120, 1, 256] - - [53, 0.0] - - - [17408, 256, 1, 256] - - [144, 0.0] - - - [18432, 10240, 1, 256] - - [53, 0.0] - - - [14592, 1280, 1, 256] - - [47, 0.0] - - - [3328, 512, 1, 256] - - [69, 0.0] - - - [14336, 1280, 1, 256] - - [47, 0.0] - - - [13616, 2865, 1, 256] - - [53, 0.0] - - - [8192, 256, 1, 256] - - [53, 0.0] - - - [10240, 1281, 1, 256] - - [45, 0.0] - - - [1840, 1841, 1, 256] - - [47, 0.0] - - - [12800, 9472, 1, 256] - - [47, 0.0] - - - [17664, 256, 1, 256] - - [53, 0.0] - - - [768, 769, 1, 256] - - [149, 0.0] - - - [19456, 2048, 1, 256] - - [53, 0.0] - - - [13056, 3329, 1, 256] - - [47, 0.0] - - - [11056, 256, 1, 256] - - [47, 0.0] - - - [7424, 6144, 1, 256] - - [53, 0.0] - - - [14848, 3328, 1, 256] - - [47, 0.0] - - - [6656, 3328, 1, 256] - - [47, 0.0] - - - [10752, 1281, 1, 256] - - [45, 0.0] - - - [9984, 2865, 1, 256] - - [47, 0.0] - - - [14080, 3329, 1, 256] - - [47, 0.0] - - - [17920, 3328, 1, 256] - - [47, 0.0] - - - [13312, 10240, 1, 256] - - [53, 0.0] - - - [16640, 3584, 1, 256] - - [53, 0.0] - - - [17408, 3840, 1, 256] - - [47, 0.0] - - - [12032, 8960, 1, 256] - - [47, 0.0] - - - [10800, 2865, 1, 256] - - [53, 0.0] - - - [3072, 2816, 1, 256] - - [47, 0.0] - - - [14128, 2816, 1, 256] - - [47, 0.0] - - - [11312, 8192, 1, 256] - - [53, 0.0] - - - [2560, 2305, 1, 256] - - [47, 0.0] - - - [16640, 3072, 1, 256] - - [53, 0.0] - - - [16128, 2048, 1, 256] - - [53, 0.0] - - - [6144, 512, 1, 256] - - [69, 0.0] - - - [18688, 4864, 1, 256] - - [47, 0.0] - - - [17200, 256, 1, 256] - - [47, 0.0] - - - [8752, 2865, 1, 256] - - [53, 0.0] - - - [18944, 1280, 1, 256] - - [47, 0.0] - - - [16640, 3328, 1, 256] - - [47, 0.0] - - - [304, 305, 1, 256] - - [60, 0.0] - - - [15104, 256, 1, 256] - - [53, 0.0] - - - [7680, 3328, 1, 256] - - [47, 0.0] - - - [12336, 9216, 1, 256] - - [53, 0.0] - - - [14080, 6144, 1, 256] - - [53, 0.0] - - - [7168, 5888, 1, 256] - - [53, 0.0] - - - [7424, 1280, 1, 256] - - [53, 0.0] - - - [4864, 3584, 1, 256] - - [53, 0.0] - - - [1280, 1025, 1, 256] - - [152, 0.0] - - - [10240, 2865, 1, 256] - - [47, 0.0] - - - [18480, 10240, 1, 256] - - [53, 0.0] - - - [7680, 256, 1, 256] - - [64, 0.0] - - - [9472, 6656, 1, 256] - - [53, 0.0] - - - [12032, 6144, 1, 256] - - [53, 0.0] - - - [5120, 3329, 1, 256] - - [47, 0.0] - - - [10752, 256, 1, 256] - - [47, 0.0] - - - [6960, 256, 1, 256] - - [47, 0.0] - - - [9008, 6144, 1, 256] - - [53, 0.0] - - - [7424, 2048, 1, 256] - - [53, 0.0] - - - [5632, 1280, 1, 256] - - [53, 0.0] - - - [19712, 10240, 1, 256] - - [53, 0.0] - - - [6400, 768, 1, 256] - - [69, 0.0] - - - [10752, 3328, 1, 256] - - [47, 0.0] - - - [18432, 5376, 1, 256] - - [47, 0.0] - - - [9520, 256, 1, 256] - - [69, 0.0] - - - [5680, 2816, 1, 256] - - [47, 0.0] - - - [11008, 3329, 1, 256] - - [47, 0.0] - - - [4608, 2865, 1, 256] - - [47, 0.0] - - - [6448, 256, 1, 256] - - [69, 0.0] - - - [3584, 256, 1, 256] - - [40, 0.0] - - - [12336, 9472, 1, 256] - - [53, 0.0] - - - [1280, 1281, 1, 256] - - [57, 0.0] - - - [7936, 6144, 1, 256] - - [53, 0.0] - - - [15152, 1792, 1, 256] - - [47, 0.0] - - - [4352, 1281, 1, 256] - - [53, 0.0] - - - [12848, 2865, 1, 256] - - [53, 0.0] - - - [16944, 3584, 1, 256] - - [53, 0.0] - - - [8752, 256, 1, 256] - - [68, 0.0] - - - [6912, 256, 1, 256] - - [53, 0.0] - - - [14336, 1281, 1, 256] - - [45, 0.0] - - - [2304, 2049, 1, 256] - - [69, 0.0] - - - [9216, 6144, 1, 256] - - [53, 0.0] - - - [1072, 1024, 1, 256] - - [66, 0.0] - - - [10752, 6144, 1, 256] - - [53, 0.0] - - - [1792, 1537, 1, 256] - - [53, 0.0] - - - [17968, 2865, 1, 256] - - [47, 0.0] - - - [8448, 4864, 1, 256] - - [47, 0.0] - - - [15408, 2048, 1, 256] - - [53, 0.0] - - - [15104, 1280, 1, 256] - - [47, 0.0] - - - [9264, 2865, 1, 256] - - [53, 0.0] - - - [15616, 2865, 1, 256] - - [47, 0.0] - - - [16896, 10240, 1, 256] - - [53, 0.0] - - - [15104, 2816, 1, 256] - - [47, 0.0] - - - [13872, 2865, 1, 256] - - [47, 0.0] - - - [13056, 1280, 1, 256] - - [47, 0.0] - - - [12288, 3328, 1, 256] - - [47, 0.0] - - - [3840, 3328, 1, 256] - - [53, 0.0] - - - [9216, 1280, 1, 256] - - [47, 0.0] - - - [8448, 1281, 1, 256] - - [50, 0.0] - - - [2560, 256, 1, 256] - - [43, 0.0] - - - [4608, 1281, 1, 256] - - [53, 0.0] - - - [6144, 2865, 1, 256] - - [47, 0.0] - - - [5888, 2816, 1, 256] - - [47, 0.0] - - - [3584, 1281, 1, 256] - - [50, 0.0] - - - [18688, 5120, 1, 256] - - [53, 0.0] - - - [12288, 2816, 1, 256] - - [53, 0.0] - - - [4864, 2865, 1, 256] - - [47, 0.0] - - - [9216, 2048, 1, 256] - - [53, 0.0] - - - [13872, 768, 1, 256] - - [47, 0.0] - - - [10496, 7424, 1, 256] - - [47, 0.0] - - - [16384, 512, 1, 256] - - [53, 0.0] - - - [14848, 10240, 1, 256] - - [53, 0.0] - - - [17920, 2048, 1, 256] - - [53, 0.0] - - - [11008, 7936, 1, 256] - - [47, 0.0] - - - [1792, 1792, 1, 256] - - [53, 0.0] - - - [7680, 4864, 1, 256] - - [47, 0.0] - - - [19760, 256, 1, 256] - - [53, 0.0] - - - [15616, 1792, 1, 256] - - [47, 0.0] - - - [1792, 1793, 1, 256] - - [50, 0.0] - - - [8192, 3329, 1, 256] - - [47, 0.0] - - - [2560, 1280, 1, 256] - - [47, 0.0] - - - [1328, 1073, 1, 256] - - [56, 0.0] - - - [16896, 2865, 1, 256] - - [47, 0.0] - - - [8960, 1280, 1, 256] - - [47, 0.0] - - - [6960, 2865, 1, 256] - - [53, 0.0] - - - [1280, 1024, 1, 256] - - [59, 0.0] - - - [6400, 2048, 1, 256] - - [53, 0.0] - - - [18480, 5376, 1, 256] - - [47, 0.0] - - - [18944, 2048, 1, 256] - - [53, 0.0] - - - [9520, 6656, 1, 256] - - [53, 0.0] - - - [4352, 1536, 1, 256] - - [47, 0.0] - - - [19712, 6144, 1, 256] - - [53, 0.0] - - - [6400, 2816, 1, 256] - - [47, 0.0] - - - [1792, 1585, 1, 256] - - [47, 0.0] - - - [13312, 6144, 1, 256] - - [53, 0.0] - - - [17408, 4096, 1, 256] - - [53, 0.0] - - - [16128, 256, 1, 256] - - [69, 0.0] - - - [15104, 2048, 1, 256] - - [53, 0.0] - - - [8704, 2865, 1, 256] - - [47, 0.0] - - - [6144, 768, 1, 256] - - [47, 0.0] - - - [10496, 1280, 1, 256] - - [47, 0.0] - - - [816, 561, 1, 256] - - [62, 0.0] - - - [6912, 3840, 1, 256] - - [47, 0.0] - - - [8704, 1281, 1, 256] - - [49, 0.0] - - - [13312, 1792, 1, 256] - - [53, 0.0] - - - [5120, 1281, 1, 256] - - [50, 0.0] - - - [10496, 1281, 1, 256] - - [45, 0.0] - - - [8448, 6144, 1, 256] - - [53, 0.0] - - - [2560, 2353, 1, 256] - - [47, 0.0] - - - [4352, 1280, 1, 256] - - [69, 0.0] - - - [12336, 256, 1, 256] - - [53, 0.0] - - - [21504, 10240, 1, 256] - - [53, 0.0] - - - [31744, 6144, 1, 256] - - [53, 0.0] - - - [27648, 1280, 1, 256] - - [47, 0.0] - - - [22272, 512, 1, 256] - - [53, 0.0] - - - [29184, 256, 1, 256] - - [53, 0.0] - - - [23808, 4096, 1, 256] - - [53, 0.0] - - - [30720, 7168, 1, 256] - - [53, 0.0] - - - [29440, 2865, 1, 256] - - [47, 0.0] - - - [25600, 5632, 1, 256] - - [53, 0.0] - - - [24832, 10240, 1, 256] - - [53, 0.0] - - - [22784, 2865, 1, 256] - - [47, 0.0] - - - [24368, 768, 1, 256] - - [53, 0.0] - - - [21760, 8192, 1, 256] - - [53, 0.0] - - - [29184, 10240, 1, 256] - - [53, 0.0] - - - [26368, 6144, 1, 256] - - [53, 0.0] - - - [23088, 10240, 1, 256] - - [53, 0.0] - - - [29952, 4096, 1, 256] - - [53, 0.0] - - - [24320, 6144, 1, 256] - - [53, 0.0] - - - [32256, 2048, 1, 256] - - [53, 0.0] - - - [29488, 2865, 1, 256] - - [47, 0.0] - - - [21808, 2816, 1, 256] - - [47, 0.0] - - - [32000, 7936, 1, 256] - - [47, 0.0] - - - [23040, 10240, 1, 256] - - [53, 0.0] - - - [31792, 10240, 1, 256] - - [53, 0.0] - - - [24320, 1281, 1, 256] - - [45, 0.0] - - - [27136, 3072, 1, 256] - - [53, 0.0] - - - [31488, 6144, 1, 256] - - [53, 0.0] - - - [34096, 2865, 1, 256] - - [53, 0.0] - - - [33024, 8960, 1, 256] - - [47, 0.0] - - - [28928, 1280, 1, 256] - - [47, 0.0] - - - [31488, 5632, 1, 256] - - [53, 0.0] - - - [27696, 4096, 1, 256] - - [53, 0.0] - - - [31488, 10240, 1, 256] - - [53, 0.0] - - - [28928, 10240, 1, 256] - - [53, 0.0] - - - [26160, 10240, 1, 256] - - [53, 0.0] - - - [26112, 3328, 1, 256] - - [47, 0.0] - - - [28928, 4864, 1, 256] - - [47, 0.0] - - - [27904, 3328, 1, 256] - - [47, 0.0] - - - [29184, 5376, 1, 256] - - [47, 0.0] - - - [29952, 1281, 1, 256] - - [45, 0.0] - - - [24832, 6144, 1, 256] - - [53, 0.0] - - - [28160, 4096, 1, 256] - - [53, 0.0] - - - [24320, 1280, 1, 256] - - [47, 0.0] - - - [34816, 768, 1, 256] - - [47, 0.0] - - - [34816, 1281, 1, 256] - - [45, 0.0] - - - [27136, 2816, 1, 256] - - [47, 0.0] - - - [32256, 8192, 1, 256] - - [53, 0.0] - - - [26624, 2865, 1, 256] - - [47, 0.0] - - - [23808, 3840, 1, 256] - - [47, 0.0] - - - [29440, 5376, 1, 256] - - [47, 0.0] - - - [30464, 10240, 1, 256] - - [53, 0.0] - - - [29232, 10240, 1, 256] - - [53, 0.0] - - - [27136, 1280, 1, 256] - - [47, 0.0] - - - [27904, 6144, 1, 256] - - [53, 0.0] - - - [33024, 2816, 1, 256] - - [47, 0.0] - - - [34816, 3329, 1, 256] - - [47, 0.0] - - - [34048, 1792, 1, 256] - - [47, 0.0] - - - [21248, 7424, 1, 256] - - [47, 0.0] - - - [29952, 256, 1, 256] - - [47, 0.0] - - - [34560, 256, 1, 256] - - [53, 0.0] - - - [26368, 3072, 1, 256] - - [53, 0.0] - - - [23600, 2865, 1, 256] - - [47, 0.0] - - - [30720, 512, 1, 256] - - [53, 0.0] - - - [30768, 10240, 1, 256] - - [53, 0.0] - - - [28928, 1024, 1, 256] - - [53, 0.0] - - - [26624, 256, 1, 256] - - [53, 0.0] - - - [26928, 10240, 1, 256] - - [53, 0.0] - - - [21248, 7936, 1, 256] - - [47, 0.0] - - - [34304, 2816, 1, 256] - - [47, 0.0] - - - [29696, 3840, 1, 256] - - [47, 0.0] - - - [27696, 10240, 1, 256] - - [53, 0.0] - - - [24064, 2048, 1, 256] - - [53, 0.0] - - - [33536, 6144, 1, 256] - - [53, 0.0] - - - [32512, 8704, 1, 256] - - [53, 0.0] - - - [21552, 2816, 1, 256] - - [47, 0.0] - - - [27648, 10240, 1, 256] - - [53, 0.0] - - - [22272, 2048, 1, 256] - - [53, 0.0] - - - [28976, 5632, 1, 256] - - [53, 0.0] - - - [30720, 10240, 1, 256] - - [53, 0.0] - - - [26112, 2816, 1, 256] - - [53, 0.0] - - - [20528, 10240, 1, 256] - - [53, 0.0] - - - [29696, 1536, 1, 256] - - [53, 0.0] - - - [31536, 2865, 1, 256] - - [53, 0.0] - - - [32000, 3328, 1, 256] - - [47, 0.0] - - - [20784, 2865, 1, 256] - - [53, 0.0] - - - [33280, 9984, 1, 256] - - [47, 0.0] - - - [25600, 3329, 1, 256] - - [47, 0.0] - - - [27904, 4096, 1, 256] - - [53, 0.0] - - - [29488, 256, 1, 256] - - [53, 0.0] - - - [32048, 10240, 1, 256] - - [53, 0.0] - - - [31280, 2865, 1, 256] - - [53, 0.0] - - - [32816, 2816, 1, 256] - - [53, 0.0] - - - [34096, 2816, 1, 256] - - [47, 0.0] - - - [20992, 3328, 1, 256] - - [47, 0.0] - - - [32768, 1281, 1, 256] - - [45, 0.0] - - - [24576, 4864, 1, 256] - - [47, 0.0] - - - [30464, 3328, 1, 256] - - [47, 0.0] - - - [28208, 256, 1, 256] - - [47, 0.0] - - - [23552, 1280, 1, 256] - - [47, 0.0] - - - [20528, 7168, 1, 256] - - [53, 0.0] - - - [34560, 2865, 1, 256] - - [47, 0.0] - - - [20736, 2816, 1, 256] - - [47, 0.0] - - - [26880, 3328, 1, 256] - - [47, 0.0] - - - [31536, 8192, 1, 256] - - [53, 0.0] - - - [31744, 8448, 1, 256] - - [47, 0.0] - - - [20224, 2865, 1, 256] - - [47, 0.0] - - - [22528, 2048, 1, 256] - - [53, 0.0] - - - [24320, 2048, 1, 256] - - [53, 0.0] - - - [32512, 8960, 1, 256] - - [47, 0.0] - - - [33072, 10240, 1, 256] - - [53, 0.0] - - - [24880, 10240, 1, 256] - - [53, 0.0] - - - [21040, 7680, 1, 256] - - [53, 0.0] - - - [26368, 10240, 1, 256] - - [53, 0.0] - - - [32304, 8704, 1, 256] - - [53, 0.0] - - - [33536, 1281, 1, 256] - - [50, 0.0] - - - [27136, 1024, 1, 256] - - [53, 0.0] - - - [33792, 1281, 1, 256] - - [45, 0.0] - - - [33584, 256, 1, 256] - - [47, 0.0] - - - [20528, 7424, 1, 256] - - [47, 0.0] - - - [28928, 2865, 1, 256] - - [47, 0.0] - - - [22016, 2048, 1, 256] - - [53, 0.0] - - - [29440, 3328, 1, 256] - - [47, 0.0] - - - [30208, 2048, 1, 256] - - [53, 0.0] - - - [20480, 2816, 1, 256] - - [47, 0.0] - - - [25904, 256, 1, 256] - - [47, 0.0] - - - [20736, 10240, 1, 256] - - [53, 0.0] - - - [32816, 256, 1, 256] - - [47, 0.0] - - - [33792, 3328, 1, 256] - - [47, 0.0] - - - [22272, 1281, 1, 256] - - [45, 0.0] - - - [25600, 1280, 1, 256] - - [47, 0.0] - - - [33280, 3329, 1, 256] - - [47, 0.0] - - - [22784, 1281, 1, 256] - - [45, 0.0] - - - [25392, 2816, 1, 256] - - [53, 0.0] - - - [33280, 3328, 1, 256] - - [47, 0.0] - - - [21760, 1280, 1, 256] - - [47, 0.0] - - - [33024, 768, 1, 256] - - [53, 0.0] - - - [25088, 1792, 1, 256] - - [47, 0.0] - - - [26368, 3329, 1, 256] - - [47, 0.0] - - - [34560, 3328, 1, 256] - - [47, 0.0] - - - [23040, 6144, 1, 256] - - [53, 0.0] - - - [30464, 2048, 1, 256] - - [53, 0.0] - - - [28672, 3328, 1, 256] - - [47, 0.0] - - - [30464, 6912, 1, 256] - - [47, 0.0] - - - [32048, 2816, 1, 256] - - [47, 0.0] - - - [33792, 9728, 1, 256] - - [53, 0.0] - - - [27392, 1536, 1, 256] - - [53, 0.0] - - - [24112, 512, 1, 256] - - [53, 0.0] - - - [28160, 256, 1, 256] - - [47, 0.0] - - - [34816, 2048, 1, 256] - - [53, 0.0] - - - [25648, 10240, 1, 256] - - [53, 0.0] - - - [20992, 10240, 1, 256] - - [53, 0.0] - - - [22528, 1281, 1, 256] - - [45, 0.0] - - - [25904, 2304, 1, 256] - - [53, 0.0] - - - [27952, 2865, 1, 256] - - [53, 0.0] - - - [30976, 768, 1, 256] - - [47, 0.0] - - - [20480, 3329, 1, 256] - - [47, 0.0] - - - [33072, 256, 1, 256] - - [47, 0.0] - - - [26624, 2560, 1, 256] - - [53, 0.0] - - - [28208, 2865, 1, 256] - - [47, 0.0] - - - [26672, 3328, 1, 256] - - [47, 0.0] - - - [26880, 2865, 1, 256] - - [47, 0.0] - - - [26112, 2304, 1, 256] - - [47, 0.0] - - - [29184, 5120, 1, 256] - - [53, 0.0] - - - [29744, 6144, 1, 256] - - [53, 0.0] - - - [30464, 3329, 1, 256] - - [47, 0.0] - - - [22272, 2560, 1, 256] - - [53, 0.0] - - - [25344, 2048, 1, 256] - - [53, 0.0] - - - [31792, 256, 1, 256] - - [53, 0.0] - - - [21248, 2816, 1, 256] - - [47, 0.0] - - - [32816, 10240, 1, 256] - - [53, 0.0] - - - [27136, 3840, 1, 256] - - [47, 0.0] - - - [34096, 10240, 1, 256] - - [53, 0.0] - - - [24576, 4608, 1, 256] - - [53, 0.0] - - - [32256, 1281, 1, 256] - - [45, 0.0] - - - [26928, 2865, 1, 256] - - [53, 0.0] - - - [20784, 7424, 1, 256] - - [53, 0.0] - - - [24112, 2816, 1, 256] - - [53, 0.0] - - - [22272, 256, 1, 256] - - [47, 0.0] - - - [30208, 1281, 1, 256] - - [45, 0.0] - - - [28720, 2816, 1, 256] - - [53, 0.0] - - - [20992, 1280, 1, 256] - - [47, 0.0] - - - [31488, 1536, 1, 256] - - [53, 0.0] - - - [21296, 8192, 1, 256] - - [53, 0.0] - - - [30512, 7168, 1, 256] - - [53, 0.0] - - - [27136, 2865, 1, 256] - - [47, 0.0] - - - [25088, 3329, 1, 256] - - [47, 0.0] - - - [29696, 3329, 1, 256] - - [47, 0.0] - - - [23040, 1280, 1, 256] - - [47, 0.0] - - - [30000, 256, 1, 256] - - [47, 0.0] - - - [20224, 3329, 1, 256] - - [47, 0.0] - - - [29232, 2816, 1, 256] - - [47, 0.0] - - - [31232, 7424, 1, 256] - - [47, 0.0] - - - [29488, 2816, 1, 256] - - [53, 0.0] - - - [25904, 2865, 1, 256] - - [47, 0.0] - - - [30512, 2816, 1, 256] - - [53, 0.0] - - - [20736, 768, 1, 256] - - [47, 0.0] - - - [20480, 256, 1, 256] - - [53, 0.0] - - - [28672, 6144, 1, 256] - - [53, 0.0] - - - [26624, 2816, 1, 256] - - [47, 0.0] - - - [28928, 768, 1, 256] - - [53, 0.0] - - - [27648, 256, 1, 256] - - [53, 0.0] - - - [32256, 6144, 1, 256] - - [53, 0.0] - - - [30720, 6144, 1, 256] - - [53, 0.0] - - - [32560, 2865, 1, 256] - - [53, 0.0] - - - [23088, 9728, 1, 256] - - [47, 0.0] - - - [22784, 9728, 1, 256] - - [53, 0.0] - - - [33024, 6144, 1, 256] - - [53, 0.0] - - - [27392, 2865, 1, 256] - - [47, 0.0] - - - [21504, 1280, 1, 256] - - [47, 0.0] - - - [30720, 6656, 1, 256] - - [53, 0.0] - - - [24880, 2865, 1, 256] - - [47, 0.0] - - - [25392, 1792, 1, 256] - - [47, 0.0] - - - [20224, 2816, 1, 256] - - [47, 0.0] - - - [20224, 256, 1, 256] - - [47, 0.0] - - - [25856, 3329, 1, 256] - - [47, 0.0] - - - [30976, 256, 1, 256] - - [53, 0.0] - - - [26880, 6144, 1, 256] - - [53, 0.0] - - - [26672, 2816, 1, 256] - - [47, 0.0] - - - [25600, 256, 1, 256] - - [47, 0.0] - - - [28160, 1281, 1, 256] - - [45, 0.0] - - - [20480, 10240, 1, 256] - - [53, 0.0] - - - [21504, 7936, 1, 256] - - [47, 0.0] - - - [20272, 7168, 1, 256] - - [53, 0.0] - - - [24880, 2816, 1, 256] - - [53, 0.0] - - - [23296, 9728, 1, 256] - - [53, 0.0] - - - [34816, 2865, 1, 256] - - [47, 0.0] - - - [31792, 2865, 1, 256] - - [53, 0.0] - - - [29488, 6144, 1, 256] - - [53, 0.0] - - - [23856, 2865, 1, 256] - - [53, 0.0] - - - [25088, 256, 1, 256] - - [53, 0.0] - - - [22016, 8960, 1, 256] - - [47, 0.0] - - - [23040, 3072, 1, 256] - - [53, 0.0] - - - [23856, 512, 1, 256] - - [53, 0.0] - - - [33792, 3329, 1, 256] - - [47, 0.0] - - - [22784, 9216, 1, 256] - - [53, 0.0] - - - [30720, 4864, 1, 256] - - [47, 0.0] - - - [32000, 8192, 1, 256] - - [53, 0.0] - - - [28160, 3329, 1, 256] - - [47, 0.0] - - - [28672, 256, 1, 256] - - [53, 0.0] - - - [27648, 1281, 1, 256] - - [45, 0.0] - - - [23808, 6144, 1, 256] - - [53, 0.0] - - - [23344, 10240, 1, 256] - - [53, 0.0] - - - [20736, 7680, 1, 256] - - [53, 0.0] - - - [33024, 9216, 1, 256] - - [53, 0.0] - - - [26160, 2816, 1, 256] - - [47, 0.0] - - - [24064, 10240, 1, 256] - - [53, 0.0] - - - [24320, 768, 1, 256] - - [47, 0.0] - - - [28208, 10240, 1, 256] - - [53, 0.0] - - - [34560, 1024, 1, 256] - - [53, 0.0] - - - [33792, 1792, 1, 256] - - [47, 0.0] - - - [30720, 2816, 1, 256] - - [47, 0.0] - - - [24624, 2816, 1, 256] - - [53, 0.0] - - - [20736, 3329, 1, 256] - - [47, 0.0] - - - [21760, 1792, 1, 256] - - [47, 0.0] - - - [21760, 8704, 1, 256] - - [53, 0.0] - - - [34608, 10240, 1, 256] - - [53, 0.0] - - - [22784, 9472, 1, 256] - - [47, 0.0] - - - [31536, 2816, 1, 256] - - [47, 0.0] - - - [27904, 4352, 1, 256] - - [47, 0.0] - - - [23552, 2865, 1, 256] - - [47, 0.0] - - - [24064, 256, 1, 256] - - [53, 0.0] - - - [34304, 2048, 1, 256] - - [53, 0.0] - - - [30464, 1280, 1, 256] - - [47, 0.0] - - - [29440, 5632, 1, 256] - - [53, 0.0] - - - [21808, 8704, 1, 256] - - [53, 0.0] - - - [30464, 6656, 1, 256] - - [53, 0.0] - - - [20736, 1024, 1, 256] - - [53, 0.0] - - - [24832, 1024, 1, 256] - - [53, 0.0] - - - [24576, 1024, 1, 256] - - [53, 0.0] - - - [29184, 2048, 1, 256] - - [53, 0.0] - - - [30976, 4864, 1, 256] - - [47, 0.0] - - - [25344, 1536, 1, 256] - - [53, 0.0] - - - [22016, 1280, 1, 256] - - [47, 0.0] - - - [32560, 8960, 1, 256] - - [47, 0.0] - - - [31536, 7936, 1, 256] - - [47, 0.0] - - - [26880, 3072, 1, 256] - - [53, 0.0] - - - [28464, 2865, 1, 256] - - [47, 0.0] - - - [20224, 6400, 1, 256] - - [47, 0.0] - - - [26624, 3328, 1, 256] - - [47, 0.0] - - - [24320, 512, 1, 256] - - [53, 0.0] - - - [34352, 768, 1, 256] - - [47, 0.0] - - - [30720, 768, 1, 256] - - [47, 0.0] - - - [34560, 10240, 1, 256] - - [53, 0.0] - - - [22016, 3328, 1, 256] - - [47, 0.0] - - - [20480, 1281, 1, 256] - - [50, 0.0] - - - [31232, 2816, 1, 256] - - [47, 0.0] - - - [31232, 6144, 1, 256] - - [53, 0.0] - - - [27136, 256, 1, 256] - - [53, 0.0] - - - [23344, 256, 1, 256] - - [53, 0.0] - - - [30208, 4352, 1, 256] - - [47, 0.0] - - - [32000, 6144, 1, 256] - - [53, 0.0] - - - [29184, 6144, 1, 256] - - [53, 0.0] - - - [29232, 5632, 1, 256] - - [53, 0.0] - - - [22576, 2816, 1, 256] - - [47, 0.0] - - - [31488, 1280, 1, 256] - - [47, 0.0] - - - [23856, 2816, 1, 256] - - [53, 0.0] - - - [29184, 2865, 1, 256] - - [47, 0.0] - - - [21248, 6144, 1, 256] - - [53, 0.0] - - - [30720, 4608, 1, 256] - - [53, 0.0] - - - [27952, 256, 1, 256] - - [53, 0.0] - - - [32512, 10240, 1, 256] - - [53, 0.0] - - - [31744, 3328, 1, 256] - - [47, 0.0] - - - [22528, 3328, 1, 256] - - [47, 0.0] - - - [34048, 3329, 1, 256] - - [47, 0.0] - - - [31744, 2816, 1, 256] - - [47, 0.0] - - - [27904, 256, 1, 256] - - [53, 0.0] - - - [21552, 256, 1, 256] - - [47, 0.0] - - - [29952, 6144, 1, 256] - - [53, 0.0] - - - [22784, 3328, 1, 256] - - [47, 0.0] - - - [20784, 256, 1, 256] - - [53, 0.0] - - - [30208, 2816, 1, 256] - - [47, 0.0] - - - [31232, 5376, 1, 256] - - [47, 0.0] - - - [30256, 256, 1, 256] - - [47, 0.0] - - - [21248, 1280, 1, 256] - - [47, 0.0] - - - [28160, 1280, 1, 256] - - [47, 0.0] - - - [30720, 3329, 1, 256] - - [47, 0.0] - - - [34560, 3329, 1, 256] - - [47, 0.0] - - - [31024, 2816, 1, 256] - - [47, 0.0] - - - [32000, 256, 1, 256] - - [53, 0.0] - - - [20528, 256, 1, 256] - - [53, 0.0] - - - [24624, 10240, 1, 256] - - [53, 0.0] - - - [21504, 7680, 1, 256] - - [53, 0.0] - - - [33536, 9728, 1, 256] - - [53, 0.0] - - - [33280, 6144, 1, 256] - - [53, 0.0] - - - [20480, 2865, 1, 256] - - [47, 0.0] - - - [30720, 1281, 1, 256] - - [50, 0.0] - - - [21760, 6144, 1, 256] - - [53, 0.0] - - - [30976, 6912, 1, 256] - - [47, 0.0] - - - [27648, 2816, 1, 256] - - [47, 0.0] - - - [20992, 3329, 1, 256] - - [47, 0.0] - - - [26672, 3072, 1, 256] - - [53, 0.0] - - - [24832, 2816, 1, 256] - - [47, 0.0] - - - [23552, 9728, 1, 256] - - [53, 0.0] - - - [26880, 1280, 1, 256] - - [47, 0.0] - - - [25088, 1280, 1, 256] - - [47, 0.0] - - - [33280, 9472, 1, 256] - - [47, 0.0] - - - [27136, 3328, 1, 256] - - [47, 0.0] - - - [28416, 2816, 1, 256] - - [47, 0.0] - - - [20480, 3328, 1, 256] - - [53, 0.0] - - - [31232, 256, 1, 256] - - [47, 0.0] - - - [33328, 9728, 1, 256] - - [53, 0.0] - - - [26416, 256, 1, 256] - - [47, 0.0] - - - [31744, 2865, 1, 256] - - [47, 0.0] - - - [22784, 6144, 1, 256] - - [53, 0.0] - - - [32000, 5888, 1, 256] - - [47, 0.0] - - - [28160, 4864, 1, 256] - - [47, 0.0] - - - [34352, 2865, 1, 256] - - [53, 0.0] - - - [29696, 256, 1, 256] - - [53, 0.0] - - - [26112, 2048, 1, 256] - - [53, 0.0] - - - [25088, 5376, 1, 256] - - [47, 0.0] - - - [29952, 3329, 1, 256] - - [47, 0.0] - - - [21296, 10240, 1, 256] - - [53, 0.0] - - - [31744, 1280, 1, 256] - - [47, 0.0] - - - [21760, 256, 1, 256] - - [47, 0.0] - - - [31488, 2048, 1, 256] - - [53, 0.0] - - - [30976, 1281, 1, 256] - - [45, 0.0] - - - [23040, 256, 1, 256] - - [47, 0.0] - - - [34304, 6144, 1, 256] - - [53, 0.0] - - - [31744, 3329, 1, 256] - - [47, 0.0] - - - [31744, 5888, 1, 256] - - [47, 0.0] - - - [29184, 1281, 1, 256] - - [50, 0.0] - - - [23856, 10240, 1, 256] - - [53, 0.0] - - - [23808, 1792, 1, 256] - - [47, 0.0] - - - [32000, 1792, 1, 256] - - [47, 0.0] - - - [26880, 2816, 1, 256] - - [47, 0.0] - - - [28416, 3328, 1, 256] - - [53, 0.0] - - - [27136, 6144, 1, 256] - - [53, 0.0] - - - [28416, 4608, 1, 256] - - [53, 0.0] - - - [33536, 1280, 1, 256] - - [47, 0.0] - - - [27440, 2865, 1, 256] - - [53, 0.0] - - - [25088, 2865, 1, 256] - - [47, 0.0] - - - [30976, 2816, 1, 256] - - [47, 0.0] - - - [26672, 10240, 1, 256] - - [53, 0.0] - - - [34048, 10240, 1, 256] - - [53, 0.0] - - - [34352, 2816, 1, 256] - - [53, 0.0] - - - [22064, 2865, 1, 256] - - [47, 0.0] - - - [28208, 4864, 1, 256] - - [53, 0.0] - - - [22528, 1280, 1, 256] - - [53, 0.0] - - - [26624, 3072, 1, 256] - - [53, 0.0] - - - [33072, 2865, 1, 256] - - [53, 0.0] - - - [22576, 256, 1, 256] - - [47, 0.0] - - - [34560, 2048, 1, 256] - - [53, 0.0] - - - [29440, 5888, 1, 256] - - [47, 0.0] - - - [34560, 1280, 1, 256] - - [47, 0.0] - - - [32000, 10240, 1, 256] - - [53, 0.0] - - - [32304, 2816, 1, 256] - - [53, 0.0] - - - [30976, 2865, 1, 256] - - [47, 0.0] - - - [30208, 6400, 1, 256] - - [47, 0.0] - - - [29232, 2865, 1, 256] - - [47, 0.0] - - - [33072, 2816, 1, 256] - - [53, 0.0] - - - [30512, 2865, 1, 256] - - [53, 0.0] - - - [20016, 2816, 1, 256] - - [47, 0.0] - - - [28416, 4352, 1, 256] - - [47, 0.0] - - - [25648, 2816, 1, 256] - - [53, 0.0] - - - [25344, 1280, 1, 256] - - [47, 0.0] - - - [24576, 10240, 1, 256] - - [53, 0.0] - - - [33024, 1281, 1, 256] - - [45, 0.0] - - - [33584, 10240, 1, 256] - - [53, 0.0] - - - [28416, 4864, 1, 256] - - [47, 0.0] - - - [23296, 3329, 1, 256] - - [47, 0.0] - - - [30464, 4352, 1, 256] - - [53, 0.0] - - - [29696, 5632, 1, 256] - - [53, 0.0] - - - [25136, 256, 1, 256] - - [47, 0.0] - - - [20528, 2865, 1, 256] - - [53, 0.0] - - - [27440, 2816, 1, 256] - - [53, 0.0] - - - [28160, 2048, 1, 256] - - [53, 0.0] - - - [24320, 2816, 1, 256] - - [47, 0.0] - - - [20736, 6144, 1, 256] - - [53, 0.0] - - - [28416, 5120, 1, 256] - - [53, 0.0] - - - [21552, 8448, 1, 256] - - [47, 0.0] - - - [20736, 1281, 1, 256] - - [45, 0.0] - - - [28464, 4864, 1, 256] - - [47, 0.0] - - - [30512, 10240, 1, 256] - - [53, 0.0] - - - [34304, 512, 1, 256] - - [53, 0.0] - - - [22784, 10240, 1, 256] - - [53, 0.0] - - - [25648, 2048, 1, 256] - - [53, 0.0] - - - [25856, 10240, 1, 256] - - [53, 0.0] - - - [32256, 8960, 1, 256] - - [47, 0.0] - - - [20736, 2865, 1, 256] - - [47, 0.0] - - - [20992, 7680, 1, 256] - - [53, 0.0] - - - [31024, 10240, 1, 256] - - [53, 0.0] - - - [26112, 256, 1, 256] - - [47, 0.0] - - - [30000, 2865, 1, 256] - - [47, 0.0] - - - [25904, 2560, 1, 256] - - [53, 0.0] - - - [24832, 768, 1, 256] - - [47, 0.0] - - - [25088, 6144, 1, 256] - - [53, 0.0] - - - [24624, 1280, 1, 256] - - [53, 0.0] - - - [22016, 8192, 1, 256] - - [53, 0.0] - - - [29952, 3328, 1, 256] - - [47, 0.0] - - - [31232, 2048, 1, 256] - - [53, 0.0] - - - [30256, 6656, 1, 256] - - [53, 0.0] - - - [20992, 2816, 1, 256] - - [47, 0.0] - - - [33792, 1536, 1, 256] - - [53, 0.0] - - - [20224, 1280, 1, 256] - - [47, 0.0] - - - [25600, 5888, 1, 256] - - [47, 0.0] - - - [26624, 768, 1, 256] - - [47, 0.0] - - - [32256, 2816, 1, 256] - - [47, 0.0] - - - [21760, 1281, 1, 256] - - [45, 0.0] - - - [25392, 10240, 1, 256] - - [53, 0.0] - - - [32768, 256, 1, 256] - - [47, 0.0] - - - [22528, 3329, 1, 256] - - [47, 0.0] - - - [23552, 3329, 1, 256] - - [47, 0.0] - - - [33024, 2865, 1, 256] - - [47, 0.0] - - - [29696, 2816, 1, 256] - - [47, 0.0] - - - [27392, 10240, 1, 256] - - [53, 0.0] - - - [23040, 2048, 1, 256] - - [53, 0.0] - - - [27648, 6144, 1, 256] - - [53, 0.0] - - - [22016, 2304, 1, 256] - - [47, 0.0] - - - [34560, 1281, 1, 256] - - [45, 0.0] - - - [27136, 1281, 1, 256] - - [50, 0.0] - - - [32000, 1281, 1, 256] - - [45, 0.0] - - - [27184, 3840, 1, 256] - - [53, 0.0] - - - [24880, 1536, 1, 256] - - [53, 0.0] - - - [28672, 768, 1, 256] - - [53, 0.0] - - - [34816, 2816, 1, 256] - - [47, 0.0] - - - [26160, 256, 1, 256] - - [53, 0.0] - - - [30464, 7168, 1, 256] - - [53, 0.0] - - - [30208, 3328, 1, 256] - - [47, 0.0] - - - [32304, 10240, 1, 256] - - [53, 0.0] - - - [26624, 1280, 1, 256] - - [47, 0.0] - - - [29696, 10240, 1, 256] - - [53, 0.0] - - - [32000, 8704, 1, 256] - - [53, 0.0] - - - [27392, 1281, 1, 256] - - [44, 0.0] - - - [26416, 2865, 1, 256] - - [47, 0.0] - - - [26160, 2560, 1, 256] - - [53, 0.0] - - - [28672, 3329, 1, 256] - - [47, 0.0] - - - [23808, 256, 1, 256] - - [53, 0.0] - - - [27184, 10240, 1, 256] - - [53, 0.0] - - - [33280, 2048, 1, 256] - - [53, 0.0] - - - [33280, 2816, 1, 256] - - [53, 0.0] - - - [23040, 9984, 1, 256] - - [47, 0.0] - - - [26112, 1280, 1, 256] - - [47, 0.0] - - - [33328, 9984, 1, 256] - - [47, 0.0] - - - [32560, 9216, 1, 256] - - [53, 0.0] - - - [22832, 9728, 1, 256] - - [53, 0.0] - - - [27904, 1280, 1, 256] - - [47, 0.0] - - - [33280, 1281, 1, 256] - - [45, 0.0] - - - [33280, 1280, 1, 256] - - [47, 0.0] - - - [32048, 256, 1, 256] - - [53, 0.0] - - - [27184, 2865, 1, 256] - - [53, 0.0] - - - [26880, 3329, 1, 256] - - [47, 0.0] - - - [20784, 7680, 1, 256] - - [53, 0.0] - - - [24832, 3329, 1, 256] - - [47, 0.0] - - - [25856, 1280, 1, 256] - - [47, 0.0] - - - [34560, 2816, 1, 256] - - [47, 0.0] - - - [20016, 256, 1, 256] - - [47, 0.0] - - - [23600, 256, 1, 256] - - [53, 0.0] - - - [22576, 9216, 1, 256] - - [53, 0.0] - - - [25344, 5632, 1, 256] - - [53, 0.0] - - - [28928, 5632, 1, 256] - - [53, 0.0] - - - [31024, 256, 1, 256] - - [53, 0.0] - - - [21552, 2865, 1, 256] - - [47, 0.0] - - - [29184, 3072, 1, 256] - - [53, 0.0] - - - [24320, 2865, 1, 256] - - [47, 0.0] - - - [20480, 6656, 1, 256] - - [53, 0.0] - - - [33536, 10240, 1, 256] - - [53, 0.0] - - - [20736, 1280, 1, 256] - - [47, 0.0] - - - [24832, 1280, 1, 256] - - [47, 0.0] - - - [29488, 10240, 1, 256] - - [53, 0.0] - - - [27392, 6144, 1, 256] - - [53, 0.0] - - - [29440, 3329, 1, 256] - - [47, 0.0] - - - [25856, 1281, 1, 256] - - [50, 0.0] - - - [34560, 768, 1, 256] - - [47, 0.0] - - - [31488, 7680, 1, 256] - - [50, 0.0] - - - [29184, 5632, 1, 256] - - [53, 0.0] - - - [32512, 512, 1, 256] - - [53, 0.0] - - - [26112, 2865, 1, 256] - - [47, 0.0] - - - [32512, 1280, 1, 256] - - [47, 0.0] - - - [20992, 1024, 1, 256] - - [53, 0.0] - - - [27904, 10240, 1, 256] - - [53, 0.0] - - - [29952, 6656, 1, 256] - - [53, 0.0] - - - [21248, 2048, 1, 256] - - [53, 0.0] - - - [34352, 256, 1, 256] - - [47, 0.0] - - - [24064, 512, 1, 256] - - [53, 0.0] - - - [32816, 2865, 1, 256] - - [53, 0.0] - - - [33840, 256, 1, 256] - - [53, 0.0] - - - [33792, 1280, 1, 256] - - [47, 0.0] - - - [21296, 7936, 1, 256] - - [53, 0.0] - - - [34096, 256, 1, 256] - - [47, 0.0] - - - [32256, 8704, 1, 256] - - [53, 0.0] - - - [30464, 1281, 1, 256] - - [45, 0.0] - - - [28464, 2816, 1, 256] - - [53, 0.0] - - - [25136, 2865, 1, 256] - - [53, 0.0] - - - [31792, 8448, 1, 256] - - [47, 0.0] - - - [24320, 4608, 1, 256] - - [53, 0.0] - - - [25088, 5120, 1, 256] - - [53, 0.0] - - - [31744, 2048, 1, 256] - - [53, 0.0] - - - [30720, 1280, 1, 256] - - [47, 0.0] - - - [34048, 256, 1, 256] - - [47, 0.0] - - - [28416, 512, 1, 256] - - [53, 0.0] - - - [22272, 10240, 1, 256] - - [53, 0.0] - - - [32512, 3328, 1, 256] - - [47, 0.0] - - - [29744, 10240, 1, 256] - - [53, 0.0] - - - [22784, 2048, 1, 256] - - [53, 0.0] - - - [23552, 2048, 1, 256] - - [53, 0.0] - - - [25344, 2816, 1, 256] - - [47, 0.0] - - - [27440, 3840, 1, 256] - - [53, 0.0] - - - [21552, 10240, 1, 256] - - [53, 0.0] - - - [21808, 256, 1, 256] - - [47, 0.0] - - - [24576, 6144, 1, 256] - - [53, 0.0] - - - [29744, 256, 1, 256] - - [47, 0.0] - - - [31488, 3328, 1, 256] - - [47, 0.0] - - - [33536, 3329, 1, 256] - - [47, 0.0] - - - [21040, 256, 1, 256] - - [47, 0.0] - - - [22272, 9216, 1, 256] - - [53, 0.0] - - - [27648, 4096, 1, 256] - - [53, 0.0] - - - [29440, 1280, 1, 256] - - [47, 0.0] - - - [31744, 7936, 1, 256] - - [47, 0.0] - - - [26624, 1281, 1, 256] - - [50, 0.0] - - - [28672, 2048, 1, 256] - - [53, 0.0] - - - [24064, 3328, 1, 256] - - [47, 0.0] - - - [25344, 3329, 1, 256] - - [47, 0.0] - - - [33280, 9728, 1, 256] - - [53, 0.0] - - - [22320, 8960, 1, 256] - - [53, 0.0] - - - [30464, 6144, 1, 256] - - [53, 0.0] - - - [34304, 2304, 1, 256] - - [47, 0.0] - - - [28928, 256, 1, 256] - - [53, 0.0] - - - [27392, 1280, 1, 256] - - [47, 0.0] - - - [26672, 2865, 1, 256] - - [53, 0.0] - - - [28720, 10240, 1, 256] - - [53, 0.0] - - - [25088, 2816, 1, 256] - - [47, 0.0] - - - [31280, 256, 1, 256] - - [47, 0.0] - - - [29488, 5888, 1, 256] - - [53, 0.0] - - - [30720, 2048, 1, 256] - - [53, 0.0] - - - [21808, 10240, 1, 256] - - [53, 0.0] - - - [24576, 2865, 1, 256] - - [53, 0.0] - - - [23808, 1280, 1, 256] - - [47, 0.0] - - - [33280, 1024, 1, 256] - - [53, 0.0] - - - [25856, 256, 1, 256] - - [53, 0.0] - - - [25648, 2304, 1, 256] - - [47, 0.0] - - - [29952, 2865, 1, 256] - - [47, 0.0] - - - [23040, 1024, 1, 256] - - [53, 0.0] - - - [34304, 3328, 1, 256] - - [47, 0.0] - - - [31792, 8192, 1, 256] - - [53, 0.0] - - - [24576, 2816, 1, 256] - - [47, 0.0] - - - [27648, 1536, 1, 256] - - [53, 0.0] - - - [23296, 9472, 1, 256] - - [53, 0.0] - - - [24624, 256, 1, 256] - - [53, 0.0] - - - [20736, 2048, 1, 256] - - [53, 0.0] - - - [28720, 5376, 1, 256] - - [53, 0.0] - - - [20480, 512, 1, 256] - - [53, 0.0] - - - [33840, 2865, 1, 256] - - [47, 0.0] - - - [24064, 2865, 1, 256] - - [47, 0.0] - - - [24064, 2816, 1, 256] - - [47, 0.0] - - - [20992, 256, 1, 256] - - [53, 0.0] - - - [33328, 256, 1, 256] - - [53, 0.0] - - - [28928, 5120, 1, 256] - - [53, 0.0] - - - [34304, 256, 1, 256] - - [53, 0.0] - - - [34304, 1281, 1, 256] - - [44, 0.0] - - - [31744, 1281, 1, 256] - - [45, 0.0] - - - [33584, 2816, 1, 256] - - [53, 0.0] - - - [24064, 4352, 1, 256] - - [47, 0.0] - - - [20224, 6912, 1, 256] - - [47, 0.0] - - - [21504, 1281, 1, 256] - - [44, 0.0] - - - [33536, 3328, 1, 256] - - [47, 0.0] - - - [34816, 3328, 1, 256] - - [47, 0.0] - - - [31024, 7680, 1, 256] - - [53, 0.0] - - - [22016, 3329, 1, 256] - - [47, 0.0] - - - [25344, 1281, 1, 256] - - [50, 0.0] - - - [31744, 7680, 1, 256] - - [53, 0.0] - - - [27952, 10240, 1, 256] - - [53, 0.0] - - - [23808, 2048, 1, 256] - - [53, 0.0] - - - [32768, 2816, 1, 256] - - [47, 0.0] - - - [34816, 256, 1, 256] - - [47, 0.0] - - - [27904, 2865, 1, 256] - - [47, 0.0] - - - [31232, 1280, 1, 256] - - [47, 0.0] - - - [22016, 1281, 1, 256] - - [45, 0.0] - - - [22528, 8704, 1, 256] - - [53, 0.0] - - - [22528, 9216, 1, 256] - - [53, 0.0] - - - [34816, 1280, 1, 256] - - [47, 0.0] - - - [23808, 10240, 1, 256] - - [53, 0.0] - - - [32512, 2048, 1, 256] - - [53, 0.0] - - - [34816, 1024, 1, 256] - - [53, 0.0] - - - [34048, 2048, 1, 256] - - [53, 0.0] - - - [30768, 2816, 1, 256] - - [53, 0.0] - - - [22272, 3329, 1, 256] - - [47, 0.0] - - - [25600, 3328, 1, 256] - - [47, 0.0] - - - [34048, 2816, 1, 256] - - [47, 0.0] - - - [22064, 8704, 1, 256] - - [53, 0.0] - - - [25648, 256, 1, 256] - - [53, 0.0] - - - [22784, 768, 1, 256] - - [47, 0.0] - - - [27904, 2048, 1, 256] - - [53, 0.0] - - - [22528, 9472, 1, 256] - - [47, 0.0] - - - [21504, 2865, 1, 256] - - [47, 0.0] - - - [28672, 5376, 1, 256] - - [47, 0.0] - - - [22576, 9472, 1, 256] - - [53, 0.0] - - - [24576, 256, 1, 256] - - [53, 0.0] - - - [28672, 5120, 1, 256] - - [53, 0.0] - - - [24576, 3328, 1, 256] - - [47, 0.0] - - - [32816, 9472, 1, 256] - - [53, 0.0] - - - [27440, 256, 1, 256] - - [53, 0.0] - - - [22272, 8704, 1, 256] - - [53, 0.0] - - - [30000, 2816, 1, 256] - - [47, 0.0] - - - [26928, 2816, 1, 256] - - [53, 0.0] - - - [22064, 2816, 1, 256] - - [53, 0.0] - - - [23552, 3328, 1, 256] - - [47, 0.0] - - - [28416, 256, 1, 256] - - [47, 0.0] - - - [28928, 6144, 1, 256] - - [53, 0.0] - - - [32768, 512, 1, 256] - - [53, 0.0] - - - [22272, 2865, 1, 256] - - [53, 0.0] - - - [26928, 256, 1, 256] - - [53, 0.0] - - - [21760, 10240, 1, 256] - - [53, 0.0] - - - [26368, 512, 1, 256] - - [53, 0.0] - - - [26672, 256, 1, 256] - - [47, 0.0] - - - [33328, 2865, 1, 256] - - [53, 0.0] - - - [30720, 3328, 1, 256] - - [47, 0.0] - - - [25856, 2865, 1, 256] - - [47, 0.0] - - - [25088, 3328, 1, 256] - - [47, 0.0] - - - [28416, 2560, 1, 256] - - [53, 0.0] - - - [33536, 9472, 1, 256] - - [47, 0.0] - - - [20480, 1280, 1, 256] - - [47, 0.0] - - - [30208, 6144, 1, 256] - - [53, 0.0] - - - [34864, 1024, 1, 256] - - [53, 0.0] - - - [33280, 256, 1, 256] - - [53, 0.0] - - - [23296, 3328, 1, 256] - - [47, 0.0] - - - [32560, 256, 1, 256] - - [47, 0.0] - - - [32560, 2816, 1, 256] - - [47, 0.0] - - - [33536, 256, 1, 256] - - [53, 0.0] - - - [34608, 768, 1, 256] - - [47, 0.0] - - - [24832, 5120, 1, 256] - - [53, 0.0] - - - [25856, 2048, 1, 256] - - [53, 0.0] - - - [30768, 256, 1, 256] - - [53, 0.0] - - - [30000, 6656, 1, 256] - - [53, 0.0] - - - [24320, 1024, 1, 256] - - [53, 0.0] - - - [33280, 9216, 1, 256] - - [53, 0.0] - - - [31488, 5376, 1, 256] - - [47, 0.0] - - - [28416, 1281, 1, 256] - - [45, 0.0] - - - [27392, 3584, 1, 256] - - [53, 0.0] - - - [26368, 2048, 1, 256] - - [53, 0.0] - - - [22528, 256, 1, 256] - - [53, 0.0] - - - [32768, 2048, 1, 256] - - [53, 0.0] - - - [30256, 6912, 1, 256] - - [47, 0.0] - - - [28672, 512, 1, 256] - - [53, 0.0] - - - [21760, 8448, 1, 256] - - [53, 0.0] - - - [34560, 6144, 1, 256] - - [53, 0.0] - - - [27696, 2816, 1, 256] - - [47, 0.0] - - - [29952, 2048, 1, 256] - - [53, 0.0] - - - [22576, 10240, 1, 256] - - [53, 0.0] - - - [25600, 1792, 1, 256] - - [47, 0.0] - - - [28976, 10240, 1, 256] - - [53, 0.0] - - - [29952, 1280, 1, 256] - - [47, 0.0] - - - [26368, 2816, 1, 256] - - [47, 0.0] - - - [26416, 3072, 1, 256] - - [53, 0.0] - - - [27648, 3329, 1, 256] - - [47, 0.0] - - - [34560, 2560, 1, 256] - - [53, 0.0] - - - [32048, 8448, 1, 256] - - [53, 0.0] - - - [30464, 2865, 1, 256] - - [47, 0.0] - - - [34048, 3328, 1, 256] - - [47, 0.0] - - - [23808, 2865, 1, 256] - - [47, 0.0] - - - [25600, 2816, 1, 256] - - [53, 0.0] - - - [20736, 6912, 1, 256] - - [47, 0.0] - - - [24576, 512, 1, 256] - - [53, 0.0] - - - [33792, 256, 1, 256] - - [53, 0.0] - - - [22576, 2865, 1, 256] - - [47, 0.0] - - - [30464, 256, 1, 256] - - [47, 0.0] - - - [24368, 2816, 1, 256] - - [53, 0.0] - - - [20224, 512, 1, 256] - - [53, 0.0] - - - [30512, 6912, 1, 256] - - [53, 0.0] - - - [20272, 2816, 1, 256] - - [53, 0.0] - - - [23296, 256, 1, 256] - - [47, 0.0] - - - [27904, 2816, 1, 256] - - [47, 0.0] - - - [29184, 1280, 1, 256] - - [47, 0.0] - - - [24112, 10240, 1, 256] - - [53, 0.0] - - - [31280, 7680, 1, 256] - - [53, 0.0] - - - [24064, 6144, 1, 256] - - [53, 0.0] - - - [26624, 6144, 1, 256] - - [53, 0.0] - - - [30768, 2865, 1, 256] - - [53, 0.0] - - - [20528, 2816, 1, 256] - - [47, 0.0] - - - [25392, 2865, 1, 256] - - [47, 0.0] - - - [22272, 6144, 1, 256] - - [53, 0.0] - - - [25088, 10240, 1, 256] - - [53, 0.0] - - - [25344, 2865, 1, 256] - - [47, 0.0] - - - [23552, 1792, 1, 256] - - [47, 0.0] - - - [23296, 3584, 1, 256] - - [53, 0.0] - - - [28160, 2816, 1, 256] - - [47, 0.0] - - - [20272, 2865, 1, 256] - - [53, 0.0] - - - [22832, 9472, 1, 256] - - [47, 0.0] - - - [21760, 7936, 1, 256] - - [47, 0.0] - - - [26928, 3328, 1, 256] - - [47, 0.0] - - - [33072, 9472, 1, 256] - - [53, 0.0] - - - [33024, 1280, 1, 256] - - [47, 0.0] - - - [34352, 512, 1, 256] - - [53, 0.0] - - - [26368, 2865, 1, 256] - - [47, 0.0] - - - [27952, 4352, 1, 256] - - [53, 0.0] - - - [21504, 8192, 1, 256] - - [53, 0.0] - - - [22320, 9216, 1, 256] - - [53, 0.0] - - - [31232, 2865, 1, 256] - - [47, 0.0] - - - [21248, 7680, 1, 256] - - [53, 0.0] - - - [24368, 256, 1, 256] - - [53, 0.0] - - - [25648, 2865, 1, 256] - - [53, 0.0] - - - [21248, 2865, 1, 256] - - [53, 0.0] - - - [28416, 2865, 1, 256] - - [47, 0.0] - - - [24320, 3329, 1, 256] - - [47, 0.0] - - - [27648, 2048, 1, 256] - - [53, 0.0] - - - [27648, 2865, 1, 256] - - [47, 0.0] - - - [26880, 2048, 1, 256] - - [53, 0.0] - - - [28672, 2560, 1, 256] - - [53, 0.0] - - - [24064, 1280, 1, 256] - - [47, 0.0] - - - [30256, 2865, 1, 256] - - [53, 0.0] - - - [22064, 10240, 1, 256] - - [47, 0.0] - - - [30464, 4608, 1, 256] - - [53, 0.0] - - - [22016, 6144, 1, 256] - - [53, 0.0] - - - [29440, 2816, 1, 256] - - [47, 0.0] - - - [25392, 2048, 1, 256] - - [53, 0.0] - - - [20992, 2048, 1, 256] - - [53, 0.0] - - - [33024, 3329, 1, 256] - - [47, 0.0] - - - [20224, 3328, 1, 256] - - [47, 0.0] - - - [28208, 4608, 1, 256] - - [53, 0.0] - - - [25344, 6144, 1, 256] - - [53, 0.0] - - - [30464, 512, 1, 256] - - [53, 0.0] - - - [21248, 3329, 1, 256] - - [47, 0.0] - - - [29696, 6144, 1, 256] - - [53, 0.0] - - - [20992, 7936, 1, 256] - - [47, 0.0] - - - [33024, 9472, 1, 256] - - [47, 0.0] - - - [32000, 3329, 1, 256] - - [47, 0.0] - - - [21248, 1281, 1, 256] - - [45, 0.0] - - - [24624, 1024, 1, 256] - - [53, 0.0] - - - [22272, 2816, 1, 256] - - [47, 0.0] - - - [29440, 1281, 1, 256] - - [44, 0.0] - - - [30464, 6400, 1, 256] - - [53, 0.0] - - - [25136, 10240, 1, 256] - - [47, 0.0] - - - [23040, 9472, 1, 256] - - [47, 0.0] - - - [33840, 2816, 1, 256] - - [47, 0.0] - - - [30976, 1024, 1, 256] - - [53, 0.0] - - - [34048, 6144, 1, 256] - - [53, 0.0] - - - [32000, 2048, 1, 256] - - [53, 0.0] - - - [32048, 2865, 1, 256] - - [47, 0.0] - - - [33328, 10240, 1, 256] - - [53, 0.0] - - - [25088, 1536, 1, 256] - - [53, 0.0] - - - [30512, 256, 1, 256] - - [47, 0.0] - - - [20480, 6912, 1, 256] - - [47, 0.0] - - - [34608, 2816, 1, 256] - - [47, 0.0] - - - [22064, 256, 1, 256] - - [53, 0.0] - - - [25600, 2865, 1, 256] - - [47, 0.0] - - - [26880, 1024, 1, 256] - - [53, 0.0] - - - [27392, 2048, 1, 256] - - [53, 0.0] - - - [30208, 10240, 1, 256] - - [53, 0.0] - - - [20016, 10240, 1, 256] - - [53, 0.0] - - - [26880, 10240, 1, 256] - - [53, 0.0] - - - [28160, 3328, 1, 256] - - [47, 0.0] - - - [33536, 2048, 1, 256] - - [53, 0.0] - - - [31232, 7936, 1, 256] - - [47, 0.0] - - - [31536, 10240, 1, 256] - - [47, 0.0] - - - [24832, 1536, 1, 256] - - [53, 0.0] - - - [32768, 768, 1, 256] - - [47, 0.0] - - - [29440, 6144, 1, 256] - - [53, 0.0] - - - [26112, 2560, 1, 256] - - [53, 0.0] - - - [33792, 6144, 1, 256] - - [53, 0.0] - - - [22528, 10240, 1, 256] - - [53, 0.0] - - - [20480, 768, 1, 256] - - [47, 0.0] - - - [22320, 256, 1, 256] - - [53, 0.0] - - - [23808, 3328, 1, 256] - - [47, 0.0] - - - [28464, 256, 1, 256] - - [47, 0.0] - - - [27136, 2048, 1, 256] - - [53, 0.0] - - - [29744, 6400, 1, 256] - - [53, 0.0] - - - [20480, 7168, 1, 256] - - [53, 0.0] - - - [22832, 256, 1, 256] - - [47, 0.0] - - - [21552, 8192, 1, 256] - - [53, 0.0] - - - [25856, 2560, 1, 256] - - [53, 0.0] - - - [28160, 6144, 1, 256] - - [53, 0.0] - - - [31280, 2816, 1, 256] - - [53, 0.0] - - - [23600, 10240, 1, 256] - - [53, 0.0] - - - [26368, 1281, 1, 256] - - [45, 0.0] - - - [24576, 1280, 1, 256] - - [47, 0.0] - - - [33536, 1536, 1, 256] - - [53, 0.0] - - - [23088, 2816, 1, 256] - - [53, 0.0] - - - [26624, 2048, 1, 256] - - [53, 0.0] - - - [29952, 2816, 1, 256] - - [47, 0.0] - - - [21760, 2048, 1, 256] - - [53, 0.0] - - - [30976, 6144, 1, 256] - - [53, 0.0] - - - [29696, 1280, 1, 256] - - [47, 0.0] - - - [30208, 4096, 1, 256] - - [53, 0.0] - - - [24832, 2865, 1, 256] - - [47, 0.0] - - - [31488, 1281, 1, 256] - - [45, 0.0] - - - [34304, 2865, 1, 256] - - [47, 0.0] - - - [32512, 256, 1, 256] - - [47, 0.0] - - - [25136, 1536, 1, 256] - - [53, 0.0] - - - [26112, 3329, 1, 256] - - [47, 0.0] - - - [24880, 1280, 1, 256] - - [53, 0.0] - - - [28208, 2816, 1, 256] - - [47, 0.0] - - - [29184, 5888, 1, 256] - - [47, 0.0] - - - [28160, 4352, 1, 256] - - [47, 0.0] - - - [34352, 10240, 1, 256] - - [53, 0.0] - - - [23856, 256, 1, 256] - - [53, 0.0] - - - [25344, 10240, 1, 256] - - [53, 0.0] - - - [20992, 1281, 1, 256] - - [50, 0.0] - - - [26624, 512, 1, 256] - - [53, 0.0] - - - [21040, 10240, 1, 256] - - [53, 0.0] - - - [23040, 3328, 1, 256] - - [47, 0.0] - - - [30976, 7168, 1, 256] - - [53, 0.0] - - - [25856, 2304, 1, 256] - - [47, 0.0] - - - [24368, 1024, 1, 256] - - [53, 0.0] - - - [33280, 2865, 1, 256] - - [47, 0.0] - - - [23296, 1536, 1, 256] - - [53, 0.0] - - - [21504, 6144, 1, 256] - - [50, 0.0] - - - [23552, 2816, 1, 256] - - [47, 0.0] - - - [30464, 2816, 1, 256] - - [47, 0.0] - - - [22832, 2865, 1, 256] - - [53, 0.0] - - - [24576, 2048, 1, 256] - - [53, 0.0] - - - [22272, 8448, 1, 256] - - [53, 0.0] - - - [32256, 1280, 1, 256] - - [53, 0.0] - - - [25856, 5888, 1, 256] - - [53, 0.0] - - - [30976, 5120, 1, 256] - - [53, 0.0] - - - [29184, 3329, 1, 256] - - [47, 0.0] - - - [24112, 2865, 1, 256] - - [53, 0.0] - - - [29744, 2816, 1, 256] - - [47, 0.0] - - - [21760, 2816, 1, 256] - - [47, 0.0] - - - [25600, 2048, 1, 256] - - [53, 0.0] - - - [32000, 1280, 1, 256] - - [47, 0.0] - - - [25856, 3328, 1, 256] - - [47, 0.0] - - - [20016, 6656, 1, 256] - - [53, 0.0] - - - [32256, 2865, 1, 256] - - [47, 0.0] - - - [22272, 3328, 1, 256] - - [47, 0.0] - - - [21504, 3328, 1, 256] - - [47, 0.0] - - - [31232, 5120, 1, 256] - - [53, 0.0] - - - [24112, 256, 1, 256] - - [53, 0.0] - - - [30208, 1280, 1, 256] - - [47, 0.0] - - - [22064, 8960, 1, 256] - - [53, 0.0] - - - [28160, 10240, 1, 256] - - [53, 0.0] - - - [21504, 1536, 1, 256] - - [53, 0.0] - - - [31744, 5632, 1, 256] - - [53, 0.0] - - - [20272, 6912, 1, 256] - - [47, 0.0] - - - [29952, 1792, 1, 256] - - [47, 0.0] - - - [25904, 10240, 1, 256] - - [53, 0.0] - - - [25344, 1792, 1, 256] - - [47, 0.0] - - - [32512, 8448, 1, 256] - - [47, 0.0] - - - [25088, 2048, 1, 256] - - [53, 0.0] - - - [23808, 9984, 1, 256] - - [47, 0.0] - - - [32768, 3329, 1, 256] - - [47, 0.0] - - - [34816, 6144, 1, 256] - - [53, 0.0] - - - [32256, 256, 1, 256] - - [53, 0.0] - - - [26368, 3328, 1, 256] - - [47, 0.0] - - - [23296, 1280, 1, 256] - - [47, 0.0] - - - [34608, 1024, 1, 256] - - [53, 0.0] - - - [30976, 1280, 1, 256] - - [47, 0.0] - - - [22528, 6144, 1, 256] - - [53, 0.0] - - - [21248, 10240, 1, 256] - - [53, 0.0] - - - [22528, 2865, 1, 256] - - [47, 0.0] - - - [22528, 768, 1, 256] - - [47, 0.0] - - - [22016, 8704, 1, 256] - - [53, 0.0] - - - [30720, 6912, 1, 256] - - [47, 0.0] - - - [33024, 2048, 1, 256] - - [53, 0.0] - - - [31232, 3329, 1, 256] - - [47, 0.0] - - - [33024, 3328, 1, 256] - - [47, 0.0] - - - [30976, 7424, 1, 256] - - [47, 0.0] - - - [27136, 3584, 1, 256] - - [53, 0.0] - - - [34048, 1280, 1, 256] - - [47, 0.0] - - - [34864, 1280, 1, 256] - - [53, 0.0] - - - [25600, 2304, 1, 256] - - [47, 0.0] - - - [21760, 3329, 1, 256] - - [47, 0.0] - - - [26928, 3584, 1, 256] - - [53, 0.0] - - - [28976, 2816, 1, 256] - - [53, 0.0] - - - [24832, 4864, 1, 256] - - [47, 0.0] - - - [21248, 1536, 1, 256] - - [53, 0.0] - - - [23808, 2816, 1, 256] - - [47, 0.0] - - - [32768, 9472, 1, 256] - - [47, 0.0] - - - [27392, 3328, 1, 256] - - [47, 0.0] - - - [26880, 3584, 1, 256] - - [53, 0.0] - - - [23552, 1281, 1, 256] - - [45, 0.0] - - - [27648, 3840, 1, 256] - - [47, 0.0] - - - [22016, 10240, 1, 256] - - [53, 0.0] - - - [34816, 2560, 1, 256] - - [53, 0.0] - - - [31536, 256, 1, 256] - - [47, 0.0] - - - [34816, 10240, 1, 256] - - [53, 0.0] - - - [27904, 1792, 1, 256] - - [47, 0.0] - - - [33792, 10240, 1, 256] - - [53, 0.0] - - - [23296, 2816, 1, 256] - - [47, 0.0] - - - [31024, 7424, 1, 256] - - [47, 0.0] - - - [22784, 1280, 1, 256] - - [47, 0.0] - - - [30976, 2048, 1, 256] - - [53, 0.0] - - - [27392, 4096, 1, 256] - - [53, 0.0] - - - [33792, 2816, 1, 256] - - [47, 0.0] - - - [32560, 10240, 1, 256] - - [53, 0.0] - - - [20736, 7424, 1, 256] - - [47, 0.0] - - - [28672, 2865, 1, 256] - - [47, 0.0] - - - [31488, 256, 1, 256] - - [53, 0.0] - - - [20992, 7424, 1, 256] - - [47, 0.0] - - - [21504, 1792, 1, 256] - - [47, 0.0] - - - [27696, 2865, 1, 256] - - [53, 0.0] - - - [33024, 1024, 1, 256] - - [53, 0.0] - - - [22016, 256, 1, 256] - - [53, 0.0] - - - [23088, 256, 1, 256] - - [47, 0.0] - - - [28976, 256, 1, 256] - - [47, 0.0] - - - [27392, 256, 1, 256] - - [53, 0.0] - - - [34304, 3329, 1, 256] - - [47, 0.0] - - - [32512, 9216, 1, 256] - - [53, 0.0] - - - [31488, 3329, 1, 256] - - [47, 0.0] - - - [20016, 2865, 1, 256] - - [47, 0.0] - - - [22016, 8448, 1, 256] - - [53, 0.0] - - - [31024, 2865, 1, 256] - - [47, 0.0] - - - [29440, 256, 1, 256] - - [53, 0.0] - - - [34608, 2865, 1, 256] - - [53, 0.0] - - - [20480, 2048, 1, 256] - - [53, 0.0] - - - [28160, 2865, 1, 256] - - [47, 0.0] - - - [28416, 2304, 1, 256] - - [47, 0.0] - - - [23552, 6144, 1, 256] - - [53, 0.0] - - - [21296, 256, 1, 256] - - [53, 0.0] - - - [28672, 4864, 1, 256] - - [47, 0.0] - - - [27648, 1792, 1, 256] - - [47, 0.0] - - - [31488, 7424, 1, 256] - - [47, 0.0] - - - [23040, 2865, 1, 256] - - [47, 0.0] - - - [30976, 3328, 1, 256] - - [47, 0.0] - - - [25856, 1792, 1, 256] - - [47, 0.0] - - - [33536, 9984, 1, 256] - - [47, 0.0] - - - [24832, 1281, 1, 256] - - [45, 0.0] - - - [29184, 3328, 1, 256] - - [47, 0.0] - - - [32000, 2816, 1, 256] - - [47, 0.0] - - - [34304, 768, 1, 256] - - [47, 0.0] - - - [24576, 1281, 1, 256] - - [45, 0.0] - - - [25088, 1281, 1, 256] - - [45, 0.0] - - - [29744, 2865, 1, 256] - - [47, 0.0] - - - [25136, 2816, 1, 256] - - [47, 0.0] - - - [29696, 1281, 1, 256] - - [50, 0.0] - - - [27392, 3329, 1, 256] - - [47, 0.0] - - - [31488, 2816, 1, 256] - - [47, 0.0] - - - [30976, 10240, 1, 256] - - [53, 0.0] - - - [26624, 3329, 1, 256] - - [47, 0.0] - - - [34304, 1280, 1, 256] - - [47, 0.0] - - - [25392, 256, 1, 256] - - [53, 0.0] - - - [26624, 10240, 1, 256] - - [53, 0.0] - - - [26112, 6144, 1, 256] - - [53, 0.0] - - - [29696, 3328, 1, 256] - - [47, 0.0] - - - [32304, 2865, 1, 256] - - [47, 0.0] - - - [24368, 2865, 1, 256] - - [53, 0.0] - - - [31488, 8192, 1, 256] - - [53, 0.0] - - - [20224, 6656, 1, 256] - - [53, 0.0] - - - [31232, 1281, 1, 256] - - [44, 0.0] - - - [21296, 2865, 1, 256] - - [53, 0.0] - - - [24112, 768, 1, 256] - - [53, 0.0] - - - [32000, 8448, 1, 256] - - [47, 0.0] - - - [23552, 1536, 1, 256] - - [53, 0.0] - - - [30976, 7680, 1, 256] - - [53, 0.0] - - - [31280, 10240, 1, 256] - - [53, 0.0] - - - [23344, 9984, 1, 256] - - [53, 0.0] - - - [21248, 8192, 1, 256] - - [53, 0.0] - - - [29696, 6400, 1, 256] - - [47, 0.0] - - - [32304, 8960, 1, 256] - - [47, 0.0] - - - [27184, 256, 1, 256] - - [53, 0.0] - - - [28464, 10240, 1, 256] - - [53, 0.0] - - - [20736, 256, 1, 256] - - [53, 0.0] - - - [31232, 10240, 1, 256] - - [53, 0.0] - - - [25856, 6144, 1, 256] - - [53, 0.0] - - - [27440, 10240, 1, 256] - - [53, 0.0] - - - [23088, 2865, 1, 256] - - [53, 0.0] - - - [29696, 3584, 1, 256] - - [53, 0.0] - - - [23040, 9728, 1, 256] - - [53, 0.0] - - - [31744, 10240, 1, 256] - - [53, 0.0] - - - [31744, 1792, 1, 256] - - [47, 0.0] - - - [24320, 256, 1, 256] - - [53, 0.0] - - - [27696, 256, 1, 256] - - [47, 0.0] - - - [29696, 2865, 1, 256] - - [47, 0.0] - - - [22784, 3072, 1, 256] - - [53, 0.0] - - - [29952, 5888, 1, 256] - - [47, 0.0] - - - [28928, 2816, 1, 256] - - [47, 0.0] - - - [30768, 7424, 1, 256] - - [53, 0.0] - - - [27440, 4096, 1, 256] - - [53, 0.0] - - - [24064, 4096, 1, 256] - - [53, 0.0] - - - [32256, 3329, 1, 256] - - [47, 0.0] - - - [30976, 3329, 1, 256] - - [47, 0.0] - - - [25600, 10240, 1, 256] - - [53, 0.0] - - - [20224, 6144, 1, 256] - - [53, 0.0] - - - [21040, 7936, 1, 256] - - [47, 0.0] - - - [26368, 2560, 1, 256] - - [53, 0.0] - - - [32512, 1281, 1, 256] - - [45, 0.0] - - - [28928, 3072, 1, 256] - - [53, 0.0] - - - [34864, 2865, 1, 256] - - [53, 0.0] - - - [23552, 9984, 1, 256] - - [47, 0.0] - - - [21040, 2865, 1, 256] - - [47, 0.0] - - - [34048, 1281, 1, 256] - - [45, 0.0] - - - [23296, 10240, 1, 256] - - [53, 0.0] - - - [32768, 6144, 1, 256] - - [53, 0.0] - - - [25904, 2816, 1, 256] - - [53, 0.0] - - - [31232, 1024, 1, 256] - - [53, 0.0] - - - [27648, 3328, 1, 256] - - [47, 0.0] - - - [34864, 256, 1, 256] - - [47, 0.0] - - - [21248, 256, 1, 256] - - [47, 0.0] - - - [26416, 10240, 1, 256] - - [53, 0.0] - - - [27184, 3584, 1, 256] - - [53, 0.0] - - - [23296, 2048, 1, 256] - - [53, 0.0] - - - [34048, 512, 1, 256] - - [53, 0.0] - - - [21760, 2865, 1, 256] - - [47, 0.0] - - - [28672, 2816, 1, 256] - - [47, 0.0] - - - [28672, 4608, 1, 256] - - [53, 0.0] - - - [34560, 512, 1, 256] - - [53, 0.0] - - - [32768, 2865, 1, 256] - - [47, 0.0] - - - [30208, 6912, 1, 256] - - [47, 0.0] - - - [32512, 6144, 1, 256] - - [53, 0.0] - - - [24832, 3328, 1, 256] - - [47, 0.0] - - - [27392, 2816, 1, 256] - - [47, 0.0] - - - [32768, 8704, 1, 256] - - [53, 0.0] - - - [23552, 10240, 1, 256] - - [53, 0.0] - - - [32816, 9216, 1, 256] - - [47, 0.0] - - - [33024, 10240, 1, 256] - - [53, 0.0] - - - [34608, 256, 1, 256] - - [53, 0.0] - - - [20736, 3328, 1, 256] - - [47, 0.0] - - - [31232, 7680, 1, 256] - - [53, 0.0] - - - [22528, 512, 1, 256] - - [53, 0.0] - - - [30208, 2865, 1, 256] - - [47, 0.0] - - - [22272, 2304, 1, 256] - - [47, 0.0] - - - [32512, 2816, 1, 256] - - [47, 0.0] - - - [31488, 7936, 1, 256] - - [47, 0.0] - - - [28416, 2048, 1, 256] - - [53, 0.0] - - - [22784, 3329, 1, 256] - - [47, 0.0] - - - [23040, 2816, 1, 256] - - [47, 0.0] - - - [24320, 3328, 1, 256] - - [47, 0.0] - - - [24064, 1281, 1, 256] - - [44, 0.0] - - - [33072, 9728, 1, 256] - - [53, 0.0] - - - [29440, 10240, 1, 256] - - [53, 0.0] - - - [30208, 6656, 1, 256] - - [53, 0.0] - - - [32768, 3328, 1, 256] - - [47, 0.0] - - - [28416, 6144, 1, 256] - - [53, 0.0] - - - [27904, 4608, 1, 256] - - [53, 0.0] - - - [27184, 2816, 1, 256] - - [47, 0.0] - - - [29184, 1024, 1, 256] - - [53, 0.0] - - - [31744, 1536, 1, 256] - - [53, 0.0] - - - [28416, 10240, 1, 256] - - [53, 0.0] - - - [24368, 10240, 1, 256] - - [53, 0.0] - - - [27904, 3329, 1, 256] - - [47, 0.0] - - - [25344, 3328, 1, 256] - - [47, 0.0] - - - [29952, 6400, 1, 256] - - [47, 0.0] - - - [29440, 2048, 1, 256] - - [53, 0.0] - - - [28928, 1281, 1, 256] - - [44, 0.0] - - - [30208, 3329, 1, 256] - - [47, 0.0] - - - [23088, 9984, 1, 256] - - [47, 0.0] - - - [29184, 2816, 1, 256] - - [47, 0.0] - - - [22528, 2560, 1, 256] - - [53, 0.0] - - - [33328, 2816, 1, 256] - - [47, 0.0] - - - [26368, 256, 1, 256] - - [53, 0.0] - - - [22832, 10240, 1, 256] - - [53, 0.0] - - - [31792, 2816, 1, 256] - - [53, 0.0] - - - [24832, 2048, 1, 256] - - [53, 0.0] - - - [24880, 256, 1, 256] - - [53, 0.0] - - - [33840, 10240, 1, 256] - - [53, 0.0] - - - [33584, 9984, 1, 256] - - [47, 0.0] - - - [28672, 10240, 1, 256] - - [53, 0.0] - - - [24832, 256, 1, 256] - - [47, 0.0] - - - [31488, 2865, 1, 256] - - [47, 0.0] - - - [30720, 7424, 1, 256] - - [47, 0.0] - - - [33536, 2816, 1, 256] - - [47, 0.0] - - - [30000, 6400, 1, 256] - - [53, 0.0] - - - [20224, 1281, 1, 256] - - [45, 0.0] - - - [22832, 2816, 1, 256] - - [47, 0.0] - - - [25600, 6144, 1, 256] - - [53, 0.0] - - - [24320, 4352, 1, 256] - - [53, 0.0] - - - [32768, 10240, 1, 256] - - [53, 0.0] - - - [26880, 768, 1, 256] - - [47, 0.0] - - - [24576, 3329, 1, 256] - - [47, 0.0] - - - [27904, 3840, 1, 256] - - [47, 0.0] - - - [30256, 2816, 1, 256] - - [47, 0.0] - - - [23296, 1281, 1, 256] - - [50, 0.0] - - - [26880, 256, 1, 256] - - [53, 0.0] - - - [23344, 2816, 1, 256] - - [53, 0.0] - - - [33792, 2048, 1, 256] - - [53, 0.0] - - - [21504, 3329, 1, 256] - - [53, 0.0] - - - [20272, 256, 1, 256] - - [53, 0.0] - - - [32768, 1280, 1, 256] - - [47, 0.0] - - - [32256, 10240, 1, 256] - - [53, 0.0] - - - [27952, 2816, 1, 256] - - [47, 0.0] - - - [28928, 5376, 1, 256] - - [47, 0.0] - - - [20992, 6144, 1, 256] - - [53, 0.0] - - - [20224, 2048, 1, 256] - - [53, 0.0] - - - [33280, 10240, 1, 256] - - [53, 0.0] - - - [24064, 3329, 1, 256] - - [47, 0.0] - - - [32768, 9216, 1, 256] - - [53, 0.0] - - - [20016, 6912, 1, 256] - - [47, 0.0] - - - [22320, 10240, 1, 256] - - [47, 0.0] - - - [22784, 256, 1, 256] - - [53, 0.0] - - - [34816, 512, 1, 256] - - [53, 0.0] - - - [32048, 8704, 1, 256] - - [53, 0.0] - - - [29232, 5888, 1, 256] - - [53, 0.0] - - - [24064, 768, 1, 256] - - [47, 0.0] - - - [33792, 9984, 1, 256] - - [47, 0.0] - - - [32512, 3329, 1, 256] - - [47, 0.0] - - - [21504, 2048, 1, 256] - - [53, 0.0] - - - [28160, 2304, 1, 256] - - [47, 0.0] - - - [20784, 10240, 1, 256] - - [53, 0.0] - - - [20224, 7168, 1, 256] - - [53, 0.0] - - - [28976, 2865, 1, 256] - - [47, 0.0] - - - [21296, 2816, 1, 256] - - [53, 0.0] - - - [23552, 256, 1, 256] - - [47, 0.0] - - - [26160, 2865, 1, 256] - - [53, 0.0] - - - [23600, 2816, 1, 256] - - [47, 0.0] - - - [20480, 7424, 1, 256] - - [47, 0.0] - - - [28928, 3329, 1, 256] - - [47, 0.0] - - - [20784, 2816, 1, 256] - - [47, 0.0] - - - [25344, 256, 1, 256] - - [47, 0.0] - - - [20224, 10240, 1, 256] - - [53, 0.0] - - - [28672, 1280, 1, 256] - - [47, 0.0] - - - [29232, 256, 1, 256] - - [47, 0.0] - - - [28720, 2865, 1, 256] - - [53, 0.0] - - - [22016, 2816, 1, 256] - - [47, 0.0] - - - [25600, 1536, 1, 256] - - [53, 0.0] - - - [26112, 10240, 1, 256] - - [53, 0.0] - - - [27136, 10240, 1, 256] - - [53, 0.0] - - - [31744, 8192, 1, 256] - - [53, 0.0] - - - [24320, 10240, 1, 256] - - [53, 0.0] - - - [29952, 10240, 1, 256] - - [53, 0.0] - - - [23296, 9984, 1, 256] - - [47, 0.0] - - - [34560, 2304, 1, 256] - - [47, 0.0] - - - [32000, 2865, 1, 256] - - [47, 0.0] - - - [25088, 1024, 1, 256] - - [53, 0.0] - - - [20272, 10240, 1, 256] - - [53, 0.0] - - - [25344, 5376, 1, 256] - - [47, 0.0] - - - [21760, 3328, 1, 256] - - [47, 0.0] - - - [32768, 8960, 1, 256] - - [47, 0.0] - - - [29952, 3840, 1, 256] - - [47, 0.0] - - - [32512, 2865, 1, 256] - - [47, 0.0] - - - [23344, 2865, 1, 256] - - [47, 0.0] - - - [24576, 768, 1, 256] - - [47, 0.0] - - - [27648, 3584, 1, 256] - - [53, 0.0] - - - [27952, 4608, 1, 256] - - [53, 0.0] - - - [29440, 3584, 1, 256] - - [53, 0.0] - - - [34096, 512, 1, 256] - - [53, 0.0] - - - [32304, 256, 1, 256] - - [53, 0.0] - - - [21040, 2816, 1, 256] - - [53, 0.0] - - - [22784, 1024, 1, 256] - - [53, 0.0] - - - [22784, 2816, 1, 256] - - [47, 0.0] - - - [25856, 2816, 1, 256] - - [47, 0.0] - - - [23296, 6144, 1, 256] - - [53, 0.0] - - - [28160, 4608, 1, 256] - - [53, 0.0] - - - [25136, 1792, 1, 256] - - [53, 0.0] - - - [30208, 256, 1, 256] - - [47, 0.0] - - - [23808, 1281, 1, 256] - - [45, 0.0] - - - [26368, 2304, 1, 256] - - [47, 0.0] - - - [27648, 4352, 1, 256] - - [47, 0.0] - - - [31280, 7936, 1, 256] - - [53, 0.0] - - - [22320, 2865, 1, 256] - - [53, 0.0] - - - [22320, 2816, 1, 256] - - [53, 0.0] - - - [28720, 5120, 1, 256] - - [53, 0.0] - - - [22272, 1280, 1, 256] - - [47, 0.0] - - - [31232, 3328, 1, 256] - - [47, 0.0] - - - [29696, 2048, 1, 256] - - [53, 0.0] - - - [34048, 9984, 1, 256] - - [47, 0.0] - - - [28416, 1280, 1, 256] - - [47, 0.0] - - - [21504, 2816, 1, 256] - - [47, 0.0] - - - [33536, 2865, 1, 256] - - [47, 0.0] - - - [23552, 3840, 1, 256] - - [47, 0.0] - - - [31744, 256, 1, 256] - - [53, 0.0] - - - [25600, 1281, 1, 256] - - [45, 0.0] - - - [30768, 7168, 1, 256] - - [47, 0.0] - - - [23808, 3329, 1, 256] - - [47, 0.0] - - - [32256, 3328, 1, 256] - - [47, 0.0] - - - [23040, 9216, 1, 256] - - [53, 0.0] - - - [33024, 256, 1, 256] - - [47, 0.0] - - - [33584, 2865, 1, 256] - - [53, 0.0] - - - [21504, 8448, 1, 256] - - [47, 0.0] - - - [27904, 1281, 1, 256] - - [45, 0.0] - - - [34304, 10240, 1, 256] - - [53, 0.0] - - - [20992, 2865, 1, 256] - - [47, 0.0] - - - [22528, 8960, 1, 256] - - [47, 0.0] - - - [28928, 3328, 1, 256] - - [47, 0.0] - - - [21808, 2865, 1, 256] - - [53, 0.0] - - - [26416, 2816, 1, 256] - - [47, 0.0] - - - [27392, 3840, 1, 256] - - [47, 0.0] - - - [26112, 1281, 1, 256] - - [45, 0.0] - - - [34864, 10240, 1, 256] - - [53, 0.0] - - - [29440, 1536, 1, 256] - - [53, 0.0] - - - [30256, 10240, 1, 256] - - [53, 0.0] - - - [22528, 2816, 1, 256] - - [53, 0.0] - - - [28928, 2048, 1, 256] - - [53, 0.0] - - - [28976, 5376, 1, 256] - - [47, 0.0] - - - [20736, 7168, 1, 256] - - [53, 0.0] - - - [22016, 2865, 1, 256] - - [47, 0.0] - - - [26368, 1280, 1, 256] - - [47, 0.0] - - - [24624, 2865, 1, 256] - - [53, 0.0] - - - [23040, 3329, 1, 256] - - [47, 0.0] - - - [23296, 2865, 1, 256] - - [47, 0.0] - - - [28416, 3329, 1, 256] - - [47, 0.0] - - - [23040, 1281, 1, 256] - - [44, 0.0] - - - [21808, 8448, 1, 256] - - [47, 0.0] - - - [30720, 2865, 1, 256] - - [47, 0.0] - - - [22272, 8960, 1, 256] - - [47, 0.0] - - - [34864, 2816, 1, 256] - - [53, 0.0] - - - [31232, 7168, 1, 256] - - [53, 0.0] - - - [27696, 4352, 1, 256] - - [53, 0.0] - - - [21504, 256, 1, 256] - - [47, 0.0] - - - [28672, 1281, 1, 256] - - [50, 0.0] - - - [29696, 1792, 1, 256] - - [47, 0.0] - - - [28464, 5120, 1, 256] - - [53, 0.0] - - - [27136, 3329, 1, 256] - - [47, 0.0] - - - [21248, 3328, 1, 256] - - [47, 0.0] - - - [26880, 1281, 1, 256] - - [45, 0.0] - - - [32256, 8448, 1, 256] - - [47, 0.0] - - - [20480, 6144, 1, 256] - - [53, 0.0] - - - [34048, 2865, 1, 256] - - [47, 0.0] - - - [29696, 5888, 1, 256] - - [47, 0.0] - - - [28720, 256, 1, 256] - - [53, 0.0] - - - [33792, 2865, 1, 256] - - [47, 0.0] - - - [22784, 8960, 1, 256] - - [47, 0.0] - - - [30720, 256, 1, 256] - - [53, 0.0] - - - [23808, 512, 1, 256] - - [53, 0.0] - - - [33024, 9728, 1, 256] - - [53, 0.0] - - - [42624, 13824, 1, 384] - - [53, 0.0] - - - [33024, 3840, 1, 384] - - [47, 0.0] - - - [33408, 15360, 1, 384] - - [53, 0.0] - - - [44160, 8832, 1, 384] - - [47, 0.0] - - - [31488, 2688, 1, 384] - - [47, 0.0] - - - [39168, 3072, 1, 384] - - [53, 0.0] - - - [31872, 5760, 1, 384] - - [47, 0.0] - - - [36096, 13440, 1, 384] - - [47, 0.0] - - - [41856, 1152, 1, 384] - - [47, 0.0] - - - [32256, 1153, 1, 384] - - [50, 0.0] - - - [44160, 1153, 1, 384] - - [45, 0.0] - - - [31488, 7296, 1, 384] - - [47, 0.0] - - - [43008, 9216, 1, 384] - - [53, 0.0] - - - [31872, 6144, 1, 384] - - [53, 0.0] - - - [32640, 7297, 1, 384] - - [47, 0.0] - - - [33792, 1152, 1, 384] - - [47, 0.0] - - - [43776, 13441, 1, 384] - - [53, 0.0] - - - [36480, 1153, 1, 384] - - [50, 0.0] - - - [37632, 1152, 1, 384] - - [53, 0.0] - - - [37248, 8448, 1, 384] - - [53, 0.0] - - - [31872, 7297, 1, 384] - - [53, 0.0] - - - [41856, 7296, 1, 384] - - [47, 0.0] - - - [39936, 7297, 1, 384] - - [53, 0.0] - - - [35712, 1153, 1, 384] - - [44, 0.0] - - - [35712, 3072, 1, 384] - - [53, 0.0] - - - [31488, 1153, 1, 384] - - [50, 0.0] - - - [36480, 1152, 1, 384] - - [47, 0.0] - - - [36864, 9216, 1, 384] - - [53, 0.0] - - - [42624, 15360, 1, 384] - - [53, 0.0] - - - [37632, 8832, 1, 384] - - [47, 0.0] - - - [32640, 1153, 1, 384] - - [50, 0.0] - - - [36864, 3072, 1, 384] - - [53, 0.0] - - - [32640, 6912, 1, 384] - - [47, 0.0] - - - [31872, 13440, 1, 384] - - [53, 0.0] - - - [39168, 3840, 1, 384] - - [53, 0.0] - - - [39168, 10368, 1, 384] - - [53, 0.0] - - - [33792, 3072, 1, 384] - - [53, 0.0] - - - [39552, 1536, 1, 384] - - [53, 0.0] - - - [38784, 7296, 1, 384] - - [47, 0.0] - - - [40320, 1153, 1, 384] - - [50, 0.0] - - - [42240, 1152, 1, 384] - - [53, 0.0] - - - [43776, 14976, 1, 384] - - [53, 0.0] - - - [38784, 9216, 1, 384] - - [53, 0.0] - - - [33024, 4224, 1, 384] - - [47, 0.0] - - - [43776, 7297, 1, 384] - - [53, 0.0] - - - [34560, 9216, 1, 384] - - [53, 0.0] - - - [43392, 8064, 1, 384] - - [53, 0.0] - - - [34944, 7296, 1, 384] - - [53, 0.0] - - - [38400, 7296, 1, 384] - - [47, 0.0] - - - [41856, 6912, 1, 384] - - [53, 0.0] - - - [40704, 3072, 1, 384] - - [53, 0.0] - - - [41472, 12672, 1, 384] - - [47, 0.0] - - - [36864, 1920, 1, 384] - - [53, 0.0] - - - [43008, 1920, 1, 384] - - [47, 0.0] - - - [43008, 13824, 1, 384] - - [53, 0.0] - - - [31104, 13441, 1, 384] - - [47, 0.0] - - - [41472, 12288, 1, 384] - - [53, 0.0] - - - [31488, 7297, 1, 384] - - [53, 0.0] - - - [35712, 6912, 1, 384] - - [53, 0.0] - - - [40704, 5376, 1, 384] - - [47, 0.0] - - - [36480, 9216, 1, 384] - - [53, 0.0] - - - [38784, 13440, 1, 384] - - [47, 0.0] - - - [36096, 15360, 1, 384] - - [53, 0.0] - - - [41856, 15360, 1, 384] - - [53, 0.0] - - - [37632, 2688, 1, 384] - - [47, 0.0] - - - [33792, 4608, 1, 384] - - [53, 0.0] - - - [38400, 13440, 1, 384] - - [53, 0.0] - - - [31104, 3072, 1, 384] - - [53, 0.0] - - - [33792, 13440, 1, 384] - - [47, 0.0] - - - [34176, 5376, 1, 384] - - [47, 0.0] - - - [31872, 3072, 1, 384] - - [53, 0.0] - - - [33792, 1920, 1, 384] - - [53, 0.0] - - - [34560, 1153, 1, 384] - - [50, 0.0] - - - [43392, 15360, 1, 384] - - [53, 0.0] - - - [39168, 4224, 1, 384] - - [47, 0.0] - - - [43776, 1153, 1, 384] - - [50, 0.0] - - - [41472, 6528, 1, 384] - - [53, 0.0] - - - [42240, 1153, 1, 384] - - [50, 0.0] - - - [36480, 13441, 1, 384] - - [53, 0.0] - - - [31488, 5760, 1, 384] - - [47, 0.0] - - - [34560, 13440, 1, 384] - - [47, 0.0] - - - [32256, 3072, 1, 384] - - [53, 0.0] - - - [37632, 15360, 1, 384] - - [53, 0.0] - - - [43776, 8448, 1, 384] - - [53, 0.0] - - - [37248, 13440, 1, 384] - - [47, 0.0] - - - [34944, 13440, 1, 384] - - [47, 0.0] - - - [41088, 3072, 1, 384] - - [53, 0.0] - - - [43008, 14208, 1, 384] - - [47, 0.0] - - - [33792, 7296, 1, 384] - - [47, 0.0] - - - [43392, 8448, 1, 384] - - [47, 0.0] - - - [31104, 7297, 1, 384] - - [47, 0.0] - - - [31104, 2304, 1, 384] - - [53, 0.0] - - - [35712, 1152, 1, 384] - - [53, 0.0] - - - [39552, 13440, 1, 384] - - [53, 0.0] - - - [37632, 2304, 1, 384] - - [47, 0.0] - - - [31872, 1153, 1, 384] - - [45, 0.0] - - - [39552, 3072, 1, 384] - - [53, 0.0] - - - [36864, 15360, 1, 384] - - [53, 0.0] - - - [33408, 4608, 1, 384] - - [53, 0.0] - - - [43392, 7297, 1, 384] - - [47, 0.0] - - - [32256, 7296, 1, 384] - - [53, 0.0] - - - [41472, 7296, 1, 384] - - [47, 0.0] - - - [38016, 9216, 1, 384] - - [53, 0.0] - - - [38784, 1153, 1, 384] - - [45, 0.0] - - - [34944, 2688, 1, 384] - - [53, 0.0] - - - [36864, 1152, 1, 384] - - [53, 0.0] - - - [39168, 7297, 1, 384] - - [47, 0.0] - - - [33024, 768, 1, 384] - - [47, 0.0] - - - [34560, 13441, 1, 384] - - [53, 0.0] - - - [33792, 7680, 1, 384] - - [53, 0.0] - - - [36864, 1153, 1, 384] - - [50, 0.0] - - - [40320, 4992, 1, 384] - - [47, 0.0] - - - [31488, 13440, 1, 384] - - [47, 0.0] - - - [39552, 10752, 1, 384] - - [53, 0.0] - - - [36096, 1152, 1, 384] - - [53, 0.0] - - - [44160, 1152, 1, 384] - - [53, 0.0] - - - [37632, 9216, 1, 384] - - [53, 0.0] - - - [37248, 15360, 1, 384] - - [53, 0.0] - - - [34944, 5760, 1, 384] - - [47, 0.0] - - - [41088, 15360, 1, 384] - - [53, 0.0] - - - [41088, 11904, 1, 384] - - [53, 0.0] - - - [35328, 6528, 1, 384] - - [53, 0.0] - - - [32640, 15360, 1, 384] - - [53, 0.0] - - - [33024, 7297, 1, 384] - - [47, 0.0] - - - [31104, 1153, 1, 384] - - [45, 0.0] - - - [40704, 1153, 1, 384] - - [45, 0.0] - - - [42240, 13440, 1, 384] - - [53, 0.0] - - - [41472, 7297, 1, 384] - - [53, 0.0] - - - [33408, 3072, 1, 384] - - [53, 0.0] - - - [40704, 13440, 1, 384] - - [47, 0.0] - - - [39168, 7296, 1, 384] - - [47, 0.0] - - - [34176, 9216, 1, 384] - - [53, 0.0] - - - [35328, 15360, 1, 384] - - [53, 0.0] - - - [38400, 1152, 1, 384] - - [53, 0.0] - - - [37248, 3072, 1, 384] - - [53, 0.0] - - - [31488, 2304, 1, 384] - - [53, 0.0] - - - [40704, 1152, 1, 384] - - [47, 0.0] - - - [39168, 768, 1, 384] - - [53, 0.0] - - - [34944, 1153, 1, 384] - - [50, 0.0] - - - [39936, 13440, 1, 384] - - [47, 0.0] - - - [43008, 7297, 1, 384] - - [47, 0.0] - - - [33024, 15360, 1, 384] - - [53, 0.0] - - - [34176, 1920, 1, 384] - - [53, 0.0] - - - [40320, 15360, 1, 384] - - [53, 0.0] - - - [37632, 3072, 1, 384] - - [53, 0.0] - - - [40320, 11136, 1, 384] - - [47, 0.0] - - - [34944, 1152, 1, 384] - - [53, 0.0] - - - [44160, 14976, 1, 384] - - [47, 0.0] - - - [33792, 1536, 1, 384] - - [53, 0.0] - - - [38016, 13441, 1, 384] - - [53, 0.0] - - - [37632, 7296, 1, 384] - - [53, 0.0] - - - [41856, 6528, 1, 384] - - [47, 0.0] - - - [36096, 6912, 1, 384] - - [47, 0.0] - - - [39936, 15360, 1, 384] - - [50, 0.0] - - - [43776, 9216, 1, 384] - - [53, 0.0] - - - [38400, 9600, 1, 384] - - [47, 0.0] - - - [39552, 15360, 1, 384] - - [53, 0.0] - - - [37248, 2304, 1, 384] - - [47, 0.0] - - - [33792, 1153, 1, 384] - - [50, 0.0] - - - [42624, 1152, 1, 384] - - [53, 0.0] - - - [35328, 3072, 1, 384] - - [53, 0.0] - - - [37632, 13440, 1, 384] - - [47, 0.0] - - - [38400, 3072, 1, 384] - - [53, 0.0] - - - [32640, 1152, 1, 384] - - [47, 0.0] - - - [31872, 1152, 1, 384] - - [53, 0.0] - - - [40320, 3072, 1, 384] - - [53, 0.0] - - - [38016, 15360, 1, 384] - - [53, 0.0] - - - [35712, 9216, 1, 384] - - [53, 0.0] - - - [33024, 13441, 1, 384] - - [47, 0.0] - - - [36096, 3072, 1, 384] - - [53, 0.0] - - - [36864, 13440, 1, 384] - - [47, 0.0] - - - [33408, 13441, 1, 384] - - [47, 0.0] - - - [37248, 9216, 1, 384] - - [53, 0.0] - - - [31488, 1152, 1, 384] - - [53, 0.0] - - - [31488, 3072, 1, 384] - - [53, 0.0] - - - [35328, 1152, 1, 384] - - [47, 0.0] - - - [37248, 7297, 1, 384] - - [47, 0.0] - - - [34944, 6144, 1, 384] - - [53, 0.0] - - - [36480, 1536, 1, 384] - - [53, 0.0] - - - [39168, 15360, 1, 384] - - [53, 0.0] - - - [43392, 13441, 1, 384] - - [47, 0.0] - - - [42624, 1536, 1, 384] - - [53, 0.0] - - - [36480, 7296, 1, 384] - - [53, 0.0] - - - [33792, 9216, 1, 384] - - [53, 0.0] - - - [36096, 768, 1, 384] - - [53, 0.0] - - - [33408, 1536, 1, 384] - - [53, 0.0] - - - [31872, 13441, 1, 384] - - [47, 0.0] - - - [43008, 13440, 1, 384] - - [53, 0.0] - - - [33024, 1152, 1, 384] - - [47, 0.0] - - - [34560, 5376, 1, 384] - - [47, 0.0] - - - [32640, 3840, 1, 384] - - [47, 0.0] - - - [33408, 1153, 1, 384] - - [50, 0.0] - - - [32256, 1152, 1, 384] - - [47, 0.0] - - - [41856, 13440, 1, 384] - - [47, 0.0] - - - [43776, 2688, 1, 384] - - [53, 0.0] - - - [34560, 8832, 1, 384] - - [47, 0.0] - - - [32256, 6528, 1, 384] - - [47, 0.0] - - - [33408, 13440, 1, 384] - - [53, 0.0] - - - [36096, 7296, 1, 384] - - [47, 0.0] - - - [43776, 3072, 1, 384] - - [53, 0.0] - - - [38784, 7297, 1, 384] - - [47, 0.0] - - - [39936, 7296, 1, 384] - - [47, 0.0] - - - [37632, 8448, 1, 384] - - [53, 0.0] - - - [43392, 9216, 1, 384] - - [53, 0.0] - - - [41856, 13056, 1, 384] - - [47, 0.0] - - - [30720, 13441, 1, 384] - - [47, 0.0] - - - [36864, 7680, 1, 384] - - [53, 0.0] - - - [41472, 1152, 1, 384] - - [47, 0.0] - - - [39168, 13440, 1, 384] - - [47, 0.0] - - - [43776, 2304, 1, 384] - - [47, 0.0] - - - [34176, 15360, 1, 384] - - [53, 0.0] - - - [36096, 7297, 1, 384] - - [47, 0.0] - - - [33792, 4992, 1, 384] - - [47, 0.0] - - - [35712, 15360, 1, 384] - - [53, 0.0] - - - [39168, 9984, 1, 384] - - [47, 0.0] - - - [36096, 9216, 1, 384] - - [53, 0.0] - - - [43008, 1536, 1, 384] - - [53, 0.0] - - - [33408, 9216, 1, 384] - - [53, 0.0] - - - [40704, 7296, 1, 384] - - [53, 0.0] - - - [38016, 2688, 1, 384] - - [53, 0.0] - - - [39168, 13441, 1, 384] - - [47, 0.0] - - - [39168, 9216, 1, 384] - - [53, 0.0] - - - [38400, 15360, 1, 384] - - [53, 0.0] - - - [43392, 2304, 1, 384] - - [47, 0.0] - - - [38400, 13441, 1, 384] - - [53, 0.0] - - - [43008, 1152, 1, 384] - - [47, 0.0] - - - [39936, 4608, 1, 384] - - [53, 0.0] - - - [43392, 14592, 1, 384] - - [53, 0.0] - - - [34176, 13441, 1, 384] - - [53, 0.0] - - - [38784, 9984, 1, 384] - - [53, 0.0] - - - [44160, 13441, 1, 384] - - [47, 0.0] - - - [31488, 5376, 1, 384] - - [53, 0.0] - - - [39936, 13441, 1, 384] - - [47, 0.0] - - - [34176, 1152, 1, 384] - - [53, 0.0] - - - [32640, 3072, 1, 384] - - [53, 0.0] - - - [34560, 15360, 1, 384] - - [53, 0.0] - - - [34944, 15360, 1, 384] - - [53, 0.0] - - - [37632, 13441, 1, 384] - - [53, 0.0] - - - [40320, 5376, 1, 384] - - [47, 0.0] - - - [41856, 12672, 1, 384] - - [47, 0.0] - - - [34176, 4992, 1, 384] - - [53, 0.0] - - - [42624, 7297, 1, 384] - - [47, 0.0] - - - [41856, 1153, 1, 384] - - [50, 0.0] - - - [41472, 9216, 1, 384] - - [53, 0.0] - - - [40704, 2304, 1, 384] - - [53, 0.0] - - - [36864, 8064, 1, 384] - - [47, 0.0] - - - [40704, 5760, 1, 384] - - [53, 0.0] - - - [41088, 7297, 1, 384] - - [47, 0.0] - - - [38784, 1152, 1, 384] - - [47, 0.0] - - - [38784, 3072, 1, 384] - - [53, 0.0] - - - [34560, 2304, 1, 384] - - [53, 0.0] - - - [36096, 1153, 1, 384] - - [50, 0.0] - - - [35712, 13440, 1, 384] - - [53, 0.0] - - - [39936, 1152, 1, 384] - - [53, 0.0] - - - [43392, 14208, 1, 384] - - [47, 0.0] - - - [39552, 1153, 1, 384] - - [45, 0.0] - - - [35712, 6528, 1, 384] - - [47, 0.0] - - - [31104, 5376, 1, 384] - - [53, 0.0] - - - [31104, 9216, 1, 384] - - [53, 0.0] - - - [33024, 9216, 1, 384] - - [53, 0.0] - - - [39936, 11136, 1, 384] - - [47, 0.0] - - - [43008, 3072, 1, 384] - - [53, 0.0] - - - [41856, 768, 1, 384] - - [53, 0.0] - - - [43776, 1152, 1, 384] - - [47, 0.0] - - - [34176, 7297, 1, 384] - - [47, 0.0] - - - [38016, 7297, 1, 384] - - [47, 0.0] - - - [36480, 7680, 1, 384] - - [53, 0.0] - - - [38400, 7297, 1, 384] - - [53, 0.0] - - - [44160, 2688, 1, 384] - - [53, 0.0] - - - [33792, 15360, 1, 384] - - [53, 0.0] - - - [40704, 2688, 1, 384] - - [47, 0.0] - - - [38784, 3840, 1, 384] - - [47, 0.0] - - - [44160, 7296, 1, 384] - - [47, 0.0] - - - [41088, 2688, 1, 384] - - [53, 0.0] - - - [38016, 3072, 1, 384] - - [53, 0.0] - - - [42240, 7296, 1, 384] - - [53, 0.0] - - - [41856, 9216, 1, 384] - - [53, 0.0] - - - [32640, 13440, 1, 384] - - [53, 0.0] - - - [40320, 13441, 1, 384] - - [53, 0.0] - - - [36480, 13440, 1, 384] - - [53, 0.0] - - - [41856, 7297, 1, 384] - - [53, 0.0] - - - [41088, 7296, 1, 384] - - [47, 0.0] - - - [33408, 1152, 1, 384] - - [53, 0.0] - - - [43392, 1920, 1, 384] - - [47, 0.0] - - - [31104, 1920, 1, 384] - - [47, 0.0] - - - [31488, 15360, 1, 384] - - [53, 0.0] - - - [31872, 7296, 1, 384] - - [53, 0.0] - - - [43008, 7680, 1, 384] - - [53, 0.0] - - - [35328, 13440, 1, 384] - - [47, 0.0] - - - [43776, 15360, 1, 384] - - [53, 0.0] - - - [34944, 3072, 1, 384] - - [53, 0.0] - - - [37248, 1153, 1, 384] - - [45, 0.0] - - - [31104, 1152, 1, 384] - - [53, 0.0] - - - [34560, 7297, 1, 384] - - [53, 0.0] - - - [43776, 14592, 1, 384] - - [47, 0.0] - - - [33408, 7296, 1, 384] - - [47, 0.0] - - - [33024, 7296, 1, 384] - - [53, 0.0] - - - [33024, 13440, 1, 384] - - [53, 0.0] - - - [31104, 7296, 1, 384] - - [47, 0.0] - - - [42240, 9216, 1, 384] - - [53, 0.0] - - - [34944, 13441, 1, 384] - - [53, 0.0] - - - [33792, 7297, 1, 384] - - [47, 0.0] - - - [35328, 13441, 1, 384] - - [47, 0.0] - - - [34176, 7296, 1, 384] - - [53, 0.0] - - - [40320, 1920, 1, 384] - - [53, 0.0] - - - [31872, 15360, 1, 384] - - [53, 0.0] - - - [39168, 1153, 1, 384] - - [45, 0.0] - - - [31104, 4992, 1, 384] - - [47, 0.0] - - - [41088, 1152, 1, 384] - - [53, 0.0] - - - [39552, 10368, 1, 384] - - [47, 0.0] - - - [40704, 11520, 1, 384] - - [47, 0.0] - - - [36864, 7297, 1, 384] - - [53, 0.0] - - - [42240, 15360, 1, 384] - - [53, 0.0] - - - [34560, 1152, 1, 384] - - [47, 0.0] - - - [31104, 13440, 1, 384] - - [47, 0.0] - - - [31488, 9216, 1, 384] - - [53, 0.0] - - - [34176, 3072, 1, 384] - - [53, 0.0] - - - [41088, 1153, 1, 384] - - [45, 0.0] - - - [43392, 1153, 1, 384] - - [45, 0.0] - - - [42240, 6912, 1, 384] - - [47, 0.0] - - - [43008, 15360, 1, 384] - - [53, 0.0] - - - [42240, 7297, 1, 384] - - [53, 0.0] - - - [43776, 7296, 1, 384] - - [53, 0.0] - - - [35712, 7296, 1, 384] - - [47, 0.0] - - - [38400, 9216, 1, 384] - - [53, 0.0] - - - [39936, 9216, 1, 384] - - [53, 0.0] - - - [32256, 6144, 1, 384] - - [53, 0.0] - - - [42624, 7680, 1, 384] - - [53, 0.0] - - - [33408, 4224, 1, 384] - - [47, 0.0] - - - [38784, 768, 1, 384] - - [53, 0.0] - - - [38016, 7296, 1, 384] - - [53, 0.0] - - - [34560, 5760, 1, 384] - - [47, 0.0] - - - [34944, 7297, 1, 384] - - [47, 0.0] - - - [38016, 8832, 1, 384] - - [47, 0.0] - - - [39936, 1920, 1, 384] - - [47, 0.0] - - - [40320, 11520, 1, 384] - - [47, 0.0] - - - [32256, 7297, 1, 384] - - [47, 0.0] - - - [33792, 13441, 1, 384] - - [47, 0.0] - - - [41472, 3072, 1, 384] - - [53, 0.0] - - - [33024, 1153, 1, 384] - - [50, 0.0] - - - [36864, 7296, 1, 384] - - [47, 0.0] - - - [38016, 1153, 1, 384] - - [50, 0.0] - - - [40320, 7297, 1, 384] - - [53, 0.0] - - - [42624, 13441, 1, 384] - - [47, 0.0] - - - [43008, 13441, 1, 384] - - [47, 0.0] - - - [39552, 9216, 1, 384] - - [53, 0.0] - - - [35328, 9216, 1, 384] - - [53, 0.0] - - - [42624, 3072, 1, 384] - - [53, 0.0] - - - [40320, 13440, 1, 384] - - [53, 0.0] - - - [42240, 13441, 1, 384] - - [53, 0.0] - - - [39936, 10752, 1, 384] - - [53, 0.0] - - - [41472, 6144, 1, 384] - - [53, 0.0] - - - [36864, 1536, 1, 384] - - [53, 0.0] - - - [33408, 7297, 1, 384] - - [47, 0.0] - - - [31872, 2688, 1, 384] - - [53, 0.0] - - - [41472, 1153, 1, 384] - - [50, 0.0] - - - [38400, 1153, 1, 384] - - [45, 0.0] - - - [38400, 3456, 1, 384] - - [47, 0.0] - - - [41856, 13441, 1, 384] - - [53, 0.0] - - - [43392, 1152, 1, 384] - - [47, 0.0] - - - [39552, 4608, 1, 384] - - [53, 0.0] - - - [40704, 15360, 1, 384] - - [53, 0.0] - - - [42240, 3072, 1, 384] - - [53, 0.0] - - - [32640, 3456, 1, 384] - - [53, 0.0] - - - [35712, 768, 1, 384] - - [47, 0.0] - - - [31104, 15360, 1, 384] - - [53, 0.0] - - - [40704, 13441, 1, 384] - - [53, 0.0] - - - [32640, 7296, 1, 384] - - [53, 0.0] - - - [34176, 8448, 1, 384] - - [47, 0.0] - - - [32640, 13441, 1, 384] - - [53, 0.0] - - - [36864, 13441, 1, 384] - - [47, 0.0] - - - [34176, 13440, 1, 384] - - [47, 0.0] - - - [37248, 1152, 1, 384] - - [53, 0.0] - - - [44160, 7297, 1, 384] - - [47, 0.0] - - - [41088, 6144, 1, 384] - - [53, 0.0] - - - [39936, 1536, 1, 384] - - [53, 0.0] - - - [44160, 15360, 1, 384] - - [53, 0.0] - - - [35712, 7297, 1, 384] - - [47, 0.0] - - - [35328, 6144, 1, 384] - - [53, 0.0] - - - [42624, 7296, 1, 384] - - [47, 0.0] - - - [33408, 7680, 1, 384] - - [53, 0.0] - - - [41472, 13441, 1, 384] - - [53, 0.0] - - - [43776, 8832, 1, 384] - - [47, 0.0] - - - [32256, 15360, 1, 384] - - [53, 0.0] - - - [32256, 9216, 1, 384] - - [53, 0.0] - - - [31872, 9216, 1, 384] - - [53, 0.0] - - - [37248, 7296, 1, 384] - - [47, 0.0] - - - [40320, 1152, 1, 384] - - [47, 0.0] - - - [34560, 8448, 1, 384] - - [47, 0.0] - - - [38784, 3456, 1, 384] - - [53, 0.0] - - - [41472, 15360, 1, 384] - - [53, 0.0] - - - [41856, 3072, 1, 384] - - [53, 0.0] - - - [41088, 13441, 1, 384] - - [53, 0.0] - - - [39936, 1153, 1, 384] - - [45, 0.0] - - - [37248, 1920, 1, 384] - - [53, 0.0] - - - [39552, 7296, 1, 384] - - [47, 0.0] - - - [40320, 2304, 1, 384] - - [47, 0.0] - - - [34560, 2688, 1, 384] - - [47, 0.0] - - - [42240, 13056, 1, 384] - - [47, 0.0] - - - [40320, 9216, 1, 384] - - [53, 0.0] - - - [40704, 7297, 1, 384] - - [47, 0.0] - - - [43776, 13440, 1, 384] - - [47, 0.0] - - - [39936, 4992, 1, 384] - - [53, 0.0] - - - [42624, 13440, 1, 384] - - [47, 0.0] - - - [37632, 1153, 1, 384] - - [50, 0.0] - - - [33024, 3072, 1, 384] - - [53, 0.0] - - - [40704, 9216, 1, 384] - - [53, 0.0] - - - [42624, 1153, 1, 384] - - [50, 0.0] - - - [43392, 13440, 1, 384] - - [47, 0.0] - - - [36480, 3072, 1, 384] - - [53, 0.0] - - - [41088, 12288, 1, 384] - - [53, 0.0] - - - [39168, 1152, 1, 384] - - [47, 0.0] - - - [39936, 3072, 1, 384] - - [53, 0.0] - - - [35712, 13441, 1, 384] - - [53, 0.0] - - - [41088, 13440, 1, 384] - - [47, 0.0] - - - [43392, 3072, 1, 384] - - [53, 0.0] - - - [33792, 8064, 1, 384] - - [53, 0.0] - - - [32256, 13440, 1, 384] - - [53, 0.0] - - - [35328, 7297, 1, 384] - - [47, 0.0] - - - [40704, 11904, 1, 384] - - [47, 0.0] - - - [33024, 6912, 1, 384] - - [53, 0.0] - - - [38784, 15360, 1, 384] - - [53, 0.0] - - - [42240, 768, 1, 384] - - [47, 0.0] - - - [44160, 13440, 1, 384] - - [53, 0.0] - - - [39552, 7297, 1, 384] - - [47, 0.0] - - - [32640, 768, 1, 384] - - [53, 0.0] - - - [44160, 9216, 1, 384] - - [53, 0.0] - - - [32640, 6528, 1, 384] - - [47, 0.0] - - - [39552, 13441, 1, 384] - - [47, 0.0] - - - [31488, 13441, 1, 384] - - [53, 0.0] - - - [43008, 7296, 1, 384] - - [53, 0.0] - - - [41088, 5760, 1, 384] - - [47, 0.0] - - - [41472, 13440, 1, 384] - - [47, 0.0] - - - [43392, 7296, 1, 384] - - [53, 0.0] - - - [34944, 9216, 1, 384] - - [53, 0.0] - - - [43008, 1153, 1, 384] - - [50, 0.0] - - - [32640, 9216, 1, 384] - - [53, 0.0] - - - [36096, 13441, 1, 384] - - [47, 0.0] - - - [39552, 1152, 1, 384] - - [53, 0.0] - - - [37632, 7297, 1, 384] - - [53, 0.0] - - - [42624, 9216, 1, 384] - - [53, 0.0] - - - [43008, 8064, 1, 384] - - [47, 0.0] - - - [38784, 9600, 1, 384] - - [47, 0.0] - - - [37248, 8064, 1, 384] - - [53, 0.0] - - - [30720, 15360, 1, 384] - - [53, 0.0] - - - [38016, 13440, 1, 384] - - [53, 0.0] - - - [34944, 8832, 1, 384] - - [53, 0.0] - - - [37248, 13441, 1, 384] - - [53, 0.0] - - - [34560, 7296, 1, 384] - - [53, 0.0] - - - [44160, 3072, 1, 384] - - [53, 0.0] - - - [40320, 7296, 1, 384] - - [47, 0.0] - - - [34176, 2304, 1, 384] - - [47, 0.0] - - - [41088, 9216, 1, 384] - - [53, 0.0] - - - [34176, 1153, 1, 384] - - [50, 0.0] - - - [39552, 4224, 1, 384] - - [47, 0.0] - - - [38784, 13441, 1, 384] - - [47, 0.0] - - - [36480, 7297, 1, 384] - - [47, 0.0] - - - [32256, 3456, 1, 384] - - [47, 0.0] - - - [34176, 8064, 1, 384] - - [47, 0.0] - - - [36480, 15360, 1, 384] - - [53, 0.0] - - - [34560, 3072, 1, 384] - - [53, 0.0] - - - [35328, 7296, 1, 384] - - [53, 0.0] - - - [32256, 13441, 1, 384] - - [53, 0.0] - - - [38016, 1152, 1, 384] - - [47, 0.0] - - - [35328, 1153, 1, 384] - - [49, 0.0] - - - [23040, 7296, 1, 384] - - [47, 0.0] - - - [12672, 7296, 1, 384] - - [53, 0.0] - - - [4224, 4225, 1, 384] - - [47, 0.0] - - - [19968, 13440, 1, 384] - - [47, 0.0] - - - [16128, 3072, 1, 384] - - [53, 0.0] - - - [19968, 9216, 1, 384] - - [53, 0.0] - - - [24576, 13440, 1, 384] - - [47, 0.0] - - - [17280, 3072, 1, 384] - - [53, 0.0] - - - [16512, 9216, 1, 384] - - [53, 0.0] - - - [21120, 1536, 1, 384] - - [53, 0.0] - - - [18432, 13441, 1, 384] - - [47, 0.0] - - - [21120, 9216, 1, 384] - - [53, 0.0] - - - [27264, 3072, 1, 384] - - [53, 0.0] - - - [12288, 4608, 1, 384] - - [53, 0.0] - - - [22272, 5376, 1, 384] - - [53, 0.0] - - - [7296, 6912, 1, 384] - - [47, 0.0] - - - [26880, 9216, 1, 384] - - [53, 0.0] - - - [3072, 2688, 1, 384] - - [47, 0.0] - - - [16512, 2688, 1, 384] - - [53, 0.0] - - - [8064, 7680, 1, 384] - - [53, 0.0] - - - [22656, 1153, 1, 384] - - [45, 0.0] - - - [24960, 8064, 1, 384] - - [53, 0.0] - - - [23808, 9216, 1, 384] - - [53, 0.0] - - - [29568, 15360, 1, 384] - - [53, 0.0] - - - [1920, 1152, 1, 384] - - [50, 0.0] - - - [11136, 10752, 1, 384] - - [53, 0.0] - - - [25728, 1152, 1, 384] - - [53, 0.0] - - - [19584, 3072, 1, 384] - - [53, 0.0] - - - [3840, 1153, 1, 384] - - [47, 0.0] - - - [15360, 7296, 1, 384] - - [53, 0.0] - - - [13056, 12673, 1, 384] - - [47, 0.0] - - - [5376, 5377, 1, 384] - - [47, 0.0] - - - [28416, 13440, 1, 384] - - [47, 0.0] - - - [11904, 4224, 1, 384] - - [47, 0.0] - - - [24576, 10752, 1, 384] - - [53, 0.0] - - - [20352, 7297, 1, 384] - - [47, 0.0] - - - [16512, 7296, 1, 384] - - [47, 0.0] - - - [17280, 13441, 1, 384] - - [53, 0.0] - - - [24192, 10368, 1, 384] - - [47, 0.0] - - - [20352, 6528, 1, 384] - - [47, 0.0] - - - [1920, 1536, 1, 384] - - [47, 0.0] - - - [15744, 8064, 1, 384] - - [53, 0.0] - - - [13056, 3072, 1, 384] - - [53, 0.0] - - - [20352, 7296, 1, 384] - - [47, 0.0] - - - [10368, 1152, 1, 384] - - [47, 0.0] - - - [16128, 1152, 1, 384] - - [47, 0.0] - - - [13440, 7297, 1, 384] - - [47, 0.0] - - - [19200, 13441, 1, 384] - - [47, 0.0] - - - [13440, 13441, 1, 384] - - [47, 0.0] - - - [7680, 7297, 1, 384] - - [53, 0.0] - - - [27648, 14208, 1, 384] - - [47, 0.0] - - - [23424, 9216, 1, 384] - - [53, 0.0] - - - [24960, 1153, 1, 384] - - [45, 0.0] - - - [28032, 2304, 1, 384] - - [47, 0.0] - - - [30720, 3072, 1, 384] - - [53, 0.0] - - - [11904, 1152, 1, 384] - - [47, 0.0] - - - [24576, 3072, 1, 384] - - [53, 0.0] - - - [26112, 1153, 1, 384] - - [50, 0.0] - - - [10368, 10369, 1, 384] - - [53, 0.0] - - - [14976, 1536, 1, 384] - - [53, 0.0] - - - [11520, 7296, 1, 384] - - [47, 0.0] - - - [5376, 5376, 1, 384] - - [53, 0.0] - - - [28800, 7296, 1, 384] - - [47, 0.0] - - - [22656, 3072, 1, 384] - - [53, 0.0] - - - [11904, 7296, 1, 384] - - [53, 0.0] - - - [13824, 3072, 1, 384] - - [53, 0.0] - - - [21504, 13440, 1, 384] - - [53, 0.0] - - - [28800, 13440, 1, 384] - - [53, 0.0] - - - [13824, 7296, 1, 384] - - [47, 0.0] - - - [28416, 13441, 1, 384] - - [53, 0.0] - - - [20736, 7296, 1, 384] - - [53, 0.0] - - - [4992, 4608, 1, 384] - - [53, 0.0] - - - [21888, 1153, 1, 384] - - [45, 0.0] - - - [6912, 3072, 1, 384] - - [53, 0.0] - - - [7680, 7680, 1, 384] - - [53, 0.0] - - - [11904, 11905, 1, 384] - - [53, 0.0] - - - [9600, 1920, 1, 384] - - [47, 0.0] - - - [25728, 2688, 1, 384] - - [53, 0.0] - - - [29568, 3840, 1, 384] - - [53, 0.0] - - - [9984, 7297, 1, 384] - - [47, 0.0] - - - [13056, 2688, 1, 384] - - [47, 0.0] - - - [3456, 1920, 1, 384] - - [45, 0.0] - - - [19200, 1152, 1, 384] - - [53, 0.0] - - - [15744, 2304, 1, 384] - - [53, 0.0] - - - [17664, 7296, 1, 384] - - [47, 0.0] - - - [3072, 3072, 1, 384] - - [53, 0.0] - - - [21888, 7296, 1, 384] - - [53, 0.0] - - - [16128, 13440, 1, 384] - - [53, 0.0] - - - [23040, 1153, 1, 384] - - [49, 0.0] - - - [21504, 9216, 1, 384] - - [53, 0.0] - - - [21120, 4608, 1, 384] - - [53, 0.0] - - - [10368, 1153, 1, 384] - - [49, 0.0] - - - [29184, 13441, 1, 384] - - [47, 0.0] - - - [8832, 1536, 1, 384] - - [53, 0.0] - - - [30336, 3072, 1, 384] - - [53, 0.0] - - - [24192, 1153, 1, 384] - - [50, 0.0] - - - [16128, 2304, 1, 384] - - [47, 0.0] - - - [20736, 13440, 1, 384] - - [47, 0.0] - - - [24960, 7297, 1, 384] - - [53, 0.0] - - - [18048, 1536, 1, 384] - - [53, 0.0] - - - [19200, 5760, 1, 384] - - [53, 0.0] - - - [13440, 13056, 1, 384] - - [53, 0.0] - - - [6144, 1152, 1, 384] - - [53, 0.0] - - - [1920, 1920, 1, 384] - - [58, 0.0] - - - [18816, 5376, 1, 384] - - [53, 0.0] - - - [28800, 2688, 1, 384] - - [53, 0.0] - - - [20352, 3840, 1, 384] - - [47, 0.0] - - - [3840, 3841, 1, 384] - - [47, 0.0] - - - [17280, 768, 1, 384] - - [53, 0.0] - - - [21888, 2304, 1, 384] - - [53, 0.0] - - - [28416, 14592, 1, 384] - - [47, 0.0] - - - [18816, 3072, 1, 384] - - [53, 0.0] - - - [25344, 13440, 1, 384] - - [53, 0.0] - - - [20736, 6912, 1, 384] - - [53, 0.0] - - - [26880, 1152, 1, 384] - - [53, 0.0] - - - [29952, 3072, 1, 384] - - [53, 0.0] - - - [24960, 8448, 1, 384] - - [47, 0.0] - - - [15360, 8064, 1, 384] - - [53, 0.0] - - - [27648, 1920, 1, 384] - - [47, 0.0] - - - [3456, 2304, 1, 384] - - [47, 0.0] - - - [23040, 6528, 1, 384] - - [47, 0.0] - - - [14208, 1153, 1, 384] - - [47, 0.0] - - - [27648, 1153, 1, 384] - - [45, 0.0] - - - [1920, 1921, 1, 384] - - [53, 0.0] - - - [19584, 13441, 1, 384] - - [47, 0.0] - - - [8448, 3072, 1, 384] - - [53, 0.0] - - - [16512, 13441, 1, 384] - - [47, 0.0] - - - [4992, 768, 1, 384] - - [53, 0.0] - - - [28416, 14976, 1, 384] - - [47, 0.0] - - - [8448, 1152, 1, 384] - - [53, 0.0] - - - [20352, 9216, 1, 384] - - [53, 0.0] - - - [19584, 1153, 1, 384] - - [50, 0.0] - - - [20736, 768, 1, 384] - - [53, 0.0] - - - [28416, 2688, 1, 384] - - [47, 0.0] - - - [27264, 13440, 1, 384] - - [47, 0.0] - - - [16128, 7296, 1, 384] - - [53, 0.0] - - - [27648, 13440, 1, 384] - - [53, 0.0] - - - [26880, 13056, 1, 384] - - [47, 0.0] - - - [6528, 1920, 1, 384] - - [53, 0.0] - - - [20352, 13441, 1, 384] - - [53, 0.0] - - - [12288, 7297, 1, 384] - - [53, 0.0] - - - [21120, 7680, 1, 384] - - [53, 0.0] - - - [13824, 13441, 1, 384] - - [47, 0.0] - - - [26112, 13440, 1, 384] - - [53, 0.0] - - - [16512, 7297, 1, 384] - - [47, 0.0] - - - [6144, 5761, 1, 384] - - [53, 0.0] - - - [24960, 1152, 1, 384] - - [47, 0.0] - - - [9600, 9216, 1, 384] - - [53, 0.0] - - - [22272, 1153, 1, 384] - - [50, 0.0] - - - [24960, 2304, 1, 384] - - [53, 0.0] - - - [11136, 7296, 1, 384] - - [53, 0.0] - - - [28800, 3072, 1, 384] - - [53, 0.0] - - - [6912, 2688, 1, 384] - - [53, 0.0] - - - [25728, 3072, 1, 384] - - [53, 0.0] - - - [15744, 13441, 1, 384] - - [53, 0.0] - - - [18816, 7296, 1, 384] - - [53, 0.0] - - - [18816, 7297, 1, 384] - - [47, 0.0] - - - [13440, 13440, 1, 384] - - [47, 0.0] - - - [29184, 3456, 1, 384] - - [47, 0.0] - - - [8064, 768, 1, 384] - - [53, 0.0] - - - [4992, 4609, 1, 384] - - [53, 0.0] - - - [26496, 13056, 1, 384] - - [53, 0.0] - - - [21504, 4608, 1, 384] - - [53, 0.0] - - - [18048, 9216, 1, 384] - - [53, 0.0] - - - [14592, 13441, 1, 384] - - [47, 0.0] - - - [22656, 1152, 1, 384] - - [53, 0.0] - - - [14976, 3072, 1, 384] - - [53, 0.0] - - - [24960, 13441, 1, 384] - - [53, 0.0] - - - [768, 768, 1, 384] - - [34, 0.0] - - - [12672, 4992, 1, 384] - - [53, 0.0] - - - [11136, 3072, 1, 384] - - [53, 0.0] - - - [19584, 1152, 1, 384] - - [47, 0.0] - - - [16896, 3456, 1, 384] - - [47, 0.0] - - - [23040, 1152, 1, 384] - - [53, 0.0] - - - [6528, 6528, 1, 384] - - [47, 0.0] - - - [25344, 3072, 1, 384] - - [53, 0.0] - - - [2688, 1536, 1, 384] - - [68, 0.0] - - - [5760, 1536, 1, 384] - - [53, 0.0] - - - [6144, 5760, 1, 384] - - [47, 0.0] - - - [21504, 8064, 1, 384] - - [53, 0.0] - - - [12288, 12288, 1, 384] - - [53, 0.0] - - - [16128, 13441, 1, 384] - - [53, 0.0] - - - [25344, 8448, 1, 384] - - [47, 0.0] - - - [23808, 7297, 1, 384] - - [53, 0.0] - - - [15744, 7296, 1, 384] - - [53, 0.0] - - - [16896, 13441, 1, 384] - - [47, 0.0] - - - [15360, 1920, 1, 384] - - [53, 0.0] - - - [21504, 1152, 1, 384] - - [53, 0.0] - - - [6912, 1152, 1, 384] - - [53, 0.0] - - - [16512, 3072, 1, 384] - - [53, 0.0] - - - [28800, 1153, 1, 384] - - [45, 0.0] - - - [21888, 8064, 1, 384] - - [47, 0.0] - - - [20736, 7297, 1, 384] - - [47, 0.0] - - - [10752, 10753, 1, 384] - - [53, 0.0] - - - [8832, 7297, 1, 384] - - [47, 0.0] - - - [28032, 7297, 1, 384] - - [47, 0.0] - - - [23424, 9600, 1, 384] - - [53, 0.0] - - - [23040, 13440, 1, 384] - - [47, 0.0] - - - [26880, 13441, 1, 384] - - [47, 0.0] - - - [4224, 4224, 1, 384] - - [53, 0.0] - - - [9600, 9600, 1, 384] - - [53, 0.0] - - - [26112, 1152, 1, 384] - - [53, 0.0] - - - [29568, 3456, 1, 384] - - [53, 0.0] - - - [28032, 9216, 1, 384] - - [53, 0.0] - - - [27648, 9216, 1, 384] - - [53, 0.0] - - - [17664, 1153, 1, 384] - - [45, 0.0] - - - [12672, 12289, 1, 384] - - [53, 0.0] - - - [21888, 1152, 1, 384] - - [53, 0.0] - - - [21888, 9216, 1, 384] - - [53, 0.0] - - - [10752, 10369, 1, 384] - - [47, 0.0] - - - [22656, 7296, 1, 384] - - [53, 0.0] - - - [13440, 13057, 1, 384] - - [53, 0.0] - - - [10752, 1153, 1, 384] - - [49, 0.0] - - - [12672, 3072, 1, 384] - - [53, 0.0] - - - [23424, 13440, 1, 384] - - [47, 0.0] - - - [29952, 3840, 1, 384] - - [53, 0.0] - - - [18432, 1920, 1, 384] - - [47, 0.0] - - - [26112, 7297, 1, 384] - - [53, 0.0] - - - [18816, 1153, 1, 384] - - [45, 0.0] - - - [17664, 4224, 1, 384] - - [47, 0.0] - - - [11520, 11521, 1, 384] - - [47, 0.0] - - - [30720, 1920, 1, 384] - - [53, 0.0] - - - [15360, 13441, 1, 384] - - [53, 0.0] - - - [17664, 13441, 1, 384] - - [47, 0.0] - - - [26496, 3072, 1, 384] - - [53, 0.0] - - - [20736, 4224, 1, 384] - - [47, 0.0] - - - [18816, 13441, 1, 384] - - [47, 0.0] - - - [18048, 13441, 1, 384] - - [47, 0.0] - - - [20352, 3072, 1, 384] - - [53, 0.0] - - - [1152, 768, 1, 384] - - [42, 0.0] - - - [16896, 7296, 1, 384] - - [47, 0.0] - - - [28800, 9216, 1, 384] - - [53, 0.0] - - - [9600, 1152, 1, 384] - - [47, 0.0] - - - [29952, 1153, 1, 384] - - [45, 0.0] - - - [20736, 1153, 1, 384] - - [50, 0.0] - - - [19584, 5760, 1, 384] - - [47, 0.0] - - - [29568, 7296, 1, 384] - - [47, 0.0] - - - [7296, 3072, 1, 384] - - [53, 0.0] - - - [27264, 1152, 1, 384] - - [47, 0.0] - - - [12288, 4992, 1, 384] - - [53, 0.0] - - - [5760, 5376, 1, 384] - - [53, 0.0] - - - [30720, 1152, 1, 384] - - [53, 0.0] - - - [14208, 13441, 1, 384] - - [53, 0.0] - - - [21504, 7296, 1, 384] - - [53, 0.0] - - - [7296, 6913, 1, 384] - - [53, 0.0] - - - [23808, 6912, 1, 384] - - [53, 0.0] - - - [20352, 768, 1, 384] - - [53, 0.0] - - - [2688, 2688, 1, 384] - - [53, 0.0] - - - [13056, 12672, 1, 384] - - [53, 0.0] - - - [29568, 13440, 1, 384] - - [47, 0.0] - - - [11904, 1153, 1, 384] - - [45, 0.0] - - - [2688, 2689, 1, 384] - - [50, 0.0] - - - [9984, 9985, 1, 384] - - [53, 0.0] - - - [22272, 13440, 1, 384] - - [47, 0.0] - - - [30336, 15360, 1, 384] - - [53, 0.0] - - - [21504, 7680, 1, 384] - - [53, 0.0] - - - [24192, 13441, 1, 384] - - [53, 0.0] - - - [15360, 1536, 1, 384] - - [53, 0.0] - - - [24576, 7297, 1, 384] - - [47, 0.0] - - - [11136, 3456, 1, 384] - - [53, 0.0] - - - [9600, 1153, 1, 384] - - [47, 0.0] - - - [18048, 7297, 1, 384] - - [47, 0.0] - - - [6144, 1153, 1, 384] - - [49, 0.0] - - - [23040, 9600, 1, 384] - - [53, 0.0] - - - [26880, 1153, 1, 384] - - [45, 0.0] - - - [10752, 7297, 1, 384] - - [53, 0.0] - - - [6912, 6529, 1, 384] - - [47, 0.0] - - - [29184, 9216, 1, 384] - - [53, 0.0] - - - [20736, 9216, 1, 384] - - [53, 0.0] - - - [23808, 1152, 1, 384] - - [53, 0.0] - - - [11136, 1153, 1, 384] - - [47, 0.0] - - - [25344, 1152, 1, 384] - - [53, 0.0] - - - [25344, 13441, 1, 384] - - [47, 0.0] - - - [14976, 7296, 1, 384] - - [53, 0.0] - - - [14592, 13440, 1, 384] - - [53, 0.0] - - - [7680, 7681, 1, 384] - - [53, 0.0] - - - [29568, 768, 1, 384] - - [47, 0.0] - - - [5760, 1152, 1, 384] - - [47, 0.0] - - - [21888, 13441, 1, 384] - - [47, 0.0] - - - [17664, 768, 1, 384] - - [47, 0.0] - - - [25728, 11904, 1, 384] - - [53, 0.0] - - - [9984, 2688, 1, 384] - - [53, 0.0] - - - [28416, 1153, 1, 384] - - [45, 0.0] - - - [17664, 3072, 1, 384] - - [53, 0.0] - - - [23040, 7297, 1, 384] - - [47, 0.0] - - - [8448, 8448, 1, 384] - - [47, 0.0] - - - [4608, 4225, 1, 384] - - [53, 0.0] - - - [4224, 2688, 1, 384] - - [53, 0.0] - - - [3072, 1152, 1, 384] - - [47, 0.0] - - - [29184, 1152, 1, 384] - - [53, 0.0] - - - [13440, 3072, 1, 384] - - [53, 0.0] - - - [6912, 6913, 1, 384] - - [47, 0.0] - - - [18432, 13440, 1, 384] - - [47, 0.0] - - - [14208, 7296, 1, 384] - - [53, 0.0] - - - [5376, 768, 1, 384] - - [144, 0.0] - - - [29184, 7296, 1, 384] - - [47, 0.0] - - - [20352, 1152, 1, 384] - - [47, 0.0] - - - [2304, 1153, 1, 384] - - [47, 0.0] - - - [23808, 9984, 1, 384] - - [53, 0.0] - - - [8448, 8065, 1, 384] - - [53, 0.0] - - - [24576, 1152, 1, 384] - - [47, 0.0] - - - [1536, 1537, 1, 384] - - [69, 0.0] - - - [4224, 3072, 1, 384] - - [53, 0.0] - - - [19968, 7296, 1, 384] - - [53, 0.0] - - - [19200, 5376, 1, 384] - - [47, 0.0] - - - [4608, 1152, 1, 384] - - [53, 0.0] - - - [18432, 4992, 1, 384] - - [47, 0.0] - - - [26880, 7297, 1, 384] - - [47, 0.0] - - - [15744, 3072, 1, 384] - - [53, 0.0] - - - [22272, 7296, 1, 384] - - [47, 0.0] - - - [20352, 6912, 1, 384] - - [53, 0.0] - - - [26880, 13440, 1, 384] - - [47, 0.0] - - - [4224, 3840, 1, 384] - - [53, 0.0] - - - [23424, 13441, 1, 384] - - [53, 0.0] - - - [16512, 13440, 1, 384] - - [47, 0.0] - - - [21120, 1152, 1, 384] - - [47, 0.0] - - - [10368, 3072, 1, 384] - - [53, 0.0] - - - [28032, 13440, 1, 384] - - [47, 0.0] - - - [14208, 6528, 1, 384] - - [47, 0.0] - - - [768, 769, 1, 384] - - [34, 0.0] - - - [3456, 1152, 1, 384] - - [47, 0.0] - - - [12672, 1152, 1, 384] - - [53, 0.0] - - - [7680, 3072, 1, 384] - - [53, 0.0] - - - [19200, 2304, 1, 384] - - [53, 0.0] - - - [13056, 1153, 1, 384] - - [45, 0.0] - - - [27264, 1153, 1, 384] - - [45, 0.0] - - - [29568, 1153, 1, 384] - - [45, 0.0] - - - [11520, 11136, 1, 384] - - [53, 0.0] - - - [9216, 9216, 1, 384] - - [53, 0.0] - - - [18048, 1153, 1, 384] - - [45, 0.0] - - - [8064, 1152, 1, 384] - - [47, 0.0] - - - [22272, 7297, 1, 384] - - [47, 0.0] - - - [22272, 13441, 1, 384] - - [53, 0.0] - - - [22656, 2688, 1, 384] - - [47, 0.0] - - - [19584, 6144, 1, 384] - - [53, 0.0] - - - [8064, 7297, 1, 384] - - [47, 0.0] - - - [8064, 7681, 1, 384] - - [53, 0.0] - - - [23808, 7296, 1, 384] - - [47, 0.0] - - - [24960, 7296, 1, 384] - - [47, 0.0] - - - [14208, 6912, 1, 384] - - [47, 0.0] - - - [19968, 6528, 1, 384] - - [53, 0.0] - - - [28416, 7296, 1, 384] - - [53, 0.0] - - - [29952, 13440, 1, 384] - - [47, 0.0] - - - [17280, 7297, 1, 384] - - [47, 0.0] - - - [1536, 1152, 1, 384] - - [53, 0.0] - - - [8832, 1153, 1, 384] - - [50, 0.0] - - - [28032, 1153, 1, 384] - - [45, 0.0] - - - [2688, 2305, 1, 384] - - [53, 0.0] - - - [8064, 3072, 1, 384] - - [53, 0.0] - - - [28032, 3072, 1, 384] - - [53, 0.0] - - - [3840, 3456, 1, 384] - - [47, 0.0] - - - [21888, 1920, 1, 384] - - [47, 0.0] - - - [11904, 11520, 1, 384] - - [53, 0.0] - - - [9600, 9601, 1, 384] - - [53, 0.0] - - - [21120, 13440, 1, 384] - - [47, 0.0] - - - [19584, 2688, 1, 384] - - [53, 0.0] - - - [6912, 6528, 1, 384] - - [47, 0.0] - - - [29568, 1152, 1, 384] - - [53, 0.0] - - - [23808, 3072, 1, 384] - - [53, 0.0] - - - [18816, 4992, 1, 384] - - [53, 0.0] - - - [29952, 9216, 1, 384] - - [53, 0.0] - - - [22656, 13440, 1, 384] - - [53, 0.0] - - - [20352, 3456, 1, 384] - - [47, 0.0] - - - [3456, 1153, 1, 384] - - [50, 0.0] - - - [3840, 3457, 1, 384] - - [53, 0.0] - - - [15744, 8448, 1, 384] - - [53, 0.0] - - - [26112, 3072, 1, 384] - - [53, 0.0] - - - [28032, 14208, 1, 384] - - [53, 0.0] - - - [21504, 1536, 1, 384] - - [53, 0.0] - - - [11520, 768, 1, 384] - - [47, 0.0] - - - [6528, 6144, 1, 384] - - [53, 0.0] - - - [18432, 1153, 1, 384] - - [45, 0.0] - - - [3072, 1920, 1, 384] - - [47, 0.0] - - - [25344, 9216, 1, 384] - - [53, 0.0] - - - [30336, 7297, 1, 384] - - [53, 0.0] - - - [8832, 1152, 1, 384] - - [53, 0.0] - - - [26112, 9216, 1, 384] - - [53, 0.0] - - - [29952, 7296, 1, 384] - - [53, 0.0] - - - [11520, 11137, 1, 384] - - [47, 0.0] - - - [16896, 13440, 1, 384] - - [47, 0.0] - - - [29568, 13441, 1, 384] - - [53, 0.0] - - - [30336, 9216, 1, 384] - - [53, 0.0] - - - [2688, 1152, 1, 384] - - [53, 0.0] - - - [10368, 10368, 1, 384] - - [53, 0.0] - - - [25344, 11520, 1, 384] - - [53, 0.0] - - - [24576, 1920, 1, 384] - - [47, 0.0] - - - [11904, 4608, 1, 384] - - [53, 0.0] - - - [12672, 5376, 1, 384] - - [53, 0.0] - - - [11520, 3072, 1, 384] - - [53, 0.0] - - - [3072, 3073, 1, 384] - - [53, 0.0] - - - [24960, 11136, 1, 384] - - [47, 0.0] - - - [9984, 9600, 1, 384] - - [47, 0.0] - - - [19200, 2688, 1, 384] - - [47, 0.0] - - - [26496, 7296, 1, 384] - - [47, 0.0] - - - [23040, 3072, 1, 384] - - [53, 0.0] - - - [5760, 5761, 1, 384] - - [53, 0.0] - - - [5760, 5377, 1, 384] - - [47, 0.0] - - - [26880, 768, 1, 384] - - [53, 0.0] - - - [13824, 7297, 1, 384] - - [47, 0.0] - - - [13440, 7296, 1, 384] - - [47, 0.0] - - - [16128, 8448, 1, 384] - - [47, 0.0] - - - [24960, 3072, 1, 384] - - [53, 0.0] - - - [6144, 6144, 1, 384] - - [53, 0.0] - - - [27648, 13441, 1, 384] - - [53, 0.0] - - - [10368, 7297, 1, 384] - - [53, 0.0] - - - [22272, 2304, 1, 384] - - [47, 0.0] - - - [30720, 1153, 1, 384] - - [45, 0.0] - - - [24192, 13440, 1, 384] - - [47, 0.0] - - - [9984, 9984, 1, 384] - - [47, 0.0] - - - [29952, 1152, 1, 384] - - [53, 0.0] - - - [26112, 12672, 1, 384] - - [47, 0.0] - - - [8448, 7296, 1, 384] - - [47, 0.0] - - - [19584, 13440, 1, 384] - - [47, 0.0] - - - [21120, 1153, 1, 384] - - [45, 0.0] - - - [8832, 8449, 1, 384] - - [53, 0.0] - - - [28032, 13441, 1, 384] - - [53, 0.0] - - - [7680, 1153, 1, 384] - - [45, 0.0] - - - [19584, 9216, 1, 384] - - [53, 0.0] - - - [28800, 1152, 1, 384] - - [53, 0.0] - - - [29952, 768, 1, 384] - - [47, 0.0] - - - [12288, 1152, 1, 384] - - [53, 0.0] - - - [9600, 9217, 1, 384] - - [53, 0.0] - - - [14976, 13441, 1, 384] - - [53, 0.0] - - - [25344, 8832, 1, 384] - - [53, 0.0] - - - [18432, 4608, 1, 384] - - [53, 0.0] - - - [2304, 1920, 1, 384] - - [45, 0.0] - - - [11520, 4224, 1, 384] - - [47, 0.0] - - - [26496, 1153, 1, 384] - - [50, 0.0] - - - [28416, 2304, 1, 384] - - [47, 0.0] - - - [19200, 3072, 1, 384] - - [53, 0.0] - - - [26112, 7296, 1, 384] - - [47, 0.0] - - - [21504, 7297, 1, 384] - - [47, 0.0] - - - [4224, 1152, 1, 384] - - [47, 0.0] - - - [17664, 3840, 1, 384] - - [47, 0.0] - - - [6144, 1536, 1, 384] - - [53, 0.0] - - - [28032, 14592, 1, 384] - - [53, 0.0] - - - [8064, 8064, 1, 384] - - [47, 0.0] - - - [11136, 1152, 1, 384] - - [47, 0.0] - - - [13056, 7297, 1, 384] - - [47, 0.0] - - - [19968, 3456, 1, 384] - - [53, 0.0] - - - [25344, 7297, 1, 384] - - [47, 0.0] - - - [17280, 3840, 1, 384] - - [47, 0.0] - - - [28416, 1152, 1, 384] - - [47, 0.0] - - - [21120, 3072, 1, 384] - - [53, 0.0] - - - [28416, 7297, 1, 384] - - [53, 0.0] - - - [6528, 6529, 1, 384] - - [47, 0.0] - - - [26496, 9216, 1, 384] - - [53, 0.0] - - - [14592, 7296, 1, 384] - - [47, 0.0] - - - [14208, 1152, 1, 384] - - [47, 0.0] - - - [24576, 1536, 1, 384] - - [53, 0.0] - - - [18048, 7296, 1, 384] - - [53, 0.0] - - - [4608, 3072, 1, 384] - - [53, 0.0] - - - [28800, 14976, 1, 384] - - [47, 0.0] - - - [17664, 1152, 1, 384] - - [47, 0.0] - - - [24576, 7680, 1, 384] - - [53, 0.0] - - - [16896, 9216, 1, 384] - - [53, 0.0] - - - [20736, 3840, 1, 384] - - [53, 0.0] - - - [27264, 9216, 1, 384] - - [53, 0.0] - - - [21888, 3072, 1, 384] - - [53, 0.0] - - - [24576, 11136, 1, 384] - - [47, 0.0] - - - [14592, 1153, 1, 384] - - [50, 0.0] - - - [23424, 7296, 1, 384] - - [53, 0.0] - - - [22272, 3072, 1, 384] - - [53, 0.0] - - - [8832, 8832, 1, 384] - - [53, 0.0] - - - [8064, 7296, 1, 384] - - [53, 0.0] - - - [22656, 8832, 1, 384] - - [53, 0.0] - - - [22272, 2688, 1, 384] - - [47, 0.0] - - - [6528, 1152, 1, 384] - - [53, 0.0] - - - [8832, 8833, 1, 384] - - [53, 0.0] - - - [28800, 15360, 1, 384] - - [53, 0.0] - - - [23424, 1153, 1, 384] - - [50, 0.0] - - - [13440, 1152, 1, 384] - - [53, 0.0] - - - [10752, 10368, 1, 384] - - [53, 0.0] - - - [3456, 3456, 1, 384] - - [47, 0.0] - - - [4608, 4608, 1, 384] - - [53, 0.0] - - - [4224, 1153, 1, 384] - - [53, 0.0] - - - [12672, 2304, 1, 384] - - [53, 0.0] - - - [25728, 7297, 1, 384] - - [53, 0.0] - - - [5376, 1153, 1, 384] - - [53, 0.0] - - - [30720, 4992, 1, 384] - - [53, 0.0] - - - [27264, 7297, 1, 384] - - [47, 0.0] - - - [21504, 1920, 1, 384] - - [47, 0.0] - - - [11136, 11136, 1, 384] - - [53, 0.0] - - - [22656, 6144, 1, 384] - - [53, 0.0] - - - [26496, 13440, 1, 384] - - [53, 0.0] - - - [9216, 7296, 1, 384] - - [53, 0.0] - - - [17280, 7296, 1, 384] - - [53, 0.0] - - - [23040, 13441, 1, 384] - - [47, 0.0] - - - [23808, 13441, 1, 384] - - [47, 0.0] - - - [30336, 4224, 1, 384] - - [47, 0.0] - - - [6144, 1920, 1, 384] - - [47, 0.0] - - - [11904, 11904, 1, 384] - - [47, 0.0] - - - [30336, 13441, 1, 384] - - [47, 0.0] - - - [11904, 1536, 1, 384] - - [53, 0.0] - - - [24576, 9216, 1, 384] - - [53, 0.0] - - - [9984, 2304, 1, 384] - - [53, 0.0] - - - [18048, 4608, 1, 384] - - [53, 0.0] - - - [18432, 7297, 1, 384] - - [53, 0.0] - - - [11136, 3840, 1, 384] - - [47, 0.0] - - - [12288, 11904, 1, 384] - - [47, 0.0] - - - [19584, 7296, 1, 384] - - [53, 0.0] - - - [3072, 2689, 1, 384] - - [47, 0.0] - - - [2304, 2305, 1, 384] - - [47, 0.0] - - - [26496, 7297, 1, 384] - - [47, 0.0] - - - [15744, 1152, 1, 384] - - [53, 0.0] - - - [6912, 6912, 1, 384] - - [53, 0.0] - - - [4992, 3072, 1, 384] - - [53, 0.0] - - - [15744, 13440, 1, 384] - - [47, 0.0] - - - [2688, 2304, 1, 384] - - [53, 0.0] - - - [8448, 7297, 1, 384] - - [53, 0.0] - - - [25344, 11904, 1, 384] - - [47, 0.0] - - - [18432, 7296, 1, 384] - - [47, 0.0] - - - [8448, 8449, 1, 384] - - [47, 0.0] - - - [30720, 1536, 1, 384] - - [53, 0.0] - - - [9216, 1153, 1, 384] - - [45, 0.0] - - - [24192, 9216, 1, 384] - - [53, 0.0] - - - [25344, 2688, 1, 384] - - [47, 0.0] - - - [24576, 1153, 1, 384] - - [49, 0.0] - - - [14208, 7297, 1, 384] - - [47, 0.0] - - - [12672, 1920, 1, 384] - - [47, 0.0] - - - [4608, 4224, 1, 384] - - [47, 0.0] - - - [27264, 1536, 1, 384] - - [53, 0.0] - - - [24576, 13441, 1, 384] - - [47, 0.0] - - - [21504, 4992, 1, 384] - - [47, 0.0] - - - [21888, 4992, 1, 384] - - [53, 0.0] - - - [18432, 3072, 1, 384] - - [53, 0.0] - - - [19968, 6144, 1, 384] - - [53, 0.0] - - - [24192, 1536, 1, 384] - - [53, 0.0] - - - [9600, 7297, 1, 384] - - [47, 0.0] - - - [13824, 6528, 1, 384] - - [47, 0.0] - - - [2304, 2304, 1, 384] - - [53, 0.0] - - - [23424, 9984, 1, 384] - - [47, 0.0] - - - [18816, 1152, 1, 384] - - [53, 0.0] - - - [1152, 769, 1, 384] - - [39, 0.0] - - - [23424, 768, 1, 384] - - [53, 0.0] - - - [17280, 1153, 1, 384] - - [50, 0.0] - - - [9600, 2304, 1, 384] - - [53, 0.0] - - - [29184, 7297, 1, 384] - - [47, 0.0] - - - [26880, 3072, 1, 384] - - [53, 0.0] - - - [11520, 11520, 1, 384] - - [47, 0.0] - - - [23040, 6144, 1, 384] - - [53, 0.0] - - - [18048, 13440, 1, 384] - - [53, 0.0] - - - [30336, 1536, 1, 384] - - [53, 0.0] - - - [14976, 7680, 1, 384] - - [53, 0.0] - - - [14976, 1152, 1, 384] - - [47, 0.0] - - - [15360, 7680, 1, 384] - - [53, 0.0] - - - [28800, 13441, 1, 384] - - [47, 0.0] - - - [28032, 1920, 1, 384] - - [47, 0.0] - - - [16128, 2688, 1, 384] - - [47, 0.0] - - - [6144, 6145, 1, 384] - - [53, 0.0] - - - [10368, 7296, 1, 384] - - [53, 0.0] - - - [5760, 3072, 1, 384] - - [53, 0.0] - - - [24960, 9216, 1, 384] - - [53, 0.0] - - - [14592, 768, 1, 384] - - [47, 0.0] - - - [14208, 768, 1, 384] - - [47, 0.0] - - - [6912, 1153, 1, 384] - - [53, 0.0] - - - [21888, 13440, 1, 384] - - [53, 0.0] - - - [13056, 5760, 1, 384] - - [47, 0.0] - - - [12288, 1920, 1, 384] - - [47, 0.0] - - - [13056, 13056, 1, 384] - - [47, 0.0] - - - [6528, 1153, 1, 384] - - [47, 0.0] - - - [22272, 8448, 1, 384] - - [53, 0.0] - - - [7296, 1153, 1, 384] - - [45, 0.0] - - - [17280, 3456, 1, 384] - - [53, 0.0] - - - [27264, 13441, 1, 384] - - [53, 0.0] - - - [9216, 7297, 1, 384] - - [47, 0.0] - - - [4992, 4992, 1, 384] - - [53, 0.0] - - - [16128, 7297, 1, 384] - - [47, 0.0] - - - [20352, 13440, 1, 384] - - [47, 0.0] - - - [30336, 1153, 1, 384] - - [45, 0.0] - - - [13056, 7296, 1, 384] - - [53, 0.0] - - - [27648, 1152, 1, 384] - - [47, 0.0] - - - [13824, 6144, 1, 384] - - [53, 0.0] - - - [9216, 1920, 1, 384] - - [47, 0.0] - - - [17280, 13440, 1, 384] - - [53, 0.0] - - - [21888, 5376, 1, 384] - - [53, 0.0] - - - [3456, 3072, 1, 384] - - [53, 0.0] - - - [13440, 1153, 1, 384] - - [45, 0.0] - - - [24192, 7680, 1, 384] - - [53, 0.0] - - - [29952, 4224, 1, 384] - - [47, 0.0] - - - [8832, 3072, 1, 384] - - [53, 0.0] - - - [5760, 5760, 1, 384] - - [53, 0.0] - - - [23424, 6912, 1, 384] - - [53, 0.0] - - - [24192, 3072, 1, 384] - - [53, 0.0] - - - [18048, 3072, 1, 384] - - [53, 0.0] - - - [27264, 7296, 1, 384] - - [53, 0.0] - - - [11520, 3840, 1, 384] - - [47, 0.0] - - - [18432, 1536, 1, 384] - - [53, 0.0] - - - [11136, 10753, 1, 384] - - [53, 0.0] - - - [9600, 7296, 1, 384] - - [53, 0.0] - - - [26496, 13441, 1, 384] - - [47, 0.0] - - - [29568, 9216, 1, 384] - - [53, 0.0] - - - [25728, 7296, 1, 384] - - [53, 0.0] - - - [6528, 3072, 1, 384] - - [53, 0.0] - - - [18816, 9216, 1, 384] - - [53, 0.0] - - - [1920, 1153, 1, 384] - - [144, 0.0] - - - [1152, 1153, 1, 384] - - [160, 0.0] - - - [16896, 1153, 1, 384] - - [45, 0.0] - - - [4992, 1153, 1, 384] - - [53, 0.0] - - - [22656, 13441, 1, 384] - - [53, 0.0] - - - [9984, 1152, 1, 384] - - [47, 0.0] - - - [26496, 768, 1, 384] - - [53, 0.0] - - - [25344, 2304, 1, 384] - - [53, 0.0] - - - [14592, 6912, 1, 384] - - [53, 0.0] - - - [9216, 8833, 1, 384] - - [47, 0.0] - - - [19584, 7297, 1, 384] - - [47, 0.0] - - - [8448, 1153, 1, 384] - - [47, 0.0] - - - [21120, 7297, 1, 384] - - [53, 0.0] - - - [11520, 7297, 1, 384] - - [47, 0.0] - - - [12288, 7296, 1, 384] - - [47, 0.0] - - - [4224, 3841, 1, 384] - - [47, 0.0] - - - [9984, 9601, 1, 384] - - [47, 0.0] - - - [2304, 1152, 1, 384] - - [58, 0.0] - - - [21120, 7296, 1, 384] - - [53, 0.0] - - - [15360, 1153, 1, 384] - - [45, 0.0] - - - [27648, 3072, 1, 384] - - [53, 0.0] - - - [19200, 1153, 1, 384] - - [45, 0.0] - - - [28032, 1152, 1, 384] - - [53, 0.0] - - - [12672, 12288, 1, 384] - - [53, 0.0] - - - [22272, 5760, 1, 384] - - [53, 0.0] - - - [26496, 1152, 1, 384] - - [53, 0.0] - - - [26880, 7296, 1, 384] - - [53, 0.0] - - - [6528, 2304, 1, 384] - - [47, 0.0] - - - [9984, 7296, 1, 384] - - [47, 0.0] - - - [19968, 1152, 1, 384] - - [47, 0.0] - - - [10368, 9984, 1, 384] - - [53, 0.0] - - - [3840, 3840, 1, 384] - - [53, 0.0] - - - [5376, 1152, 1, 384] - - [47, 0.0] - - - [24192, 7296, 1, 384] - - [47, 0.0] - - - [14592, 3072, 1, 384] - - [53, 0.0] - - - [27648, 7297, 1, 384] - - [47, 0.0] - - - [23424, 1152, 1, 384] - - [53, 0.0] - - - [3456, 3457, 1, 384] - - [45, 0.0] - - - [13056, 2304, 1, 384] - - [53, 0.0] - - - [23808, 768, 1, 384] - - [53, 0.0] - - - [18048, 1152, 1, 384] - - [53, 0.0] - - - [28416, 9216, 1, 384] - - [53, 0.0] - - - [21888, 7297, 1, 384] - - [47, 0.0] - - - [25728, 12288, 1, 384] - - [53, 0.0] - - - [21120, 4224, 1, 384] - - [47, 0.0] - - - [20736, 3072, 1, 384] - - [53, 0.0] - - - [3840, 2688, 1, 384] - - [47, 0.0] - - - [29568, 7297, 1, 384] - - [53, 0.0] - - - [13824, 1153, 1, 384] - - [50, 0.0] - - - [15744, 1153, 1, 384] - - [47, 0.0] - - - [11136, 768, 1, 384] - - [53, 0.0] - - - [17664, 7297, 1, 384] - - [53, 0.0] - - - [24192, 7297, 1, 384] - - [53, 0.0] - - - [25344, 1153, 1, 384] - - [50, 0.0] - - - [30720, 4608, 1, 384] - - [53, 0.0] - - - [25728, 9216, 1, 384] - - [53, 0.0] - - - [29184, 1153, 1, 384] - - [45, 0.0] - - - [30336, 1152, 1, 384] - - [53, 0.0] - - - [24960, 13440, 1, 384] - - [47, 0.0] - - - [18432, 9216, 1, 384] - - [53, 0.0] - - - [15360, 13440, 1, 384] - - [53, 0.0] - - - [12288, 1536, 1, 384] - - [53, 0.0] - - - [8832, 8448, 1, 384] - - [53, 0.0] - - - [19968, 7297, 1, 384] - - [47, 0.0] - - - [19968, 3072, 1, 384] - - [53, 0.0] - - - [24960, 1920, 1, 384] - - [53, 0.0] - - - [15360, 1152, 1, 384] - - [47, 0.0] - - - [30720, 7296, 1, 384] - - [53, 0.0] - - - [14976, 1153, 1, 384] - - [44, 0.0] - - - [25344, 7296, 1, 384] - - [53, 0.0] - - - [16512, 8832, 1, 384] - - [53, 0.0] - - - [26112, 13441, 1, 384] - - [47, 0.0] - - - [22272, 1152, 1, 384] - - [53, 0.0] - - - [27648, 1536, 1, 384] - - [53, 0.0] - - - [15744, 1920, 1, 384] - - [53, 0.0] - - - [5760, 1153, 1, 384] - - [49, 0.0] - - - [29952, 13441, 1, 384] - - [47, 0.0] - - - [12672, 1153, 1, 384] - - [53, 0.0] - - - [13440, 2688, 1, 384] - - [47, 0.0] - - - [18816, 13440, 1, 384] - - [47, 0.0] - - - [22656, 9216, 1, 384] - - [53, 0.0] - - - [9216, 1152, 1, 384] - - [47, 0.0] - - - [20736, 1152, 1, 384] - - [47, 0.0] - - - [8832, 7296, 1, 384] - - [53, 0.0] - - - [15744, 7297, 1, 384] - - [53, 0.0] - - - [16512, 1153, 1, 384] - - [50, 0.0] - - - [29952, 7297, 1, 384] - - [47, 0.0] - - - [11136, 7297, 1, 384] - - [47, 0.0] - - - [9600, 3072, 1, 384] - - [53, 0.0] - - - [28800, 7297, 1, 384] - - [47, 0.0] - - - [27648, 13824, 1, 384] - - [53, 0.0] - - - [23808, 10368, 1, 384] - - [47, 0.0] - - - [13824, 13440, 1, 384] - - [47, 0.0] - - - [9216, 1536, 1, 384] - - [53, 0.0] - - - [23808, 1153, 1, 384] - - [45, 0.0] - - - [15360, 3072, 1, 384] - - [53, 0.0] - - - [12288, 3072, 1, 384] - - [53, 0.0] - - - [28416, 3072, 1, 384] - - [53, 0.0] - - - [30336, 13440, 1, 384] - - [53, 0.0] - - - [1152, 1152, 1, 384] - - [35, 0.0] - - - [21504, 3072, 1, 384] - - [53, 0.0] - - - [23040, 9216, 1, 384] - - [53, 0.0] - - - [22656, 7297, 1, 384] - - [47, 0.0] - - - [22656, 5760, 1, 384] - - [53, 0.0] - - - [12288, 11905, 1, 384] - - [47, 0.0] - - - [28032, 7296, 1, 384] - - [53, 0.0] - - - [29184, 3072, 1, 384] - - [53, 0.0] - - - [7680, 1152, 1, 384] - - [53, 0.0] - - - [16896, 7297, 1, 384] - - [47, 0.0] - - - [13056, 5376, 1, 384] - - [47, 0.0] - - - [5376, 4993, 1, 384] - - [53, 0.0] - - - [17280, 9216, 1, 384] - - [53, 0.0] - - - [8448, 8064, 1, 384] - - [47, 0.0] - - - [4608, 1153, 1, 384] - - [44, 0.0] - - - [19200, 9216, 1, 384] - - [53, 0.0] - - - [30720, 7297, 1, 384] - - [47, 0.0] - - - [13440, 5760, 1, 384] - - [53, 0.0] - - - [9984, 3072, 1, 384] - - [53, 0.0] - - - [29952, 15360, 1, 384] - - [53, 0.0] - - - [3840, 1152, 1, 384] - - [47, 0.0] - - - [10368, 9985, 1, 384] - - [47, 0.0] - - - [14592, 7297, 1, 384] - - [47, 0.0] - - - [3456, 3073, 1, 384] - - [53, 0.0] - - - [22272, 9216, 1, 384] - - [53, 0.0] - - - [8064, 8065, 1, 384] - - [47, 0.0] - - - [1536, 1536, 1, 384] - - [57, 0.0] - - - [30336, 4608, 1, 384] - - [53, 0.0] - - - [26112, 12288, 1, 384] - - [53, 0.0] - - - [11904, 11521, 1, 384] - - [47, 0.0] - - - [13440, 6144, 1, 384] - - [53, 0.0] - - - [19200, 13440, 1, 384] - - [47, 0.0] - - - [17280, 1152, 1, 384] - - [53, 0.0] - - - [23424, 3072, 1, 384] - - [53, 0.0] - - - [2304, 1921, 1, 384] - - [47, 0.0] - - - [12672, 7297, 1, 384] - - [47, 0.0] - - - [16896, 1152, 1, 384] - - [47, 0.0] - - - [18432, 1152, 1, 384] - - [47, 0.0] - - - [27264, 13824, 1, 384] - - [53, 0.0] - - - [10752, 1152, 1, 384] - - [47, 0.0] - - - [30336, 7296, 1, 384] - - [47, 0.0] - - - [11904, 3072, 1, 384] - - [53, 0.0] - - - [2304, 768, 1, 384] - - [47, 0.0] - - - [14592, 1152, 1, 384] - - [47, 0.0] - - - [20736, 13441, 1, 384] - - [53, 0.0] - - - [10752, 10752, 1, 384] - - [53, 0.0] - - - [23808, 13440, 1, 384] - - [53, 0.0] - - - [5376, 4992, 1, 384] - - [53, 0.0] - - - [10752, 3072, 1, 384] - - [53, 0.0] - - - [24576, 7296, 1, 384] - - [47, 0.0] - - - [7296, 7296, 1, 384] - - [53, 0.0] - - - [19200, 7296, 1, 384] - - [53, 0.0] - - - [25728, 8832, 1, 384] - - [47, 0.0] - - - [18048, 4224, 1, 384] - - [53, 0.0] - - - [4992, 1152, 1, 384] - - [47, 0.0] - - - [22272, 8832, 1, 384] - - [47, 0.0] - - - [21504, 1153, 1, 384] - - [45, 0.0] - - - [14208, 13440, 1, 384] - - [53, 0.0] - - - [10752, 7296, 1, 384] - - [47, 0.0] - - - [24192, 1152, 1, 384] - - [47, 0.0] - - - [7296, 1152, 1, 384] - - [47, 0.0] - - - [16128, 1153, 1, 384] - - [50, 0.0] - - - [19200, 7297, 1, 384] - - [53, 0.0] - - - [4992, 4993, 1, 384] - - [47, 0.0] - - - [12672, 12673, 1, 384] - - [47, 0.0] - - - [14208, 3072, 1, 384] - - [53, 0.0] - - - [23424, 6528, 1, 384] - - [53, 0.0] - - - [24576, 8064, 1, 384] - - [47, 0.0] - - - [6528, 6145, 1, 384] - - [53, 0.0] - - - [1920, 1537, 1, 384] - - [50, 0.0] - - - [21888, 8448, 1, 384] - - [47, 0.0] - - - [3072, 1536, 1, 384] - - [53, 0.0] - - - [7680, 7296, 1, 384] - - [53, 0.0] - - - [16896, 3072, 1, 384] - - [53, 0.0] - - - [24960, 11520, 1, 384] - - [53, 0.0] - - - [13824, 1152, 1, 384] - - [53, 0.0] - - - [25728, 1153, 1, 384] - - [45, 0.0] - - - [19968, 13441, 1, 384] - - [47, 0.0] - - - [13056, 13057, 1, 384] - - [53, 0.0] - - - [29184, 13440, 1, 384] - - [47, 0.0] - - - [23424, 7297, 1, 384] - - [47, 0.0] - - - [9216, 8832, 1, 384] - - [53, 0.0] - - - [11520, 1153, 1, 384] - - [47, 0.0] - - - [19968, 1153, 1, 384] - - [45, 0.0] - - - [14976, 13440, 1, 384] - - [53, 0.0] - - - [9216, 3072, 1, 384] - - [53, 0.0] - - - [24192, 10752, 1, 384] - - [53, 0.0] - - - [16128, 8832, 1, 384] - - [47, 0.0] - - - [9984, 1153, 1, 384] - - [50, 0.0] - - - [8064, 1153, 1, 384] - - [53, 0.0] - - - [12672, 12672, 1, 384] - - [53, 0.0] - - - [25728, 13441, 1, 384] - - [53, 0.0] - - - [11520, 1152, 1, 384] - - [53, 0.0] - - - [26496, 12672, 1, 384] - - [47, 0.0] - - - [1920, 768, 1, 384] - - [160, 0.0] - - - [20352, 1153, 1, 384] - - [45, 0.0] - - - [10368, 2688, 1, 384] - - [47, 0.0] - - - [6912, 2304, 1, 384] - - [47, 0.0] - - - [17664, 13440, 1, 384] - - [53, 0.0] - - - [17664, 9216, 1, 384] - - [53, 0.0] - - - [25728, 13440, 1, 384] - - [47, 0.0] - - - [10752, 3456, 1, 384] - - [53, 0.0] - - - [6144, 3072, 1, 384] - - [53, 0.0] - - - [9216, 9217, 1, 384] - - [53, 0.0] - - - [3840, 2304, 1, 384] - - [53, 0.0] - - - [12288, 12289, 1, 384] - - [53, 0.0] - - - [11136, 11137, 1, 384] - - [47, 0.0] - - - [11904, 7297, 1, 384] - - [53, 0.0] - - - [29568, 3072, 1, 384] - - [53, 0.0] - - - [12288, 1153, 1, 384] - - [50, 0.0] - - - [18816, 1920, 1, 384] - - [47, 0.0] - - - [13056, 1152, 1, 384] - - [47, 0.0] - - - [8448, 768, 1, 384] - - [47, 0.0] - - - [18816, 2304, 1, 384] - - [53, 0.0] - - - [5376, 3072, 1, 384] - - [53, 0.0] - - - [16512, 1152, 1, 384] - - [53, 0.0] - - - [27648, 7296, 1, 384] - - [47, 0.0] - - - [7296, 2688, 1, 384] - - [53, 0.0] - - - [29184, 15360, 1, 384] - - [53, 0.0] - - - [4608, 4609, 1, 384] - - [53, 0.0] - - - [7296, 7297, 1, 384] - - [53, 0.0] - - - [30720, 9216, 1, 384] - - [53, 0.0] - - - [16384, 3072, 1, 256] - - [53, 0.0] - - - [42496, 10240, 1, 256] - - [53, 0.0] - - - [20992, 7168, 1, 256] - - [53, 0.0] - - - [8960, 5632, 1, 256] - - [53, 0.0] - - - [4864, 256, 1, 256] - - [68, 0.0] - - - [23552, 3584, 1, 256] - - [53, 0.0] - - - [2560, 1281, 1, 256] - - [36, 0.0] - - - [7168, 1280, 1, 256] - - [53, 0.0] - - - [1536, 1153, 1, 384] - - [162, 0.0] - - - [18224, 256, 1, 256] - - [47, 0.0] - - - [13441, 128, 1, 384] - - [47, 0.0] - - - [10753, 128, 1, 384] - - [164, 0.0] - - - [12289, 128, 1, 384] - - [68, 0.0] - - - [385, 128, 1, 384] - - [158, 0.0] - - - [11136, 128, 1, 384] - - [170, 0.0] - - - [13440, 128, 1, 384] - - [68, 0.0] - - - [1153, 128, 1, 384] - - [65, 0.0] - - - [6145, 128, 1, 384] - - [132, 0.0] - - - [4225, 128, 1, 384] - - [139, 0.0] - - - [1537, 128, 1, 384] - - [133, 0.0] - - - [8064, 128, 1, 384] - - [153, 0.0] - - - [3072, 128, 1, 384] - - [56, 0.0] - - - [3457, 128, 1, 384] - - [62, 0.0] - - - [5760, 128, 1, 384] - - [165, 0.0] - - - [8449, 128, 1, 384] - - [53, 0.0] - - - [2305, 128, 1, 384] - - [65, 0.0] - - - [11520, 128, 1, 384] - - [141, 0.0] - - - [11521, 128, 1, 384] - - [144, 0.0] - - - [6528, 128, 1, 384] - - [41, 0.0] - - - [14208, 128, 1, 384] - - [47, 0.0] - - - [768, 128, 1, 384] - - [65, 0.0] - - - [12672, 128, 1, 384] - - [66, 0.0] - - - [9216, 128, 1, 384] - - [56, 0.0] - - - [8448, 128, 1, 384] - - [153, 0.0] - - - [6144, 128, 1, 384] - - [132, 0.0] - - - [2689, 128, 1, 384] - - [54, 0.0] - - - [4224, 128, 1, 384] - - [127, 0.0] - - - [9601, 128, 1, 384] - - [172, 0.0] - - - [13056, 128, 1, 384] - - [64, 0.0] - - - [8065, 128, 1, 384] - - [153, 0.0] - - - [2304, 128, 1, 384] - - [163, 0.0] - - - [8833, 128, 1, 384] - - [144, 0.0] - - - [13824, 128, 1, 384] - - [47, 0.0] - - - [7680, 128, 1, 384] - - [132, 0.0] - - - [3840, 128, 1, 384] - - [56, 0.0] - - - [1920, 128, 1, 384] - - [133, 0.0] - - - [5761, 128, 1, 384] - - [40, 0.0] - - - [7681, 128, 1, 384] - - [53, 0.0] - - - [4608, 128, 1, 384] - - [138, 0.0] - - - [10369, 128, 1, 384] - - [172, 0.0] - - - [3841, 128, 1, 384] - - [127, 0.0] - - - [7296, 128, 1, 384] - - [167, 0.0] - - - [7297, 128, 1, 384] - - [166, 0.0] - - - [10752, 128, 1, 384] - - [52, 0.0] - - - [1536, 128, 1, 384] - - [55, 0.0] - - - [11137, 128, 1, 384] - - [141, 0.0] - - - [2688, 128, 1, 384] - - [145, 0.0] - - - [4609, 128, 1, 384] - - [149, 0.0] - - - [6529, 128, 1, 384] - - [167, 0.0] - - - [11905, 128, 1, 384] - - [62, 0.0] - - - [6912, 128, 1, 384] - - [132, 0.0] - - - [769, 128, 1, 384] - - [54, 0.0] - - - [12288, 128, 1, 384] - - [67, 0.0] - - - [15360, 128, 1, 384] - - [40, 0.0] - - - [9600, 128, 1, 384] - - [68, 0.0] - - - [13057, 128, 1, 384] - - [64, 0.0] - - - [10368, 128, 1, 384] - - [172, 0.0] - - - [12673, 128, 1, 384] - - [144, 0.0] - - - [9217, 128, 1, 384] - - [58, 0.0] - - - [4993, 128, 1, 384] - - [168, 0.0] - - - [9984, 128, 1, 384] - - [136, 0.0] - - - [6913, 128, 1, 384] - - [42, 0.0] - - - [8832, 128, 1, 384] - - [66, 0.0] - - - [3073, 128, 1, 384] - - [56, 0.0] - - - [14976, 128, 1, 384] - - [169, 0.0] - - - [384, 128, 1, 384] - - [146, 0.0] - - - [5377, 128, 1, 384] - - [128, 0.0] - - - [1152, 128, 1, 384] - - [156, 0.0] - - - [9985, 128, 1, 384] - - [171, 0.0] - - - [14592, 128, 1, 384] - - [38, 0.0] - - - [4992, 128, 1, 384] - - [136, 0.0] - - - [3456, 128, 1, 384] - - [56, 0.0] - - - [1921, 128, 1, 384] - - [58, 0.0] - - - [5376, 128, 1, 384] - - [128, 0.0] - - - [11904, 128, 1, 384] - - [144, 0.0] - - - [44544, 2048, 1, 384] - - [53, 0.0] - - - [39552, 512, 1, 384] - - [53, 0.0] - - - [38016, 22145, 1, 384] - - [53, 0.0] - - - [39552, 23297, 1, 384] - - [47, 0.0] - - - [39552, 23681, 1, 384] - - [47, 0.0] - - - [36864, 2048, 1, 384] - - [53, 0.0] - - - [44544, 28673, 1, 384] - - [53, 0.0] - - - [43776, 512, 1, 384] - - [53, 0.0] - - - [43392, 1024, 1, 384] - - [53, 0.0] - - - [42240, 4096, 1, 384] - - [53, 0.0] - - - [42624, 26369, 1, 384] - - [47, 0.0] - - - [35328, 1024, 1, 384] - - [53, 0.0] - - - [36096, 384, 1, 384] - - [47, 0.0] - - - [38784, 4096, 1, 384] - - [53, 0.0] - - - [39552, 384, 1, 384] - - [47, 0.0] - - - [42240, 8192, 1, 384] - - [53, 0.0] - - - [42240, 25985, 1, 384] - - [53, 0.0] - - - [38016, 4096, 1, 384] - - [53, 0.0] - - - [39168, 4096, 1, 384] - - [53, 0.0] - - - [35328, 19457, 1, 384] - - [53, 0.0] - - - [43392, 2048, 1, 384] - - [53, 0.0] - - - [38400, 4096, 1, 384] - - [53, 0.0] - - - [35712, 1024, 1, 384] - - [53, 0.0] - - - [36480, 2048, 1, 384] - - [53, 0.0] - - - [40704, 512, 1, 384] - - [53, 0.0] - - - [36864, 20609, 1, 384] - - [53, 0.0] - - - [37632, 21761, 1, 384] - - [47, 0.0] - - - [38016, 2048, 1, 384] - - [53, 0.0] - - - [44160, 2048, 1, 384] - - [53, 0.0] - - - [35328, 384, 1, 384] - - [47, 0.0] - - - [43392, 384, 1, 384] - - [47, 0.0] - - - [39168, 512, 1, 384] - - [53, 0.0] - - - [38784, 1024, 1, 384] - - [53, 0.0] - - - [35328, 2048, 1, 384] - - [53, 0.0] - - - [44544, 8192, 1, 384] - - [53, 0.0] - - - [40704, 384, 1, 384] - - [53, 0.0] - - - [39936, 512, 1, 384] - - [53, 0.0] - - - [41472, 25217, 1, 384] - - [47, 0.0] - - - [42240, 2048, 1, 384] - - [53, 0.0] - - - [37632, 512, 1, 384] - - [53, 0.0] - - - [37248, 1024, 1, 384] - - [53, 0.0] - - - [42240, 26369, 1, 384] - - [47, 0.0] - - - [43776, 384, 1, 384] - - [53, 0.0] - - - [44160, 8192, 1, 384] - - [53, 0.0] - - - [39936, 1024, 1, 384] - - [53, 0.0] - - - [43392, 27137, 1, 384] - - [47, 0.0] - - - [39936, 384, 1, 384] - - [47, 0.0] - - - [41472, 25601, 1, 384] - - [53, 0.0] - - - [36864, 4096, 1, 384] - - [53, 0.0] - - - [43392, 8192, 1, 384] - - [53, 0.0] - - - [36096, 512, 1, 384] - - [53, 0.0] - - - [36480, 4096, 1, 384] - - [53, 0.0] - - - [40320, 512, 1, 384] - - [53, 0.0] - - - [41088, 4096, 1, 384] - - [53, 0.0] - - - [43776, 27521, 1, 384] - - [53, 0.0] - - - [35328, 19073, 1, 384] - - [47, 0.0] - - - [44160, 384, 1, 384] - - [53, 0.0] - - - [36864, 8192, 1, 384] - - [53, 0.0] - - - [41088, 2048, 1, 384] - - [53, 0.0] - - - [38016, 21761, 1, 384] - - [47, 0.0] - - - [41856, 1024, 1, 384] - - [53, 0.0] - - - [39552, 8192, 1, 384] - - [53, 0.0] - - - [37632, 4096, 1, 384] - - [53, 0.0] - - - [41856, 384, 1, 384] - - [53, 0.0] - - - [44160, 28289, 1, 384] - - [53, 0.0] - - - [43008, 26753, 1, 384] - - [53, 0.0] - - - [38400, 512, 1, 384] - - [53, 0.0] - - - [39168, 384, 1, 384] - - [53, 0.0] - - - [37632, 1024, 1, 384] - - [53, 0.0] - - - [44544, 4096, 1, 384] - - [53, 0.0] - - - [42240, 512, 1, 384] - - [53, 0.0] - - - [43008, 2048, 1, 384] - - [53, 0.0] - - - [36480, 20609, 1, 384] - - [47, 0.0] - - - [36864, 512, 1, 384] - - [53, 0.0] - - - [43008, 384, 1, 384] - - [47, 0.0] - - - [43392, 4096, 1, 384] - - [53, 0.0] - - - [38400, 22145, 1, 384] - - [47, 0.0] - - - [39936, 23681, 1, 384] - - [47, 0.0] - - - [36096, 19841, 1, 384] - - [53, 0.0] - - - [44544, 512, 1, 384] - - [53, 0.0] - - - [38400, 2048, 1, 384] - - [53, 0.0] - - - [41856, 25985, 1, 384] - - [47, 0.0] - - - [42624, 2048, 1, 384] - - [53, 0.0] - - - [38400, 1024, 1, 384] - - [53, 0.0] - - - [36480, 512, 1, 384] - - [53, 0.0] - - - [42624, 26753, 1, 384] - - [53, 0.0] - - - [43776, 27905, 1, 384] - - [47, 0.0] - - - [37248, 2048, 1, 384] - - [53, 0.0] - - - [35712, 19841, 1, 384] - - [53, 0.0] - - - [43392, 27521, 1, 384] - - [47, 0.0] - - - [43008, 1024, 1, 384] - - [53, 0.0] - - - [42624, 512, 1, 384] - - [53, 0.0] - - - [41472, 384, 1, 384] - - [47, 0.0] - - - [40704, 2048, 1, 384] - - [53, 0.0] - - - [36096, 2048, 1, 384] - - [53, 0.0] - - - [39936, 4096, 1, 384] - - [53, 0.0] - - - [40320, 2048, 1, 384] - - [53, 0.0] - - - [41088, 8192, 1, 384] - - [53, 0.0] - - - [35328, 8192, 1, 384] - - [53, 0.0] - - - [40320, 4096, 1, 384] - - [53, 0.0] - - - [41856, 512, 1, 384] - - [53, 0.0] - - - [39552, 4096, 1, 384] - - [53, 0.0] - - - [35712, 2048, 1, 384] - - [53, 0.0] - - - [39936, 24065, 1, 384] - - [53, 0.0] - - - [36480, 20225, 1, 384] - - [47, 0.0] - - - [38016, 1024, 1, 384] - - [53, 0.0] - - - [43008, 512, 1, 384] - - [53, 0.0] - - - [40704, 24833, 1, 384] - - [47, 0.0] - - - [37248, 4096, 1, 384] - - [53, 0.0] - - - [41856, 4096, 1, 384] - - [53, 0.0] - - - [41472, 512, 1, 384] - - [53, 0.0] - - - [39552, 2048, 1, 384] - - [53, 0.0] - - - [41088, 384, 1, 384] - - [53, 0.0] - - - [36480, 8192, 1, 384] - - [53, 0.0] - - - [37632, 2048, 1, 384] - - [53, 0.0] - - - [40704, 8192, 1, 384] - - [53, 0.0] - - - [36864, 20993, 1, 384] - - [53, 0.0] - - - [35328, 512, 1, 384] - - [53, 0.0] - - - [40320, 384, 1, 384] - - [53, 0.0] - - - [36096, 1024, 1, 384] - - [53, 0.0] - - - [42624, 8192, 1, 384] - - [53, 0.0] - - - [38784, 22529, 1, 384] - - [47, 0.0] - - - [44160, 4096, 1, 384] - - [53, 0.0] - - - [41472, 4096, 1, 384] - - [53, 0.0] - - - [36480, 1024, 1, 384] - - [53, 0.0] - - - [38784, 2048, 1, 384] - - [53, 0.0] - - - [44544, 1024, 1, 384] - - [53, 0.0] - - - [41088, 24833, 1, 384] - - [47, 0.0] - - - [36864, 384, 1, 384] - - [47, 0.0] - - - [43392, 512, 1, 384] - - [53, 0.0] - - - [39168, 8192, 1, 384] - - [53, 0.0] - - - [42624, 4096, 1, 384] - - [53, 0.0] - - - [40320, 24065, 1, 384] - - [53, 0.0] - - - [44160, 512, 1, 384] - - [53, 0.0] - - - [38016, 384, 1, 384] - - [53, 0.0] - - - [38016, 512, 1, 384] - - [53, 0.0] - - - [37248, 512, 1, 384] - - [53, 0.0] - - - [43776, 2048, 1, 384] - - [53, 0.0] - - - [35712, 8192, 1, 384] - - [53, 0.0] - - - [38400, 384, 1, 384] - - [47, 0.0] - - - [42240, 1024, 1, 384] - - [53, 0.0] - - - [35712, 19457, 1, 384] - - [53, 0.0] - - - [41856, 2048, 1, 384] - - [53, 0.0] - - - [41472, 1024, 1, 384] - - [53, 0.0] - - - [37632, 384, 1, 384] - - [53, 0.0] - - - [40704, 1024, 1, 384] - - [53, 0.0] - - - [43008, 27137, 1, 384] - - [53, 0.0] - - - [40704, 4096, 1, 384] - - [53, 0.0] - - - [36096, 20225, 1, 384] - - [53, 0.0] - - - [39936, 8192, 1, 384] - - [53, 0.0] - - - [38784, 384, 1, 384] - - [53, 0.0] - - - [38784, 8192, 1, 384] - - [53, 0.0] - - - [42624, 384, 1, 384] - - [53, 0.0] - - - [35712, 4096, 1, 384] - - [53, 0.0] - - - [37632, 8192, 1, 384] - - [53, 0.0] - - - [38784, 22913, 1, 384] - - [47, 0.0] - - - [36864, 1024, 1, 384] - - [53, 0.0] - - - [37248, 384, 1, 384] - - [47, 0.0] - - - [39168, 23297, 1, 384] - - [47, 0.0] - - - [40704, 24449, 1, 384] - - [47, 0.0] - - - [41472, 2048, 1, 384] - - [53, 0.0] - - - [44160, 27905, 1, 384] - - [53, 0.0] - - - [44160, 1024, 1, 384] - - [53, 0.0] - - - [36480, 384, 1, 384] - - [47, 0.0] - - - [42240, 384, 1, 384] - - [47, 0.0] - - - [44544, 28289, 1, 384] - - [47, 0.0] - - - [37248, 21377, 1, 384] - - [47, 0.0] - - - [36096, 4096, 1, 384] - - [53, 0.0] - - - [38784, 512, 1, 384] - - [53, 0.0] - - - [35712, 384, 1, 384] - - [47, 0.0] - - - [43776, 1024, 1, 384] - - [53, 0.0] - - - [41088, 25217, 1, 384] - - [47, 0.0] - - - [40320, 8192, 1, 384] - - [53, 0.0] - - - [39168, 22913, 1, 384] - - [47, 0.0] - - - [38400, 8192, 1, 384] - - [53, 0.0] - - - [41088, 512, 1, 384] - - [53, 0.0] - - - [42624, 1024, 1, 384] - - [53, 0.0] - - - [39168, 2048, 1, 384] - - [53, 0.0] - - - [43008, 4096, 1, 384] - - [53, 0.0] - - - [35712, 512, 1, 384] - - [53, 0.0] - - - [41856, 8192, 1, 384] - - [53, 0.0] - - - [43008, 8192, 1, 384] - - [53, 0.0] - - - [41472, 8192, 1, 384] - - [53, 0.0] - - - [41088, 1024, 1, 384] - - [53, 0.0] - - - [37248, 20993, 1, 384] - - [53, 0.0] - - - [44544, 384, 1, 384] - - [53, 0.0] - - - [36096, 8192, 1, 384] - - [53, 0.0] - - - [43776, 8192, 1, 384] - - [53, 0.0] - - - [41856, 25601, 1, 384] - - [47, 0.0] - - - [37632, 21377, 1, 384] - - [47, 0.0] - - - [40320, 24449, 1, 384] - - [47, 0.0] - - - [43776, 4096, 1, 384] - - [53, 0.0] - - - [35328, 4096, 1, 384] - - [53, 0.0] - - - [39552, 1024, 1, 384] - - [53, 0.0] - - - [38016, 8192, 1, 384] - - [53, 0.0] - - - [38400, 22529, 1, 384] - - [53, 0.0] - - - [39936, 2048, 1, 384] - - [53, 0.0] - - - [39168, 1024, 1, 384] - - [53, 0.0] - - - [37248, 8192, 1, 384] - - [53, 0.0] - - - [40320, 1024, 1, 384] - - [53, 0.0] - - - [26112, 1024, 1, 384] - - [53, 0.0] - - - [24192, 2048, 1, 384] - - [53, 0.0] - - - [13440, 5761, 1, 384] - - [53, 0.0] - - - [3456, 384, 1, 384] - - [150, 0.0] - - - [21888, 4096, 1, 384] - - [53, 0.0] - - - [384, 384, 1, 384] - - [65, 0.0] - - - [21120, 1024, 1, 384] - - [53, 0.0] - - - [30336, 4096, 1, 384] - - [53, 0.0] - - - [31488, 512, 1, 384] - - [53, 0.0] - - - [2304, 1793, 1, 384] - - [47, 0.0] - - - [16896, 9217, 1, 384] - - [53, 0.0] - - - [9216, 1024, 1, 384] - - [53, 0.0] - - - [29568, 1024, 1, 384] - - [53, 0.0] - - - [27264, 11393, 1, 384] - - [47, 0.0] - - - [33408, 17537, 1, 384] - - [47, 0.0] - - - [18816, 1024, 1, 384] - - [53, 0.0] - - - [5760, 1024, 1, 384] - - [47, 0.0] - - - [31104, 14849, 1, 384] - - [53, 0.0] - - - [18816, 4096, 1, 384] - - [53, 0.0] - - - [11136, 1024, 1, 384] - - [53, 0.0] - - - [17664, 9985, 1, 384] - - [47, 0.0] - - - [9216, 512, 1, 384] - - [47, 0.0] - - - [17664, 1024, 1, 384] - - [53, 0.0] - - - [17664, 512, 1, 384] - - [53, 0.0] - - - [31488, 384, 1, 384] - - [47, 0.0] - - - [15744, 8065, 1, 384] - - [53, 0.0] - - - [5760, 3841, 1, 384] - - [53, 0.0] - - - [24192, 1024, 1, 384] - - [53, 0.0] - - - [20352, 384, 1, 384] - - [53, 0.0] - - - [21888, 2048, 1, 384] - - [53, 0.0] - - - [7680, 2048, 1, 384] - - [53, 0.0] - - - [2688, 512, 1, 384] - - [141, 0.0] - - - [13056, 1024, 1, 384] - - [53, 0.0] - - - [22656, 14977, 1, 384] - - [53, 0.0] - - - [10752, 6785, 1, 384] - - [53, 0.0] - - - [6912, 2048, 1, 384] - - [53, 0.0] - - - [15360, 512, 1, 384] - - [53, 0.0] - - - [31104, 384, 1, 384] - - [53, 0.0] - - - [30720, 14465, 1, 384] - - [47, 0.0] - - - [17280, 2048, 1, 384] - - [53, 0.0] - - - [34176, 1024, 1, 384] - - [53, 0.0] - - - [16896, 2048, 1, 384] - - [53, 0.0] - - - [17664, 384, 1, 384] - - [53, 0.0] - - - [21504, 512, 1, 384] - - [53, 0.0] - - - [18048, 10369, 1, 384] - - [47, 0.0] - - - [15744, 1024, 1, 384] - - [53, 0.0] - - - [33408, 4096, 1, 384] - - [53, 0.0] - - - [11904, 4096, 1, 384] - - [53, 0.0] - - - [18816, 512, 1, 384] - - [47, 0.0] - - - [34944, 4096, 1, 384] - - [53, 0.0] - - - [13824, 2048, 1, 384] - - [53, 0.0] - - - [3840, 512, 1, 384] - - [173, 0.0] - - - [4992, 1024, 1, 384] - - [53, 0.0] - - - [11136, 7553, 1, 384] - - [47, 0.0] - - - [16512, 1024, 1, 384] - - [53, 0.0] - - - [17280, 9217, 1, 384] - - [53, 0.0] - - - [29184, 1024, 1, 384] - - [53, 0.0] - - - [18048, 512, 1, 384] - - [47, 0.0] - - - [6528, 384, 1, 384] - - [50, 0.0] - - - [28416, 1024, 1, 384] - - [53, 0.0] - - - [2688, 1153, 1, 384] - - [45, 0.0] - - - [34560, 18305, 1, 384] - - [47, 0.0] - - - [20736, 384, 1, 384] - - [47, 0.0] - - - [11520, 512, 1, 384] - - [53, 0.0] - - - [26112, 8192, 1, 384] - - [53, 0.0] - - - [31872, 384, 1, 384] - - [53, 0.0] - - - [24192, 512, 1, 384] - - [53, 0.0] - - - [19968, 2048, 1, 384] - - [53, 0.0] - - - [32256, 8192, 1, 384] - - [53, 0.0] - - - [11520, 384, 1, 384] - - [47, 0.0] - - - [1920, 1409, 1, 384] - - [53, 0.0] - - - [25728, 9857, 1, 384] - - [53, 0.0] - - - [9216, 5633, 1, 384] - - [53, 0.0] - - - [28032, 12161, 1, 384] - - [53, 0.0] - - - [28800, 8192, 1, 384] - - [53, 0.0] - - - [28416, 12161, 1, 384] - - [53, 0.0] - - - [23040, 15361, 1, 384] - - [53, 0.0] - - - [31488, 15617, 1, 384] - - [47, 0.0] - - - [22272, 14209, 1, 384] - - [53, 0.0] - - - [1536, 512, 1, 384] - - [41, 0.0] - - - [1152, 257, 1, 384] - - [133, 0.0] - - - [21120, 2048, 1, 384] - - [53, 0.0] - - - [32256, 16001, 1, 384] - - [47, 0.0] - - - [9600, 6017, 1, 384] - - [47, 0.0] - - - [32640, 384, 1, 384] - - [53, 0.0] - - - [34176, 512, 1, 384] - - [53, 0.0] - - - [10368, 512, 1, 384] - - [53, 0.0] - - - [21120, 384, 1, 384] - - [47, 0.0] - - - [29568, 4096, 1, 384] - - [53, 0.0] - - - [31872, 2048, 1, 384] - - [53, 0.0] - - - [8832, 384, 1, 384] - - [66, 0.0] - - - [4224, 384, 1, 384] - - [58, 0.0] - - - [33408, 8192, 1, 384] - - [53, 0.0] - - - [768, 257, 1, 384] - - [65, 0.0] - - - [10368, 6401, 1, 384] - - [53, 0.0] - - - [13824, 384, 1, 384] - - [47, 0.0] - - - [29568, 512, 1, 384] - - [53, 0.0] - - - [28032, 1024, 1, 384] - - [53, 0.0] - - - [19200, 384, 1, 384] - - [47, 0.0] - - - [23040, 2048, 1, 384] - - [53, 0.0] - - - [8448, 4481, 1, 384] - - [47, 0.0] - - - [22272, 14593, 1, 384] - - [47, 0.0] - - - [26496, 10241, 1, 384] - - [53, 0.0] - - - [19584, 384, 1, 384] - - [53, 0.0] - - - [4992, 3457, 1, 384] - - [53, 0.0] - - - [22656, 384, 1, 384] - - [47, 0.0] - - - [15360, 1024, 1, 384] - - [53, 0.0] - - - [7296, 2048, 1, 384] - - [53, 0.0] - - - [30720, 384, 1, 384] - - [53, 0.0] - - - [6144, 2177, 1, 384] - - [45, 0.0] - - - [30720, 14849, 1, 384] - - [53, 0.0] - - - [23424, 2048, 1, 384] - - [53, 0.0] - - - [5760, 384, 1, 384] - - [53, 0.0] - - - [6144, 2561, 1, 384] - - [53, 0.0] - - - [12672, 384, 1, 384] - - [53, 0.0] - - - [16128, 8065, 1, 384] - - [47, 0.0] - - - [10752, 7169, 1, 384] - - [53, 0.0] - - - [2304, 384, 1, 384] - - [161, 0.0] - - - [18816, 2048, 1, 384] - - [53, 0.0] - - - [22272, 4096, 1, 384] - - [53, 0.0] - - - [12672, 4993, 1, 384] - - [47, 0.0] - - - [12288, 512, 1, 384] - - [47, 0.0] - - - [13056, 4993, 1, 384] - - [53, 0.0] - - - [19584, 512, 1, 384] - - [53, 0.0] - - - [30336, 14465, 1, 384] - - [53, 0.0] - - - [5376, 3841, 1, 384] - - [47, 0.0] - - - [17664, 9601, 1, 384] - - [47, 0.0] - - - [29952, 2048, 1, 384] - - [53, 0.0] - - - [8832, 512, 1, 384] - - [53, 0.0] - - - [9984, 512, 1, 384] - - [53, 0.0] - - - [19200, 1024, 1, 384] - - [53, 0.0] - - - [24192, 8321, 1, 384] - - [47, 0.0] - - - [26112, 10241, 1, 384] - - [47, 0.0] - - - [17280, 9601, 1, 384] - - [47, 0.0] - - - [7296, 384, 1, 384] - - [47, 0.0] - - - [16512, 8449, 1, 384] - - [47, 0.0] - - - [11904, 4225, 1, 384] - - [47, 0.0] - - - [24576, 4096, 1, 384] - - [53, 0.0] - - - [6912, 2945, 1, 384] - - [50, 0.0] - - - [33024, 16769, 1, 384] - - [47, 0.0] - - - [24576, 8705, 1, 384] - - [53, 0.0] - - - [16128, 2048, 1, 384] - - [53, 0.0] - - - [13824, 6145, 1, 384] - - [53, 0.0] - - - [28800, 512, 1, 384] - - [53, 0.0] - - - [33792, 8192, 1, 384] - - [53, 0.0] - - - [27648, 11393, 1, 384] - - [47, 0.0] - - - [21888, 384, 1, 384] - - [53, 0.0] - - - [12672, 4096, 1, 384] - - [53, 0.0] - - - [23040, 14977, 1, 384] - - [47, 0.0] - - - [11904, 384, 1, 384] - - [47, 0.0] - - - [7680, 3713, 1, 384] - - [53, 0.0] - - - [24576, 8192, 1, 384] - - [53, 0.0] - - - [34176, 384, 1, 384] - - [47, 0.0] - - - [17664, 2048, 1, 384] - - [53, 0.0] - - - [29952, 4096, 1, 384] - - [53, 0.0] - - - [9984, 6017, 1, 384] - - [53, 0.0] - - - [33408, 2048, 1, 384] - - [153, 0.0] - - - [21120, 4096, 1, 384] - - [53, 0.0] - - - [34560, 4096, 1, 384] - - [53, 0.0] - - - [19200, 11521, 1, 384] - - [47, 0.0] - - - [21120, 13057, 1, 384] - - [47, 0.0] - - - [25728, 384, 1, 384] - - [47, 0.0] - - - [28800, 12929, 1, 384] - - [47, 0.0] - - - [20736, 1024, 1, 384] - - [53, 0.0] - - - [18816, 10753, 1, 384] - - [53, 0.0] - - - [34560, 8192, 1, 384] - - [53, 0.0] - - - [23040, 512, 1, 384] - - [53, 0.0] - - - [30336, 2048, 1, 384] - - [53, 0.0] - - - [17280, 512, 1, 384] - - [53, 0.0] - - - [19200, 2048, 1, 384] - - [53, 0.0] - - - [12288, 4225, 1, 384] - - [53, 0.0] - - - [15744, 7681, 1, 384] - - [53, 0.0] - - - [30720, 4096, 1, 384] - - [53, 0.0] - - - [10752, 384, 1, 384] - - [144, 0.0] - - - [15744, 512, 1, 384] - - [53, 0.0] - - - [24960, 384, 1, 384] - - [53, 0.0] - - - [768, 384, 1, 384] - - [133, 0.0] - - - [6912, 3329, 1, 384] - - [47, 0.0] - - - [8064, 512, 1, 384] - - [68, 0.0] - - - [26496, 384, 1, 384] - - [53, 0.0] - - - [24960, 4096, 1, 384] - - [53, 0.0] - - - [19584, 11905, 1, 384] - - [47, 0.0] - - - [16512, 8833, 1, 384] - - [47, 0.0] - - - [18816, 384, 1, 384] - - [47, 0.0] - - - [23808, 1024, 1, 384] - - [53, 0.0] - - - [16512, 384, 1, 384] - - [47, 0.0] - - - [8448, 4865, 1, 384] - - [53, 0.0] - - - [34944, 1024, 1, 384] - - [53, 0.0] - - - [29184, 4096, 1, 384] - - [53, 0.0] - - - [8832, 2048, 1, 384] - - [53, 0.0] - - - [9984, 1024, 1, 384] - - [53, 0.0] - - - [22272, 1024, 1, 384] - - [53, 0.0] - - - [14592, 6913, 1, 384] - - [53, 0.0] - - - [9216, 2048, 1, 384] - - [53, 0.0] - - - [7296, 1024, 1, 384] - - [53, 0.0] - - - [26880, 8192, 1, 384] - - [53, 0.0] - - - [26880, 10625, 1, 384] - - [47, 0.0] - - - [28800, 12545, 1, 384] - - [47, 0.0] - - - [18048, 1024, 1, 384] - - [53, 0.0] - - - [27264, 11009, 1, 384] - - [47, 0.0] - - - [12288, 2048, 1, 384] - - [53, 0.0] - - - [19200, 4096, 1, 384] - - [53, 0.0] - - - [32256, 384, 1, 384] - - [47, 0.0] - - - [9216, 5249, 1, 384] - - [47, 0.0] - - - [29952, 14081, 1, 384] - - [47, 0.0] - - - [7680, 384, 1, 384] - - [53, 0.0] - - - [19200, 11137, 1, 384] - - [47, 0.0] - - - [14976, 1024, 1, 384] - - [53, 0.0] - - - [25728, 1024, 1, 384] - - [53, 0.0] - - - [3456, 1921, 1, 384] - - [47, 0.0] - - - [21120, 13441, 1, 384] - - [47, 0.0] - - - [15360, 2048, 1, 384] - - [53, 0.0] - - - [34560, 512, 1, 384] - - [53, 0.0] - - - [31872, 8192, 1, 384] - - [53, 0.0] - - - [32640, 16769, 1, 384] - - [47, 0.0] - - - [26496, 1024, 1, 384] - - [53, 0.0] - - - [12672, 1024, 1, 384] - - [53, 0.0] - - - [3072, 384, 1, 384] - - [58, 0.0] - - - [31104, 4096, 1, 384] - - [53, 0.0] - - - [25344, 4096, 1, 384] - - [53, 0.0] - - - [4224, 2689, 1, 384] - - [53, 0.0] - - - [24576, 1024, 1, 384] - - [53, 0.0] - - - [8448, 512, 1, 384] - - [53, 0.0] - - - [1536, 1025, 1, 384] - - [69, 0.0] - - - [14208, 6145, 1, 384] - - [53, 0.0] - - - [27264, 384, 1, 384] - - [47, 0.0] - - - [34560, 1024, 1, 384] - - [53, 0.0] - - - [14976, 6913, 1, 384] - - [53, 0.0] - - - [21504, 2048, 1, 384] - - [53, 0.0] - - - [14208, 4096, 1, 384] - - [53, 0.0] - - - [14592, 4096, 1, 384] - - [53, 0.0] - - - [6528, 2561, 1, 384] - - [50, 0.0] - - - [34176, 18305, 1, 384] - - [53, 0.0] - - - [19968, 384, 1, 384] - - [47, 0.0] - - - [30720, 8192, 1, 384] - - [53, 0.0] - - - [14592, 512, 1, 384] - - [53, 0.0] - - - [25728, 2048, 1, 384] - - [53, 0.0] - - - [23424, 4096, 1, 384] - - [53, 0.0] - - - [27264, 2048, 1, 384] - - [53, 0.0] - - - [21504, 1024, 1, 384] - - [53, 0.0] - - - [30336, 384, 1, 384] - - [47, 0.0] - - - [2688, 1024, 1, 384] - - [53, 0.0] - - - [22656, 4096, 1, 384] - - [53, 0.0] - - - [20352, 2048, 1, 384] - - [53, 0.0] - - - [33408, 384, 1, 384] - - [47, 0.0] - - - [15360, 4096, 1, 384] - - [53, 0.0] - - - [22272, 512, 1, 384] - - [53, 0.0] - - - [14208, 384, 1, 384] - - [47, 0.0] - - - [32640, 512, 1, 384] - - [53, 0.0] - - - [23808, 512, 1, 384] - - [53, 0.0] - - - [24960, 1024, 1, 384] - - [53, 0.0] - - - [4608, 512, 1, 384] - - [66, 0.0] - - - [25344, 2048, 1, 384] - - [53, 0.0] - - - [11904, 1024, 1, 384] - - [53, 0.0] - - - [28416, 12545, 1, 384] - - [47, 0.0] - - - [14208, 6529, 1, 384] - - [47, 0.0] - - - [13824, 5761, 1, 384] - - [47, 0.0] - - - [26112, 9857, 1, 384] - - [47, 0.0] - - - [9600, 2048, 1, 384] - - [53, 0.0] - - - [33024, 1024, 1, 384] - - [53, 0.0] - - - [34944, 18689, 1, 384] - - [53, 0.0] - - - [13824, 512, 1, 384] - - [53, 0.0] - - - [26880, 384, 1, 384] - - [53, 0.0] - - - [15744, 384, 1, 384] - - [53, 0.0] - - - [29568, 8192, 1, 384] - - [53, 0.0] - - - [24960, 9089, 1, 384] - - [47, 0.0] - - - [28032, 2048, 1, 384] - - [53, 0.0] - - - [19968, 11905, 1, 384] - - [47, 0.0] - - - [6528, 2945, 1, 384] - - [45, 0.0] - - - [20352, 12289, 1, 384] - - [53, 0.0] - - - [5376, 512, 1, 384] - - [53, 0.0] - - - [5376, 3457, 1, 384] - - [47, 0.0] - - - [21504, 384, 1, 384] - - [47, 0.0] - - - [11520, 1024, 1, 384] - - [53, 0.0] - - - [3840, 1921, 1, 384] - - [53, 0.0] - - - [18432, 4096, 1, 384] - - [53, 0.0] - - - [28416, 2048, 1, 384] - - [53, 0.0] - - - [3456, 512, 1, 384] - - [53, 0.0] - - - [2688, 384, 1, 384] - - [36, 0.0] - - - [28032, 4096, 1, 384] - - [53, 0.0] - - - [16128, 384, 1, 384] - - [53, 0.0] - - - [33792, 17537, 1, 384] - - [47, 0.0] - - - [2688, 1793, 1, 384] - - [53, 0.0] - - - [27648, 1024, 1, 384] - - [53, 0.0] - - - [13440, 1024, 1, 384] - - [53, 0.0] - - - [28032, 8192, 1, 384] - - [53, 0.0] - - - [34560, 18689, 1, 384] - - [53, 0.0] - - - [16896, 512, 1, 384] - - [53, 0.0] - - - [13056, 2048, 1, 384] - - [53, 0.0] - - - [3072, 1537, 1, 384] - - [173, 0.0] - - - [3072, 512, 1, 384] - - [68, 0.0] - - - [25344, 9089, 1, 384] - - [47, 0.0] - - - [9600, 384, 1, 384] - - [53, 0.0] - - - [26880, 512, 1, 384] - - [53, 0.0] - - - [33024, 512, 1, 384] - - [53, 0.0] - - - [21888, 1024, 1, 384] - - [53, 0.0] - - - [18048, 384, 1, 384] - - [47, 0.0] - - - [16896, 4096, 1, 384] - - [53, 0.0] - - - [23808, 384, 1, 384] - - [47, 0.0] - - - [26496, 4096, 1, 384] - - [53, 0.0] - - - [20736, 13057, 1, 384] - - [53, 0.0] - - - [24576, 512, 1, 384] - - [53, 0.0] - - - [14592, 6529, 1, 384] - - [53, 0.0] - - - [6528, 512, 1, 384] - - [68, 0.0] - - - [22656, 14593, 1, 384] - - [47, 0.0] - - - [26112, 2048, 1, 384] - - [53, 0.0] - - - [25728, 9473, 1, 384] - - [47, 0.0] - - - [15744, 2048, 1, 384] - - [53, 0.0] - - - [31488, 1024, 1, 384] - - [53, 0.0] - - - [11136, 2048, 1, 384] - - [53, 0.0] - - - [4608, 2689, 1, 384] - - [47, 0.0] - - - [30720, 1024, 1, 384] - - [53, 0.0] - - - [1920, 512, 1, 384] - - [41, 0.0] - - - [25728, 8192, 1, 384] - - [53, 0.0] - - - [31104, 2048, 1, 384] - - [53, 0.0] - - - [3456, 1024, 1, 384] - - [47, 0.0] - - - [25344, 384, 1, 384] - - [53, 0.0] - - - [27264, 8192, 1, 384] - - [53, 0.0] - - - [16128, 4096, 1, 384] - - [53, 0.0] - - - [20736, 12673, 1, 384] - - [47, 0.0] - - - [4224, 2305, 1, 384] - - [50, 0.0] - - - [27648, 11777, 1, 384] - - [53, 0.0] - - - [6144, 512, 1, 384] - - [53, 0.0] - - - [24576, 2048, 1, 384] - - [53, 0.0] - - - [15360, 384, 1, 384] - - [53, 0.0] - - - [34944, 19073, 1, 384] - - [47, 0.0] - - - [33792, 384, 1, 384] - - [53, 0.0] - - - [15360, 7681, 1, 384] - - [53, 0.0] - - - [34176, 17921, 1, 384] - - [53, 0.0] - - - [10368, 1024, 1, 384] - - [53, 0.0] - - - [34176, 8192, 1, 384] - - [53, 0.0] - - - [34176, 2048, 1, 384] - - [53, 0.0] - - - [7680, 4097, 1, 384] - - [53, 0.0] - - - [10752, 1024, 1, 384] - - [53, 0.0] - - - [9984, 2048, 1, 384] - - [53, 0.0] - - - [5760, 2048, 1, 384] - - [53, 0.0] - - - [30336, 1024, 1, 384] - - [53, 0.0] - - - [23424, 384, 1, 384] - - [53, 0.0] - - - [13440, 5377, 1, 384] - - [47, 0.0] - - - [14592, 2048, 1, 384] - - [53, 0.0] - - - [31872, 4096, 1, 384] - - [53, 0.0] - - - [6528, 2048, 1, 384] - - [53, 0.0] - - - [8064, 384, 1, 384] - - [68, 0.0] - - - [31872, 16001, 1, 384] - - [47, 0.0] - - - [16896, 1024, 1, 384] - - [53, 0.0] - - - [15360, 7297, 1, 384] - - [53, 0.0] - - - [33792, 4096, 1, 384] - - [53, 0.0] - - - [16896, 384, 1, 384] - - [53, 0.0] - - - [29952, 1024, 1, 384] - - [53, 0.0] - - - [768, 512, 1, 384] - - [56, 0.0] - - - [24576, 384, 1, 384] - - [53, 0.0] - - - [9984, 384, 1, 384] - - [53, 0.0] - - - [28416, 4096, 1, 384] - - [53, 0.0] - - - [11904, 7937, 1, 384] - - [47, 0.0] - - - [22656, 512, 1, 384] - - [53, 0.0] - - - [32640, 16385, 1, 384] - - [47, 0.0] - - - [14592, 1024, 1, 384] - - [53, 0.0] - - - [29952, 13697, 1, 384] - - [47, 0.0] - - - [32640, 1024, 1, 384] - - [53, 0.0] - - - [24960, 512, 1, 384] - - [53, 0.0] - - - [24192, 384, 1, 384] - - [53, 0.0] - - - [10752, 512, 1, 384] - - [53, 0.0] - - - [25344, 8192, 1, 384] - - [53, 0.0] - - - [32256, 16385, 1, 384] - - [53, 0.0] - - - [18432, 10753, 1, 384] - - [53, 0.0] - - - [27648, 512, 1, 384] - - [53, 0.0] - - - [28800, 4096, 1, 384] - - [53, 0.0] - - - [13440, 512, 1, 384] - - [53, 0.0] - - - [22272, 2048, 1, 384] - - [53, 0.0] - - - [29184, 2048, 1, 384] - - [53, 0.0] - - - [29952, 8192, 1, 384] - - [53, 0.0] - - - [384, 385, 1, 384] - - [65, 0.0] - - - [33408, 17153, 1, 384] - - [47, 0.0] - - - [27264, 512, 1, 384] - - [53, 0.0] - - - [33792, 1024, 1, 384] - - [53, 0.0] - - - [12288, 384, 1, 384] - - [53, 0.0] - - - [4224, 1024, 1, 384] - - [47, 0.0] - - - [13056, 5377, 1, 384] - - [53, 0.0] - - - [9600, 5633, 1, 384] - - [53, 0.0] - - - [30336, 512, 1, 384] - - [53, 0.0] - - - [7680, 1024, 1, 384] - - [53, 0.0] - - - [14976, 384, 1, 384] - - [53, 0.0] - - - [11904, 512, 1, 384] - - [53, 0.0] - - - [16128, 512, 1, 384] - - [53, 0.0] - - - [16128, 8449, 1, 384] - - [53, 0.0] - - - [18432, 2048, 1, 384] - - [53, 0.0] - - - [32256, 1024, 1, 384] - - [53, 0.0] - - - [16896, 8833, 1, 384] - - [47, 0.0] - - - [11136, 7169, 1, 384] - - [53, 0.0] - - - [8832, 4865, 1, 384] - - [47, 0.0] - - - [13440, 4096, 1, 384] - - [53, 0.0] - - - [10752, 2048, 1, 384] - - [53, 0.0] - - - [27264, 1024, 1, 384] - - [53, 0.0] - - - [1536, 384, 1, 384] - - [34, 0.0] - - - [20352, 1024, 1, 384] - - [53, 0.0] - - - [30720, 512, 1, 384] - - [53, 0.0] - - - [16512, 512, 1, 384] - - [53, 0.0] - - - [20736, 4096, 1, 384] - - [53, 0.0] - - - [23424, 15745, 1, 384] - - [53, 0.0] - - - [24960, 2048, 1, 384] - - [53, 0.0] - - - [32256, 2048, 1, 384] - - [53, 0.0] - - - [10368, 384, 1, 384] - - [47, 0.0] - - - [14976, 7297, 1, 384] - - [53, 0.0] - - - [23040, 4096, 1, 384] - - [53, 0.0] - - - [16512, 4096, 1, 384] - - [53, 0.0] - - - [20736, 512, 1, 384] - - [53, 0.0] - - - [34560, 384, 1, 384] - - [53, 0.0] - - - [23040, 1024, 1, 384] - - [53, 0.0] - - - [5376, 384, 1, 384] - - [53, 0.0] - - - [11136, 512, 1, 384] - - [53, 0.0] - - - [19200, 512, 1, 384] - - [53, 0.0] - - - [19584, 11521, 1, 384] - - [47, 0.0] - - - [21504, 4096, 1, 384] - - [53, 0.0] - - - [25728, 4096, 1, 384] - - [53, 0.0] - - - [4992, 512, 1, 384] - - [68, 0.0] - - - [26880, 4096, 1, 384] - - [53, 0.0] - - - [31488, 15233, 1, 384] - - [47, 0.0] - - - [2304, 1409, 1, 384] - - [36, 0.0] - - - [28800, 1024, 1, 384] - - [53, 0.0] - - - [25344, 9473, 1, 384] - - [53, 0.0] - - - [13824, 4096, 1, 384] - - [53, 0.0] - - - [18048, 2048, 1, 384] - - [53, 0.0] - - - [13056, 512, 1, 384] - - [47, 0.0] - - - [31104, 8192, 1, 384] - - [53, 0.0] - - - [1152, 641, 1, 384] - - [155, 0.0] - - - [8064, 1024, 1, 384] - - [53, 0.0] - - - [7296, 512, 1, 384] - - [53, 0.0] - - - [12672, 4609, 1, 384] - - [53, 0.0] - - - [27264, 4096, 1, 384] - - [53, 0.0] - - - [11520, 2048, 1, 384] - - [53, 0.0] - - - [15744, 4096, 1, 384] - - [53, 0.0] - - - [19968, 512, 1, 384] - - [53, 0.0] - - - [5760, 2177, 1, 384] - - [47, 0.0] - - - [3840, 384, 1, 384] - - [160, 0.0] - - - [30336, 8192, 1, 384] - - [53, 0.0] - - - [28416, 8192, 1, 384] - - [53, 0.0] - - - [25344, 512, 1, 384] - - [53, 0.0] - - - [7296, 3713, 1, 384] - - [53, 0.0] - - - [28416, 384, 1, 384] - - [47, 0.0] - - - [19584, 2048, 1, 384] - - [53, 0.0] - - - [10368, 2048, 1, 384] - - [53, 0.0] - - - [33024, 4096, 1, 384] - - [53, 0.0] - - - [4224, 512, 1, 384] - - [53, 0.0] - - - [26496, 8192, 1, 384] - - [53, 0.0] - - - [768, 385, 1, 384] - - [55, 0.0] - - - [23040, 384, 1, 384] - - [47, 0.0] - - - [11520, 7937, 1, 384] - - [47, 0.0] - - - [28800, 384, 1, 384] - - [53, 0.0] - - - [8064, 4481, 1, 384] - - [47, 0.0] - - - [28032, 384, 1, 384] - - [53, 0.0] - - - [31104, 512, 1, 384] - - [53, 0.0] - - - [23808, 16129, 1, 384] - - [47, 0.0] - - - [29184, 384, 1, 384] - - [47, 0.0] - - - [9600, 512, 1, 384] - - [53, 0.0] - - - [26112, 512, 1, 384] - - [53, 0.0] - - - [31488, 8192, 1, 384] - - [53, 0.0] - - - [8448, 384, 1, 384] - - [53, 0.0] - - - [34944, 8192, 1, 384] - - [53, 0.0] - - - [4608, 3073, 1, 384] - - [53, 0.0] - - - [30720, 2048, 1, 384] - - [53, 0.0] - - - [34944, 512, 1, 384] - - [53, 0.0] - - - [27648, 8192, 1, 384] - - [53, 0.0] - - - [33024, 2048, 1, 384] - - [53, 0.0] - - - [26112, 4096, 1, 384] - - [53, 0.0] - - - [17280, 384, 1, 384] - - [53, 0.0] - - - [33024, 17153, 1, 384] - - [47, 0.0] - - - [14208, 2048, 1, 384] - - [53, 0.0] - - - [13440, 2048, 1, 384] - - [53, 0.0] - - - [1536, 641, 1, 384] - - [36, 0.0] - - - [8064, 4097, 1, 384] - - [53, 0.0] - - - [26496, 10625, 1, 384] - - [53, 0.0] - - - [33024, 384, 1, 384] - - [47, 0.0] - - - [26112, 384, 1, 384] - - [53, 0.0] - - - [23424, 15361, 1, 384] - - [53, 0.0] - - - [34944, 2048, 1, 384] - - [53, 0.0] - - - [32256, 512, 1, 384] - - [53, 0.0] - - - [23808, 15745, 1, 384] - - [53, 0.0] - - - [5760, 512, 1, 384] - - [53, 0.0] - - - [16128, 1024, 1, 384] - - [53, 0.0] - - - [31488, 4096, 1, 384] - - [53, 0.0] - - - [29568, 13313, 1, 384] - - [53, 0.0] - - - [18816, 11137, 1, 384] - - [53, 0.0] - - - [26496, 2048, 1, 384] - - [53, 0.0] - - - [1920, 384, 1, 384] - - [159, 0.0] - - - [31872, 1024, 1, 384] - - [53, 0.0] - - - [12672, 512, 1, 384] - - [53, 0.0] - - - [13056, 4096, 1, 384] - - [53, 0.0] - - - [17280, 1024, 1, 384] - - [53, 0.0] - - - [12288, 1024, 1, 384] - - [53, 0.0] - - - [1152, 512, 1, 384] - - [34, 0.0] - - - [31104, 15233, 1, 384] - - [53, 0.0] - - - [4608, 384, 1, 384] - - [47, 0.0] - - - [21888, 512, 1, 384] - - [53, 0.0] - - - [33408, 1024, 1, 384] - - [53, 0.0] - - - [8448, 2048, 1, 384] - - [53, 0.0] - - - [7296, 3329, 1, 384] - - [45, 0.0] - - - [10368, 6785, 1, 384] - - [47, 0.0] - - - [8832, 1024, 1, 384] - - [53, 0.0] - - - [31104, 1024, 1, 384] - - [53, 0.0] - - - [11520, 7553, 1, 384] - - [53, 0.0] - - - [34176, 4096, 1, 384] - - [53, 0.0] - - - [20352, 512, 1, 384] - - [53, 0.0] - - - [18432, 512, 1, 384] - - [53, 0.0] - - - [31488, 2048, 1, 384] - - [53, 0.0] - - - [9984, 6401, 1, 384] - - [53, 0.0] - - - [6144, 2048, 1, 384] - - [53, 0.0] - - - [22656, 2048, 1, 384] - - [53, 0.0] - - - [2304, 512, 1, 384] - - [66, 0.0] - - - [21504, 13441, 1, 384] - - [47, 0.0] - - - [1920, 1025, 1, 384] - - [49, 0.0] - - - [24960, 8705, 1, 384] - - [53, 0.0] - - - [16512, 2048, 1, 384] - - [53, 0.0] - - - [26880, 11009, 1, 384] - - [53, 0.0] - - - [32256, 4096, 1, 384] - - [53, 0.0] - - - [14976, 2048, 1, 384] - - [53, 0.0] - - - [21120, 512, 1, 384] - - [53, 0.0] - - - [31872, 512, 1, 384] - - [53, 0.0] - - - [8064, 2048, 1, 384] - - [53, 0.0] - - - [3072, 1024, 1, 384] - - [68, 0.0] - - - [23808, 2048, 1, 384] - - [53, 0.0] - - - [12672, 2048, 1, 384] - - [53, 0.0] - - - [19968, 4096, 1, 384] - - [53, 0.0] - - - [14976, 512, 1, 384] - - [53, 0.0] - - - [25344, 1024, 1, 384] - - [53, 0.0] - - - [31872, 15617, 1, 384] - - [53, 0.0] - - - [20352, 12673, 1, 384] - - [47, 0.0] - - - [11136, 384, 1, 384] - - [50, 0.0] - - - [32640, 8192, 1, 384] - - [53, 0.0] - - - [28800, 2048, 1, 384] - - [53, 0.0] - - - [22656, 1024, 1, 384] - - [53, 0.0] - - - [17280, 4096, 1, 384] - - [53, 0.0] - - - [17664, 4096, 1, 384] - - [53, 0.0] - - - [32640, 2048, 1, 384] - - [53, 0.0] - - - [28032, 11777, 1, 384] - - [53, 0.0] - - - [20352, 4096, 1, 384] - - [53, 0.0] - - - [33792, 512, 1, 384] - - [53, 0.0] - - - [24192, 4096, 1, 384] - - [53, 0.0] - - - [9216, 384, 1, 384] - - [47, 0.0] - - - [6912, 512, 1, 384] - - [53, 0.0] - - - [14208, 1024, 1, 384] - - [53, 0.0] - - - [26496, 512, 1, 384] - - [53, 0.0] - - - [4992, 384, 1, 384] - - [151, 0.0] - - - [33408, 512, 1, 384] - - [53, 0.0] - - - [3456, 1537, 1, 384] - - [50, 0.0] - - - [21888, 14209, 1, 384] - - [53, 0.0] - - - [24576, 8321, 1, 384] - - [47, 0.0] - - - [33792, 17921, 1, 384] - - [47, 0.0] - - - [13440, 384, 1, 384] - - [47, 0.0] - - - [18432, 384, 1, 384] - - [47, 0.0] - - - [6912, 1024, 1, 384] - - [53, 0.0] - - - [22272, 384, 1, 384] - - [53, 0.0] - - - [3840, 2305, 1, 384] - - [53, 0.0] - - - [6144, 1024, 1, 384] - - [47, 0.0] - - - [7680, 512, 1, 384] - - [53, 0.0] - - - [19584, 4096, 1, 384] - - [53, 0.0] - - - [23808, 4096, 1, 384] - - [53, 0.0] - - - [29568, 384, 1, 384] - - [53, 0.0] - - - [29184, 512, 1, 384] - - [53, 0.0] - - - [13056, 384, 1, 384] - - [53, 0.0] - - - [28032, 512, 1, 384] - - [53, 0.0] - - - [26880, 2048, 1, 384] - - [53, 0.0] - - - [18048, 9985, 1, 384] - - [47, 0.0] - - - [29952, 512, 1, 384] - - [53, 0.0] - - - [27648, 2048, 1, 384] - - [53, 0.0] - - - [29568, 13697, 1, 384] - - [47, 0.0] - - - [19584, 1024, 1, 384] - - [53, 0.0] - - - [27648, 384, 1, 384] - - [47, 0.0] - - - [6912, 384, 1, 384] - - [68, 0.0] - - - [26880, 1024, 1, 384] - - [53, 0.0] - - - [24960, 8192, 1, 384] - - [53, 0.0] - - - [13824, 1024, 1, 384] - - [53, 0.0] - - - [11904, 2048, 1, 384] - - [53, 0.0] - - - [34560, 2048, 1, 384] - - [53, 0.0] - - - [12288, 4609, 1, 384] - - [53, 0.0] - - - [21504, 13825, 1, 384] - - [53, 0.0] - - - [29184, 8192, 1, 384] - - [53, 0.0] - - - [12288, 4096, 1, 384] - - [53, 0.0] - - - [23424, 1024, 1, 384] - - [53, 0.0] - - - [14208, 512, 1, 384] - - [53, 0.0] - - - [25728, 512, 1, 384] - - [53, 0.0] - - - [29568, 2048, 1, 384] - - [53, 0.0] - - - [9600, 1024, 1, 384] - - [53, 0.0] - - - [29952, 384, 1, 384] - - [53, 0.0] - - - [18048, 4096, 1, 384] - - [53, 0.0] - - - [30336, 14081, 1, 384] - - [47, 0.0] - - - [24192, 8192, 1, 384] - - [53, 0.0] - - - [33792, 2048, 1, 384] - - [53, 0.0] - - - [6144, 384, 1, 384] - - [68, 0.0] - - - [8448, 1024, 1, 384] - - [53, 0.0] - - - [6528, 1024, 1, 384] - - [47, 0.0] - - - [18432, 10369, 1, 384] - - [47, 0.0] - - - [19968, 1024, 1, 384] - - [53, 0.0] - - - [23424, 512, 1, 384] - - [53, 0.0] - - - [20736, 2048, 1, 384] - - [53, 0.0] - - - [29184, 12929, 1, 384] - - [47, 0.0] - - - [3072, 1153, 1, 384] - - [160, 0.0] - - - [28416, 512, 1, 384] - - [53, 0.0] - - - [14592, 384, 1, 384] - - [53, 0.0] - - - [18432, 1024, 1, 384] - - [53, 0.0] - - - [29184, 13313, 1, 384] - - [53, 0.0] - - - [32640, 4096, 1, 384] - - [53, 0.0] - - - [21888, 13825, 1, 384] - - [53, 0.0] - - - [5376, 1024, 1, 384] - - [53, 0.0] - - - [4608, 1024, 1, 384] - - [53, 0.0] - - - [8832, 5249, 1, 384] - - [53, 0.0] - - - [14976, 4096, 1, 384] - - [53, 0.0] - - - [3840, 1024, 1, 384] - - [47, 0.0] - - - [24192, 16129, 1, 384] - - [47, 0.0] - - - [19968, 12289, 1, 384] - - [53, 0.0] - - - [1152, 384, 1, 384] - - [56, 0.0] - - - [27648, 4096, 1, 384] - - [53, 0.0] - - - [4992, 3073, 1, 384] - - [53, 0.0] - - - [33024, 8192, 1, 384] - - [53, 0.0] - - - [34944, 384, 1, 384] - - [47, 0.0] - - - [32, 28672, 1, 32] - - [180, 23.634] - - - [32, 24576, 1, 32] - - [179, 23.263] - - - [32, 16384, 1, 32] - - [176, 22.424] - - - [32, 20480, 1, 32] - - [178, 22.915] - - - [32, 12288, 1, 32] - - [176, 20.484] - - - [32, 8192, 1, 32] - - [177, 16.315] - - - [32, 4096, 1, 32] - - [175, 8.373] - - - [32, 32768, 1, 32] - - [174, 23.586] - - - [4224, 3840, 1, 4096] - - [211, 98.172] - - - [5376, 4096, 1, 4096] - - [212, 94.215] - - - [7040, 4096, 1, 384] - - [213, 86.756] - - - [7040, 4096, 1, 768] - - [211, 93.117] - - - [7040, 4096, 1, 1536] - - [214, 96.418] - - - [3840, 4224, 1, 4096] - - [181, 79.968] - - - [3840, 4224, 1, 4224] - - [182, 79.975] - - - [3840, 4224, 1, 4320] - - [183, 79.983] - - - [7680, 8448, 1, 8192] - - [184, 81.886] - - - [7680, 8448, 1, 8448] - - [184, 81.886] - - - [7680, 8448, 1, 8640] - - [184, 81.886] - - - [4096, 7169, 1, 512] - - [221, 84.266] - - - [4096, 7681, 1, 512] - - [214, 86.881] - - - [4096, 8193, 1, 512] - - [192, 72.003] - - - [4608, 512, 1, 512] - - [225, 61.385] - - - [4608, 8193, 1, 512] - - [214, 85.027] - - - [4608, 8705, 1, 512] - - [214, 87.168] - - - [4608, 9217, 1, 512] - - [188, 72.347] - - - [5120, 512, 1, 512] - - [225, 66.118] - - - [5120, 9217, 1, 512] - - [214, 86.807] - - - [5120, 9729, 1, 512] - - [214, 89.066] - - - [5120, 10241, 1, 512] - - [192, 75.207] - - - [5632, 512, 1, 512] - - [231, 60.997] - - - [5632, 10241, 1, 512] - - [236, 87.256] - - - [5632, 10753, 1, 512] - - [214, 89.955] - - - [5632, 11265, 1, 512] - - [192, 74.781] - - - [6144, 512, 1, 512] - - [190, 59.74] - - - [6144, 11265, 1, 512] - - [214, 88.199] - - - [6144, 11777, 1, 512] - - [214, 89.303] - - - [6144, 12289, 1, 512] - - [196, 76.384] - - - [6656, 512, 1, 512] - - [235, 70.472] - - - [6656, 12289, 1, 512] - - [221, 88.4] - - - [6656, 12801, 1, 512] - - [214, 89.997] - - - [6656, 13313, 1, 512] - - [192, 77.745] - - - [7168, 512, 1, 512] - - [234, 63.113] - - - [7168, 13313, 1, 512] - - [221, 88.748] - - - [7168, 13825, 1, 512] - - [214, 89.926] - - - [7168, 14337, 1, 512] - - [192, 77.707] - - - [7680, 512, 1, 512] - - [185, 50.744] - - - [7680, 14337, 1, 512] - - [221, 89.136] - - - [7680, 14849, 1, 512] - - [214, 90.715] - - - [7680, 15361, 1, 512] - - [191, 78.686] - - - [8192, 512, 1, 512] - - [225, 70.193] - - - [8192, 15361, 1, 512] - - [221, 89.749] - - - [8192, 15873, 1, 512] - - [214, 90.691] - - - [8192, 16385, 1, 512] - - [192, 81.698] - - - [8704, 512, 1, 512] - - [227, 72.992] - - - [8704, 16385, 1, 512] - - [233, 89.771] - - - [8704, 16897, 1, 512] - - [232, 90.561] - - - [8704, 17409, 1, 512] - - [191, 80.333] - - - [9216, 512, 1, 512] - - [231, 69.449] - - - [9216, 17409, 1, 512] - - [221, 89.961] - - - [9216, 17921, 1, 512] - - [214, 90.635] - - - [9216, 18433, 1, 512] - - [191, 80.66] - - - [9728, 512, 1, 512] - - [230, 72.797] - - - [9728, 18433, 1, 512] - - [221, 90.251] - - - [9728, 18945, 1, 512] - - [214, 91.185] - - - [9728, 19457, 1, 512] - - [192, 81.229] - - - [10240, 512, 1, 512] - - [223, 75.717] - - - [10240, 19457, 1, 512] - - [221, 90.202] - - - [10240, 19969, 1, 512] - - [221, 90.809] - - - [10240, 20481, 1, 512] - - [192, 82.047] - - - [10752, 512, 1, 512] - - [229, 69.5] - - - [10752, 20481, 1, 512] - - [221, 90.317] - - - [10752, 20993, 1, 512] - - [214, 91.332] - - - [10752, 21505, 1, 512] - - [191, 82.497] - - - [11264, 512, 1, 512] - - [225, 72.139] - - - [11264, 21505, 1, 512] - - [221, 89.945] - - - [11264, 22017, 1, 512] - - [228, 91.061] - - - [11264, 22529, 1, 512] - - [198, 80.014] - - - [11776, 512, 1, 512] - - [189, 58.925] - - - [11776, 22529, 1, 512] - - [221, 90.5] - - - [11776, 23041, 1, 512] - - [214, 91.233] - - - [11776, 23553, 1, 512] - - [191, 81.812] - - - [12288, 512, 1, 512] - - [190, 60.868] - - - [12288, 23553, 1, 512] - - [221, 90.7] - - - [12288, 24065, 1, 512] - - [214, 91.263] - - - [12288, 24577, 1, 512] - - [192, 83.558] - - - [12800, 512, 1, 512] - - [224, 74.505] - - - [12800, 24577, 1, 512] - - [221, 90.642] - - - [12800, 25089, 1, 512] - - [221, 91.231] - - - [12800, 25601, 1, 512] - - [191, 82.218] - - - [13312, 512, 1, 512] - - [223, 76.879] - - - [13312, 25601, 1, 512] - - [221, 90.815] - - - [13312, 26113, 1, 512] - - [214, 91.276] - - - [13312, 26625, 1, 512] - - [191, 83.155] - - - [13824, 512, 1, 512] - - [223, 79.563] - - - [13824, 26625, 1, 512] - - [221, 90.806] - - - [13824, 27137, 1, 512] - - [221, 91.445] - - - [13824, 27649, 1, 512] - - [191, 82.737] - - - [14336, 512, 1, 512] - - [187, 56.776] - - - [14336, 27649, 1, 512] - - [221, 90.756] - - - [14336, 28161, 1, 512] - - [221, 91.422] - - - [14336, 28673, 1, 512] - - [191, 82.851] - - - [14848, 512, 1, 512] - - [226, 75.655] - - - [14848, 28673, 1, 512] - - [221, 90.927] - - - [14848, 29185, 1, 512] - - [214, 91.506] - - - [14848, 29697, 1, 512] - - [192, 82.656] - - - [15360, 512, 1, 512] - - [225, 77.528] - - - [15360, 29697, 1, 512] - - [221, 91.059] - - - [15360, 30209, 1, 512] - - [214, 91.665] - - - [15360, 30721, 1, 512] - - [191, 82.881] - - - [15872, 512, 1, 512] - - [223, 75.258] - - - [15872, 30721, 1, 512] - - [221, 91.154] - - - [15872, 31233, 1, 512] - - [221, 91.635] - - - [15872, 31745, 1, 512] - - [192, 82.7] - - - [16384, 512, 1, 512] - - [223, 77.779] - - - [16384, 31745, 1, 512] - - [221, 91.133] - - - [16384, 32257, 1, 512] - - [214, 91.736] - - - [16384, 32769, 1, 512] - - [195, 85.459] - - - [16896, 512, 1, 512] - - [223, 79.634] - - - [16896, 32769, 1, 512] - - [221, 90.626] - - - [16896, 33281, 1, 512] - - [219, 91.642] - - - [16896, 33793, 1, 512] - - [194, 82.857] - - - [17408, 512, 1, 512] - - [224, 81.592] - - - [17408, 33793, 1, 512] - - [221, 91.33] - - - [17408, 34305, 1, 512] - - [221, 91.738] - - - [17408, 34817, 1, 512] - - [197, 84.577] - - - [17920, 512, 1, 512] - - [211, 72.341] - - - [17920, 34817, 1, 512] - - [221, 91.248] - - - [17920, 35329, 1, 512] - - [221, 91.794] - - - [17920, 35841, 1, 512] - - [193, 84.812] - - - [18432, 512, 1, 512] - - [211, 74.155] - - - [18432, 35841, 1, 512] - - [221, 91.331] - - - [18432, 36353, 1, 512] - - [221, 91.808] - - - [18432, 36865, 1, 512] - - [197, 84.884] - - - [18944, 512, 1, 512] - - [223, 76.111] - - - [18944, 36865, 1, 512] - - [221, 91.179] - - - [18944, 37377, 1, 512] - - [221, 91.875] - - - [18944, 37889, 1, 512] - - [198, 84.397] - - - [19456, 512, 1, 512] - - [222, 78.142] - - - [19456, 37889, 1, 512] - - [221, 91.415] - - - [19456, 38401, 1, 512] - - [221, 91.847] - - - [19456, 38913, 1, 512] - - [192, 84.243] - - - [19968, 512, 1, 512] - - [222, 79.818] - - - [19968, 38913, 1, 512] - - [221, 91.304] - - - [19968, 39425, 1, 512] - - [221, 91.892] - - - [19968, 39937, 1, 512] - - [191, 83.572] - - - [20480, 512, 1, 512] - - [214, 81.698] - - - [20480, 39937, 1, 512] - - [221, 91.491] - - - [20480, 40449, 1, 512] - - [221, 91.965] - - - [20480, 40961, 1, 512] - - [194, 85.334] - - - [20992, 512, 1, 512] - - [211, 83.685] - - - [20992, 40961, 1, 512] - - [221, 91.379] - - - [20992, 41473, 1, 512] - - [221, 91.976] - - - [20992, 41985, 1, 512] - - [191, 83.79] - - - [21504, 512, 1, 512] - - [211, 75.353] - - - [21504, 41985, 1, 512] - - [221, 91.429] - - - [21504, 42497, 1, 512] - - [221, 91.991] - - - [21504, 43009, 1, 512] - - [198, 83.487] - - - [22016, 512, 1, 512] - - [188, 62.253] - - - [22016, 43009, 1, 512] - - [221, 91.527] - - - [22016, 43521, 1, 512] - - [221, 92.009] - - - [22016, 44033, 1, 512] - - [194, 84.676] - - - [22528, 512, 1, 512] - - [186, 63.381] - - - [22528, 44033, 1, 512] - - [221, 91.443] - - - [22528, 44545, 1, 512] - - [219, 91.907] - - - [22528, 45057, 1, 512] - - [194, 85.559] - - - [23040, 512, 1, 512] - - [188, 64.394] - - - [23040, 45057, 1, 512] - - [221, 90.527] - - - [23040, 45569, 1, 512] - - [221, 92.045] - - - [23040, 46081, 1, 512] - - [191, 84.393] - - - [23552, 512, 1, 512] - - [187, 65.205] - - - [23552, 46081, 1, 512] - - [221, 91.562] - - - [23552, 46593, 1, 512] - - [221, 92.076] - - - [23552, 47105, 1, 512] - - [194, 85.301] - - - [24064, 512, 1, 512] - - [215, 83.359] - - - [24064, 47105, 1, 512] - - [221, 91.552] - - - [24064, 47617, 1, 512] - - [221, 92.066] - - - [24064, 48129, 1, 512] - - [198, 84.846] - - - [24576, 512, 1, 512] - - [211, 85.009] - - - [24576, 48129, 1, 512] - - [221, 91.695] - - - [24576, 48641, 1, 512] - - [221, 92.107] - - - [24576, 49153, 1, 512] - - [194, 85.545] - - - [25088, 512, 1, 512] - - [211, 77.48] - - - [25088, 49153, 1, 512] - - [219, 85.449] - - - [25088, 49665, 1, 512] - - [221, 92.123] - - - [25088, 50177, 1, 512] - - [197, 84.948] - - - [25600, 512, 1, 512] - - [211, 79.144] - - - [25600, 50177, 1, 512] - - [221, 90.784] - - - [25600, 50689, 1, 512] - - [221, 92.031] - - - [25600, 51201, 1, 512] - - [191, 84.823] - - - [26112, 512, 1, 512] - - [190, 64.603] - - - [26112, 51201, 1, 512] - - [221, 91.688] - - - [26112, 51713, 1, 512] - - [221, 92.122] - - - [26112, 52225, 1, 512] - - [197, 85.637] - - - [26624, 512, 1, 512] - - [211, 81.967] - - - [26624, 52225, 1, 512] - - [221, 91.617] - - - [26624, 52737, 1, 512] - - [221, 92.135] - - - [26624, 53249, 1, 512] - - [194, 84.865] - - - [27136, 512, 1, 512] - - [211, 83.218] - - - [27136, 53249, 1, 512] - - [221, 91.664] - - - [27136, 53761, 1, 512] - - [221, 92.151] - - - [27136, 54273, 1, 512] - - [198, 84.654] - - - [27648, 512, 1, 512] - - [190, 67.069] - - - [27648, 54273, 1, 512] - - [221, 91.75] - - - [27648, 54785, 1, 512] - - [221, 92.157] - - - [27648, 55297, 1, 512] - - [197, 84.534] - - - [28160, 512, 1, 512] - - [187, 67.372] - - - [28160, 55297, 1, 512] - - [221, 90.497] - - - [28160, 55809, 1, 512] - - [221, 92.093] - - - [28160, 56321, 1, 512] - - [197, 84.571] - - - [28672, 512, 1, 512] - - [215, 79.267] - - - [28672, 56321, 1, 512] - - [221, 91.763] - - - [28672, 56833, 1, 512] - - [221, 92.177] - - - [28672, 57345, 1, 512] - - [197, 85.47] - - - [29184, 512, 1, 512] - - [214, 80.494] - - - [29184, 57345, 1, 512] - - [221, 91.592] - - - [29184, 57857, 1, 512] - - [221, 92.18] - - - [29184, 58369, 1, 512] - - [194, 85.478] - - - [29696, 512, 1, 512] - - [211, 81.983] - - - [29696, 58369, 1, 512] - - [221, 91.78] - - - [29696, 58881, 1, 512] - - [221, 92.176] - - - [29696, 59393, 1, 512] - - [191, 84.807] - - - [30208, 512, 1, 512] - - [211, 83.183] - - - [30208, 59393, 1, 512] - - [221, 91.831] - - - [30208, 59905, 1, 512] - - [221, 92.196] - - - [30208, 60417, 1, 512] - - [193, 85.176] - - - [30720, 512, 1, 512] - - [214, 84.494] - - - [30720, 60417, 1, 512] - - [221, 91.767] - - - [30720, 60929, 1, 512] - - [221, 92.201] - - - [30720, 61441, 1, 512] - - [194, 85.323] - - - [31232, 512, 1, 512] - - [188, 67.379] - - - [31232, 61441, 1, 512] - - [221, 90.844] - - - [31232, 61953, 1, 512] - - [219, 92.133] - - - [31232, 62465, 1, 512] - - [194, 84.989] - - - [31744, 512, 1, 512] - - [215, 79.585] - - - [31744, 62465, 1, 512] - - [221, 91.769] - - - [31744, 62977, 1, 512] - - [221, 92.218] - - - [31744, 63489, 1, 512] - - [194, 84.774] - - - [32256, 512, 1, 512] - - [215, 80.749] - - - [32256, 63489, 1, 512] - - [221, 91.833] - - - [32256, 64001, 1, 512] - - [221, 92.212] - - - [32256, 64513, 1, 512] - - [197, 85.646] - - - [32768, 512, 1, 512] - - [214, 81.909] - - - [32768, 64513, 1, 512] - - [221, 91.357] - - - [32768, 65025, 1, 512] - - [221, 92.224] - - - [32768, 65537, 1, 512] - - [197, 85.69] - - - [33280, 512, 1, 512] - - [211, 83.172] - - - [33280, 65537, 1, 512] - - [219, 85.532] - - - [33280, 66049, 1, 512] - - [221, 92.233] - - - [33280, 66561, 1, 512] - - [197, 85.308] - - - [33792, 512, 1, 512] - - [211, 84.451] - - - [33792, 66561, 1, 512] - - [221, 90.835] - - - [33792, 67073, 1, 512] - - [221, 92.149] - - - [33792, 67585, 1, 512] - - [194, 85.355] - - - [34304, 512, 1, 512] - - [211, 85.586] - - - [34304, 67585, 1, 512] - - [221, 91.736] - - - [34304, 68097, 1, 512] - - [221, 92.193] - - - [34304, 68609, 1, 512] - - [194, 85.568] - - - [34816, 512, 1, 512] - - [211, 86.681] - - - [34816, 68609, 1, 512] - - [220, 90.306] - - - [34816, 69121, 1, 512] - - [221, 92.213] - - - [34816, 69633, 1, 512] - - [194, 85.524] - - - [35328, 512, 1, 512] - - [215, 80.742] - - - [35328, 69633, 1, 512] - - [218, 89.741] - - - [35328, 70145, 1, 512] - - [221, 92.207] - - - [35328, 70657, 1, 512] - - [197, 85.596] - - - [35840, 512, 1, 512] - - [211, 82.007] - - - [35840, 70657, 1, 512] - - [221, 91.292] - - - [35840, 71169, 1, 512] - - [221, 92.229] - - - [35840, 71681, 1, 512] - - [197, 85.763] - - - [36352, 512, 1, 512] - - [214, 83.002] - - - [36352, 71681, 1, 512] - - [221, 90.974] - - - [36352, 72193, 1, 512] - - [221, 92.24] - - - [36352, 72705, 1, 512] - - [197, 85.374] - - - [36864, 512, 1, 512] - - [211, 84.236] - - - [36864, 72705, 1, 512] - - [221, 90.37] - - - [36864, 73217, 1, 512] - - [219, 92.162] - - - [36864, 73729, 1, 512] - - [197, 85.465] - - - [37376, 512, 1, 512] - - [211, 85.258] - - - [37376, 73729, 1, 512] - - [221, 89.017] - - - [37376, 74241, 1, 512] - - [221, 92.2] - - - [37376, 74753, 1, 512] - - [197, 85.196] - - - [37888, 512, 1, 512] - - [211, 86.425] - - - [37888, 74753, 1, 512] - - [220, 90.364] - - - [37888, 75265, 1, 512] - - [221, 92.208] - - - [37888, 75777, 1, 512] - - [194, 85.489] - - - [38400, 512, 1, 512] - - [211, 87.374] - - - [38400, 75777, 1, 512] - - [220, 90.276] - - - [38400, 76289, 1, 512] - - [221, 92.209] - - - [38400, 76801, 1, 512] - - [194, 85.368] - - - [38912, 512, 1, 512] - - [211, 82.06] - - - [38912, 76801, 1, 512] - - [220, 90.287] - - - [38912, 77313, 1, 512] - - [221, 92.198] - - - [38912, 77825, 1, 512] - - [194, 85.45] - - - [39424, 512, 1, 512] - - [211, 82.977] - - - [39424, 77825, 1, 512] - - [221, 89.64] - - - [39424, 78337, 1, 512] - - [219, 92.16] - - - [39424, 78849, 1, 512] - - [197, 85.348] - - - [39936, 512, 1, 512] - - [214, 83.98] - - - [39936, 78849, 1, 512] - - [220, 90.232] - - - [39936, 79361, 1, 512] - - [219, 92.182] - - - [39936, 79873, 1, 512] - - [193, 84.762] - - - [40448, 512, 1, 512] - - [211, 85.011] - - - [40448, 79873, 1, 512] - - [220, 90.277] - - - [40448, 80385, 1, 512] - - [221, 92.212] - - - [40448, 80897, 1, 512] - - [197, 85.097] - - - [40960, 512, 1, 512] - - [211, 85.996] - - - [40960, 80897, 1, 512] - - [220, 90.293] - - - [40960, 81409, 1, 512] - - [219, 92.192] - - - [40960, 81921, 1, 512] - - [194, 85.27] - - - [41472, 512, 1, 512] - - [211, 87.024] - - - [41472, 81921, 1, 512] - - [219, 85.572] - - - [41472, 82433, 1, 512] - - [219, 92.203] - - - [41472, 82945, 1, 512] - - [197, 84.868] - - - [41984, 512, 1, 512] - - [211, 88.005] - - - [41984, 82945, 1, 512] - - [220, 90.291] - - - [41984, 83457, 1, 512] - - [219, 92.19] - - - [41984, 83969, 1, 512] - - [193, 84.871] - - - [42496, 512, 1, 512] - - [214, 82.942] - - - [42496, 83969, 1, 512] - - [220, 90.02] - - - [42496, 84481, 1, 512] - - [219, 92.11] - - - [42496, 84993, 1, 512] - - [197, 85.396] - - - [43008, 512, 1, 512] - - [211, 83.845] - - - [43008, 84993, 1, 512] - - [220, 90.16] - - - [43008, 85505, 1, 512] - - [219, 92.164] - - - [43008, 86017, 1, 512] - - [197, 85.435] - - - [43520, 512, 1, 512] - - [211, 84.899] - - - [43520, 86017, 1, 512] - - [220, 89.802] - - - [43520, 86529, 1, 512] - - [219, 92.172] - - - [43520, 87041, 1, 512] - - [197, 85.381] - - - [44032, 512, 1, 512] - - [211, 85.742] - - - [44032, 87041, 1, 512] - - [220, 90.025] - - - [44032, 87553, 1, 512] - - [219, 92.168] - - - [44032, 88065, 1, 512] - - [194, 85.288] - - - [44544, 512, 1, 512] - - [211, 86.81] - - - [44544, 88065, 1, 512] - - [220, 89.883] - - - [44544, 88577, 1, 512] - - [219, 92.129] - - - [44544, 89089, 1, 512] - - [197, 85.114] - - - [45056, 512, 1, 512] - - [214, 87.534] - - - [45056, 89089, 1, 512] - - [220, 89.934] - - - [45056, 89601, 1, 512] - - [219, 92.112] - - - [45056, 90113, 1, 512] - - [194, 85.132] - - - [45568, 512, 1, 512] - - [211, 88.529] - - - [45568, 90113, 1, 512] - - [218, 89.378] - - - [45568, 90625, 1, 512] - - [219, 92.104] - - - [45568, 91137, 1, 512] - - [197, 85.215] - - - [46080, 512, 1, 512] - - [211, 83.779] - - - [46080, 91137, 1, 512] - - [218, 89.848] - - - [46080, 91649, 1, 512] - - [219, 91.674] - - - [46080, 92161, 1, 512] - - [197, 85.271] - - - [46592, 512, 1, 512] - - [211, 84.697] - - - [46592, 92161, 1, 512] - - [218, 89.86] - - - [46592, 92673, 1, 512] - - [219, 91.634] - - - [46592, 93185, 1, 512] - - [197, 85.093] - - - [47104, 512, 1, 512] - - [211, 85.444] - - - [47104, 93185, 1, 512] - - [218, 89.813] - - - [47104, 93697, 1, 512] - - [219, 91.474] - - - [47104, 94209, 1, 512] - - [197, 85.18] - - - [47616, 512, 1, 512] - - [211, 86.399] - - - [47616, 94209, 1, 512] - - [218, 89.612] - - - [47616, 94721, 1, 512] - - [214, 90.878] - - - [47616, 95233, 1, 512] - - [194, 85.144] - - - [48128, 512, 1, 512] - - [211, 87.248] - - - [48128, 95233, 1, 512] - - [218, 89.879] - - - [48128, 95745, 1, 512] - - [214, 90.972] - - - [48128, 96257, 1, 512] - - [197, 85.061] - - - [48640, 512, 1, 512] - - [214, 88.068] - - - [48640, 96257, 1, 512] - - [218, 89.804] - - - [48640, 96769, 1, 512] - - [214, 91.017] - - - [48640, 97281, 1, 512] - - [194, 84.928] - - - [49152, 512, 1, 512] - - [211, 88.927] - - - [49152, 97281, 1, 512] - - [218, 89.796] - - - [49152, 97793, 1, 512] - - [214, 90.938] - - - [49152, 98305, 1, 512] - - [194, 84.87] - - - [49664, 512, 1, 512] - - [211, 84.521] - - - [49664, 98305, 1, 512] - - [219, 85.468] - - - [49664, 98817, 1, 512] - - [214, 90.984] - - - [49664, 99329, 1, 512] - - [194, 84.773] - - - [50176, 512, 1, 512] - - [211, 85.318] - - - [50176, 99329, 1, 512] - - [218, 89.7] - - - [50176, 99841, 1, 512] - - [214, 90.547] - - - [50176, 100353, 1, 512] - - [197, 84.94] - - - [50688, 512, 1, 512] - - [211, 86.189] - - - [50688, 100353, 1, 512] - - [218, 89.558] - - - [50688, 100865, 1, 512] - - [214, 90.628] - - - [50688, 101377, 1, 512] - - [194, 84.962] - - - [51200, 512, 1, 512] - - [211, 86.997] - - - [51200, 101377, 1, 512] - - [218, 89.438] - - - [51200, 101889, 1, 512] - - [214, 90.328] - - - [51200, 102401, 1, 512] - - [197, 84.994] - - - [51712, 512, 1, 512] - - [211, 87.728] - - - [51712, 102401, 1, 512] - - [218, 87.961] - - - [51712, 102913, 1, 512] - - [214, 90.029] - - - [51712, 103425, 1, 512] - - [194, 84.604] - - - [52224, 512, 1, 512] - - [215, 88.533] - - - [52224, 103425, 1, 512] - - [218, 89.057] - - - [52224, 103937, 1, 512] - - [214, 90.138] - - - [52224, 104449, 1, 512] - - [197, 84.94] - - - [52736, 512, 1, 512] - - [214, 89.235] - - - [52736, 104449, 1, 512] - - [218, 88.834] - - - [52736, 104961, 1, 512] - - [214, 89.95] - - - [52736, 105473, 1, 512] - - [197, 84.85] - - - [53248, 512, 1, 512] - - [211, 85.231] - - - [53248, 105473, 1, 512] - - [218, 88.678] - - - [53248, 105985, 1, 512] - - [214, 89.674] - - - [53248, 106497, 1, 512] - - [194, 84.815] - - - [53760, 512, 1, 512] - - [211, 86.004] - - - [53760, 106497, 1, 512] - - [219, 86.065] - - - [53760, 107009, 1, 512] - - [214, 89.576] - - - [53760, 107521, 1, 512] - - [194, 84.895] - - - [54272, 512, 1, 512] - - [211, 86.683] - - - [54272, 107521, 1, 512] - - [217, 88.377] - - - [54272, 108033, 1, 512] - - [214, 89.336] - - - [54272, 108545, 1, 512] - - [197, 84.744] - - - [54784, 512, 1, 512] - - [211, 87.335] - - - [54784, 108545, 1, 512] - - [217, 87.922] - - - [54784, 109057, 1, 512] - - [214, 89.197] - - - [54784, 109569, 1, 512] - - [194, 84.751] - - - [55296, 512, 1, 512] - - [215, 88.163] - - - [55296, 109569, 1, 512] - - [217, 88.111] - - - [55296, 110081, 1, 512] - - [214, 89.255] - - - [55296, 110593, 1, 512] - - [197, 84.749] - - - [55808, 512, 1, 512] - - [211, 88.944] - - - [55808, 110593, 1, 512] - - [216, 86.787] - - - [55808, 111105, 1, 512] - - [214, 88.791] - - - [55808, 111617, 1, 512] - - [194, 84.715] - - - [56320, 512, 1, 512] - - [211, 89.699] - - - [56320, 111617, 1, 512] - - [217, 88.022] - - - [56320, 112129, 1, 512] - - [214, 88.449] - - - [56320, 112641, 1, 512] - - [194, 84.635] - - - [56832, 512, 1, 512] - - [211, 85.676] - - - [56832, 112641, 1, 512] - - [217, 87.572] - - - [56832, 113153, 1, 512] - - [214, 88.617] - - - [56832, 113665, 1, 512] - - [197, 84.496] - - - [57344, 512, 1, 512] - - [215, 86.337] - - - [57344, 113665, 1, 512] - - [217, 87.727] - - - [57344, 114177, 1, 512] - - [214, 87.984] - - - [57344, 114689, 1, 512] - - [197, 84.476] - - - [57856, 512, 1, 512] - - [211, 87.142] - - - [57856, 114689, 1, 512] - - [194, 84.365] - - - [57856, 115201, 1, 512] - - [218, 87.771] - - - [57856, 115713, 1, 512] - - [197, 84.499] - - - [58368, 512, 1, 512] - - [211, 87.801] - - - [58368, 115713, 1, 512] - - [217, 87.395] - - - [58368, 116225, 1, 512] - - [216, 87.822] - - - [58368, 116737, 1, 512] - - [194, 84.555] - - - [58880, 512, 1, 512] - - [211, 88.56] - - - [58880, 116737, 1, 512] - - [216, 87.112] - - - [58880, 117249, 1, 512] - - [216, 87.626] - - - [58880, 117761, 1, 512] - - [197, 84.526] - - - [59392, 512, 1, 512] - - [211, 89.194] - - - [59392, 117761, 1, 512] - - [217, 87.114] - - - [59392, 118273, 1, 512] - - [216, 87.433] - - - [59392, 118785, 1, 512] - - [194, 84.499] - - - [59904, 512, 1, 512] - - [211, 85.538] - - - [59904, 118785, 1, 512] - - [216, 86.052] - - - [59904, 119297, 1, 512] - - [216, 87.416] - - - [59904, 119809, 1, 512] - - [197, 84.498] - - - [60416, 512, 1, 512] - - [211, 86.092] - - - [60416, 119809, 1, 512] - - [216, 86.906] - - - [60416, 120321, 1, 512] - - [216, 87.243] - - - [60416, 120833, 1, 512] - - [197, 84.436] - - - [60928, 512, 1, 512] - - [211, 86.874] - - - [60928, 120833, 1, 512] - - [216, 86.678] - - - [60928, 121345, 1, 512] - - [216, 87.119] - - - [60928, 121857, 1, 512] - - [194, 84.373] - - - [61440, 512, 1, 512] - - [211, 87.551] - - - [61440, 121857, 1, 512] - - [216, 86.737] - - - [61440, 122369, 1, 512] - - [216, 87.048] - - - [61440, 122881, 1, 512] - - [194, 84.312] - - - [61952, 512, 1, 512] - - [211, 88.249] - - - [61952, 122881, 1, 512] - - [197, 84.364] - - - [61952, 123393, 1, 512] - - [216, 86.828] - - - [61952, 123905, 1, 512] - - [194, 84.256] - - - [62464, 512, 1, 512] - - [211, 88.818] - - - [62464, 123905, 1, 512] - - [216, 86.514] - - - [62464, 124417, 1, 512] - - [216, 86.752] - - - [62464, 124929, 1, 512] - - [194, 84.363] - - - [62976, 512, 1, 512] - - [211, 89.481] - - - [62976, 124929, 1, 512] - - [216, 86.248] - - - [62976, 125441, 1, 512] - - [216, 86.681] - - - [62976, 125953, 1, 512] - - [197, 84.354] - - - [63488, 512, 1, 512] - - [211, 85.97] - - - [63488, 125953, 1, 512] - - [216, 86.312] - - - [63488, 126465, 1, 512] - - [216, 86.502] - - - [63488, 126977, 1, 512] - - [197, 84.135] - - - [64000, 512, 1, 512] - - [211, 86.643] - - - [64000, 126977, 1, 512] - - [216, 85.249] - - - [64000, 127489, 1, 512] - - [197, 84.172] - - - [64000, 128001, 1, 512] - - [194, 84.304] - - - [64512, 512, 1, 512] - - [201, 73.112] - - - [64512, 128001, 1, 512] - - [197, 84.344] - - - [4096, 4096, 1, 4128] - - [200, 87.241] - - - [25600, 25600, 1, 512] - - [202, 85.988] - - - [512, 512, 1, 512] - - [244, 21.302] - - - [1024, 512, 1, 512] - - [243, 28.74] - - - [1536, 512, 1, 512] - - [241, 42.307] - - - [1536, 1024, 1, 512] - - [199, 48.976] - - - [2048, 512, 1, 512] - - [240, 40.467] - - - [2048, 1024, 1, 512] - - [204, 46.216] - - - [2560, 512, 1, 512] - - [239, 49.588] - - - [2560, 1024, 1, 512] - - [205, 55.747] - - - [2560, 1536, 1, 512] - - [202, 50.237] - - - [3072, 512, 1, 512] - - [238, 57.893] - - - [3072, 1024, 1, 512] - - [206, 60.637] - - - [3072, 1536, 1, 512] - - [203, 58.883] - - - [3584, 512, 1, 512] - - [237, 48.732] - - - [3584, 1536, 1, 512] - - [207, 55.13] - - - [3584, 2048, 1, 512] - - [199, 56.359] - - - [4096, 512, 1, 512] - - [225, 54.564] - - - [4096, 1536, 1, 512] - - [207, 58.771] - - - [4096, 2048, 1, 512] - - [203, 60.943] - - - [4608, 2048, 1, 512] - - [203, 60.553] - - - [4608, 2560, 1, 512] - - [199, 64.545] - - - [5120, 2048, 1, 512] - - [203, 66.724] - - - [5120, 2560, 1, 512] - - [199, 65.066] - - - [5632, 2560, 1, 512] - - [202, 68.634] - - - [5632, 3072, 1, 512] - - [201, 68.816] - - - [6144, 2560, 1, 512] - - [201, 69.537] - - - [6144, 3072, 1, 512] - - [201, 70.207] - - - [6656, 3072, 1, 512] - - [201, 70.839] - - - [6656, 3584, 1, 512] - - [202, 72.129] - - - [7168, 3072, 1, 512] - - [201, 71.894] - - - [7168, 3584, 1, 512] - - [199, 71.702] - - - [7680, 3584, 1, 512] - - [202, 72.491] - - - [7680, 4096, 1, 512] - - [201, 75.567] - - - [8192, 3584, 1, 512] - - [201, 74.321] - - - [8192, 4096, 1, 512] - - [202, 77.538] - - - [8704, 4096, 1, 512] - - [199, 73.842] - - - [8704, 4608, 1, 512] - - [202, 76.742] - - - [9216, 4096, 1, 512] - - [202, 76.115] - - - [9216, 4608, 1, 512] - - [202, 77.313] - - - [9728, 4608, 1, 512] - - [201, 78.031] - - - [9728, 5120, 1, 512] - - [202, 78.019] - - - [10240, 4608, 1, 512] - - [199, 76.024] - - - [10240, 5120, 1, 512] - - [199, 79.933] - - - [10752, 5120, 1, 512] - - [202, 77.607] - - - [10752, 5632, 1, 512] - - [199, 76.711] - - - [11264, 5120, 1, 512] - - [202, 80.152] - - - [11264, 5632, 1, 512] - - [201, 79.638] - - - [11776, 5632, 1, 512] - - [201, 78.593] - - - [11776, 6144, 1, 512] - - [203, 80.044] - - - [12288, 5632, 1, 512] - - [203, 80.999] - - - [12288, 6144, 1, 512] - - [202, 80.958] - - - [12800, 6144, 1, 512] - - [199, 79.796] - - - [12800, 6656, 1, 512] - - [199, 80.542] - - - [13312, 6144, 1, 512] - - [202, 81.716] - - - [13312, 6656, 1, 512] - - [199, 80.581] - - - [13824, 6656, 1, 512] - - [199, 81.209] - - - [13824, 7168, 1, 512] - - [203, 81.62] - - - [14336, 6656, 1, 512] - - [203, 81.561] - - - [14336, 7168, 1, 512] - - [201, 82.537] - - - [14848, 7168, 1, 512] - - [203, 82.172] - - - [14848, 7680, 1, 512] - - [201, 82.338] - - - [15360, 7168, 1, 512] - - [202, 82.038] - - - [15360, 7680, 1, 512] - - [201, 82.45] - - - [15872, 7680, 1, 512] - - [202, 82.442] - - - [15872, 8192, 1, 512] - - [202, 83.202] - - - [16384, 7680, 1, 512] - - [202, 84.271] - - - [16384, 8192, 1, 512] - - [202, 84.799] - - - [16896, 8192, 1, 512] - - [202, 83.133] - - - [16896, 8704, 1, 512] - - [202, 82.906] - - - [17408, 8192, 1, 512] - - [202, 83.515] - - - [17408, 8704, 1, 512] - - [202, 82.388] - - - [17920, 8704, 1, 512] - - [203, 83.796] - - - [17920, 9216, 1, 512] - - [199, 83.806] - - - [18432, 8704, 1, 512] - - [202, 83.579] - - - [18432, 9216, 1, 512] - - [202, 84.415] - - - [18944, 9216, 1, 512] - - [201, 84.22] - - - [18944, 9728, 1, 512] - - [202, 83.678] - - - [19456, 9216, 1, 512] - - [199, 84.014] - - - [19456, 9728, 1, 512] - - [203, 84.282] - - - [19968, 9728, 1, 512] - - [202, 83.353] - - - [19968, 10240, 1, 512] - - [202, 84.377] - - - [20480, 9728, 1, 512] - - [199, 84.963] - - - [20480, 10240, 1, 512] - - [203, 84.922] - - - [20992, 10240, 1, 512] - - [203, 84.203] - - - [20992, 10752, 1, 512] - - [207, 84.344] - - - [21504, 10240, 1, 512] - - [203, 84.949] - - - [21504, 10752, 1, 512] - - [203, 85.001] - - - [22016, 10752, 1, 512] - - [205, 83.843] - - - [22016, 11264, 1, 512] - - [203, 84.536] - - - [22528, 10752, 1, 512] - - [199, 85.051] - - - [22528, 11264, 1, 512] - - [203, 84.828] - - - [23040, 11264, 1, 512] - - [202, 84.877] - - - [23040, 11776, 1, 512] - - [203, 84.515] - - - [23552, 11264, 1, 512] - - [202, 84.992] - - - [23552, 11776, 1, 512] - - [199, 84.827] - - - [24064, 11776, 1, 512] - - [199, 83.983] - - - [24064, 12288, 1, 512] - - [201, 84.334] - - - [24576, 11776, 1, 512] - - [202, 85.132] - - - [24576, 12288, 1, 512] - - [203, 85.381] - - - [25088, 12288, 1, 512] - - [202, 84.609] - - - [25088, 12800, 1, 512] - - [199, 84.494] - - - [25600, 12288, 1, 512] - - [199, 85.114] - - - [25600, 12800, 1, 512] - - [203, 84.848] - - - [26112, 12800, 1, 512] - - [201, 84.694] - - - [26112, 13312, 1, 512] - - [202, 84.9] - - - [26624, 12800, 1, 512] - - [203, 85.171] - - - [26624, 13312, 1, 512] - - [203, 85.545] - - - [27136, 13312, 1, 512] - - [201, 84.876] - - - [27136, 13824, 1, 512] - - [201, 84.834] - - - [27648, 13312, 1, 512] - - [202, 85.458] - - - [27648, 13824, 1, 512] - - [199, 85.171] - - - [28160, 13824, 1, 512] - - [199, 84.792] - - - [28160, 14336, 1, 512] - - [209, 84.979] - - - [28672, 13824, 1, 512] - - [203, 85.128] - - - [28672, 14336, 1, 512] - - [205, 85.269] - - - [29184, 14336, 1, 512] - - [203, 85.16] - - - [29184, 14848, 1, 512] - - [205, 85.104] - - - [29696, 14336, 1, 512] - - [202, 85.199] - - - [29696, 14848, 1, 512] - - [201, 85.365] - - - [30208, 14848, 1, 512] - - [202, 85.312] - - - [30208, 15360, 1, 512] - - [202, 85.595] - - - [30720, 14848, 1, 512] - - [203, 85.427] - - - [30720, 15360, 1, 512] - - [208, 85.648] - - - [31232, 15360, 1, 512] - - [205, 84.798] - - - [31232, 15872, 1, 512] - - [203, 85.074] - - - [31744, 15360, 1, 512] - - [199, 85.469] - - - [31744, 15872, 1, 512] - - [207, 85.178] - - - [32256, 15872, 1, 512] - - [203, 84.703] - - - [32256, 16384, 1, 512] - - [202, 85.478] - - - [32768, 15872, 1, 512] - - [202, 85.751] - - - [32768, 16384, 1, 512] - - [202, 85.819] - - - [33280, 16384, 1, 512] - - [202, 85.458] - - - [33280, 16896, 1, 512] - - [205, 85.254] - - - [33792, 16384, 1, 512] - - [201, 85.316] - - - [33792, 16896, 1, 512] - - [203, 85.303] - - - [34304, 16896, 1, 512] - - [202, 84.943] - - - [34304, 17408, 1, 512] - - [199, 85.379] - - - [34816, 16896, 1, 512] - - [202, 85.402] - - - [34816, 17408, 1, 512] - - [207, 85.719] - - - [35328, 17408, 1, 512] - - [209, 85.086] - - - [35328, 17920, 1, 512] - - [203, 85.43] - - - [35840, 17408, 1, 512] - - [208, 85.502] - - - [35840, 17920, 1, 512] - - [201, 85.402] - - - [36352, 17920, 1, 512] - - [202, 85.437] - - - [36352, 18432, 1, 512] - - [202, 85.502] - - - [36864, 17920, 1, 512] - - [203, 85.493] - - - [36864, 18432, 1, 512] - - [209, 85.65] - - - [37376, 18432, 1, 512] - - [207, 85.432] - - - [37376, 18944, 1, 512] - - [202, 85.666] - - - [37888, 18432, 1, 512] - - [205, 85.516] - - - [37888, 18944, 1, 512] - - [203, 85.567] - - - [38400, 18944, 1, 512] - - [199, 85.183] - - - [38400, 19456, 1, 512] - - [202, 85.305] - - - [38912, 18944, 1, 512] - - [209, 85.729] - - - [38912, 19456, 1, 512] - - [202, 85.377] - - - [39424, 19456, 1, 512] - - [203, 85.534] - - - [39424, 19968, 1, 512] - - [209, 85.169] - - - [39936, 19456, 1, 512] - - [207, 85.312] - - - [39936, 19968, 1, 512] - - [202, 85.146] - - - [40448, 19968, 1, 512] - - [209, 85.246] - - - [40448, 20480, 1, 512] - - [203, 85.533] - - - [40960, 19968, 1, 512] - - [208, 85.436] - - - [40960, 20480, 1, 512] - - [205, 85.375] - - - [41472, 20480, 1, 512] - - [208, 85.495] - - - [41472, 20992, 1, 512] - - [207, 85.372] - - - [41984, 20480, 1, 512] - - [205, 85.546] - - - [41984, 20992, 1, 512] - - [208, 85.556] - - - [42496, 20992, 1, 512] - - [209, 85.065] - - - [42496, 21504, 1, 512] - - [209, 85.608] - - - [43008, 20992, 1, 512] - - [205, 85.703] - - - [43008, 21504, 1, 512] - - [208, 85.66] - - - [43520, 21504, 1, 512] - - [207, 85.379] - - - [43520, 22016, 1, 512] - - [202, 85.395] - - - [44032, 21504, 1, 512] - - [205, 85.443] - - - [44032, 22016, 1, 512] - - [209, 85.521] - - - [44544, 22016, 1, 512] - - [205, 84.962] - - - [44544, 22528, 1, 512] - - [205, 85.371] - - - [45056, 22016, 1, 512] - - [201, 85.168] - - - [45056, 22528, 1, 512] - - [202, 85.295] - - - [45568, 22528, 1, 512] - - [202, 85.07] - - - [45568, 23040, 1, 512] - - [205, 85.337] - - - [46080, 22528, 1, 512] - - [207, 85.264] - - - [46080, 23040, 1, 512] - - [208, 85.478] - - - [46592, 23040, 1, 512] - - [208, 85.684] - - - [46592, 23552, 1, 512] - - [205, 85.613] - - - [47104, 23040, 1, 512] - - [205, 85.402] - - - [47104, 23552, 1, 512] - - [208, 85.513] - - - [47616, 23552, 1, 512] - - [207, 85.578] - - - [47616, 24064, 1, 512] - - [205, 85.314] - - - [48128, 23552, 1, 512] - - [208, 85.39] - - - [48128, 24064, 1, 512] - - [205, 85.251] - - - [48640, 24064, 1, 512] - - [208, 85.233] - - - [48640, 24576, 1, 512] - - [208, 85.024] - - - [49152, 24064, 1, 512] - - [207, 85.112] - - - [49152, 24576, 1, 512] - - [205, 85.169] - - - [49664, 24576, 1, 512] - - [202, 84.879] - - - [49664, 25088, 1, 512] - - [208, 84.797] - - - [50176, 24576, 1, 512] - - [205, 85.138] - - - [50176, 25088, 1, 512] - - [205, 85.168] - - - [50688, 25088, 1, 512] - - [208, 85.152] - - - [50688, 25600, 1, 512] - - [208, 85.168] - - - [51200, 25088, 1, 512] - - [208, 84.848] - - - [51200, 25600, 1, 512] - - [208, 85.068] - - - [51712, 25600, 1, 512] - - [207, 84.782] - - - [51712, 26112, 1, 512] - - [205, 84.817] - - - [52224, 25600, 1, 512] - - [205, 85.059] - - - [52224, 26112, 1, 512] - - [208, 84.826] - - - [52736, 26112, 1, 512] - - [205, 85.035] - - - [52736, 26624, 1, 512] - - [205, 85.104] - - - [53248, 26112, 1, 512] - - [208, 85.047] - - - [53248, 26624, 1, 512] - - [208, 85.131] - - - [53760, 26624, 1, 512] - - [207, 85.097] - - - [53760, 27136, 1, 512] - - [205, 84.998] - - - [54272, 26624, 1, 512] - - [205, 85.013] - - - [54272, 27136, 1, 512] - - [207, 84.717] - - - [54784, 27136, 1, 512] - - [205, 85.042] - - - [54784, 27648, 1, 512] - - [208, 84.886] - - - [55296, 27136, 1, 512] - - [207, 84.754] - - - [55296, 27648, 1, 512] - - [205, 84.825] - - - [55808, 27648, 1, 512] - - [208, 84.761] - - - [55808, 28160, 1, 512] - - [205, 84.62] - - - [56320, 27648, 1, 512] - - [205, 84.893] - - - [56320, 28160, 1, 512] - - [208, 84.806] - - - [56832, 28160, 1, 512] - - [205, 84.718] - - - [56832, 28672, 1, 512] - - [205, 84.685] - - - [57344, 28160, 1, 512] - - [208, 84.627] - - - [57344, 28672, 1, 512] - - [205, 84.729] - - - [57856, 28672, 1, 512] - - [208, 84.537] - - - [57856, 29184, 1, 512] - - [205, 84.424] - - - [58368, 28672, 1, 512] - - [208, 84.657] - - - [58368, 29184, 1, 512] - - [205, 84.653] - - - [58880, 29184, 1, 512] - - [207, 84.359] - - - [58880, 29696, 1, 512] - - [208, 84.759] - - - [59392, 29184, 1, 512] - - [209, 84.657] - - - [59392, 29696, 1, 512] - - [208, 84.797] - - - [59904, 29696, 1, 512] - - [208, 84.575] - - - [59904, 30208, 1, 512] - - [205, 84.615] - - - [60416, 29696, 1, 512] - - [208, 84.599] - - - [60416, 30208, 1, 512] - - [205, 84.651] - - - [60928, 30208, 1, 512] - - [205, 84.553] - - - [60928, 30720, 1, 512] - - [208, 84.488] - - - [61440, 30208, 1, 512] - - [208, 84.498] - - - [61440, 30720, 1, 512] - - [208, 84.577] - - - [61952, 30720, 1, 512] - - [208, 84.527] - - - [61952, 31232, 1, 512] - - [205, 84.456] - - - [62464, 30720, 1, 512] - - [207, 84.607] - - - [62464, 31232, 1, 512] - - [207, 84.326] - - - [62976, 31232, 1, 512] - - [205, 84.358] - - - [62976, 31744, 1, 512] - - [207, 84.63] - - - [63488, 31232, 1, 512] - - [207, 84.313] - - - [63488, 31744, 1, 512] - - [208, 84.622] - - - [64000, 31744, 1, 512] - - [205, 84.507] - - - [64000, 32256, 1, 512] - - [208, 84.66] - - - [64512, 31744, 1, 512] - - [205, 84.679] - - - [64512, 32256, 1, 512] - - [208, 84.597] - - - [65024, 512, 1, 512] - - [202, 73.819] - - - [65024, 32256, 1, 512] - - [205, 84.503] - - - [65024, 32768, 1, 512] - - [205, 84.493] - - - [65536, 512, 1, 512] - - [199, 73.413] - - - [65536, 32256, 1, 512] - - [208, 84.57] - - - [65536, 32768, 1, 512] - - [208, 84.455] - - - [66048, 512, 1, 512] - - [203, 73.298] - - - [66048, 32768, 1, 512] - - [205, 84.617] - - - [66048, 33280, 1, 512] - - [208, 84.537] - - - [66560, 512, 1, 512] - - [202, 73.742] - - - [66560, 32768, 1, 512] - - [205, 84.528] - - - [66560, 33280, 1, 512] - - [205, 84.58] - - - [67072, 512, 1, 512] - - [199, 74.358] - - - [67072, 33280, 1, 512] - - [205, 84.298] - - - [67072, 33792, 1, 512] - - [205, 84.411] - - - [67584, 512, 1, 512] - - [199, 74.319] - - - [67584, 33280, 1, 512] - - [208, 84.365] - - - [67584, 33792, 1, 512] - - [205, 84.417] - - - [68096, 512, 1, 512] - - [202, 75.008] - - - [68096, 33792, 1, 512] - - [208, 84.353] - - - [68096, 34304, 1, 512] - - [208, 84.533] - - - [68608, 512, 1, 512] - - [201, 73.965] - - - [68608, 33792, 1, 512] - - [208, 84.46] - - - [68608, 34304, 1, 512] - - [208, 84.381] - - - [69120, 512, 1, 512] - - [201, 74.004] - - - [69120, 34304, 1, 512] - - [205, 84.41] - - - [69120, 34816, 1, 512] - - [205, 84.364] - - - [69632, 512, 1, 512] - - [199, 72.304] - - - [69632, 34304, 1, 512] - - [207, 84.194] - - - [69632, 34816, 1, 512] - - [205, 84.281] - - - [70144, 512, 1, 512] - - [201, 73.934] - - - [70144, 34816, 1, 512] - - [205, 84.278] - - - [70144, 35328, 1, 512] - - [208, 84.352] - - - [70656, 512, 1, 512] - - [203, 74.654] - - - [70656, 34816, 1, 512] - - [205, 84.424] - - - [70656, 35328, 1, 512] - - [208, 84.342] - - - [71168, 512, 1, 512] - - [199, 75.027] - - - [71168, 35328, 1, 512] - - [205, 84.414] - - - [71168, 35840, 1, 512] - - [205, 84.377] - - - [71680, 512, 1, 512] - - [201, 74.502] - - - [71680, 35328, 1, 512] - - [205, 84.299] - - - [71680, 35840, 1, 512] - - [205, 84.449] - - - [72192, 512, 1, 512] - - [201, 74.491] - - - [72192, 35840, 1, 512] - - [205, 84.279] - - - [72192, 36352, 1, 512] - - [208, 84.482] - - - [72704, 512, 1, 512] - - [203, 74.408] - - - [72704, 35840, 1, 512] - - [208, 84.391] - - - [72704, 36352, 1, 512] - - [208, 84.494] - - - [73216, 512, 1, 512] - - [201, 73.581] - - - [73216, 36352, 1, 512] - - [208, 84.448] - - - [73216, 36864, 1, 512] - - [207, 84.36] - - - [73728, 512, 1, 512] - - [203, 73.569] - - - [73728, 36352, 1, 512] - - [208, 84.543] - - - [73728, 36864, 1, 512] - - [205, 84.469] - - - [74240, 512, 1, 512] - - [202, 75.153] - - - [74240, 36864, 1, 512] - - [205, 84.412] - - - [74240, 37376, 1, 512] - - [208, 84.234] - - - [74752, 512, 1, 512] - - [199, 74.902] - - - [74752, 36864, 1, 512] - - [205, 84.533] - - - [74752, 37376, 1, 512] - - [208, 84.332] - - - [75264, 512, 1, 512] - - [199, 74.774] - - - [75264, 37376, 1, 512] - - [205, 84.374] - - - [75264, 37888, 1, 512] - - [208, 84.383] - - - [75776, 512, 1, 512] - - [203, 74.265] - - - [75776, 37376, 1, 512] - - [208, 84.348] - - - [75776, 37888, 1, 512] - - [205, 84.471] - - - [76288, 512, 1, 512] - - [202, 74.55] - - - [76288, 37888, 1, 512] - - [208, 84.359] - - - [76288, 38400, 1, 512] - - [205, 84.272] - - - [76800, 512, 1, 512] - - [203, 74.982] - - - [76800, 37888, 1, 512] - - [208, 84.512] - - - [76800, 38400, 1, 512] - - [205, 84.482] - - - [77312, 512, 1, 512] - - [201, 75.413] - - - [77312, 38400, 1, 512] - - [208, 84.36] - - - [77312, 38912, 1, 512] - - [205, 84.437] - - - [77824, 512, 1, 512] - - [201, 74.323] - - - [77824, 38400, 1, 512] - - [208, 84.423] - - - [77824, 38912, 1, 512] - - [208, 84.371] - - - [78336, 512, 1, 512] - - [201, 75.012] - - - [78336, 38912, 1, 512] - - [208, 84.416] - - - [78336, 39424, 1, 512] - - [208, 84.513] - - - [78848, 512, 1, 512] - - [201, 75.773] - - - [78848, 38912, 1, 512] - - [205, 84.521] - - - [78848, 39424, 1, 512] - - [208, 84.224] - - - [79360, 512, 1, 512] - - [202, 75.931] - - - [79360, 39424, 1, 512] - - [208, 84.446] - - - [79360, 39936, 1, 512] - - [205, 84.419] - - - [79872, 512, 1, 512] - - [202, 75.88] - - - [79872, 39424, 1, 512] - - [208, 84.524] - - - [79872, 39936, 1, 512] - - [208, 84.488] - - - [80384, 512, 1, 512] - - [199, 75.128] - - - [80384, 39936, 1, 512] - - [205, 84.443] - - - [80384, 40448, 1, 512] - - [205, 84.432] - - - [80896, 512, 1, 512] - - [201, 75.261] - - - [80896, 39936, 1, 512] - - [208, 84.555] - - - [80896, 40448, 1, 512] - - [208, 84.472] - - - [81408, 512, 1, 512] - - [203, 74.941] - - - [81408, 40448, 1, 512] - - [208, 84.477] - - - [81408, 40960, 1, 512] - - [208, 84.414] - - - [81920, 512, 1, 512] - - [202, 75.202] - - - [81920, 40448, 1, 512] - - [205, 84.404] - - - [81920, 40960, 1, 512] - - [205, 84.409] - - - [82432, 512, 1, 512] - - [199, 75.595] - - - [82432, 40960, 1, 512] - - [208, 84.334] - - - [82432, 41472, 1, 512] - - [208, 84.507] - - - [82944, 512, 1, 512] - - [201, 75.689] - - - [82944, 40960, 1, 512] - - [205, 84.471] - - - [82944, 41472, 1, 512] - - [208, 84.454] - - - [83456, 512, 1, 512] - - [202, 76.074] - - - [83456, 41472, 1, 512] - - [208, 84.373] - - - [83456, 41984, 1, 512] - - [208, 84.484] - - - [83968, 512, 1, 512] - - [203, 75.423] - - - [83968, 41472, 1, 512] - - [205, 84.372] - - - [83968, 41984, 1, 512] - - [205, 84.505] - - - [84480, 512, 1, 512] - - [202, 76.161] - - - [84480, 41984, 1, 512] - - [205, 84.452] - - - [84480, 42496, 1, 512] - - [208, 84.49] - - - [84992, 512, 1, 512] - - [203, 76.5] - - - [84992, 41984, 1, 512] - - [208, 84.473] - - - [84992, 42496, 1, 512] - - [208, 84.406] - - - [85504, 512, 1, 512] - - [203, 75.499] - - - [85504, 42496, 1, 512] - - [208, 84.347] - - - [85504, 43008, 1, 512] - - [208, 84.388] - - - [86016, 512, 1, 512] - - [203, 74.558] - - - [86016, 42496, 1, 512] - - [205, 84.4] - - - [86016, 43008, 1, 512] - - [205, 84.438] - - - [86528, 512, 1, 512] - - [203, 75.848] - - - [86528, 43008, 1, 512] - - [208, 84.484] - - - [86528, 43520, 1, 512] - - [208, 84.487] - - - [87040, 512, 1, 512] - - [203, 75.707] - - - [87040, 43008, 1, 512] - - [205, 84.486] - - - [87040, 43520, 1, 512] - - [209, 84.333] - - - [87552, 512, 1, 512] - - [199, 75.668] - - - [87552, 43520, 1, 512] - - [205, 84.451] - - - [87552, 44032, 1, 512] - - [208, 84.454] - - - [88064, 512, 1, 512] - - [201, 76.351] - - - [88064, 43520, 1, 512] - - [205, 84.503] - - - [88064, 44032, 1, 512] - - [205, 84.476] - - - [88576, 512, 1, 512] - - [202, 76.379] - - - [88576, 44032, 1, 512] - - [208, 84.413] - - - [88576, 44544, 1, 512] - - [208, 84.387] - - - [89088, 512, 1, 512] - - [202, 76.436] - - - [89088, 44032, 1, 512] - - [208, 84.447] - - - [89088, 44544, 1, 512] - - [205, 84.45] - - - [89600, 512, 1, 512] - - [201, 76.409] - - - [89600, 44544, 1, 512] - - [205, 84.419] - - - [89600, 45056, 1, 512] - - [208, 84.455] - - - [90112, 512, 1, 512] - - [201, 76.088] - - - [90112, 44544, 1, 512] - - [205, 84.473] - - - [90112, 45056, 1, 512] - - [205, 84.406] - - - [90624, 512, 1, 512] - - [199, 76.166] - - - [90624, 45056, 1, 512] - - [208, 84.355] - - - [90624, 45568, 1, 512] - - [207, 84.347] - - - [91136, 512, 1, 512] - - [202, 76.141] - - - [91136, 45056, 1, 512] - - [208, 84.537] - - - [91136, 45568, 1, 512] - - [208, 84.484] - - - [91648, 512, 1, 512] - - [199, 76.765] - - - [91648, 45568, 1, 512] - - [205, 84.392] - - - [91648, 46080, 1, 512] - - [205, 84.476] - - - [92160, 512, 1, 512] - - [201, 77.298] - - - [92160, 45568, 1, 512] - - [205, 84.514] - - - [92160, 46080, 1, 512] - - [205, 84.534] - - - [92672, 512, 1, 512] - - [202, 76.42] - - - [92672, 46080, 1, 512] - - [205, 84.419] - - - [92672, 46592, 1, 512] - - [208, 84.398] - - - [93184, 512, 1, 512] - - [203, 76.601] - - - [93184, 46080, 1, 512] - - [208, 84.534] - - - [93184, 46592, 1, 512] - - [208, 84.522] - - - [93696, 512, 1, 512] - - [199, 76.601] - - - [93696, 46592, 1, 512] - - [205, 84.466] - - - [93696, 47104, 1, 512] - - [208, 84.455] - - - [94208, 512, 1, 512] - - [199, 76.004] - - - [94208, 46592, 1, 512] - - [208, 84.396] - - - [94208, 47104, 1, 512] - - [208, 84.454] - - - [94720, 512, 1, 512] - - [202, 76.438] - - - [94720, 47104, 1, 512] - - [205, 84.486] - - - [94720, 47616, 1, 512] - - [208, 84.516] - - - [95232, 512, 1, 512] - - [199, 76.972] - - - [95232, 47104, 1, 512] - - [208, 84.557] - - - [95232, 47616, 1, 512] - - [205, 84.552] - - - [95744, 512, 1, 512] - - [199, 76.61] - - - [95744, 47616, 1, 512] - - [205, 84.403] - - - [95744, 48128, 1, 512] - - [205, 84.408] - - - [96256, 512, 1, 512] - - [202, 77.293] - - - [96256, 47616, 1, 512] - - [208, 84.577] - - - [96256, 48128, 1, 512] - - [208, 84.555] - - - [96768, 512, 1, 512] - - [203, 76.477] - - - [96768, 48128, 1, 512] - - [205, 84.523] - - - [96768, 48640, 1, 512] - - [205, 84.398] - - - [97280, 512, 1, 512] - - [203, 77.021] - - - [97280, 48128, 1, 512] - - [208, 84.507] - - - [97280, 48640, 1, 512] - - [208, 84.482] - - - [97792, 512, 1, 512] - - [199, 76.941] - - - [97792, 48640, 1, 512] - - [208, 84.406] - - - [97792, 49152, 1, 512] - - [208, 84.38] - - - [98304, 512, 1, 512] - - [199, 75.274] - - - [98304, 48640, 1, 512] - - [208, 84.371] - - - [98304, 49152, 1, 512] - - [208, 84.317] - - - [98816, 512, 1, 512] - - [203, 77.36] - - - [98816, 49152, 1, 512] - - [208, 84.337] - - - [98816, 49664, 1, 512] - - [208, 84.464] - - - [99328, 512, 1, 512] - - [201, 77.234] - - - [99328, 49152, 1, 512] - - [208, 84.555] - - - [99328, 49664, 1, 512] - - [205, 84.453] - - - [99840, 512, 1, 512] - - [199, 77.617] - - - [99840, 49664, 1, 512] - - [205, 84.445] - - - [99840, 50176, 1, 512] - - [208, 84.434] - - - [100352, 512, 1, 512] - - [201, 77.515] - - - [100352, 49664, 1, 512] - - [205, 84.56] - - - [100352, 50176, 1, 512] - - [205, 84.554] - - - [100864, 512, 1, 512] - - [202, 77.374] - - - [100864, 50176, 1, 512] - - [208, 84.41] - - - [100864, 50688, 1, 512] - - [205, 84.459] - - - [101376, 512, 1, 512] - - [202, 76.864] - - - [101376, 50176, 1, 512] - - [205, 84.513] - - - [101376, 50688, 1, 512] - - [208, 84.52] - - - [101888, 512, 1, 512] - - [201, 76.921] - - - [101888, 50688, 1, 512] - - [205, 84.323] - - - [101888, 51200, 1, 512] - - [208, 84.46] - - - [102400, 512, 1, 512] - - [199, 76.414] - - - [102400, 50688, 1, 512] - - [205, 84.426] - - - [102400, 51200, 1, 512] - - [205, 84.483] - - - [102912, 512, 1, 512] - - [199, 77.034] - - - [102912, 51200, 1, 512] - - [208, 84.463] - - - [102912, 51712, 1, 512] - - [205, 84.456] - - - [103424, 512, 1, 512] - - [202, 77.418] - - - [103424, 51200, 1, 512] - - [205, 84.522] - - - [103424, 51712, 1, 512] - - [208, 84.497] - - - [103936, 512, 1, 512] - - [199, 77.374] - - - [103936, 51712, 1, 512] - - [208, 84.378] - - - [103936, 52224, 1, 512] - - [208, 84.45] - - - [104448, 512, 1, 512] - - [203, 77.585] - - - [104448, 51712, 1, 512] - - [205, 84.496] - - - [104448, 52224, 1, 512] - - [205, 84.534] - - - [104960, 512, 1, 512] - - [202, 76.982] - - - [104960, 52224, 1, 512] - - [208, 84.456] - - - [104960, 52736, 1, 512] - - [208, 84.415] - - - [105472, 512, 1, 512] - - [203, 77.18] - - - [105472, 52224, 1, 512] - - [205, 84.522] - - - [105472, 52736, 1, 512] - - [208, 84.554] - - - [105984, 512, 1, 512] - - [203, 77.363] - - - [105984, 52736, 1, 512] - - [205, 84.424] - - - [105984, 53248, 1, 512] - - [208, 84.417] - - - [106496, 512, 1, 512] - - [203, 76.517] - - - [106496, 52736, 1, 512] - - [205, 84.451] - - - [106496, 53248, 1, 512] - - [208, 84.467] - - - [107008, 512, 1, 512] - - [203, 77.252] - - - [107008, 53248, 1, 512] - - [205, 84.405] - - - [107008, 53760, 1, 512] - - [205, 84.41] - - - [107520, 512, 1, 512] - - [203, 78.2] - - - [107520, 53248, 1, 512] - - [205, 84.536] - - - [107520, 53760, 1, 512] - - [208, 84.436] - - - [108032, 512, 1, 512] - - [203, 76.961] - - - [108032, 53760, 1, 512] - - [205, 84.46] - - - [108032, 54272, 1, 512] - - [208, 84.491] - - - [108544, 512, 1, 512] - - [199, 77.396] - - - [108544, 53760, 1, 512] - - [208, 84.506] - - - [108544, 54272, 1, 512] - - [208, 84.538] - - - [109056, 512, 1, 512] - - [201, 77.299] - - - [109056, 54272, 1, 512] - - [208, 84.49] - - - [109056, 54784, 1, 512] - - [208, 84.494] - - - [109568, 512, 1, 512] - - [201, 77.807] - - - [109568, 54272, 1, 512] - - [208, 84.592] - - - [109568, 54784, 1, 512] - - [205, 84.481] - - - [110080, 512, 1, 512] - - [203, 77.453] - - - [110080, 54784, 1, 512] - - [205, 84.552] - - - [110080, 55296, 1, 512] - - [205, 84.521] - - - [110592, 512, 1, 512] - - [202, 76.902] - - - [110592, 54784, 1, 512] - - [205, 84.528] - - - [110592, 55296, 1, 512] - - [205, 84.46] - - - [111104, 512, 1, 512] - - [199, 77.381] - - - [111104, 55296, 1, 512] - - [208, 84.516] - - - [111104, 55808, 1, 512] - - [205, 84.508] - - - [111616, 512, 1, 512] - - [199, 77.016] - - - [111616, 55296, 1, 512] - - [208, 84.557] - - - [111616, 55808, 1, 512] - - [208, 84.513] - - - [112128, 512, 1, 512] - - [203, 77.446] - - - [112128, 55808, 1, 512] - - [205, 84.472] - - - [112128, 56320, 1, 512] - - [205, 84.524] - - - [112640, 512, 1, 512] - - [199, 77.275] - - - [112640, 55808, 1, 512] - - [205, 84.545] - - - [112640, 56320, 1, 512] - - [208, 84.547] - - - [113152, 512, 1, 512] - - [203, 78.135] - - - [113152, 56320, 1, 512] - - [205, 84.498] - - - [113152, 56832, 1, 512] - - [205, 84.517] - - - [113664, 512, 1, 512] - - [203, 78.105] - - - [113664, 56320, 1, 512] - - [208, 84.557] - - - [113664, 56832, 1, 512] - - [208, 84.512] - - - [114176, 512, 1, 512] - - [202, 78.134] - - - [114176, 56832, 1, 512] - - [205, 84.47] - - - [114176, 57344, 1, 512] - - [205, 84.431] - - - [114688, 512, 1, 512] - - [202, 76.538] - - - [114688, 56832, 1, 512] - - [205, 84.436] - - - [114688, 57344, 1, 512] - - [208, 84.407] - - - [115200, 512, 1, 512] - - [203, 77.242] - - - [115200, 57344, 1, 512] - - [208, 84.397] - - - [115200, 57856, 1, 512] - - [208, 84.455] - - - [115712, 512, 1, 512] - - [202, 77.957] - - - [115712, 57344, 1, 512] - - [208, 84.509] - - - [115712, 57856, 1, 512] - - [208, 84.518] - - - [116224, 512, 1, 512] - - [201, 77.748] - - - [116224, 57856, 1, 512] - - [205, 84.515] - - - [116224, 58368, 1, 512] - - [205, 84.465] - - - [116736, 512, 1, 512] - - [203, 78.214] - - - [116736, 57856, 1, 512] - - [205, 84.516] - - - [116736, 58368, 1, 512] - - [208, 84.534] - - - [117248, 512, 1, 512] - - [199, 78.189] - - - [117248, 58368, 1, 512] - - [208, 84.508] - - - [117248, 58880, 1, 512] - - [208, 84.476] - - - [117760, 512, 1, 512] - - [203, 77.984] - - - [117760, 58368, 1, 512] - - [208, 84.528] - - - [117760, 58880, 1, 512] - - [205, 84.514] - - - [118272, 512, 1, 512] - - [203, 77.843] - - - [118272, 58880, 1, 512] - - [208, 84.458] - - - [118272, 59392, 1, 512] - - [208, 84.486] - - - [118784, 512, 1, 512] - - [202, 76.849] - - - [118784, 58880, 1, 512] - - [208, 84.454] - - - [118784, 59392, 1, 512] - - [205, 84.493] - - - [119296, 512, 1, 512] - - [203, 78.065] - - - [119296, 59392, 1, 512] - - [205, 84.489] - - - [119296, 59904, 1, 512] - - [205, 84.443] - - - [119808, 512, 1, 512] - - [199, 77.93] - - - [119808, 59392, 1, 512] - - [205, 84.479] - - - [119808, 59904, 1, 512] - - [208, 84.539] - - - [120320, 512, 1, 512] - - [203, 78.323] - - - [120320, 59904, 1, 512] - - [208, 84.503] - - - [120320, 60416, 1, 512] - - [205, 84.506] - - - [120832, 512, 1, 512] - - [203, 78.669] - - - [120832, 59904, 1, 512] - - [205, 84.543] - - - [120832, 60416, 1, 512] - - [205, 84.577] - - - [121344, 512, 1, 512] - - [202, 78.466] - - - [121344, 60416, 1, 512] - - [208, 84.47] - - - [121344, 60928, 1, 512] - - [208, 84.429] - - - [121856, 512, 1, 512] - - [203, 78.249] - - - [121856, 60416, 1, 512] - - [205, 84.525] - - - [121856, 60928, 1, 512] - - [208, 84.472] - - - [122368, 512, 1, 512] - - [201, 77.96] - - - [122368, 60928, 1, 512] - - [205, 84.459] - - - [122368, 61440, 1, 512] - - [205, 84.434] - - - [122880, 512, 1, 512] - - [203, 77.203] - - - [122880, 60928, 1, 512] - - [205, 84.433] - - - [122880, 61440, 1, 512] - - [205, 84.423] - - - [123392, 512, 1, 512] - - [203, 77.819] - - - [123392, 61440, 1, 512] - - [208, 84.437] - - - [123392, 61952, 1, 512] - - [205, 84.473] - - - [123904, 512, 1, 512] - - [199, 77.888] - - - [123904, 61440, 1, 512] - - [205, 84.526] - - - [123904, 61952, 1, 512] - - [208, 84.542] - - - [124416, 512, 1, 512] - - [203, 78.389] - - - [124416, 61952, 1, 512] - - [208, 84.54] - - - [124416, 62464, 1, 512] - - [205, 84.455] - - - [124928, 512, 1, 512] - - [203, 78.352] - - - [124928, 61952, 1, 512] - - [205, 84.543] - - - [124928, 62464, 1, 512] - - [208, 84.506] - - - [125440, 512, 1, 512] - - [202, 78.241] - - - [125440, 62464, 1, 512] - - [208, 84.456] - - - [125440, 62976, 1, 512] - - [205, 84.495] - - - [125952, 512, 1, 512] - - [201, 78.124] - - - [125952, 62464, 1, 512] - - [208, 84.551] - - - [125952, 62976, 1, 512] - - [205, 84.513] - - - [126464, 512, 1, 512] - - [199, 77.776] - - - [126464, 62976, 1, 512] - - [208, 84.493] - - - [126464, 63488, 1, 512] - - [208, 84.479] - - - [126976, 512, 1, 512] - - [201, 77.478] - - - [126976, 62976, 1, 512] - - [208, 84.481] - - - [126976, 63488, 1, 512] - - [205, 84.443] - - - [127488, 512, 1, 512] - - [199, 78.129] - - - [127488, 63488, 1, 512] - - [208, 84.437] - - - [127488, 64000, 1, 512] - - [205, 84.446] - - - [128000, 512, 1, 512] - - [202, 78.81] - - - [128000, 63488, 1, 512] - - [208, 84.518] - - - [128000, 64000, 1, 512] - - [208, 84.512] - - - [4096, 1537, 1, 512] - - [203, 56.965] - - - [4096, 2049, 1, 512] - - [199, 60.102] - - - [4608, 2049, 1, 512] - - [199, 59.312] - - - [5120, 2049, 1, 512] - - [210, 60.259] - - - [5120, 2561, 1, 512] - - [199, 64.18] - - - [5632, 2561, 1, 512] - - [199, 63.27] - - - [6144, 2561, 1, 512] - - [201, 64.076] - - - [6144, 3073, 1, 512] - - [201, 67.052] - - - [6656, 3073, 1, 512] - - [203, 66.368] - - - [7168, 3073, 1, 512] - - [199, 67.451] - - - [7168, 3585, 1, 512] - - [201, 68.396] - - - [7680, 3585, 1, 512] - - [202, 68.39] - - - [8192, 3585, 1, 512] - - [199, 70.039] - - - [8192, 4097, 1, 512] - - [201, 73.424] - - - [8704, 4097, 1, 512] - - [199, 71.405] - - - [9216, 4097, 1, 512] - - [201, 72.618] - - - [9216, 4609, 1, 512] - - [201, 72.08] - - - [9728, 4609, 1, 512] - - [199, 73.974] - - - [10240, 4609, 1, 512] - - [201, 72.797] - - - [10240, 5121, 1, 512] - - [201, 76.207] - - - [10752, 5121, 1, 512] - - [201, 74.785] - - - [11264, 5121, 1, 512] - - [201, 73.871] - - - [11264, 5633, 1, 512] - - [199, 75.663] - - - [11776, 5633, 1, 512] - - [201, 74.877] - - - [12288, 5633, 1, 512] - - [199, 76.802] - - - [12288, 6145, 1, 512] - - [201, 78.555] - - - [12800, 6145, 1, 512] - - [201, 77.139] - - - [13312, 6145, 1, 512] - - [201, 77.862] - - - [13312, 6657, 1, 512] - - [201, 76.86] - - - [13824, 6657, 1, 512] - - [199, 76.819] - - - [14336, 6657, 1, 512] - - [199, 77.92] - - - [14336, 7169, 1, 512] - - [199, 78.882] - - - [14848, 7169, 1, 512] - - [199, 78.606] - - - [15360, 7169, 1, 512] - - [201, 76.953] - - - [15360, 7681, 1, 512] - - [199, 78.092] - - - [15872, 7681, 1, 512] - - [199, 79.504] - - - [16384, 7681, 1, 512] - - [201, 81.752] - - - [16384, 8193, 1, 512] - - [201, 81.803] - - - [16896, 8193, 1, 512] - - [201, 80.207] - - - [17408, 8193, 1, 512] - - [199, 80.476] - - - [17408, 8705, 1, 512] - - [201, 79.606] - - - [17920, 8705, 1, 512] - - [201, 79.421] - - - [18432, 8705, 1, 512] - - [201, 80.548] - - - [18432, 9217, 1, 512] - - [199, 80.984] - - - [18944, 9217, 1, 512] - - [199, 81.199] - - - [19456, 9217, 1, 512] - - [201, 81.591] - - - [19456, 9729, 1, 512] - - [201, 81.603] - - - [19968, 9729, 1, 512] - - [201, 79.514] - - - [20480, 9729, 1, 512] - - [201, 81.67] - - - [20480, 10241, 1, 512] - - [202, 81.669] - - - [20992, 10241, 1, 512] - - [199, 81.649] - - - [21504, 10241, 1, 512] - - [199, 81.53] - - - [21504, 10753, 1, 512] - - [201, 80.46] - - - [22016, 10753, 1, 512] - - [199, 80.263] - - - [22528, 10753, 1, 512] - - [199, 80.785] - - - [22528, 11265, 1, 512] - - [201, 82.576] - - - [23040, 11265, 1, 512] - - [201, 81.682] - - - [23552, 11265, 1, 512] - - [199, 82.218] - - - [23552, 11777, 1, 512] - - [201, 81.835] - - - [24064, 11777, 1, 512] - - [199, 81.988] - - - [24576, 11777, 1, 512] - - [201, 82.666] - - - [24576, 12289, 1, 512] - - [201, 83.301] - - - [25088, 12289, 1, 512] - - [201, 82.447] - - - [25600, 12289, 1, 512] - - [201, 82.674] - - - [25600, 12801, 1, 512] - - [201, 81.397] - - - [26112, 12801, 1, 512] - - [201, 80.87] - - - [26624, 12801, 1, 512] - - [199, 81.737] - - - [26624, 13313, 1, 512] - - [201, 82.949] - - - [27136, 13313, 1, 512] - - [201, 82.904] - - - [27648, 13313, 1, 512] - - [201, 83.081] - - - [27648, 13825, 1, 512] - - [208, 83.267] - - - [28160, 13825, 1, 512] - - [205, 82.508] - - - [28672, 13825, 1, 512] - - [201, 82.727] - - - [28672, 14337, 1, 512] - - [209, 83.377] - - - [29184, 14337, 1, 512] - - [201, 82.621] - - - [29696, 14337, 1, 512] - - [208, 84.371] - - - [29696, 14849, 1, 512] - - [201, 81.521] - - - [30208, 14849, 1, 512] - - [205, 84.292] - - - [30720, 14849, 1, 512] - - [207, 83.971] - - - [30720, 15361, 1, 512] - - [209, 82.731] - - - [31232, 15361, 1, 512] - - [199, 82.586] - - - [31744, 15361, 1, 512] - - [205, 83.832] - - - [31744, 15873, 1, 512] - - [199, 83.69] - - - [32256, 15873, 1, 512] - - [201, 83.39] - - - [32768, 15873, 1, 512] - - [201, 83.769] - - - [32768, 16385, 1, 512] - - [199, 84.379] - - - [33280, 16385, 1, 512] - - [207, 83.991] - - - [33792, 16385, 1, 512] - - [208, 83.593] - - - [33792, 16897, 1, 512] - - [199, 82.453] - - - [34304, 16897, 1, 512] - - [208, 82.902] - - - [34816, 16897, 1, 512] - - [201, 83.07] - - - [34816, 17409, 1, 512] - - [201, 83.76] - - - [35328, 17409, 1, 512] - - [202, 82.928] - - - [35840, 17409, 1, 512] - - [207, 83.981] - - - [35840, 17921, 1, 512] - - [205, 84.455] - - - [36352, 17921, 1, 512] - - [205, 83.351] - - - [36864, 17921, 1, 512] - - [207, 83.903] - - - [36864, 18433, 1, 512] - - [209, 83.834] - - - [37376, 18433, 1, 512] - - [199, 83.174] - - - [37888, 18433, 1, 512] - - [208, 84.623] - - - [37888, 18945, 1, 512] - - [201, 82.701] - - - [38400, 18945, 1, 512] - - [199, 83.07] - - - [38912, 18945, 1, 512] - - [205, 83.553] - - - [38912, 19457, 1, 512] - - [202, 83.679] - - - [39424, 19457, 1, 512] - - [199, 83.502] - - - [39936, 19457, 1, 512] - - [207, 84.459] - - - [39936, 19969, 1, 512] - - [199, 83.852] - - - [40448, 19969, 1, 512] - - [205, 83.896] - - - [40960, 19969, 1, 512] - - [207, 83.812] - - - [40960, 20481, 1, 512] - - [205, 83.882] - - - [41472, 20481, 1, 512] - - [201, 83.708] - - - [41984, 20481, 1, 512] - - [208, 83.92] - - - [41984, 20993, 1, 512] - - [205, 83.677] - - - [42496, 20993, 1, 512] - - [201, 82.994] - - - [43008, 20993, 1, 512] - - [209, 82.944] - - - [43008, 21505, 1, 512] - - [205, 83.762] - - - [43520, 21505, 1, 512] - - [201, 83.13] - - - [44032, 21505, 1, 512] - - [208, 84.005] - - - [44032, 22017, 1, 512] - - [201, 83.295] - - - [44544, 22017, 1, 512] - - [199, 83.469] - - - [45056, 22017, 1, 512] - - [205, 84.194] - - - [45056, 22529, 1, 512] - - [207, 84.169] - - - [45568, 22529, 1, 512] - - [199, 83.61] - - - [46080, 22529, 1, 512] - - [205, 84.851] - - - [46080, 23041, 1, 512] - - [207, 83.567] - - - [46592, 23041, 1, 512] - - [199, 83.68] - - - [47104, 23041, 1, 512] - - [205, 84.531] - - - [47104, 23553, 1, 512] - - [207, 84.612] - - - [47616, 23553, 1, 512] - - [208, 84.355] - - - [48128, 23553, 1, 512] - - [205, 84.046] - - - [48128, 24065, 1, 512] - - [209, 83.801] - - - [48640, 24065, 1, 512] - - [208, 84.247] - - - [49152, 24065, 1, 512] - - [208, 83.899] - - - [49152, 24577, 1, 512] - - [209, 84.506] - - - [49664, 24577, 1, 512] - - [208, 83.95] - - - [50176, 24577, 1, 512] - - [208, 84.071] - - - [50176, 25089, 1, 512] - - [205, 84.003] - - - [50688, 25089, 1, 512] - - [201, 83.161] - - - [51200, 25089, 1, 512] - - [209, 83.153] - - - [51200, 25601, 1, 512] - - [205, 84.442] - - - [51712, 25601, 1, 512] - - [208, 83.989] - - - [52224, 25601, 1, 512] - - [207, 83.97] - - - [52224, 26113, 1, 512] - - [208, 84.059] - - - [52736, 26113, 1, 512] - - [205, 83.975] - - - [53248, 26113, 1, 512] - - [209, 83.362] - - - [53248, 26625, 1, 512] - - [209, 84.125] - - - [53760, 26625, 1, 512] - - [208, 83.618] - - - [54272, 26625, 1, 512] - - [208, 84.238] - - - [54272, 27137, 1, 512] - - [208, 84.125] - - - [54784, 27137, 1, 512] - - [208, 83.73] - - - [55296, 27137, 1, 512] - - [207, 83.141] - - - [55296, 27649, 1, 512] - - [209, 83.726] - - - [55808, 27649, 1, 512] - - [208, 83.416] - - - [56320, 27649, 1, 512] - - [208, 84.326] - - - [56320, 28161, 1, 512] - - [208, 84.294] - - - [56832, 28161, 1, 512] - - [208, 83.397] - - - [57344, 28161, 1, 512] - - [201, 82.943] - - - [57344, 28673, 1, 512] - - [205, 84.131] - - - [57856, 28673, 1, 512] - - [205, 83.948] - - - [58368, 28673, 1, 512] - - [209, 83.787] - - - [58368, 29185, 1, 512] - - [205, 83.812] - - - [58880, 29185, 1, 512] - - [207, 83.575] - - - [59392, 29185, 1, 512] - - [209, 83.18] - - - [59392, 29697, 1, 512] - - [207, 83.63] - - - [59904, 29697, 1, 512] - - [208, 83.544] - - - [60416, 29697, 1, 512] - - [208, 83.837] - - - [60416, 30209, 1, 512] - - [208, 83.89] - - - [60928, 30209, 1, 512] - - [208, 83.022] - - - [61440, 30209, 1, 512] - - [205, 83.407] - - - [61440, 30721, 1, 512] - - [205, 83.915] - - - [61952, 30721, 1, 512] - - [205, 83.048] - - - [62464, 30721, 1, 512] - - [205, 83.65] - - - [62464, 31233, 1, 512] - - [209, 83.209] - - - [62976, 31233, 1, 512] - - [209, 83.473] - - - [63488, 31233, 1, 512] - - [205, 83.755] - - - [63488, 31745, 1, 512] - - [208, 83.897] - - - [64000, 31745, 1, 512] - - [209, 83.118] - - - [64512, 31745, 1, 512] - - [208, 83.963] - - - [64512, 32257, 1, 512] - - [207, 83.197] - - - [65024, 32257, 1, 512] - - [208, 83.297] - - - [65536, 32257, 1, 512] - - [208, 83.074] - - - [65536, 32769, 1, 512] - - [205, 83.81] - - - [66048, 32769, 1, 512] - - [205, 83.81] - - - [66560, 32769, 1, 512] - - [209, 83.596] - - - [66560, 33281, 1, 512] - - [208, 83.066] - - - [67072, 33281, 1, 512] - - [208, 83.643] - - - [67584, 33281, 1, 512] - - [208, 83.838] - - - [67584, 33793, 1, 512] - - [208, 83.726] - - - [68096, 33793, 1, 512] - - [208, 83.923] - - - [68608, 33793, 1, 512] - - [207, 83.622] - - - [68608, 34305, 1, 512] - - [205, 83.83] - - - [69120, 34305, 1, 512] - - [205, 83.076] - - - [69632, 34305, 1, 512] - - [205, 83.53] - - - [69632, 34817, 1, 512] - - [207, 83.395] - - - [70144, 34817, 1, 512] - - [208, 83.648] - - - [70656, 34817, 1, 512] - - [205, 83.858] - - - [70656, 35329, 1, 512] - - [208, 83.426] - - - [71168, 35329, 1, 512] - - [205, 83.473] - - - [71680, 35329, 1, 512] - - [205, 83.723] - - - [71680, 35841, 1, 512] - - [208, 83.852] - - - [72192, 35841, 1, 512] - - [208, 83.373] - - - [72704, 35841, 1, 512] - - [205, 84.012] - - - [72704, 36353, 1, 512] - - [208, 83.663] - - - [73216, 36353, 1, 512] - - [205, 83.568] - - - [73728, 36353, 1, 512] - - [207, 83.135] - - - [73728, 36865, 1, 512] - - [208, 83.846] - - - [74240, 36865, 1, 512] - - [205, 83.845] - - - [74752, 36865, 1, 512] - - [208, 84.04] - - - [74752, 37377, 1, 512] - - [205, 83.473] - - - [75264, 37377, 1, 512] - - [205, 83.775] - - - [75776, 37377, 1, 512] - - [208, 84.042] - - - [75776, 37889, 1, 512] - - [205, 84.031] - - - [76288, 37889, 1, 512] - - [208, 83.783] - - - [76800, 37889, 1, 512] - - [205, 83.84] - - - [76800, 38401, 1, 512] - - [205, 83.991] - - - [77312, 38401, 1, 512] - - [205, 83.14] - - - [77824, 38401, 1, 512] - - [208, 83.691] - - - [77824, 38913, 1, 512] - - [208, 83.821] - - - [78336, 38913, 1, 512] - - [208, 83.825] - - - [78848, 38913, 1, 512] - - [208, 83.888] - - - [78848, 39425, 1, 512] - - [207, 83.773] - - - [79360, 39425, 1, 512] - - [208, 83.669] - - - [79872, 39425, 1, 512] - - [205, 83.871] - - - [79872, 39937, 1, 512] - - [208, 83.765] - - - [80384, 39937, 1, 512] - - [205, 84.012] - - - [80896, 39937, 1, 512] - - [208, 83.973] - - - [80896, 40449, 1, 512] - - [208, 83.961] - - - [81408, 40449, 1, 512] - - [208, 83.847] - - - [81920, 40449, 1, 512] - - [205, 83.161] - - - [81920, 40961, 1, 512] - - [207, 83.72] - - - [82432, 40961, 1, 512] - - [209, 83.55] - - - [82944, 40961, 1, 512] - - [208, 83.859] - - - [82944, 41473, 1, 512] - - [205, 83.366] - - - [83456, 41473, 1, 512] - - [208, 83.749] - - - [83968, 41473, 1, 512] - - [208, 83.465] - - - [83968, 41985, 1, 512] - - [208, 84.114] - - - [84480, 41985, 1, 512] - - [205, 83.794] - - - [84992, 41985, 1, 512] - - [208, 84.042] - - - [84992, 42497, 1, 512] - - [208, 83.915] - - - [85504, 42497, 1, 512] - - [208, 83.55] - - - [86016, 42497, 1, 512] - - [208, 83.836] - - - [86016, 43009, 1, 512] - - [205, 83.841] - - - [86528, 43009, 1, 512] - - [205, 84.079] - - - [87040, 43009, 1, 512] - - [208, 84.007] - - - [87040, 43521, 1, 512] - - [205, 83.645] - - - [87552, 43521, 1, 512] - - [208, 83.583] - - - [88064, 43521, 1, 512] - - [205, 83.674] - - - [88064, 44033, 1, 512] - - [205, 83.944] - - - [88576, 44033, 1, 512] - - [205, 83.592] - - - [89088, 44033, 1, 512] - - [208, 84.101] - - - [89088, 44545, 1, 512] - - [208, 83.409] - - - [89600, 44545, 1, 512] - - [208, 83.352] - - - [90112, 44545, 1, 512] - - [205, 83.856] - - - [90112, 45057, 1, 512] - - [205, 83.921] - - - [90624, 45057, 1, 512] - - [205, 83.694] - - - [91136, 45057, 1, 512] - - [208, 83.762] - - - [91136, 45569, 1, 512] - - [208, 83.946] - - - [91648, 45569, 1, 512] - - [208, 84.085] - - - [92160, 45569, 1, 512] - - [207, 83.714] - - - [92160, 46081, 1, 512] - - [205, 84.088] - - - [92672, 46081, 1, 512] - - [208, 84.071] - - - [93184, 46081, 1, 512] - - [205, 83.624] - - - [93184, 46593, 1, 512] - - [209, 83.615] - - - [93696, 46593, 1, 512] - - [208, 83.525] - - - [94208, 46593, 1, 512] - - [208, 83.962] - - - [94208, 47105, 1, 512] - - [205, 84.067] - - - [94720, 47105, 1, 512] - - [205, 83.8] - - - [95232, 47105, 1, 512] - - [205, 83.7] - - - [95232, 47617, 1, 512] - - [208, 83.435] - - - [95744, 47617, 1, 512] - - [208, 83.806] - - - [96256, 47617, 1, 512] - - [205, 83.776] - - - [96256, 48129, 1, 512] - - [205, 84.155] - - - [96768, 48129, 1, 512] - - [205, 83.723] - - - [97280, 48129, 1, 512] - - [208, 84.07] - - - [97280, 48641, 1, 512] - - [208, 83.991] - - - [97792, 48641, 1, 512] - - [205, 83.671] - - - [98304, 48641, 1, 512] - - [208, 83.85] - - - [98304, 49153, 1, 512] - - [205, 83.852] - - - [98816, 49153, 1, 512] - - [205, 83.818] - - - [99328, 49153, 1, 512] - - [205, 84.054] - - - [99328, 49665, 1, 512] - - [205, 84.065] - - - [99840, 49665, 1, 512] - - [208, 84.02] - - - [100352, 49665, 1, 512] - - [205, 84.128] - - - [100352, 50177, 1, 512] - - [205, 84.046] - - - [100864, 50177, 1, 512] - - [205, 83.688] - - - [101376, 50177, 1, 512] - - [209, 83.827] - - - [101376, 50689, 1, 512] - - [208, 83.852] - - - [101888, 50689, 1, 512] - - [205, 83.759] - - - [102400, 50689, 1, 512] - - [205, 84.054] - - - [102400, 51201, 1, 512] - - [208, 84.116] - - - [102912, 51201, 1, 512] - - [205, 84.119] - - - [103424, 51201, 1, 512] - - [208, 84.122] - - - [103424, 51713, 1, 512] - - [208, 84.122] - - - [103936, 51713, 1, 512] - - [208, 84.105] - - - [104448, 51713, 1, 512] - - [205, 84.219] - - - [104448, 52225, 1, 512] - - [205, 84.152] - - - [104960, 52225, 1, 512] - - [205, 84.061] - - - [105472, 52225, 1, 512] - - [205, 84.15] - - - [105472, 52737, 1, 512] - - [205, 83.797] - - - [105984, 52737, 1, 512] - - [208, 83.554] - - - [106496, 52737, 1, 512] - - [205, 84.033] - - - [106496, 53249, 1, 512] - - [205, 84.085] - - - [107008, 53249, 1, 512] - - [205, 84.038] - - - [107520, 53249, 1, 512] - - [205, 84.142] - - - [107520, 53761, 1, 512] - - [205, 83.908] - - - [108032, 53761, 1, 512] - - [208, 83.821] - - - [108544, 53761, 1, 512] - - [205, 83.913] - - - [108544, 54273, 1, 512] - - [205, 84.07] - - - [109056, 54273, 1, 512] - - [208, 84.044] - - - [109568, 54273, 1, 512] - - [208, 84.135] - - - [109568, 54785, 1, 512] - - [205, 83.938] - - - [110080, 54785, 1, 512] - - [208, 84.017] - - - [110592, 54785, 1, 512] - - [205, 83.952] - - - [110592, 55297, 1, 512] - - [205, 84.014] - - - [111104, 55297, 1, 512] - - [205, 84.092] - - - [111616, 55297, 1, 512] - - [205, 84.143] - - - [111616, 55809, 1, 512] - - [205, 83.89] - - - [112128, 55809, 1, 512] - - [205, 84.157] - - - [112640, 55809, 1, 512] - - [208, 84.188] - - - [112640, 56321, 1, 512] - - [205, 84.228] - - - [113152, 56321, 1, 512] - - [208, 84.168] - - - [113664, 56321, 1, 512] - - [208, 84.247] - - - [113664, 56833, 1, 512] - - [208, 84.246] - - - [114176, 56833, 1, 512] - - [208, 84.248] - - - [114688, 56833, 1, 512] - - [205, 84.072] - - - [114688, 57345, 1, 512] - - [208, 84.029] - - - [115200, 57345, 1, 512] - - [205, 84.093] - - - [115712, 57345, 1, 512] - - [208, 83.952] - - - [115712, 57857, 1, 512] - - [205, 83.908] - - - [116224, 57857, 1, 512] - - [205, 84.259] - - - [116736, 57857, 1, 512] - - [205, 84.238] - - - [116736, 58369, 1, 512] - - [208, 84.177] - - - [117248, 58369, 1, 512] - - [205, 84.072] - - - [117760, 58369, 1, 512] - - [205, 84.237] - - - [117760, 58881, 1, 512] - - [205, 84.195] - - - [118272, 58881, 1, 512] - - [208, 84.166] - - - [118784, 58881, 1, 512] - - [205, 84.217] - - - [118784, 59393, 1, 512] - - [205, 84.208] - - - [119296, 59393, 1, 512] - - [205, 84.183] - - - [119808, 59393, 1, 512] - - [205, 84.255] - - - [119808, 59905, 1, 512] - - [205, 84.265] - - - [120320, 59905, 1, 512] - - [205, 84.132] - - - [120832, 59905, 1, 512] - - [205, 84.036] - - - [120832, 60417, 1, 512] - - [208, 84.272] - - - [121344, 60417, 1, 512] - - [205, 84.278] - - - [121856, 60417, 1, 512] - - [205, 84.25] - - - [121856, 60929, 1, 512] - - [205, 84.272] - - - [122368, 60929, 1, 512] - - [208, 83.905] - - - [122880, 60929, 1, 512] - - [205, 84.104] - - - [122880, 61441, 1, 512] - - [208, 83.837] - - - [123392, 61441, 1, 512] - - [208, 83.936] - - - [123904, 61441, 1, 512] - - [208, 84.185] - - - [123904, 61953, 1, 512] - - [208, 84.231] - - - [124416, 61953, 1, 512] - - [205, 84.239] - - - [124928, 61953, 1, 512] - - [205, 84.299] - - - [124928, 62465, 1, 512] - - [208, 84.25] - - - [125440, 62465, 1, 512] - - [208, 84.25] - - - [125952, 62465, 1, 512] - - [205, 84.205] - - - [125952, 62977, 1, 512] - - [208, 84.263] - - - [126464, 62977, 1, 512] - - [205, 84.274] - - - [126976, 62977, 1, 512] - - [205, 84.178] - - - [126976, 63489, 1, 512] - - [208, 84.182] - - - [127488, 63489, 1, 512] - - [205, 84.188] - - - [128000, 63489, 1, 512] - - [208, 84.22] - - - [3584, 6657, 1, 512] - - [214, 83.642] - - - [3584, 6145, 1, 512] - - [214, 82.244] - - - [3072, 5633, 1, 512] - - [214, 83.582] - - - [3072, 5121, 1, 512] - - [214, 82.701] - - - [2560, 4609, 1, 512] - - [214, 79.643] - - - [2560, 4097, 1, 512] - - [214, 80.496] - - - [2048, 3585, 1, 512] - - [212, 70.837] - - - [2048, 3073, 1, 512] - - [223, 70.403] - - - [1536, 2561, 1, 512] - - [242, 65.594] - - - [1536, 2049, 1, 512] - - [231, 64.851] - - - [1024, 1537, 1, 512] - - [243, 56.842] - - - [1024, 1025, 1, 512] - - [239, 40.237] - - - [512, 513, 1, 512] - - [245, 21.269] - - - [512, 1, 1, 512] - - [246, 0.033] - - - [6656, 4096, 1, 512] - - [248, 87.137] - - - [6144, 4992, 1, 512] - - [248, 87.63] - - - [8192, 3328, 1, 512] - - [248, 87.21] - - - [8320, 4096, 1, 512] - - [248, 87.851] - - - [7040, 4096, 1, 512] - - [247, 91.956] - - - [7040, 4096, 1, 512] - - [211, 90.075] - - - [8448, 3840, 1, 512] - - [247, 92.567] - - - [8448, 3840, 1, 512] - - [215, 90.593] - - - [7680, 4224, 1, 512] - - [247, 92.491] - - - [7680, 4224, 1, 512] - - [215, 90.538] - - - [38144, 38144, 1, 256] - - [249, 74.893] - - - [30848, 128, 1, 256] - - [250, 41.422] - - - [32256, 32256, 1, 256] - - [251, 75.584] - - - [41984, 41984, 1, 256] - - [252, 75.197] - - - [25728, 128, 1, 256] - - [253, 49.099] - - - [17280, 17280, 1, 384] - - [254, 90.896] - - - [27392, 27392, 1, 256] - - [255, 74.926] - - - [18432, 18432, 1, 384] - - [256, 89.766] - - - [38400, 38400, 1, 384] - - [252, 90.583] - - - [9472, 9472, 1, 256] - - [254, 74.227] - - - [38656, 38656, 1, 256] - - [257, 74.685] - - - [8960, 8960, 1, 256] - - [258, 73.31] - - - [36864, 36864, 1, 384] - - [259, 87.926] - - - [20608, 128, 1, 384] - - [260, 48.729] - - - [4864, 4864, 1, 256] - - [261, 68.048] - - - [25600, 25600, 1, 256] - - [256, 75.967] - - - [19200, 19200, 1, 384] - - [256, 90.96] - - - [41728, 41728, 1, 256] - - [255, 73.787] - - - [34560, 34560, 1, 256] - - [254, 75.1] - - - [5888, 5888, 1, 256] - - [262, 70.595] - - - [42496, 42496, 1, 256] - - [255, 75.137] - - - [38400, 38400, 1, 256] - - [252, 75.384] - - - [32000, 32000, 1, 256] - - [254, 75.005] - - - [10496, 10496, 1, 256] - - [252, 74.439] - - - [24960, 24960, 1, 384] - - [252, 91.209] - - - [22272, 22272, 1, 384] - - [263, 91.067] - - - [28416, 28416, 1, 384] - - [252, 90.867] - - - [19968, 19968, 1, 384] - - [252, 90.738] - - - [35072, 35072, 1, 256] - - [252, 75.071] - - - [7168, 7168, 1, 256] - - [264, 72.544] - - - [38016, 38016, 1, 384] - - [252, 91.017] - - - [38784, 38784, 1, 384] - - [252, 90.979] - - - [27264, 27264, 1, 384] - - [256, 90.776] - - - [41088, 128, 1, 384] - - [263, 60.328] - - - [10752, 10752, 1, 256] - - [254, 75.284] - - - [44928, 128, 1, 384] - - [263, 54.276] - - - [11776, 11776, 1, 256] - - [252, 75.132] - - - [33408, 128, 1, 256] - - [262, 44.177] - - - [43520, 43520, 1, 256] - - [265, 75.193] - - - [41472, 41472, 1, 256] - - [252, 74.99] - - - [34688, 128, 1, 256] - - [266, 45.463] - - - [15744, 15744, 1, 384] - - [252, 90.68] - - - [23808, 23808, 1, 256] - - [254, 75.555] - - - [1152, 3072, 1, 384] - - [267, 63.809] - - - [34816, 34816, 1, 256] - - [256, 75.387] - - - [36096, 36096, 1, 256] - - [268, 74.203] - - - [12544, 12544, 1, 256] - - [254, 74.955] - - - [29568, 29568, 1, 384] - - [251, 90.297] - - - [30720, 30720, 1, 256] - - [256, 75.552] - - - [34048, 34048, 1, 256] - - [269, 74.525] - - - [33024, 33024, 1, 256] - - [270, 75.213] - - - [39808, 128, 1, 384] - - [271, 59.233] - - - [21504, 21504, 1, 256] - - [254, 76.075] - - - [27008, 128, 1, 384] - - [272, 59.129] - - - [24448, 128, 1, 384] - - [273, 55.216] - - - [37632, 37632, 1, 256] - - [256, 74.955] - - - [23552, 23552, 1, 256] - - [256, 76.111] - - - [13056, 13056, 1, 256] - - [256, 75.313] - - - [19584, 19584, 1, 384] - - [264, 91.095] - - - [22784, 22784, 1, 256] - - [274, 75.492] - - - [28160, 28160, 1, 256] - - [275, 75.558] - - - [20992, 20992, 1, 256] - - [264, 76.172] - - - [21760, 21760, 1, 256] - - [254, 75.734] - - - [4608, 4608, 1, 256] - - [276, 69.948] - - - [30464, 30464, 1, 256] - - [277, 74.371] - - - [22272, 22272, 1, 256] - - [278, 75.594] - - - [20480, 20480, 1, 256] - - [256, 75.812] - - - [32768, 32768, 1, 256] - - [279, 57.634] - - - [34944, 34944, 1, 384] - - [264, 90.977] - - - [34560, 34560, 1, 384] - - [264, 90.826] - - - [32128, 128, 1, 384] - - [276, 51.321] - - - [42752, 42752, 1, 256] - - [254, 74.735] - - - [18048, 128, 1, 256] - - [266, 37.6] - - - [37120, 37120, 1, 256] - - [254, 74.921] - - - [20736, 20736, 1, 384] - - [280, 91.075] - - - [26112, 26112, 1, 256] - - [275, 76.0] - - - [40704, 40704, 1, 384] - - [252, 90.834] - - - [5632, 5632, 1, 256] - - [261, 71.658] - - - [22016, 22016, 1, 256] - - [257, 76.148] - - - [13568, 13568, 1, 256] - - [252, 75.309] - - - [5760, 5760, 1, 5760] - - [280, 95.994] - - - [1920, 3072, 1, 384] - - [262, 62.558] - - - [2816, 2048, 1, 256] - - [262, 51.272] - - - [19968, 19968, 1, 256] - - [255, 76.018] - - - [14976, 14976, 1, 384] - - [271, 90.548] - - - [43008, 43008, 1, 384] - - [281, 89.03] - - - [16128, 16128, 1, 384] - - [252, 90.362] - - - [25856, 25856, 1, 256] - - [256, 75.426] - - - [6400, 6400, 1, 256] - - [282, 72.272] - - - [19712, 19712, 1, 256] - - [259, 75.0] - - - [31104, 31104, 1, 384] - - [254, 90.907] - - - [39808, 128, 1, 256] - - [267, 49.04] - - - [42624, 42624, 1, 384] - - [255, 87.986] - - - [23296, 23296, 1, 256] - - [252, 75.634] - - - [33408, 128, 1, 384] - - [262, 52.984] - - - [19328, 128, 1, 384] - - [283, 46.245] - - - [31488, 31488, 1, 256] - - [264, 74.936] - - - [6144, 6144, 1, 256] - - [264, 71.929] - - - [23168, 128, 1, 384] - - [266, 53.248] - - - [32640, 32640, 1, 384] - - [256, 90.087] - - - [19328, 128, 1, 256] - - [267, 40.04] - - - [24192, 24192, 1, 384] - - [254, 91.241] - - - [15360, 15360, 1, 256] - - [252, 76.063] - - - [37376, 37376, 1, 256] - - [268, 75.399] - - - [26880, 26880, 1, 256] - - [252, 75.31] - - - [24448, 128, 1, 256] - - [284, 46.959] - - - [38528, 128, 1, 384] - - [285, 58.149] - - - [39552, 39552, 1, 384] - - [252, 90.918] - - - [28288, 128, 1, 384] - - [261, 46.221] - - - [4608, 4608, 1, 4608] - - [271, 96.81] - - - [17664, 17664, 1, 256] - - [257, 74.999] - - - [37248, 128, 1, 384] - - [282, 56.876] - - - [42368, 128, 1, 384] - - [254, 52.092] - - - [12800, 12800, 1, 256] - - [254, 75.345] - - - [36864, 36864, 1, 256] - - [256, 74.612] - - - [12032, 12032, 1, 256] - - [254, 74.709] - - - [28800, 28800, 1, 384] - - [280, 91.023] - - - [41472, 41472, 1, 384] - - [254, 90.471] - - - [44032, 44032, 1, 256] - - [264, 75.129] - - - [42240, 42240, 1, 384] - - [254, 90.742] - - - [23040, 23040, 1, 384] - - [254, 91.024] - - - [39168, 39168, 1, 256] - - [268, 74.682] - - - [30208, 30208, 1, 256] - - [269, 75.621] - - - [15872, 15872, 1, 256] - - [252, 76.103] - - - [11520, 11520, 1, 256] - - [256, 74.674] - - - [23040, 23040, 1, 256] - - [254, 76.194] - - - [11264, 11264, 1, 256] - - [252, 75.369] - - - [30976, 30976, 1, 256] - - [255, 74.406] - - - [28928, 28928, 1, 256] - - [257, 74.911] - - - [29440, 29440, 1, 256] - - [257, 75.408] - - - [32256, 32256, 1, 384] - - [264, 90.64] - - - [37248, 128, 1, 256] - - [260, 47.064] - - - [35968, 128, 1, 256] - - [286, 46.038] - - - [39168, 39168, 1, 384] - - [252, 90.701] - - - [29568, 128, 1, 256] - - [287, 39.885] - - - [28288, 128, 1, 256] - - [287, 38.51] - - - [30720, 30720, 1, 384] - - [259, 89.477] - - - [29696, 29696, 1, 256] - - [264, 75.577] - - - [32512, 32512, 1, 256] - - [251, 75.093] - - - [6656, 6656, 1, 256] - - [261, 72.68] - - - [20608, 128, 1, 256] - - [274, 41.527] - - - [41856, 41856, 1, 384] - - [256, 90.931] - - - [7744, 7744, 1, 7744] - - [252, 97.287] - - - [39424, 39424, 1, 256] - - [252, 75.347] - - - [14208, 14208, 1, 384] - - [264, 89.943] - - - [44544, 44544, 1, 384] - - [264, 90.389] - - - [4096, 4096, 1, 256] - - [262, 66.031] - - - [17920, 17920, 1, 256] - - [256, 76.073] - - - [8192, 8192, 1, 256] - - [252, 72.824] - - - [21888, 128, 1, 384] - - [274, 50.833] - - - [24064, 24064, 1, 256] - - [251, 76.017] - - - [27648, 27648, 1, 384] - - [288, 89.139] - - - [10240, 10240, 1, 256] - - [254, 75.096] - - - [18432, 2688, 1, 384] - - [254, 85.949] - - - [3840, 3072, 1, 384] - - [282, 76.792] - - - [33792, 2688, 1, 384] - - [264, 87.542] - - - [33408, 2304, 1, 384] - - [256, 88.628] - - - [41088, 1920, 1, 384] - - [288, 86.524] - - - [5376, 1536, 1, 384] - - [261, 71.681] - - - [36480, 2688, 1, 384] - - [263, 88.977] - - - [42624, 768, 1, 384] - - [280, 82.371] - - - [29952, 1536, 1, 384] - - [252, 86.156] - - - [42240, 2688, 1, 384] - - [280, 89.835] - - - [26496, 1536, 1, 384] - - [252, 85.909] - - - [17664, 2688, 1, 384] - - [263, 86.621] - - - [16896, 2304, 1, 384] - - [264, 86.299] - - - [26880, 2688, 1, 384] - - [263, 87.772] - - - [32640, 2304, 1, 384] - - [252, 85.785] - - - [11904, 2304, 1, 384] - - [254, 82.809] - - - [33024, 1536, 1, 384] - - [264, 84.384] - - - [20352, 2688, 1, 384] - - [271, 87.058] - - - [16128, 1920, 1, 384] - - [254, 82.946] - - - [35712, 1920, 1, 384] - - [264, 87.275] - - - [23808, 2688, 1, 384] - - [280, 88.104] - - - [14592, 2688, 1, 384] - - [252, 86.006] - - - [14976, 2688, 1, 384] - - [263, 85.749] - - - [11520, 2304, 1, 384] - - [252, 84.796] - - - [4608, 768, 1, 384] - - [262, 65.306] - - - [13824, 2688, 1, 384] - - [280, 86.081] - - - [7680, 2688, 1, 384] - - [254, 81.415] - - - [27648, 2688, 1, 384] - - [264, 86.702] - - - [24192, 1920, 1, 384] - - [254, 87.591] - - - [34176, 2688, 1, 384] - - [283, 88.958] - - - [8832, 2688, 1, 384] - - [280, 81.421] - - - [31488, 768, 1, 384] - - [280, 82.286] - - - [11136, 2688, 1, 384] - - [280, 84.703] - - - [4608, 2688, 1, 384] - - [276, 79.851] - - - [34176, 768, 1, 384] - - [280, 83.561] - - - [43392, 1536, 1, 384] - - [252, 87.946] - - - [38016, 1536, 1, 384] - - [256, 86.692] - - - [15744, 1536, 1, 384] - - [252, 82.237] - - - [3072, 2304, 1, 384] - - [276, 73.318] - - - [38400, 2304, 1, 384] - - [252, 88.2] - - - [29952, 2688, 1, 384] - - [285, 88.86] - - - [13440, 1920, 1, 384] - - [254, 82.655] - - - [2304, 1536, 1, 384] - - [276, 64.822] - - - [13440, 1536, 1, 384] - - [254, 81.341] - - - [30336, 2304, 1, 384] - - [263, 88.619] - - - [35328, 768, 1, 384] - - [280, 81.961] - - - [29952, 2304, 1, 384] - - [254, 87.87] - - - [14976, 768, 1, 384] - - [285, 74.424] - - - [32640, 2688, 1, 384] - - [255, 85.433] - - - [6912, 768, 1, 384] - - [287, 69.191] - - - [32256, 1920, 1, 384] - - [254, 87.539] - - - [24576, 2688, 1, 384] - - [264, 84.613] - - - [10752, 1920, 1, 384] - - [254, 81.453] - - - [32256, 768, 1, 384] - - [280, 83.799] - - - [11520, 2688, 1, 384] - - [263, 83.45] - - - [29952, 1920, 1, 384] - - [252, 88.372] - - - [33408, 2688, 1, 384] - - [285, 89.067] - - - [21120, 2688, 1, 384] - - [289, 86.9] - - - [17664, 1536, 1, 384] - - [256, 81.463] - - - [14592, 1536, 1, 384] - - [252, 81.728] - - - [9600, 2688, 1, 384] - - [256, 82.609] - - - [21120, 768, 1, 384] - - [263, 81.735] - - - [41472, 768, 1, 384] - - [271, 85.184] - - - [38784, 2304, 1, 384] - - [264, 88.806] - - - [36864, 2304, 1, 384] - - [264, 87.729] - - - [39552, 2688, 1, 384] - - [280, 89.664] - - - [29184, 1536, 1, 384] - - [252, 86.987] - - - [5376, 2688, 1, 384] - - [276, 75.025] - - - [13824, 2304, 1, 384] - - [254, 85.541] - - - [30336, 1920, 1, 384] - - [256, 87.334] - - - [7680, 1536, 1, 384] - - [252, 76.276] - - - [36096, 2688, 1, 384] - - [255, 87.086] - - - [42240, 1536, 1, 384] - - [254, 87.779] - - - [32256, 1536, 1, 384] - - [264, 86.303] - - - [30336, 2688, 1, 384] - - [271, 88.332] - - - [39168, 1536, 1, 384] - - [252, 86.962] - - - [8064, 1536, 1, 384] - - [282, 78.707] - - - [22656, 2304, 1, 384] - - [271, 88.383] - - - [10752, 768, 1, 384] - - [290, 71.538] - - - [14976, 1920, 1, 384] - - [285, 86.02] - - - [33024, 2304, 1, 384] - - [271, 87.324] - - - [38016, 2304, 1, 384] - - [280, 88.859] - - - [31104, 768, 1, 384] - - [280, 81.555] - - - [23424, 2688, 1, 384] - - [256, 87.378] - - - [16512, 2304, 1, 384] - - [255, 80.998] - - - [39552, 768, 1, 384] - - [280, 85.633] - - - [14208, 2304, 1, 384] - - [254, 83.895] - - - [36480, 768, 1, 384] - - [285, 83.786] - - - [34560, 1536, 1, 384] - - [256, 86.554] - - - [26112, 2688, 1, 384] - - [280, 88.934] - - - [19200, 1920, 1, 384] - - [254, 85.398] - - - [8448, 1536, 1, 384] - - [261, 74.895] - - - [9600, 768, 1, 384] - - [250, 64.683] - - - [20736, 1536, 1, 384] - - [252, 85.105] - - - [10368, 1536, 1, 384] - - [254, 81.171] - - - [38784, 2688, 1, 384] - - [263, 89.453] - - - [11136, 1920, 1, 384] - - [254, 83.769] - - - [23040, 2688, 1, 384] - - [263, 87.36] - - - [7680, 768, 1, 384] - - [250, 62.177] - - - [35712, 2304, 1, 384] - - [271, 88.776] - - - [29568, 1920, 1, 384] - - [278, 86.846] - - - [25728, 768, 1, 384] - - [285, 83.582] - - - [43776, 1920, 1, 384] - - [257, 86.47] - - - [27264, 1920, 1, 384] - - [256, 85.812] - - - [8832, 2304, 1, 384] - - [254, 80.583] - - - [35328, 1920, 1, 384] - - [254, 88.337] - - - [29184, 768, 1, 384] - - [280, 81.9] - - - [32256, 2688, 1, 384] - - [263, 88.127] - - - [13056, 1536, 1, 384] - - [252, 78.981] - - - [38400, 1920, 1, 384] - - [264, 88.679] - - - [27264, 2688, 1, 384] - - [254, 87.083] - - - [30720, 2304, 1, 384] - - [264, 87.779] - - - [13824, 1920, 1, 384] - - [252, 84.703] - - - [31872, 768, 1, 384] - - [263, 83.012] - - - [7296, 768, 1, 384] - - [290, 58.889] - - - [12288, 768, 1, 384] - - [263, 69.334] - - - [21888, 2688, 1, 384] - - [291, 83.957] - - - [19968, 768, 1, 384] - - [271, 78.744] - - - [12288, 2304, 1, 384] - - [252, 84.682] - - - [34560, 768, 1, 384] - - [280, 84.329] - - - [8064, 1920, 1, 384] - - [264, 79.126] - - - [28032, 2688, 1, 384] - - [263, 89.022] - - - [29568, 2688, 1, 384] - - [280, 87.448] - - - [5376, 1920, 1, 384] - - [276, 76.125] - - - [9984, 768, 1, 384] - - [287, 67.304] - - - [43392, 2688, 1, 384] - - [271, 89.587] - - - [29568, 1536, 1, 384] - - [274, 84.24] - - - [12672, 768, 1, 384] - - [250, 72.052] - - - [40960, 27648, 1, 256] - - [249, 67.36] - - - [13056, 1792, 1, 256] - - [282, 68.961] - - - [28160, 15872, 1, 256] - - [292, 75.611] - - - [23808, 11265, 1, 256] - - [259, 71.85] - - - [38912, 26624, 1, 256] - - [264, 75.182] - - - [32768, 1792, 1, 256] - - [249, 58.928] - - - [15872, 3584, 1, 256] - - [258, 72.985] - - - [39680, 27393, 1, 256] - - [281, 72.507] - - - [26112, 1536, 1, 256] - - [276, 70.393] - - - [16896, 4353, 1, 256] - - [257, 71.01] - - - [3840, 3072, 1, 256] - - [261, 63.197] - - - [6656, 1536, 1, 256] - - [276, 61.995] - - - [43776, 3072, 1, 256] - - [255, 71.088] - - - [11264, 7937, 1, 256] - - [259, 72.469] - - - [6912, 3841, 1, 256] - - [261, 66.587] - - - [40448, 9216, 1, 256] - - [269, 74.661] - - - [23296, 3072, 1, 256] - - [252, 72.423] - - - [20736, 8448, 1, 256] - - [256, 74.596] - - - [28160, 1792, 1, 256] - - [258, 71.494] - - - [42752, 1792, 1, 256] - - [256, 72.151] - - - [13056, 3072, 1, 256] - - [254, 71.139] - - - [19456, 3072, 1, 256] - - [254, 72.97] - - - [30976, 1792, 1, 256] - - [284, 71.133] - - - [35328, 23041, 1, 256] - - [270, 73.686] - - - [25856, 13568, 1, 256] - - [264, 75.047] - - - [2048, 1024, 1, 256] - - [286, 34.871] - - - [11520, 1536, 1, 256] - - [282, 67.434] - - - [15360, 3072, 1, 256] - - [254, 72.025] - - - [25344, 13056, 1, 256] - - [268, 74.792] - - - [39424, 768, 1, 256] - - [290, 68.846] - - - [28672, 3072, 1, 256] - - [252, 73.07] - - - [39680, 768, 1, 256] - - [271, 68.165] - - - [17408, 4865, 1, 256] - - [252, 71.47] - - - [29184, 768, 1, 256] - - [266, 67.284] - - - [33536, 9216, 1, 256] - - [268, 74.284] - - - [32768, 20480, 1, 256] - - [293, 58.823] - - - [40960, 1792, 1, 256] - - [265, 66.694] - - - [20992, 1792, 1, 256] - - [254, 71.07] - - - [19200, 6912, 1, 256] - - [254, 74.327] - - - [42496, 1536, 1, 256] - - [264, 71.485] - - - [20480, 8192, 1, 256] - - [252, 74.921] - - - [26880, 14337, 1, 256] - - [281, 72.433] - - - [15360, 2817, 1, 256] - - [254, 67.942] - - - [7936, 1536, 1, 256] - - [261, 63.735] - - - [8448, 3072, 1, 256] - - [276, 68.909] - - - [1792, 2048, 1, 256] - - [262, 40.334] - - - [20480, 3072, 1, 256] - - [256, 72.732] - - - [37376, 25088, 1, 256] - - [264, 75.229] - - - [21504, 768, 1, 256] - - [280, 64.581] - - - [40960, 1025, 1, 256] - - [256, 57.587] - - - [9216, 768, 1, 256] - - [287, 59.444] - - - [32512, 1792, 1, 256] - - [261, 71.363] - - - [25600, 13057, 1, 256] - - [281, 73.924] - - - [25088, 12800, 1, 256] - - [274, 75.683] - - - [38400, 26113, 1, 256] - - [255, 73.709] - - - [20736, 1792, 1, 256] - - [282, 70.596] - - - [5376, 2305, 1, 256] - - [276, 58.991] - - - [43520, 3584, 1, 256] - - [264, 74.091] - - - [12800, 1536, 1, 256] - - [294, 67.084] - - - [39680, 9216, 1, 256] - - [254, 74.129] - - - [1280, 3072, 1, 256] - - [267, 43.215] - - - [22016, 768, 1, 256] - - [285, 64.586] - - - [26880, 9216, 1, 256] - - [254, 74.278] - - - [7680, 768, 1, 256] - - [267, 51.917] - - - [26880, 14593, 1, 256] - - [281, 72.979] - - - [39424, 26881, 1, 256] - - [281, 73.642] - - - [10752, 1792, 1, 256] - - [282, 67.663] - - - [34560, 3072, 1, 256] - - [252, 72.869] - - - [29696, 17153, 1, 256] - - [256, 74.326] - - - [13568, 1536, 1, 256] - - [261, 67.459] - - - [36608, 1536, 1, 256] - - [282, 70.529] - - - [7680, 1792, 1, 256] - - [262, 65.222] - - - [37376, 1536, 1, 256] - - [256, 70.914] - - - [43520, 1536, 1, 256] - - [252, 71.431] - - - [18176, 3072, 1, 256] - - [254, 72.178] - - - [17408, 9216, 1, 256] - - [254, 74.837] - - - [20992, 8705, 1, 256] - - [259, 72.643] - - - [38144, 768, 1, 256] - - [250, 67.922] - - - [4096, 3072, 1, 256] - - [278, 64.649] - - - [10240, 6913, 1, 256] - - [259, 71.858] - - - [38656, 26113, 1, 256] - - [268, 72.926] - - - [38400, 1536, 1, 256] - - [252, 70.995] - - - [13056, 769, 1, 256] - - [271, 54.241] - - - [17664, 9216, 1, 256] - - [251, 73.709] - - - [17920, 1792, 1, 256] - - [286, 70.638] - - - [8960, 768, 1, 256] - - [287, 58.995] - - - [12544, 257, 1, 256] - - [263, 35.771] - - - [11008, 1792, 1, 256] - - [295, 65.563] - - - [34560, 22272, 1, 256] - - [252, 74.996] - - - [14336, 3072, 1, 256] - - [282, 71.953] - - - [13312, 9216, 1, 256] - - [264, 74.894] - - - [24064, 1792, 1, 256] - - [258, 71.538] - - - [27904, 15616, 1, 256] - - [254, 75.078] - - - [13056, 513, 1, 256] - - [263, 48.579] - - - [40704, 3072, 1, 256] - - [254, 73.142] - - - [24576, 12288, 1, 256] - - [281, 69.898] - - - [17152, 4864, 1, 256] - - [262, 73.546] - - - [32256, 768, 1, 256] - - [266, 68.635] - - - [5632, 768, 1, 256] - - [266, 46.604] - - - [19712, 768, 1, 256] - - [296, 61.993] - - - [2816, 3072, 1, 256] - - [282, 61.168] - - - [4608, 1537, 1, 256] - - [262, 51.08] - - - [33792, 21249, 1, 256] - - [259, 74.025] - - - [38656, 9216, 1, 256] - - [268, 74.324] - - - [19456, 7169, 1, 256] - - [259, 71.093] - - - [29440, 17152, 1, 256] - - [257, 75.486] - - - [41472, 1792, 1, 256] - - [256, 72.408] - - - [35072, 9216, 1, 256] - - [252, 74.214] - - - [35072, 22785, 1, 256] - - [259, 73.107] - - - [42752, 2817, 1, 256] - - [281, 69.038] - - - [24832, 12289, 1, 256] - - [270, 72.414] - - - [6400, 1536, 1, 256] - - [261, 60.956] - - - [30976, 3072, 1, 256] - - [258, 71.129] - - - [19968, 1792, 1, 256] - - [282, 71.033] - - - [43776, 3840, 1, 256] - - [255, 72.616] - - - [8192, 1792, 1, 256] - - [262, 63.092] - - - [8704, 3072, 1, 256] - - [261, 70.142] - - - [22272, 9729, 1, 256] - - [297, 72.315] - - - [3072, 2048, 1, 256] - - [262, 54.959] - - - [9984, 1536, 1, 256] - - [282, 64.234] - - - [38400, 25857, 1, 256] - - [255, 73.786] - - - [34304, 21761, 1, 256] - - [270, 73.859] - - - [31744, 9216, 1, 256] - - [254, 74.725] - - - [34048, 9216, 1, 256] - - [270, 73.879] - - - [36096, 768, 1, 256] - - [266, 66.163] - - - [35584, 9216, 1, 256] - - [255, 74.453] - - - [34816, 1792, 1, 256] - - [256, 72.664] - - - [40448, 3072, 1, 256] - - [278, 73.246] - - - [20480, 1536, 1, 256] - - [254, 69.236] - - - [20224, 768, 1, 256] - - [287, 64.775] - - - [36608, 24320, 1, 256] - - [265, 74.884] - - - [41216, 27648, 1, 256] - - [255, 74.634] - - - [7424, 3072, 1, 256] - - [262, 69.039] - - - [41984, 1793, 1, 256] - - [256, 67.812] - - - [13824, 1792, 1, 256] - - [262, 70.014] - - - [33280, 3072, 1, 256] - - [278, 73.418] - - - [43520, 768, 1, 256] - - [263, 69.457] - - - [29696, 3072, 1, 256] - - [264, 73.642] - - - [40960, 9216, 1, 256] - - [265, 67.099] - - - [42240, 2305, 1, 256] - - [259, 68.068] - - - [27904, 9216, 1, 256] - - [268, 74.405] - - - [11776, 8705, 1, 256] - - [278, 72.055] - - - [25856, 9216, 1, 256] - - [252, 74.358] - - - [41472, 9216, 1, 256] - - [269, 74.605] - - - [44288, 4352, 1, 256] - - [265, 73.828] - - - [9984, 6657, 1, 256] - - [256, 71.101] - - - [36352, 24065, 1, 256] - - [268, 73.734] - - - [29184, 16641, 1, 256] - - [259, 73.633] - - - [42240, 27648, 1, 256] - - [255, 74.587] - - - [44288, 3072, 1, 256] - - [281, 72.705] - - - [28672, 16129, 1, 256] - - [281, 73.603] - - - [25088, 12801, 1, 256] - - [270, 73.417] - - - [31744, 19457, 1, 256] - - [252, 72.863] - - - [42496, 1792, 1, 256] - - [252, 72.873] - - - [8960, 1792, 1, 256] - - [261, 66.553] - - - [38144, 25601, 1, 256] - - [255, 72.464] - - - [31744, 19456, 1, 256] - - [254, 75.314] - - - [37888, 9216, 1, 256] - - [254, 74.723] - - - [42240, 2049, 1, 256] - - [285, 66.103] - - - [37120, 9216, 1, 256] - - [268, 74.336] - - - [38400, 1792, 1, 256] - - [264, 72.333] - - - [14336, 2049, 1, 256] - - [256, 63.371] - - - [30976, 18433, 1, 256] - - [255, 72.375] - - - [34560, 1792, 1, 256] - - [282, 71.797] - - - [21760, 9217, 1, 256] - - [259, 71.598] - - - [7936, 3072, 1, 256] - - [276, 68.835] - - - [9216, 6145, 1, 256] - - [254, 69.56] - - - [39936, 27648, 1, 256] - - [252, 74.967] - - - [5376, 2049, 1, 256] - - [254, 56.9] - - - [13312, 769, 1, 256] - - [285, 55.124] - - - [39424, 27136, 1, 256] - - [265, 75.313] - - - [34048, 21505, 1, 256] - - [255, 72.296] - - - [23296, 768, 1, 256] - - [290, 67.076] - - - [18432, 9216, 1, 256] - - [264, 75.027] - - - [30464, 17921, 1, 256] - - [268, 72.452] - - - [9472, 3072, 1, 256] - - [256, 69.062] - - - [8960, 3072, 1, 256] - - [261, 68.93] - - - [11008, 3072, 1, 256] - - [295, 69.116] - - - [23552, 9216, 1, 256] - - [252, 74.931] - - - [27136, 9216, 1, 256] - - [270, 74.972] - - - [31488, 18945, 1, 256] - - [259, 73.064] - - - [14592, 3072, 1, 256] - - [286, 70.802] - - - [4096, 769, 1, 256] - - [272, 35.775] - - - [33024, 1792, 1, 256] - - [286, 70.613] - - - [44544, 4353, 1, 256] - - [259, 71.136] - - - [5632, 2305, 1, 256] - - [261, 60.921] - - - [10240, 7169, 1, 256] - - [264, 70.424] - - - [10752, 7681, 1, 256] - - [254, 71.73] - - - [14336, 9216, 1, 256] - - [264, 74.678] - - - [35840, 23553, 1, 256] - - [259, 73.019] - - - [19712, 7169, 1, 256] - - [281, 69.862] - - - [27136, 768, 1, 256] - - [266, 67.459] - - - [15616, 1536, 1, 256] - - [256, 67.46] - - - [42240, 2304, 1, 256] - - [264, 72.628] - - - [27136, 1536, 1, 256] - - [258, 70.093] - - - [32512, 20224, 1, 256] - - [274, 75.143] - - - [28160, 15873, 1, 256] - - [259, 73.603] - - - [14592, 9216, 1, 256] - - [298, 72.836] - - - [22272, 3072, 1, 256] - - [284, 72.388] - - - [20992, 8704, 1, 256] - - [264, 75.323] - - - [7936, 4865, 1, 256] - - [284, 69.997] - - - [37888, 25345, 1, 256] - - [281, 73.968] - - - [28416, 15873, 1, 256] - - [297, 73.027] - - - [39424, 1536, 1, 256] - - [258, 70.892] - - - [19456, 9216, 1, 256] - - [252, 74.881] - - - [40448, 1536, 1, 256] - - [257, 71.11] - - - [19456, 6913, 1, 256] - - [254, 72.86] - - - [41984, 9216, 1, 256] - - [256, 74.61] - - - [20992, 8449, 1, 256] - - [259, 72.683] - - - [10496, 7169, 1, 256] - - [281, 69.631] - - - [13568, 1025, 1, 256] - - [267, 59.023] - - - [27392, 15104, 1, 256] - - [270, 74.726] - - - [9472, 6401, 1, 256] - - [252, 70.416] - - - [41728, 768, 1, 256] - - [299, 67.967] - - - [25344, 3072, 1, 256] - - [284, 72.1] - - - [27392, 14849, 1, 256] - - [268, 72.861] - - - [28160, 3072, 1, 256] - - [254, 73.026] - - - [11776, 8449, 1, 256] - - [281, 72.092] - - - [24576, 12289, 1, 256] - - [259, 67.331] - - - [13824, 9216, 1, 256] - - [288, 74.326] - - - [27904, 15361, 1, 256] - - [268, 72.692] - - - [14848, 3072, 1, 256] - - [264, 71.769] - - - [34816, 1536, 1, 256] - - [254, 71.375] - - - [22528, 9985, 1, 256] - - [256, 73.686] - - - [4096, 1792, 1, 256] - - [261, 53.676] - - - [37120, 24833, 1, 256] - - [268, 73.124] - - - [2560, 1536, 1, 256] - - [282, 43.756] - - - [44544, 27648, 1, 256] - - [255, 74.734] - - - [24064, 1536, 1, 256] - - [276, 69.843] - - - [44032, 3841, 1, 256] - - [252, 70.954] - - - [26624, 14337, 1, 256] - - [259, 72.796] - - - [42240, 1536, 1, 256] - - [254, 70.765] - - - [6144, 1792, 1, 256] - - [252, 60.128] - - - [35584, 23297, 1, 256] - - [259, 73.151] - - - [23040, 1536, 1, 256] - - [294, 69.287] - - - [18432, 6145, 1, 256] - - [264, 70.988] - - - [39168, 3072, 1, 256] - - [257, 72.818] - - - [41728, 27648, 1, 256] - - [268, 73.797] - - - [24320, 12032, 1, 256] - - [251, 75.433] - - - [34816, 22528, 1, 256] - - [252, 75.054] - - - [25856, 13569, 1, 256] - - [255, 72.951] - - - [31232, 18689, 1, 256] - - [255, 73.883] - - - [34304, 9216, 1, 256] - - [269, 75.021] - - - [8704, 5633, 1, 256] - - [264, 70.581] - - - [41984, 3072, 1, 256] - - [256, 73.928] - - - [9728, 6401, 1, 256] - - [256, 70.891] - - - [38400, 9216, 1, 256] - - [255, 74.802] - - - [1536, 3072, 1, 256] - - [262, 50.841] - - - [37632, 1792, 1, 256] - - [294, 71.968] - - - [18944, 6656, 1, 256] - - [257, 74.667] - - - [23296, 11008, 1, 256] - - [254, 74.913] - - - [38656, 1792, 1, 256] - - [258, 71.937] - - - [29952, 17665, 1, 256] - - [300, 73.321] - - - [24576, 1536, 1, 256] - - [265, 66.688] - - - [42752, 768, 1, 256] - - [263, 68.776] - - - [35840, 768, 1, 256] - - [287, 68.381] - - - [33536, 1792, 1, 256] - - [261, 71.782] - - - [38656, 26368, 1, 256] - - [255, 74.56] - - - [44288, 4097, 1, 256] - - [281, 69.82] - - - [36096, 3072, 1, 256] - - [251, 71.493] - - - [31232, 18944, 1, 256] - - [257, 75.689] - - - [26112, 9216, 1, 256] - - [268, 74.995] - - - [12544, 3072, 1, 256] - - [262, 71.024] - - - [40704, 1792, 1, 256] - - [284, 72.126] - - - [33792, 9216, 1, 256] - - [254, 74.743] - - - [38912, 1792, 1, 256] - - [256, 72.82] - - - [28672, 16385, 1, 256] - - [281, 72.361] - - - [33280, 20993, 1, 256] - - [270, 73.865] - - - [43520, 1792, 1, 256] - - [254, 72.876] - - - [41472, 1536, 1, 256] - - [264, 71.132] - - - [23296, 11009, 1, 256] - - [281, 72.525] - - - [29184, 9216, 1, 256] - - [301, 74.75] - - - [37632, 25344, 1, 256] - - [256, 74.856] - - - [16640, 9216, 1, 256] - - [268, 74.324] - - - [14080, 1792, 1, 256] - - [258, 68.979] - - - [34048, 21760, 1, 256] - - [291, 74.55] - - - [40192, 1536, 1, 256] - - [252, 71.09] - - - [35328, 9216, 1, 256] - - [270, 74.937] - - - [14592, 2049, 1, 256] - - [266, 63.564] - - - [23808, 3072, 1, 256] - - [252, 72.789] - - - [42496, 2560, 1, 256] - - [252, 73.708] - - - [39680, 27392, 1, 256] - - [264, 74.582] - - - [25600, 13313, 1, 256] - - [259, 72.762] - - - [20480, 8193, 1, 256] - - [281, 71.705] - - - [15104, 2561, 1, 256] - - [267, 67.093] - - - [1280, 2048, 1, 256] - - [284, 42.367] - - - [18688, 3072, 1, 256] - - [252, 72.245] - - - [18944, 6401, 1, 256] - - [300, 71.873] - - - [29952, 1536, 1, 256] - - [264, 70.233] - - - [36608, 1792, 1, 256] - - [261, 71.769] - - - [26112, 13824, 1, 256] - - [292, 75.824] - - - [40704, 9216, 1, 256] - - [264, 74.131] - - - [21248, 3072, 1, 256] - - [276, 72.225] - - - [28672, 1792, 1, 256] - - [254, 71.832] - - - [29952, 9216, 1, 256] - - [274, 74.232] - - - [34304, 22017, 1, 256] - - [270, 73.821] - - - [19968, 9216, 1, 256] - - [255, 74.953] - - - [19200, 1792, 1, 256] - - [286, 69.134] - - - [23040, 10497, 1, 256] - - [297, 73.426] - - - [32256, 1792, 1, 256] - - [284, 72.227] - - - [12288, 8961, 1, 256] - - [264, 72.875] - - - [19968, 7425, 1, 256] - - [281, 72.526] - - - [33792, 3072, 1, 256] - - [256, 73.595] - - - [8704, 5377, 1, 256] - - [254, 70.194] - - - [25856, 1536, 1, 256] - - [262, 69.702] - - - [32256, 3072, 1, 256] - - [252, 73.426] - - - [22784, 10241, 1, 256] - - [268, 71.799] - - - [36096, 9216, 1, 256] - - [255, 73.759] - - - [38400, 26112, 1, 256] - - [254, 75.376] - - - [41472, 27648, 1, 256] - - [270, 74.643] - - - [15360, 3073, 1, 256] - - [254, 67.924] - - - [16128, 1792, 1, 256] - - [261, 68.836] - - - [23040, 10752, 1, 256] - - [274, 75.65] - - - [11008, 7681, 1, 256] - - [302, 70.59] - - - [28416, 16128, 1, 256] - - [278, 74.891] - - - [41728, 1792, 1, 256] - - [291, 71.002] - - - [18944, 1792, 1, 256] - - [286, 70.622] - - - [31488, 19201, 1, 256] - - [259, 73.037] - - - [42752, 27648, 1, 256] - - [256, 74.278] - - - [25088, 12545, 1, 256] - - [251, 73.619] - - - [33024, 20737, 1, 256] - - [300, 73.586] - - - [31232, 1536, 1, 256] - - [252, 70.863] - - - [44544, 4609, 1, 256] - - [281, 71.087] - - - [30208, 17921, 1, 256] - - [269, 73.75] - - - [22016, 9729, 1, 256] - - [251, 73.212] - - - [29184, 1792, 1, 256] - - [286, 71.58] - - - [22528, 1536, 1, 256] - - [261, 69.808] - - - [41216, 1025, 1, 256] - - [296, 63.189] - - - [5888, 1792, 1, 256] - - [261, 64.218] - - - [22272, 9985, 1, 256] - - [300, 72.608] - - - [22016, 3072, 1, 256] - - [274, 72.804] - - - [39680, 1792, 1, 256] - - [256, 72.047] - - - [34560, 9216, 1, 256] - - [264, 74.255] - - - [10752, 7425, 1, 256] - - [278, 71.821] - - - [16128, 3840, 1, 256] - - [282, 72.828] - - - [21248, 1792, 1, 256] - - [261, 70.279] - - - [19968, 7680, 1, 256] - - [256, 75.125] - - - [38400, 3072, 1, 256] - - [264, 73.324] - - - [20224, 9216, 1, 256] - - [264, 74.585] - - - [28928, 9216, 1, 256] - - [252, 74.117] - - - [31488, 19200, 1, 256] - - [254, 75.016] - - - [21248, 8705, 1, 256] - - [300, 72.196] - - - [34560, 22273, 1, 256] - - [259, 73.128] - - - [40448, 27648, 1, 256] - - [269, 74.661] - - - [34816, 22529, 1, 256] - - [259, 72.704] - - - [27136, 14593, 1, 256] - - [297, 73.833] - - - [43008, 3072, 1, 256] - - [254, 73.831] - - - [16384, 4097, 1, 256] - - [254, 59.138] - - - [31488, 1792, 1, 256] - - [261, 71.727] - - - [22272, 9984, 1, 256] - - [284, 75.013] - - - [26880, 1792, 1, 256] - - [254, 71.327] - - - [2816, 1792, 1, 256] - - [262, 54.473] - - - [43008, 27648, 1, 256] - - [252, 74.892] - - - [24320, 1792, 1, 256] - - [258, 70.912] - - - [12800, 1792, 1, 256] - - [258, 68.573] - - - [15872, 1792, 1, 256] - - [262, 70.342] - - - [16896, 4608, 1, 256] - - [252, 73.401] - - - [21248, 8960, 1, 256] - - [254, 74.939] - - - [24832, 12544, 1, 256] - - [258, 75.359] - - - [31744, 19201, 1, 256] - - [254, 74.065] - - - [40192, 3072, 1, 256] - - [264, 73.329] - - - [32256, 9216, 1, 256] - - [269, 74.881] - - - [6656, 1792, 1, 256] - - [286, 63.224] - - - [20736, 3072, 1, 256] - - [254, 72.45] - - - [41728, 9216, 1, 256] - - [255, 73.108] - - - [37632, 768, 1, 256] - - [250, 68.035] - - - [19968, 3072, 1, 256] - - [256, 72.482] - - - [29184, 16897, 1, 256] - - [259, 73.625] - - - [40704, 769, 1, 256] - - [260, 60.707] - - - [34304, 1792, 1, 256] - - [286, 72.252] - - - [20992, 1536, 1, 256] - - [254, 69.992] - - - [43264, 3072, 1, 256] - - [252, 73.263] - - - [11264, 768, 1, 256] - - [280, 60.064] - - - [30208, 17665, 1, 256] - - [300, 73.922] - - - [29952, 17409, 1, 256] - - [300, 72.467] - - - [41216, 9216, 1, 256] - - [268, 74.596] - - - [36352, 768, 1, 256] - - [290, 68.459] - - - [7168, 3072, 1, 256] - - [262, 67.869] - - - [36608, 768, 1, 256] - - [290, 68.316] - - - [37376, 24833, 1, 256] - - [255, 73.814] - - - [26880, 14592, 1, 256] - - [256, 75.253] - - - [23808, 768, 1, 256] - - [290, 64.822] - - - [5120, 1793, 1, 256] - - [276, 55.962] - - - [43008, 2817, 1, 256] - - [252, 69.9] - - - [33536, 3072, 1, 256] - - [278, 72.881] - - - [2048, 3072, 1, 256] - - [276, 55.21] - - - [5632, 2561, 1, 256] - - [254, 61.714] - - - [20224, 7936, 1, 256] - - [254, 74.62] - - - [28416, 9216, 1, 256] - - [259, 73.936] - - - [23552, 11265, 1, 256] - - [252, 72.46] - - - [32000, 19457, 1, 256] - - [259, 72.725] - - - [37120, 1792, 1, 256] - - [254, 71.806] - - - [44544, 4608, 1, 256] - - [255, 73.899] - - - [18944, 9216, 1, 256] - - [269, 74.773] - - - [25856, 3072, 1, 256] - - [276, 72.641] - - - [5120, 2049, 1, 256] - - [303, 55.943] - - - [37632, 25089, 1, 256] - - [259, 73.058] - - - [16384, 9216, 1, 256] - - [249, 63.2] - - - [20480, 7937, 1, 256] - - [259, 72.788] - - - [12800, 768, 1, 256] - - [267, 60.244] - - - [32000, 9216, 1, 256] - - [254, 74.163] - - - [5376, 3072, 1, 256] - - [252, 64.581] - - - [23808, 11521, 1, 256] - - [259, 72.524] - - - [18176, 9216, 1, 256] - - [264, 74.483] - - - [27648, 3072, 1, 256] - - [252, 73.527] - - - [22016, 9216, 1, 256] - - [268, 75.131] - - - [6144, 1536, 1, 256] - - [261, 58.455] - - - [19968, 768, 1, 256] - - [267, 64.87] - - - [27904, 3072, 1, 256] - - [286, 72.351] - - - [26624, 14081, 1, 256] - - [264, 73.889] - - - [14080, 3072, 1, 256] - - [286, 70.439] - - - [21504, 9216, 1, 256] - - [254, 74.988] - - - [34304, 3072, 1, 256] - - [274, 73.279] - - - [17152, 4865, 1, 256] - - [252, 70.61] - - - [41216, 1792, 1, 256] - - [252, 71.942] - - - [30720, 18433, 1, 256] - - [259, 72.945] - - - [5632, 1792, 1, 256] - - [261, 61.719] - - - [31232, 9216, 1, 256] - - [255, 74.854] - - - [18688, 9216, 1, 256] - - [264, 74.553] - - - [10240, 1792, 1, 256] - - [262, 66.547] - - - [25856, 13313, 1, 256] - - [255, 72.38] - - - [5888, 768, 1, 256] - - [282, 48.723] - - - [17152, 1792, 1, 256] - - [261, 69.112] - - - [8960, 5889, 1, 256] - - [284, 70.343] - - - [30464, 3072, 1, 256] - - [284, 71.363] - - - [25344, 13057, 1, 256] - - [268, 72.612] - - - [25600, 3072, 1, 256] - - [256, 73.346] - - - [12800, 257, 1, 256] - - [287, 36.642] - - - [20736, 8193, 1, 256] - - [281, 71.591] - - - [40192, 27648, 1, 256] - - [252, 74.477] - - - [36352, 23809, 1, 256] - - [255, 73.868] - - - [11776, 1792, 1, 256] - - [284, 68.461] - - - [14592, 2304, 1, 256] - - [286, 69.777] - - - [10240, 1536, 1, 256] - - [276, 65.762] - - - [15104, 3072, 1, 256] - - [261, 71.582] - - - [36096, 23553, 1, 256] - - [255, 72.256] - - - [18432, 5889, 1, 256] - - [252, 72.062] - - - [22528, 10241, 1, 256] - - [281, 72.313] - - - [39680, 3072, 1, 256] - - [264, 72.868] - - - [30720, 1792, 1, 256] - - [282, 72.468] - - - [22784, 1536, 1, 256] - - [261, 69.341] - - - [7680, 4609, 1, 256] - - [252, 68.021] - - - [1792, 3072, 1, 256] - - [282, 49.169] - - - [29440, 17153, 1, 256] - - [300, 73.313] - - - [30720, 3072, 1, 256] - - [256, 73.577] - - - [40192, 9216, 1, 256] - - [256, 74.318] - - - [37888, 25601, 1, 256] - - [281, 72.878] - - - [30208, 3072, 1, 256] - - [256, 72.997] - - - [35072, 22784, 1, 256] - - [252, 74.903] - - - [22784, 1792, 1, 256] - - [261, 70.849] - - - [33024, 20481, 1, 256] - - [269, 72.827] - - - [9216, 3072, 1, 256] - - [261, 70.425] - - - [25088, 768, 1, 256] - - [250, 66.789] - - - [42240, 26369, 1, 128] - - [249, 45.705] - - - [39168, 512, 1, 128] - - [276, 41.06] - - - [41728, 8192, 1, 128] - - [291, 45.865] - - - [35072, 512, 1, 128] - - [285, 41.862] - - - [33280, 2048, 1, 128] - - [266, 44.567] - - - [35456, 1024, 1, 128] - - [263, 42.971] - - - [36992, 20993, 1, 128] - - [249, 45.407] - - - [31488, 15489, 1, 128] - - [249, 45.828] - - - [36864, 128, 1, 128] - - [280, 30.903] - - - [35200, 2048, 1, 128] - - [280, 44.094] - - - [34688, 2048, 1, 128] - - [267, 44.433] - - - [34304, 4096, 1, 128] - - [294, 46.126] - - - [33408, 128, 1, 128] - - [267, 28.689] - - - [43264, 27393, 1, 128] - - [265, 45.698] - - - [30336, 4096, 1, 128] - - [294, 45.255] - - - [34816, 512, 1, 128] - - [250, 42.118] - - - [39808, 23937, 1, 128] - - [291, 43.271] - - - [40448, 512, 1, 128] - - [287, 41.853] - - - [44544, 28545, 1, 128] - - [277, 45.737] - - - [34688, 18689, 1, 128] - - [256, 45.523] - - - [38912, 512, 1, 128] - - [261, 41.394] - - - [34048, 1024, 1, 128] - - [261, 43.28] - - - [40320, 24321, 1, 128] - - [249, 45.408] - - - [36736, 1024, 1, 128] - - [262, 42.751] - - - [32000, 128, 1, 128] - - [267, 27.968] - - - [38144, 1024, 1, 128] - - [256, 43.025] - - - [38144, 8192, 1, 128] - - [264, 46.834] - - - [43136, 1024, 1, 128] - - [261, 42.913] - - - [43264, 2048, 1, 128] - - [285, 44.133] - - - [37120, 128, 1, 128] - - [258, 31.012] - - - [29952, 13953, 1, 128] - - [264, 45.983] - - - [37504, 21505, 1, 128] - - [278, 44.546] - - - [41856, 2048, 1, 128] - - [287, 44.478] - - - [41856, 8192, 1, 128] - - [264, 46.566] - - - [33664, 8192, 1, 128] - - [274, 46.232] - - - [33280, 512, 1, 128] - - [258, 40.592] - - - [31744, 128, 1, 128] - - [261, 27.614] - - - [29952, 8192, 1, 128] - - [278, 46.599] - - - [34176, 8192, 1, 128] - - [274, 46.412] - - - [41984, 2048, 1, 128] - - [288, 44.722] - - - [33920, 4096, 1, 128] - - [254, 45.581] - - - [39808, 8192, 1, 128] - - [249, 44.228] - - - [30592, 2048, 1, 128] - - [287, 43.833] - - - [30336, 2048, 1, 128] - - [267, 43.833] - - - [42368, 1024, 1, 128] - - [267, 43.084] - - - [38784, 22913, 1, 128] - - [249, 45.544] - - - [40576, 512, 1, 128] - - [254, 41.519] - - - [41600, 2048, 1, 128] - - [249, 43.878] - - - [41856, 4096, 1, 128] - - [254, 45.762] - - - [30848, 14849, 1, 128] - - [264, 45.765] - - - [43776, 2048, 1, 128] - - [287, 43.542] - - - [37376, 128, 1, 128] - - [290, 31.226] - - - [36224, 20353, 1, 128] - - [265, 45.672] - - - [34176, 1024, 1, 128] - - [261, 42.871] - - - [42624, 8192, 1, 128] - - [254, 41.712] - - - [42624, 26753, 1, 128] - - [304, 40.481] - - - [40064, 4096, 1, 128] - - [249, 45.152] - - - [42112, 26241, 1, 128] - - [249, 45.551] - - - [40960, 1024, 1, 128] - - [254, 40.673] - - - [32384, 8192, 1, 128] - - [256, 46.529] - - - [33024, 8192, 1, 128] - - [256, 47.007] - - - [33664, 17665, 1, 128] - - [265, 45.702] - - - [40704, 128, 1, 128] - - [260, 33.254] - - - [37120, 512, 1, 128] - - [287, 40.735] - - - [39680, 8192, 1, 128] - - [264, 46.567] - - - [33024, 17153, 1, 128] - - [249, 46.343] - - - [35328, 1024, 1, 128] - - [262, 43.01] - - - [40320, 8192, 1, 128] - - [252, 46.483] - - - [36608, 20737, 1, 128] - - [249, 45.83] - - - [43520, 2048, 1, 128] - - [260, 44.598] - - - [35328, 19329, 1, 128] - - [277, 46.24] - - - [29696, 13697, 1, 128] - - [254, 46.562] - - - [35584, 19713, 1, 128] - - [249, 45.7] - - - [42368, 2048, 1, 128] - - [285, 44.078] - - - [30336, 8192, 1, 128] - - [262, 46.227] - - - [37120, 1024, 1, 128] - - [267, 42.986] - - - [33152, 4096, 1, 128] - - [261, 45.818] - - - [37248, 21249, 1, 128] - - [265, 45.316] - - - [39680, 4096, 1, 128] - - [252, 45.776] - - - [35712, 8192, 1, 128] - - [252, 46.366] - - - [35328, 512, 1, 128] - - [261, 40.342] - - - [30208, 128, 1, 128] - - [287, 26.654] - - - [44288, 2048, 1, 128] - - [263, 44.467] - - - [30464, 14465, 1, 128] - - [301, 43.511] - - - [33792, 17793, 1, 128] - - [256, 46.448] - - - [35968, 2048, 1, 128] - - [287, 44.208] - - - [32512, 4096, 1, 128] - - [254, 45.842] - - - [39424, 128, 1, 128] - - [284, 32.351] - - - [35968, 20097, 1, 128] - - [252, 45.717] - - - [34944, 18945, 1, 128] - - [254, 45.624] - - - [38656, 22657, 1, 128] - - [277, 45.798] - - - [31872, 16001, 1, 128] - - [265, 45.704] - - - [38016, 512, 1, 128] - - [267, 40.69] - - - [35200, 1024, 1, 128] - - [267, 42.748] - - - [44672, 28801, 1, 128] - - [265, 45.545] - - - [33408, 8192, 1, 128] - - [254, 46.532] - - - [39680, 23681, 1, 128] - - [265, 45.518] - - - [42368, 26497, 1, 128] - - [278, 45.346] - - - [31872, 128, 1, 128] - - [290, 27.758] - - - [39296, 2048, 1, 128] - - [265, 43.934] - - - [36736, 20865, 1, 128] - - [249, 45.681] - - - [30848, 14977, 1, 128] - - [264, 45.811] - - - [42880, 27009, 1, 128] - - [291, 44.895] - - - [38400, 4096, 1, 128] - - [264, 45.966] - - - [35072, 4096, 1, 128] - - [264, 45.875] - - - [39296, 8192, 1, 128] - - [256, 46.525] - - - [32768, 8192, 1, 128] - - [264, 35.565] - - - [43392, 128, 1, 128] - - [286, 31.053] - - - [35584, 4096, 1, 128] - - [254, 45.526] - - - [40576, 24705, 1, 128] - - [265, 45.497] - - - [36992, 1024, 1, 128] - - [299, 42.438] - - - [43904, 128, 1, 128] - - [274, 31.389] - - - [42112, 128, 1, 128] - - [267, 33.659] - - - [38016, 128, 1, 128] - - [287, 31.37] - - - [43392, 512, 1, 128] - - [250, 41.032] - - - [42240, 2048, 1, 128] - - [280, 44.379] - - - [36608, 8192, 1, 128] - - [265, 46.612] - - - [29568, 512, 1, 128] - - [258, 38.598] - - - [41984, 4096, 1, 128] - - [252, 46.4] - - - [41216, 2048, 1, 128] - - [263, 44.506] - - - [44800, 4096, 1, 128] - - [256, 45.493] - - - [39680, 2048, 1, 128] - - [256, 43.919] - - - [33664, 128, 1, 128] - - [250, 29.078] - - - [43776, 1024, 1, 128] - - [290, 41.689] - - - [44160, 128, 1, 128] - - [266, 31.541] - - - [34816, 18817, 1, 128] - - [256, 46.801] - - - [40192, 24321, 1, 128] - - [265, 45.492] - - - [29312, 8192, 1, 128] - - [254, 46.211] - - - [37632, 21633, 1, 128] - - [265, 45.804] - - - [35968, 1024, 1, 128] - - [250, 43.0] - - - [35968, 512, 1, 128] - - [282, 40.217] - - - [38400, 512, 1, 128] - - [266, 40.674] - - - [38528, 4096, 1, 128] - - [276, 45.377] - - - [33152, 1024, 1, 128] - - [250, 42.341] - - - [30848, 8192, 1, 128] - - [264, 46.564] - - - [38912, 2048, 1, 128] - - [256, 44.728] - - - [31488, 4096, 1, 128] - - [252, 45.665] - - - [39552, 2048, 1, 128] - - [267, 44.397] - - - [34304, 18305, 1, 128] - - [277, 46.325] - - - [44416, 8192, 1, 128] - - [252, 46.505] - - - [30208, 14337, 1, 128] - - [251, 46.144] - - - [40448, 128, 1, 128] - - [305, 33.044] - - - [39424, 2048, 1, 128] - - [263, 44.584] - - - [36224, 2048, 1, 128] - - [287, 43.974] - - - [44160, 512, 1, 128] - - [276, 40.882] - - - [42752, 1024, 1, 128] - - [267, 43.363] - - - [32640, 2048, 1, 128] - - [252, 43.065] - - - [32256, 8192, 1, 128] - - [278, 47.12] - - - [34816, 128, 1, 128] - - [290, 29.795] - - - [44800, 28929, 1, 128] - - [249, 45.322] - - - [43776, 512, 1, 128] - - [260, 37.791] - - - [34432, 18561, 1, 128] - - [301, 44.536] - - - [36224, 512, 1, 128] - - [267, 40.29] - - - [30592, 128, 1, 128] - - [267, 26.897] - - - [34048, 18177, 1, 128] - - [291, 45.732] - - - [29184, 13313, 1, 128] - - [249, 45.766] - - - [40960, 128, 1, 128] - - [280, 32.703] - - - [37248, 128, 1, 128] - - [261, 31.154] - - - [43264, 128, 1, 128] - - [284, 31.142] - - - [36736, 20737, 1, 128] - - [265, 45.629] - - - [32768, 1024, 1, 128] - - [252, 36.303] - - - [31104, 1024, 1, 128] - - [276, 43.24] - - - [41216, 25217, 1, 128] - - [291, 45.833] - - - [38528, 8192, 1, 128] - - [251, 46.201] - - - [32512, 1024, 1, 128] - - [250, 42.329] - - - [31360, 15361, 1, 128] - - [265, 45.526] - - - [35712, 128, 1, 128] - - [261, 30.005] - - - [33920, 128, 1, 128] - - [262, 29.163] - - - [42368, 128, 1, 128] - - [253, 30.858] - - - [34176, 512, 1, 128] - - [252, 41.65] - - - [31488, 1024, 1, 128] - - [261, 43.165] - - - [41856, 128, 1, 128] - - [267, 33.455] - - - [43520, 128, 1, 128] - - [263, 31.296] - - - [40448, 2048, 1, 128] - - [260, 44.652] - - - [36864, 20865, 1, 128] - - [265, 46.566] - - - [34176, 18177, 1, 128] - - [291, 45.62] - - - [33792, 1024, 1, 128] - - [262, 43.346] - - - [40704, 512, 1, 128] - - [261, 41.912] - - - [32640, 1024, 1, 128] - - [252, 41.437] - - - [39424, 8192, 1, 128] - - [252, 46.96] - - - [35840, 128, 1, 128] - - [254, 30.25] - - - [35712, 19841, 1, 128] - - [265, 45.634] - - - [34944, 4096, 1, 128] - - [254, 45.826] - - - [39296, 128, 1, 128] - - [254, 32.103] - - - [39040, 1024, 1, 128] - - [285, 43.154] - - - [33536, 2048, 1, 128] - - [252, 44.356] - - - [37888, 1024, 1, 128] - - [267, 43.644] - - - [30592, 512, 1, 128] - - [261, 40.13] - - - [29184, 512, 1, 128] - - [261, 39.349] - - - [43392, 27393, 1, 128] - - [265, 45.405] - - - [43648, 2048, 1, 128] - - [256, 43.517] - - - [35456, 128, 1, 128] - - [294, 30.098] - - - [40320, 1024, 1, 128] - - [267, 43.195] - - - [29696, 512, 1, 128] - - [261, 39.605] - - - [37504, 2048, 1, 128] - - [290, 43.735] - - - [42752, 4096, 1, 128] - - [264, 45.734] - - - [44800, 1024, 1, 128] - - [262, 43.392] - - - [43904, 4096, 1, 128] - - [254, 45.454] - - - [37632, 2048, 1, 128] - - [287, 44.242] - - - [30080, 512, 1, 128] - - [266, 39.423] - - - [43520, 27521, 1, 128] - - [291, 46.018] - - - [28928, 12929, 1, 128] - - [264, 45.873] - - - [35840, 19969, 1, 128] - - [265, 46.357] - - - [42496, 8192, 1, 128] - - [252, 46.936] - - - [33408, 1024, 1, 128] - - [252, 42.705] - - - [35712, 19713, 1, 128] - - [265, 45.682] - - - [32000, 16129, 1, 128] - - [254, 46.012] - - - [35584, 128, 1, 128] - - [267, 30.103] - - - [37888, 21889, 1, 128] - - [265, 46.38] - - - [33664, 512, 1, 128] - - [287, 41.523] - - - [38400, 1024, 1, 128] - - [299, 43.355] - - - [43648, 1024, 1, 128] - - [262, 42.242] - - - [32768, 128, 1, 128] - - [267, 28.982] - - - [35200, 19329, 1, 128] - - [249, 45.615] - - - [36864, 8192, 1, 128] - - [256, 47.323] - - - [42624, 1024, 1, 128] - - [264, 41.156] - - - [43648, 27777, 1, 128] - - [249, 44.988] - - - [41472, 25601, 1, 128] - - [291, 45.782] - - - [35456, 4096, 1, 128] - - [276, 45.661] - - - [43136, 4096, 1, 128] - - [264, 45.517] - - - [36480, 8192, 1, 128] - - [256, 46.305] - - - [39040, 8192, 1, 128] - - [254, 46.156] - - - [34176, 2048, 1, 128] - - [267, 44.049] - - - [35456, 2048, 1, 128] - - [250, 44.265] - - - [33024, 128, 1, 128] - - [287, 28.36] - - - [33408, 17409, 1, 128] - - [265, 45.715] - - - [29312, 13313, 1, 128] - - [254, 45.515] - - - [42496, 512, 1, 128] - - [262, 41.174] - - - [41088, 2048, 1, 128] - - [285, 43.811] - - - [35072, 2048, 1, 128] - - [280, 44.333] - - - [33280, 4096, 1, 128] - - [294, 46.23] - - - [34816, 8192, 1, 128] - - [254, 47.493] - - - [33152, 128, 1, 128] - - [286, 28.403] - - - [40832, 24961, 1, 128] - - [265, 45.111] - - - [41472, 25473, 1, 128] - - [277, 46.035] - - - [31616, 15617, 1, 128] - - [249, 45.318] - - - [35328, 8192, 1, 128] - - [257, 47.021] - - - [37248, 512, 1, 128] - - [267, 40.754] - - - [41344, 8192, 1, 128] - - [264, 46.237] - - - [35968, 19969, 1, 128] - - [254, 45.615] - - - [35840, 8192, 1, 128] - - [264, 47.132] - - - [31872, 2048, 1, 128] - - [264, 43.727] - - - [43776, 128, 1, 128] - - [285, 31.207] - - - [34432, 128, 1, 128] - - [262, 29.398] - - - [32384, 16385, 1, 128] - - [252, 45.9] - - - [44032, 4096, 1, 128] - - [264, 46.373] - - - [44160, 1024, 1, 128] - - [290, 43.268] - - - [44032, 2048, 1, 128] - - [285, 44.796] - - - [41984, 25985, 1, 128] - - [265, 46.352] - - - [41344, 25473, 1, 128] - - [256, 45.432] - - - [32640, 128, 1, 128] - - [287, 28.095] - - - [33536, 4096, 1, 128] - - [256, 45.86] - - - [35072, 8192, 1, 128] - - [256, 46.675] - - - [43520, 1024, 1, 128] - - [260, 43.744] - - - [30848, 128, 1, 128] - - [267, 27.219] - - - [44160, 2048, 1, 128] - - [280, 44.248] - - - [40448, 24449, 1, 128] - - [277, 46.019] - - - [32128, 1024, 1, 128] - - [287, 42.82] - - - [40064, 1024, 1, 128] - - [264, 42.512] - - - [37248, 2048, 1, 128] - - [264, 43.739] - - - [39040, 23041, 1, 128] - - [249, 45.415] - - - [39040, 2048, 1, 128] - - [280, 44.221] - - - [36992, 4096, 1, 128] - - [264, 45.424] - - - [30976, 2048, 1, 128] - - [258, 42.926] - - - [29824, 13953, 1, 128] - - [291, 45.133] - - - [30720, 8192, 1, 128] - - [256, 47.523] - - - [38016, 4096, 1, 128] - - [249, 45.516] - - - [36992, 21121, 1, 128] - - [249, 45.411] - - - [37888, 128, 1, 128] - - [287, 31.44] - - - [41856, 25985, 1, 128] - - [249, 45.662] - - - [38784, 1024, 1, 128] - - [267, 42.968] - - - [32000, 4096, 1, 128] - - [252, 45.811] - - - [37504, 512, 1, 128] - - [284, 40.171] - - - [37248, 21377, 1, 128] - - [265, 45.342] - - - [36096, 128, 1, 128] - - [263, 30.362] - - - [36736, 8192, 1, 128] - - [256, 46.469] - - - [44288, 8192, 1, 128] - - [254, 46.275] - - - [41344, 25345, 1, 128] - - [256, 45.426] - - - [38272, 4096, 1, 128] - - [288, 41.461] - - - [33152, 8192, 1, 128] - - [256, 46.687] - - - [34560, 18689, 1, 128] - - [265, 45.867] - - - [34944, 512, 1, 128] - - [264, 41.997] - - - [31616, 128, 1, 128] - - [266, 27.502] - - - [40576, 24577, 1, 128] - - [249, 45.04] - - - [40192, 1024, 1, 128] - - [250, 43.302] - - - [33664, 1024, 1, 128] - - [261, 43.004] - - - [44416, 1024, 1, 128] - - [267, 43.311] - - - [43648, 8192, 1, 128] - - [264, 46.058] - - - [40448, 1024, 1, 128] - - [290, 43.246] - - - [33024, 512, 1, 128] - - [267, 40.032] - - - [29696, 4096, 1, 128] - - [264, 46.382] - - - [38016, 8192, 1, 128] - - [252, 46.349] - - - [43648, 128, 1, 128] - - [280, 31.115] - - - [43008, 8192, 1, 128] - - [254, 47.387] - - - [38528, 128, 1, 128] - - [267, 31.686] - - - [38272, 22401, 1, 128] - - [301, 40.701] - - - [30720, 512, 1, 128] - - [287, 40.496] - - - [43136, 512, 1, 128] - - [287, 40.856] - - - [33536, 1024, 1, 128] - - [254, 43.074] - - - [36352, 20353, 1, 128] - - [249, 46.128] - - - [44800, 128, 1, 128] - - [286, 32.154] - - - [44416, 28545, 1, 128] - - [249, 45.558] - - - [32128, 16257, 1, 128] - - [264, 45.982] - - - [34432, 18433, 1, 128] - - [301, 44.145] - - - [32128, 2048, 1, 128] - - [285, 44.016] - - - [39936, 4096, 1, 128] - - [264, 46.381] - - - [37760, 512, 1, 128] - - [261, 40.533] - - - [38016, 22017, 1, 128] - - [265, 45.593] - - - [38912, 8192, 1, 128] - - [254, 47.448] - - - [36480, 2048, 1, 128] - - [267, 44.212] - - - [30976, 14977, 1, 128] - - [301, 44.608] - - - [40832, 1024, 1, 128] - - [264, 42.57] - - - [43392, 4096, 1, 128] - - [265, 45.466] - - - [38272, 22273, 1, 128] - - [305, 40.665] - - - [32896, 2048, 1, 128] - - [255, 41.278] - - - [41856, 1024, 1, 128] - - [290, 43.584] - - - [29568, 1024, 1, 128] - - [285, 41.624] - - - [33280, 8192, 1, 128] - - [278, 47.149] - - - [30592, 14593, 1, 128] - - [256, 45.761] - - - [30336, 14465, 1, 128] - - [256, 45.698] - - - [40832, 512, 1, 128] - - [271, 40.792] - - - [30080, 4096, 1, 128] - - [305, 43.362] - - - [37632, 21761, 1, 128] - - [249, 45.775] - - - [35328, 19457, 1, 128] - - [291, 46.076] - - - [33792, 17921, 1, 128] - - [264, 46.337] - - - [42752, 512, 1, 128] - - [267, 40.648] - - - [39168, 4096, 1, 128] - - [291, 45.615] - - - [43520, 8192, 1, 128] - - [252, 46.897] - - - [30720, 4096, 1, 128] - - [252, 46.713] - - - [32768, 16769, 1, 128] - - [264, 34.684] - - - [35712, 4096, 1, 128] - - [265, 45.542] - - - [39424, 4096, 1, 128] - - [254, 45.949] - - - [34432, 2048, 1, 128] - - [286, 42.941] - - - [35200, 512, 1, 128] - - [250, 41.728] - - - [31488, 128, 1, 128] - - [262, 27.718] - - - [33024, 4096, 1, 128] - - [252, 46.064] - - - [37376, 8192, 1, 128] - - [264, 46.86] - - - [36352, 2048, 1, 128] - - [267, 44.472] - - - [44416, 28417, 1, 128] - - [265, 45.544] - - - [28928, 2048, 1, 128] - - [287, 43.673] - - - [43136, 27137, 1, 128] - - [249, 45.344] - - - [40960, 512, 1, 128] - - [271, 40.455] - - - [31232, 15361, 1, 128] - - [252, 45.928] - - - [31232, 1024, 1, 128] - - [261, 43.531] - - - [44416, 512, 1, 128] - - [262, 41.196] - - - [31232, 8192, 1, 128] - - [254, 46.787] - - - [29568, 128, 1, 128] - - [262, 26.027] - - - [35712, 2048, 1, 128] - - [290, 44.276] - - - [40704, 1024, 1, 128] - - [261, 43.238] - - - [33536, 17665, 1, 128] - - [291, 46.186] - - - [34304, 2048, 1, 128] - - [299, 44.427] - - - [41600, 1024, 1, 128] - - [252, 43.256] - - - [42368, 512, 1, 128] - - [284, 40.466] - - - [29696, 2048, 1, 128] - - [271, 44.525] - - - [38784, 22785, 1, 128] - - [249, 45.55] - - - [32128, 128, 1, 128] - - [262, 27.849] - - - [38400, 2048, 1, 128] - - [283, 44.604] - - - [31616, 4096, 1, 128] - - [291, 45.071] - - - [42624, 4096, 1, 128] - - [256, 41.862] - - - [37760, 2048, 1, 128] - - [285, 44.239] - - - [41472, 128, 1, 128] - - [267, 33.401] - - - [43520, 4096, 1, 128] - - [254, 46.119] - - - [38912, 22913, 1, 128] - - [265, 46.708] - - - [33280, 128, 1, 128] - - [267, 28.747] - - - [36352, 4096, 1, 128] - - [252, 45.942] - - - [33792, 128, 1, 128] - - [290, 29.019] - - - [38784, 128, 1, 128] - - [250, 32.111] - - - [34432, 4096, 1, 128] - - [277, 44.362] - - - [31616, 15745, 1, 128] - - [249, 45.341] - - - [41344, 128, 1, 128] - - [250, 33.298] - - - [32896, 512, 1, 128] - - [298, 32.523] - - - [42368, 26369, 1, 128] - - [274, 45.288] - - - [37376, 512, 1, 128] - - [267, 41.061] - - - [29184, 13185, 1, 128] - - [249, 45.997] - - - [36608, 20609, 1, 128] - - [254, 45.82] - - - [42880, 8192, 1, 128] - - [274, 45.728] - - - [34944, 19073, 1, 128] - - [256, 45.75] - - - [40064, 512, 1, 128] - - [254, 41.616] - - - [30592, 8192, 1, 128] - - [254, 46.473] - - - [41088, 128, 1, 128] - - [306, 33.019] - - - [32384, 128, 1, 128] - - [266, 28.038] - - - [30976, 4096, 1, 128] - - [278, 44.37] - - - [42112, 512, 1, 128] - - [256, 41.873] - - - [35200, 8192, 1, 128] - - [254, 46.357] - - - [32640, 16641, 1, 128] - - [254, 45.367] - - - [30080, 128, 1, 128] - - [267, 26.605] - - - [44800, 2048, 1, 128] - - [265, 44.034] - - - [35712, 1024, 1, 128] - - [276, 42.79] - - - [33280, 1024, 1, 128] - - [258, 42.848] - - - [40704, 8192, 1, 128] - - [264, 46.688] - - - [39936, 512, 1, 128] - - [261, 41.614] - - - [40832, 8192, 1, 128] - - [256, 46.421] - - - [36736, 512, 1, 128] - - [282, 40.645] - - - [33920, 1024, 1, 128] - - [290, 42.96] - - - [36608, 2048, 1, 128] - - [287, 44.236] - - - [35200, 19201, 1, 128] - - [265, 45.538] - - - [30720, 128, 1, 128] - - [267, 26.818] - - - [38144, 22145, 1, 128] - - [249, 45.754] - - - [40704, 24705, 1, 128] - - [249, 45.747] - - - [43776, 8192, 1, 128] - - [288, 45.549] - - - [43008, 2048, 1, 128] - - [252, 44.887] - - - [33536, 512, 1, 128] - - [282, 41.487] - - - [39040, 23169, 1, 128] - - [249, 45.446] - - - [44672, 8192, 1, 128] - - [254, 46.483] - - - [34944, 2048, 1, 128] - - [267, 44.314] - - - [41600, 25729, 1, 128] - - [265, 45.529] - - - [38912, 128, 1, 128] - - [267, 32.145] - - - [43904, 2048, 1, 128] - - [280, 44.341] - - - [31232, 4096, 1, 128] - - [264, 45.845] - - - [28928, 128, 1, 128] - - [267, 25.434] - - - [44032, 1024, 1, 128] - - [250, 43.694] - - - [35456, 19585, 1, 128] - - [256, 45.703] - - - [31360, 128, 1, 128] - - [263, 27.409] - - - [33920, 8192, 1, 128] - - [254, 46.43] - - - [30080, 14209, 1, 128] - - [272, 43.283] - - - [32512, 512, 1, 128] - - [299, 39.476] - - - [43392, 1024, 1, 128] - - [282, 42.75] - - - [43264, 4096, 1, 128] - - [256, 45.807] - - - [31488, 512, 1, 128] - - [264, 40.673] - - - [44032, 8192, 1, 128] - - [254, 47.089] - - - [31360, 8192, 1, 128] - - [254, 46.446] - - - [40192, 512, 1, 128] - - [254, 41.544] - - - [41984, 26113, 1, 128] - - [249, 46.343] - - - [42880, 2048, 1, 128] - - [260, 43.617] - - - [40192, 128, 1, 128] - - [262, 32.619] - - - [37760, 21761, 1, 128] - - [249, 45.57] - - - [44288, 512, 1, 128] - - [276, 40.758] - - - [43904, 512, 1, 128] - - [282, 40.392] - - - [36224, 8192, 1, 128] - - [254, 46.49] - - - [44032, 28161, 1, 128] - - [265, 46.258] - - - [38784, 512, 1, 128] - - [299, 40.788] - - - [33920, 18049, 1, 128] - - [265, 45.703] - - - [29056, 128, 1, 128] - - [267, 25.546] - - - [36992, 8192, 1, 128] - - [264, 46.344] - - - [43904, 27905, 1, 128] - - [252, 45.391] - - - [29568, 8192, 1, 128] - - [256, 46.022] - - - [29184, 8192, 1, 128] - - [262, 46.616] - - - [38656, 22785, 1, 128] - - [291, 45.782] - - - [38656, 2048, 1, 128] - - [290, 44.176] - - - [44672, 128, 1, 128] - - [252, 31.601] - - - [33792, 8192, 1, 128] - - [254, 47.181] - - - [40576, 1024, 1, 128] - - [280, 43.164] - - - [34432, 1024, 1, 128] - - [254, 42.711] - - - [40576, 2048, 1, 128] - - [280, 44.334] - - - [36480, 128, 1, 128] - - [287, 30.685] - - - [32896, 17025, 1, 128] - - [301, 44.944] - - - [29440, 2048, 1, 128] - - [282, 43.835] - - - [32640, 8192, 1, 128] - - [264, 46.225] - - - [33152, 512, 1, 128] - - [267, 40.22] - - - [34944, 8192, 1, 128] - - [254, 46.479] - - - [33536, 17537, 1, 128] - - [277, 46.122] - - - [31488, 2048, 1, 128] - - [267, 44.242] - - - [31232, 15233, 1, 128] - - [277, 46.235] - - - [31232, 2048, 1, 128] - - [296, 44.357] - - - [36352, 128, 1, 128] - - [263, 30.612] - - - [32000, 8192, 1, 128] - - [256, 46.687] - - - [30080, 1024, 1, 128] - - [258, 41.65] - - - [29696, 1024, 1, 128] - - [307, 42.71] - - - [44416, 128, 1, 128] - - [284, 31.571] - - - [31872, 512, 1, 128] - - [267, 40.288] - - - [34560, 1024, 1, 128] - - [267, 43.417] - - - [44288, 28289, 1, 128] - - [249, 45.366] - - - [32384, 2048, 1, 128] - - [290, 44.101] - - - [39296, 1024, 1, 128] - - [261, 43.051] - - - [40320, 2048, 1, 128] - - [290, 44.54] - - - [39680, 512, 1, 128] - - [271, 41.26] - - - [30720, 1024, 1, 128] - - [262, 43.411] - - - [32896, 8192, 1, 128] - - [301, 45.437] - - - [37632, 4096, 1, 128] - - [249, 45.643] - - - [37248, 4096, 1, 128] - - [254, 45.415] - - - [36096, 512, 1, 128] - - [271, 39.045] - - - [32896, 128, 1, 128] - - [250, 27.925] - - - [43392, 2048, 1, 128] - - [264, 43.776] - - - [29056, 13057, 1, 128] - - [249, 45.668] - - - [37888, 4096, 1, 128] - - [264, 46.418] - - - [39168, 128, 1, 128] - - [261, 32.284] - - - [34560, 512, 1, 128] - - [287, 42.049] - - - [34688, 8192, 1, 128] - - [256, 46.367] - - - [33152, 17281, 1, 128] - - [264, 46.072] - - - [43648, 27649, 1, 128] - - [265, 44.631] - - - [41984, 128, 1, 128] - - [252, 33.557] - - - [32256, 16385, 1, 128] - - [257, 46.375] - - - [32000, 16001, 1, 128] - - [249, 46.01] - - - [32256, 1024, 1, 128] - - [252, 43.039] - - - [43008, 4096, 1, 128] - - [252, 46.736] - - - [38656, 8192, 1, 128] - - [291, 46.561] - - - [38528, 22657, 1, 128] - - [249, 45.514] - - - [37376, 21377, 1, 128] - - [265, 46.05] - - - [43264, 27265, 1, 128] - - [249, 45.734] - - - [30848, 4096, 1, 128] - - [262, 45.587] - - - [35456, 512, 1, 128] - - [307, 39.962] - - - [39936, 23937, 1, 128] - - [249, 46.264] - - - [32128, 8192, 1, 128] - - [249, 46.465] - - - [31744, 2048, 1, 128] - - [263, 44.56] - - - [40064, 8192, 1, 128] - - [249, 45.948] - - - [42624, 2048, 1, 128] - - [254, 40.989] - - - [40192, 8192, 1, 128] - - [256, 46.619] - - - [37760, 21889, 1, 128] - - [249, 45.545] - - - [44544, 512, 1, 128] - - [284, 41.173] - - - [39168, 23169, 1, 128] - - [277, 45.761] - - - [20480, 12673, 1, 128] - - [256, 47.23] - - - [9344, 512, 1, 128] - - [261, 32.285] - - - [20352, 12673, 1, 128] - - [256, 45.916] - - - [20608, 4096, 1, 128] - - [254, 44.987] - - - [26496, 4096, 1, 128] - - [294, 44.965] - - - [20352, 1024, 1, 128] - - [256, 42.206] - - - [22144, 14465, 1, 128] - - [254, 45.898] - - - [14720, 6913, 1, 128] - - [252, 44.874] - - - [13056, 5377, 1, 128] - - [252, 44.678] - - - [22656, 4096, 1, 128] - - [282, 45.081] - - - [20224, 4096, 1, 128] - - [262, 45.205] - - - [13184, 4096, 1, 128] - - [263, 43.952] - - - [18176, 128, 1, 128] - - [260, 26.582] - - - [26112, 4096, 1, 128] - - [284, 45.885] - - - [10880, 7297, 1, 128] - - [264, 44.655] - - - [9216, 5633, 1, 128] - - [254, 44.514] - - - [8832, 5121, 1, 128] - - [256, 43.392] - - - [15488, 512, 1, 128] - - [262, 35.007] - - - [16128, 8449, 1, 128] - - [254, 45.56] - - - [24960, 9089, 1, 128] - - [264, 45.317] - - - [25472, 8192, 1, 128] - - [252, 46.584] - - - [9472, 512, 1, 128] - - [254, 33.193] - - - [10624, 6913, 1, 128] - - [264, 44.484] - - - [27008, 11137, 1, 128] - - [277, 45.459] - - - [7936, 1024, 1, 128] - - [282, 36.263] - - - [26240, 8192, 1, 128] - - [256, 42.362] - - - [24576, 1024, 1, 128] - - [254, 41.641] - - - [25600, 9601, 1, 128] - - [265, 46.181] - - - [19328, 128, 1, 128] - - [261, 27.717] - - - [11136, 512, 1, 128] - - [263, 33.263] - - - [12544, 4737, 1, 128] - - [254, 44.272] - - - [25344, 8192, 1, 128] - - [291, 46.021] - - - [21760, 512, 1, 128] - - [261, 37.072] - - - [13824, 2048, 1, 128] - - [267, 43.654] - - - [19584, 11905, 1, 128] - - [252, 45.771] - - - [4608, 3073, 1, 128] - - [262, 38.869] - - - [17664, 1024, 1, 128] - - [254, 40.529] - - - [2048, 1537, 1, 128] - - [286, 33.885] - - - [21760, 13953, 1, 128] - - [249, 46.306] - - - [21504, 13697, 1, 128] - - [249, 46.635] - - - [15616, 1024, 1, 128] - - [276, 40.824] - - - [25600, 2048, 1, 128] - - [255, 44.245] - - - [23168, 15361, 1, 128] - - [252, 45.518] - - - [19200, 512, 1, 128] - - [262, 38.611] - - - [23168, 2048, 1, 128] - - [280, 43.985] - - - [13568, 1024, 1, 128] - - [254, 41.514] - - - [20608, 1024, 1, 128] - - [274, 42.014] - - - [16256, 4096, 1, 128] - - [256, 44.813] - - - [19456, 128, 1, 128] - - [261, 28.119] - - - [14464, 6785, 1, 128] - - [252, 44.794] - - - [8704, 2048, 1, 128] - - [254, 42.448] - - - [19200, 11521, 1, 128] - - [254, 45.806] - - - [14976, 7297, 1, 128] - - [254, 44.964] - - - [25984, 128, 1, 128] - - [261, 34.207] - - - [15232, 2048, 1, 128] - - [256, 42.804] - - - [19456, 2048, 1, 128] - - [250, 43.9] - - - [25856, 9857, 1, 128] - - [264, 45.618] - - - [13568, 5761, 1, 128] - - [252, 44.662] - - - [8832, 2048, 1, 128] - - [252, 40.482] - - - [25856, 2048, 1, 128] - - [267, 43.873] - - - [20736, 12929, 1, 128] - - [254, 46.039] - - - [14208, 2048, 1, 128] - - [267, 42.396] - - - [5760, 2048, 1, 128] - - [264, 38.982] - - - [5248, 3713, 1, 128] - - [261, 40.632] - - - [10752, 512, 1, 128] - - [287, 32.745] - - - [27392, 8192, 1, 128] - - [301, 45.808] - - - [11136, 7553, 1, 128] - - [256, 44.898] - - - [16000, 8193, 1, 128] - - [254, 45.091] - - - [18304, 10497, 1, 128] - - [252, 45.496] - - - [19840, 128, 1, 128] - - [284, 28.341] - - - [18688, 11009, 1, 128] - - [252, 45.921] - - - [12544, 1024, 1, 128] - - [254, 39.443] - - - [19072, 4096, 1, 128] - - [261, 45.091] - - - [23680, 16001, 1, 128] - - [291, 45.198] - - - [15872, 128, 1, 128] - - [284, 23.922] - - - [9856, 6273, 1, 128] - - [264, 42.549] - - - [16256, 2048, 1, 128] - - [254, 42.32] - - - [28672, 8192, 1, 128] - - [252, 47.54] - - - [11008, 1024, 1, 128] - - [295, 35.775] - - - [26496, 2048, 1, 128] - - [266, 43.274] - - - [9216, 2048, 1, 128] - - [254, 41.282] - - - [18048, 128, 1, 128] - - [286, 26.342] - - - [9344, 5761, 1, 128] - - [252, 44.358] - - - [12288, 1024, 1, 128] - - [252, 39.436] - - - [22784, 128, 1, 128] - - [285, 31.335] - - - [17408, 1024, 1, 128] - - [254, 41.997] - - - [13312, 5633, 1, 128] - - [256, 45.186] - - - [19328, 2048, 1, 128] - - [290, 43.728] - - - [22784, 15105, 1, 128] - - [291, 46.187] - - - [14208, 512, 1, 128] - - [280, 32.868] - - - [27776, 11777, 1, 128] - - [264, 45.416] - - - [14208, 4096, 1, 128] - - [252, 44.998] - - - [17536, 512, 1, 128] - - [280, 38.659] - - - [26624, 512, 1, 128] - - [261, 40.688] - - - [11008, 512, 1, 128] - - [261, 32.65] - - - [22016, 128, 1, 128] - - [308, 30.621] - - - [28544, 1024, 1, 128] - - [256, 42.309] - - - [21632, 13825, 1, 128] - - [265, 45.844] - - - [24192, 512, 1, 128] - - [250, 39.572] - - - [20864, 13185, 1, 128] - - [256, 45.554] - - - [8704, 5121, 1, 128] - - [264, 43.442] - - - [15616, 128, 1, 128] - - [267, 23.393] - - - [13440, 512, 1, 128] - - [261, 38.697] - - - [8960, 5377, 1, 128] - - [254, 43.666] - - - [13440, 5761, 1, 128] - - [284, 43.687] - - - [7296, 3585, 1, 128] - - [262, 41.586] - - - [6912, 512, 1, 128] - - [261, 36.398] - - - [18176, 10369, 1, 128] - - [252, 45.848] - - - [19712, 4096, 1, 128] - - [267, 43.38] - - - [26240, 1024, 1, 128] - - [254, 41.079] - - - [16768, 2048, 1, 128] - - [267, 42.748] - - - [4864, 512, 1, 128] - - [266, 27.847] - - - [18560, 128, 1, 128] - - [261, 26.93] - - - [27520, 512, 1, 128] - - [299, 40.386] - - - [7808, 4225, 1, 128] - - [276, 42.956] - - - [25728, 1024, 1, 128] - - [261, 42.049] - - - [23424, 2048, 1, 128] - - [259, 42.136] - - - [27136, 512, 1, 128] - - [258, 41.025] - - - [15616, 4096, 1, 128] - - [264, 45.338] - - - [28416, 512, 1, 128] - - [303, 38.809] - - - [3200, 1537, 1, 128] - - [264, 32.96] - - - [16000, 512, 1, 128] - - [296, 34.729] - - - [28288, 4096, 1, 128] - - [249, 45.457] - - - [18048, 10369, 1, 128] - - [251, 44.334] - - - [17792, 10113, 1, 128] - - [251, 45.432] - - - [17024, 2048, 1, 128] - - [271, 43.476] - - - [16896, 4096, 1, 128] - - [284, 45.64] - - - [6912, 3201, 1, 128] - - [276, 41.76] - - - [24448, 128, 1, 128] - - [267, 32.828] - - - [12032, 512, 1, 128] - - [261, 35.58] - - - [16128, 8321, 1, 128] - - [256, 45.623] - - - [24064, 8192, 1, 128] - - [284, 47.018] - - - [10112, 6401, 1, 128] - - [258, 43.844] - - - [16512, 2048, 1, 128] - - [309, 39.086] - - - [8192, 4481, 1, 128] - - [262, 42.962] - - - [13568, 512, 1, 128] - - [261, 38.952] - - - [13440, 4096, 1, 128] - - [252, 44.134] - - - [28288, 128, 1, 128] - - [267, 25.262] - - - [17280, 512, 1, 128] - - [262, 38.266] - - - [14848, 7041, 1, 128] - - [252, 45.163] - - - [2944, 1409, 1, 128] - - [262, 28.732] - - - [21888, 512, 1, 128] - - [285, 35.508] - - - [20224, 512, 1, 128] - - [254, 40.45] - - - [5632, 1024, 1, 128] - - [254, 34.129] - - - [2048, 1409, 1, 128] - - [284, 31.533] - - - [9088, 5505, 1, 128] - - [264, 43.99] - - - [18048, 4096, 1, 128] - - [264, 44.168] - - - [20864, 4096, 1, 128] - - [282, 45.018] - - - [16640, 128, 1, 128] - - [299, 24.927] - - - [24320, 1024, 1, 128] - - [264, 42.944] - - - [26240, 4096, 1, 128] - - [254, 42.717] - - - [26240, 128, 1, 128] - - [250, 34.73] - - - [11776, 2048, 1, 128] - - [264, 42.593] - - - [24960, 128, 1, 128] - - [250, 33.334] - - - [11776, 8065, 1, 128] - - [294, 45.272] - - - [8192, 4609, 1, 128] - - [261, 43.552] - - - [14976, 2048, 1, 128] - - [271, 42.838] - - - [14080, 1024, 1, 128] - - [254, 42.357] - - - [23424, 15745, 1, 128] - - [256, 43.83] - - - [6144, 2433, 1, 128] - - [256, 39.949] - - - [28416, 128, 1, 128] - - [290, 25.438] - - - [12544, 4096, 1, 128] - - [252, 45.013] - - - [23936, 16257, 1, 128] - - [254, 46.159] - - - [13824, 6017, 1, 128] - - [264, 45.055] - - - [9984, 6273, 1, 128] - - [254, 44.695] - - - [7936, 4353, 1, 128] - - [252, 43.529] - - - [8448, 512, 1, 128] - - [258, 29.71] - - - [22656, 14977, 1, 128] - - [249, 45.731] - - - [24960, 8961, 1, 128] - - [252, 45.169] - - - [9344, 2048, 1, 128] - - [254, 41.138] - - - [17024, 9217, 1, 128] - - [256, 45.351] - - - [21632, 2048, 1, 128] - - [280, 43.672] - - - [20736, 4096, 1, 128] - - [254, 45.481] - - - [26752, 2048, 1, 128] - - [267, 43.698] - - - [4736, 1024, 1, 128] - - [254, 33.114] - - - [27392, 11521, 1, 128] - - [301, 44.51] - - - [28672, 12801, 1, 128] - - [252, 46.648] - - - [25088, 2048, 1, 128] - - [296, 43.926] - - - [5120, 3585, 1, 128] - - [261, 40.938] - - - [8576, 1024, 1, 128] - - [310, 38.069] - - - [25728, 9729, 1, 128] - - [254, 45.32] - - - [27264, 1024, 1, 128] - - [262, 41.537] - - - [19840, 2048, 1, 128] - - [280, 43.568] - - - [4992, 3329, 1, 128] - - [276, 41.312] - - - [10112, 512, 1, 128] - - [285, 34.534] - - - [16896, 128, 1, 128] - - [296, 25.26] - - - [9600, 512, 1, 128] - - [276, 33.131] - - - [27008, 128, 1, 128] - - [261, 35.055] - - - [15872, 8065, 1, 128] - - [252, 45.84] - - - [15360, 2048, 1, 128] - - [280, 43.411] - - - [8192, 2048, 1, 128] - - [264, 40.844] - - - [4096, 2433, 1, 128] - - [286, 39.249] - - - [12672, 512, 1, 128] - - [305, 36.955] - - - [19712, 1024, 1, 128] - - [276, 40.481] - - - [13440, 2048, 1, 128] - - [282, 42.115] - - - [7552, 3969, 1, 128] - - [276, 42.069] - - - [18432, 512, 1, 128] - - [276, 37.578] - - - [14976, 512, 1, 128] - - [263, 34.617] - - - [27904, 128, 1, 128] - - [261, 36.155] - - - [15872, 4096, 1, 128] - - [264, 45.739] - - - [14336, 128, 1, 128] - - [258, 21.74] - - - [23040, 1024, 1, 128] - - [258, 41.885] - - - [14592, 512, 1, 128] - - [261, 33.889] - - - [22528, 2048, 1, 128] - - [254, 44.374] - - - [17536, 4096, 1, 128] - - [252, 45.101] - - - [20096, 1024, 1, 128] - - [254, 41.66] - - - [18816, 2048, 1, 128] - - [285, 43.333] - - - [18432, 4096, 1, 128] - - [254, 46.42] - - - [7168, 2048, 1, 128] - - [282, 39.766] - - - [7296, 3713, 1, 128] - - [252, 42.233] - - - [8576, 2048, 1, 128] - - [267, 42.014] - - - [14848, 7169, 1, 128] - - [254, 44.814] - - - [25856, 4096, 1, 128] - - [252, 45.649] - - - [22272, 128, 1, 128] - - [286, 30.86] - - - [6016, 2433, 1, 128] - - [261, 39.116] - - - [7040, 1024, 1, 128] - - [262, 39.056] - - - [16000, 1024, 1, 128] - - [264, 39.786] - - - [8448, 4865, 1, 128] - - [254, 43.535] - - - [23296, 4096, 1, 128] - - [254, 45.463] - - - [23552, 4096, 1, 128] - - [264, 46.138] - - - [7424, 3841, 1, 128] - - [264, 42.525] - - - [18816, 11137, 1, 128] - - [252, 45.576] - - - [16896, 9217, 1, 128] - - [252, 45.874] - - - [8064, 2048, 1, 128] - - [252, 40.104] - - - [8320, 4737, 1, 128] - - [302, 42.111] - - - [3712, 512, 1, 128] - - [262, 22.02] - - - [5504, 1024, 1, 128] - - [254, 32.134] - - - [20480, 1024, 1, 128] - - [261, 41.912] - - - [28544, 12545, 1, 128] - - [264, 45.641] - - - [27904, 8192, 1, 128] - - [277, 46.305] - - - [25344, 128, 1, 128] - - [296, 33.786] - - - [7680, 512, 1, 128] - - [254, 27.702] - - - [12544, 512, 1, 128] - - [263, 36.617] - - - [28544, 128, 1, 128] - - [290, 25.216] - - - [17664, 4096, 1, 128] - - [261, 45.009] - - - [16384, 8577, 1, 128] - - [256, 37.574] - - - [28032, 12033, 1, 128] - - [264, 45.692] - - - [12928, 5249, 1, 128] - - [264, 44.052] - - - [27264, 2048, 1, 128] - - [264, 42.91] - - - [19584, 4096, 1, 128] - - [254, 45.273] - - - [20352, 4096, 1, 128] - - [282, 45.354] - - - [24704, 8192, 1, 128] - - [288, 45.87] - - - [18304, 512, 1, 128] - - [262, 37.603] - - - [13568, 4096, 1, 128] - - [256, 45.258] - - - [17664, 2048, 1, 128] - - [256, 42.904] - - - [24960, 4096, 1, 128] - - [264, 45.456] - - - [27648, 11649, 1, 128] - - [252, 46.394] - - - [17664, 512, 1, 128] - - [263, 37.061] - - - [19840, 512, 1, 128] - - [264, 39.521] - - - [25984, 512, 1, 128] - - [305, 40.063] - - - [10752, 7041, 1, 128] - - [252, 45.241] - - - [7680, 2048, 1, 128] - - [264, 41.027] - - - [6784, 3201, 1, 128] - - [262, 40.973] - - - [22656, 1024, 1, 128] - - [254, 41.7] - - - [23936, 1024, 1, 128] - - [254, 42.686] - - - [15488, 2048, 1, 128] - - [287, 43.276] - - - [15744, 1024, 1, 128] - - [261, 41.287] - - - [5888, 2305, 1, 128] - - [282, 40.7] - - - [1664, 1153, 1, 128] - - [286, 22.184] - - - [20992, 13313, 1, 128] - - [256, 46.058] - - - [13056, 1024, 1, 128] - - [254, 40.26] - - - [12160, 512, 1, 128] - - [280, 36.03] - - - [11520, 2048, 1, 128] - - [264, 41.898] - - - [28544, 8192, 1, 128] - - [264, 46.393] - - - [10752, 7169, 1, 128] - - [252, 45.01] - - - [25984, 8192, 1, 128] - - [251, 46.384] - - - [7296, 1024, 1, 128] - - [252, 33.942] - - - [22400, 2048, 1, 128] - - [271, 43.599] - - - [13312, 4096, 1, 128] - - [252, 45.594] - - - [10752, 2048, 1, 128] - - [287, 42.028] - - - [9216, 1024, 1, 128] - - [262, 37.865] - - - [24320, 2048, 1, 128] - - [267, 44.178] - - - [3072, 1409, 1, 128] - - [256, 29.52] - - - [4096, 512, 1, 128] - - [260, 24.897] - - - [9472, 5761, 1, 128] - - [252, 43.936] - - - [6528, 1024, 1, 128] - - [262, 37.555] - - - [5120, 512, 1, 128] - - [303, 29.255] - - - [21120, 13441, 1, 128] - - [277, 44.906] - - - [11520, 512, 1, 128] - - [311, 34.41] - - - [18432, 128, 1, 128] - - [285, 27.171] - - - [14208, 128, 1, 128] - - [267, 21.502] - - - [26496, 1024, 1, 128] - - [262, 42.264] - - - [12032, 2048, 1, 128] - - [285, 42.76] - - - [11776, 512, 1, 128] - - [261, 35.139] - - - [19456, 11777, 1, 128] - - [264, 46.512] - - - [7040, 3329, 1, 128] - - [262, 41.787] - - - [14720, 7041, 1, 128] - - [256, 44.828] - - - [21376, 13569, 1, 128] - - [256, 45.75] - - - [21248, 128, 1, 128] - - [267, 29.72] - - - [18688, 10881, 1, 128] - - [256, 45.714] - - - [20096, 2048, 1, 128] - - [267, 43.167] - - - [28160, 12161, 1, 128] - - [277, 45.981] - - - [12416, 4096, 1, 128] - - [252, 44.607] - - - [18304, 1024, 1, 128] - - [250, 40.763] - - - [22272, 512, 1, 128] - - [287, 37.987] - - - [26496, 10625, 1, 128] - - [291, 45.111] - - - [14336, 4096, 1, 128] - - [254, 46.045] - - - [16768, 8961, 1, 128] - - [256, 45.569] - - - [21504, 2048, 1, 128] - - [288, 44.054] - - - [11264, 512, 1, 128] - - [254, 34.059] - - - [10368, 6785, 1, 128] - - [262, 44.491] - - - [19968, 2048, 1, 128] - - [290, 43.687] - - - [24576, 8577, 1, 128] - - [265, 42.526] - - - [12288, 4609, 1, 128] - - [252, 44.544] - - - [14848, 128, 1, 128] - - [262, 22.334] - - - [2560, 1024, 1, 128] - - [286, 29.199] - - - [17920, 2048, 1, 128] - - [296, 43.624] - - - [16640, 1024, 1, 128] - - [254, 41.308] - - - [5504, 3969, 1, 128] - - [295, 37.291] - - - [18816, 128, 1, 128] - - [267, 27.194] - - - [11520, 7937, 1, 128] - - [254, 44.746] - - - [5376, 1024, 1, 128] - - [252, 32.116] - - - [9856, 1024, 1, 128] - - [252, 39.055] - - - [20352, 128, 1, 128] - - [286, 28.794] - - - [24192, 128, 1, 128] - - [287, 32.783] - - - [26752, 8192, 1, 128] - - [249, 46.339] - - - [11264, 2048, 1, 128] - - [256, 42.067] - - - [20224, 2048, 1, 128] - - [267, 43.332] - - - [15104, 512, 1, 128] - - [252, 34.75] - - - [6784, 3073, 1, 128] - - [282, 40.31] - - - [8448, 1024, 1, 128] - - [294, 37.138] - - - [17536, 128, 1, 128] - - [261, 25.85] - - - [28032, 128, 1, 128] - - [250, 36.13] - - - [27136, 11265, 1, 128] - - [251, 46.115] - - - [27648, 128, 1, 128] - - [250, 35.822] - - - [26752, 10753, 1, 128] - - [265, 45.477] - - - [5376, 3841, 1, 128] - - [254, 42.384] - - - [17024, 512, 1, 128] - - [262, 37.559] - - - [14080, 128, 1, 128] - - [287, 25.612] - - - [19200, 128, 1, 128] - - [261, 27.968] - - - [22912, 4096, 1, 128] - - [261, 45.339] - - - [17792, 128, 1, 128] - - [260, 26.123] - - - [19712, 512, 1, 128] - - [254, 39.479] - - - [3712, 2049, 1, 128] - - [254, 34.365] - - - [8192, 512, 1, 128] - - [276, 28.879] - - - [25088, 512, 1, 128] - - [254, 39.464] - - - [24320, 8192, 1, 128] - - [284, 46.734] - - - [4992, 3457, 1, 128] - - [262, 41.978] - - - [20864, 1024, 1, 128] - - [264, 42.728] - - - [17280, 9601, 1, 128] - - [256, 45.382] - - - [7424, 2048, 1, 128] - - [254, 40.427] - - - [28288, 12289, 1, 128] - - [256, 45.397] - - - [24832, 8192, 1, 128] - - [286, 46.933] - - - [8960, 2048, 1, 128] - - [276, 40.69] - - - [25088, 4096, 1, 128] - - [294, 45.937] - - - [15104, 7297, 1, 128] - - [252, 45.225] - - - [26112, 8192, 1, 128] - - [284, 47.055] - - - [22912, 2048, 1, 128] - - [262, 43.139] - - - [22528, 4096, 1, 128] - - [256, 46.425] - - - [22016, 14209, 1, 128] - - [251, 46.54] - - - [16384, 128, 1, 128] - - [287, 24.795] - - - [24832, 8833, 1, 128] - - [277, 45.868] - - - [27520, 2048, 1, 128] - - [290, 43.992] - - - [2176, 1537, 1, 128] - - [262, 25.126] - - - [12160, 4353, 1, 128] - - [254, 43.944] - - - [26752, 1024, 1, 128] - - [267, 42.239] - - - [16896, 512, 1, 128] - - [299, 36.62] - - - [9600, 2048, 1, 128] - - [254, 41.598] - - - [9984, 2048, 1, 128] - - [252, 42.255] - - - [19968, 1024, 1, 128] - - [256, 42.029] - - - [10496, 2048, 1, 128] - - [256, 42.812] - - - [19072, 11393, 1, 128] - - [265, 45.652] - - - [27520, 11649, 1, 128] - - [254, 45.664] - - - [14592, 6785, 1, 128] - - [286, 44.037] - - - [13056, 5249, 1, 128] - - [264, 44.423] - - - [28416, 8192, 1, 128] - - [254, 46.394] - - - [18048, 1024, 1, 128] - - [252, 40.637] - - - [24704, 512, 1, 128] - - [286, 36.765] - - - [16640, 8961, 1, 128] - - [252, 45.941] - - - [12928, 2048, 1, 128] - - [267, 42.458] - - - [27904, 11905, 1, 128] - - [278, 45.645] - - - [9728, 6017, 1, 128] - - [261, 44.661] - - - [3456, 1921, 1, 128] - - [262, 37.335] - - - [4224, 1024, 1, 128] - - [254, 29.888] - - - [13696, 6017, 1, 128] - - [312, 42.514] - - - [18944, 512, 1, 128] - - [261, 38.516] - - - [1792, 1153, 1, 128] - - [278, 23.607] - - - [13184, 5505, 1, 128] - - [258, 43.612] - - - [23040, 15361, 1, 128] - - [264, 46.101] - - - [20864, 13057, 1, 128] - - [249, 45.465] - - - [28288, 8192, 1, 128] - - [265, 46.469] - - - [17536, 2048, 1, 128] - - [290, 43.799] - - - [23040, 512, 1, 128] - - [296, 38.562] - - - [14720, 128, 1, 128] - - [286, 22.186] - - - [1920, 1281, 1, 128] - - [261, 27.133] - - - [28800, 8192, 1, 128] - - [256, 46.396] - - - [28800, 12929, 1, 128] - - [264, 45.572] - - - [25600, 1024, 1, 128] - - [261, 42.273] - - - [24448, 1024, 1, 128] - - [276, 41.88] - - - [20480, 512, 1, 128] - - [254, 39.934] - - - [27264, 11393, 1, 128] - - [256, 45.662] - - - [24960, 1024, 1, 128] - - [267, 41.824] - - - [6272, 2561, 1, 128] - - [254, 40.26] - - - [20864, 512, 1, 128] - - [262, 40.683] - - - [15744, 8065, 1, 128] - - [252, 45.33] - - - [15104, 4096, 1, 128] - - [252, 45.379] - - - [25472, 9601, 1, 128] - - [256, 45.504] - - - [19968, 512, 1, 128] - - [261, 40.046] - - - [28416, 12545, 1, 128] - - [264, 45.589] - - - [13056, 4096, 1, 128] - - [276, 45.349] - - - [7424, 3713, 1, 128] - - [261, 43.114] - - - [24832, 2048, 1, 128] - - [266, 43.926] - - - [26368, 512, 1, 128] - - [303, 40.028] - - - [13824, 4096, 1, 128] - - [261, 45.2] - - - [21760, 1024, 1, 128] - - [262, 41.232] - - - [27136, 8192, 1, 128] - - [274, 47.092] - - - [7552, 512, 1, 128] - - [285, 27.24] - - - [17152, 2048, 1, 128] - - [287, 43.822] - - - [20224, 12417, 1, 128] - - [249, 45.968] - - - [13312, 512, 1, 128] - - [305, 37.959] - - - [23296, 15617, 1, 128] - - [256, 46.156] - - - [26240, 512, 1, 128] - - [261, 39.854] - - - [10624, 2048, 1, 128] - - [276, 41.133] - - - [17536, 9857, 1, 128] - - [252, 45.598] - - - [14848, 512, 1, 128] - - [262, 34.702] - - - [15104, 2048, 1, 128] - - [287, 43.131] - - - [23808, 4096, 1, 128] - - [265, 45.544] - - - [10112, 6529, 1, 128] - - [262, 43.982] - - - [20992, 128, 1, 128] - - [267, 29.53] - - - [3584, 1024, 1, 128] - - [282, 26.305] - - - [21376, 13697, 1, 128] - - [254, 45.838] - - - [11904, 1024, 1, 128] - - [252, 39.736] - - - [8960, 512, 1, 128] - - [282, 31.586] - - - [18560, 512, 1, 128] - - [254, 37.97] - - - [21632, 4096, 1, 128] - - [264, 45.217] - - - [18048, 10241, 1, 128] - - [278, 44.043] - - - [26112, 512, 1, 128] - - [250, 40.26] - - - [4480, 1024, 1, 128] - - [254, 31.324] - - - [23552, 2048, 1, 128] - - [271, 44.203] - - - [22912, 128, 1, 128] - - [290, 31.394] - - - [24064, 16257, 1, 128] - - [277, 46.571] - - - [27776, 128, 1, 128] - - [287, 35.926] - - - [15488, 7681, 1, 128] - - [252, 44.725] - - - [24832, 4096, 1, 128] - - [254, 45.784] - - - [11904, 4225, 1, 128] - - [264, 43.69] - - - [6400, 2817, 1, 128] - - [254, 40.365] - - - [17152, 4096, 1, 128] - - [262, 45.392] - - - [27392, 2048, 1, 128] - - [313, 42.968] - - - [8960, 5249, 1, 128] - - [286, 43.23] - - - [4736, 3073, 1, 128] - - [254, 39.057] - - - [14592, 2048, 1, 128] - - [262, 41.79] - - - [14464, 2048, 1, 128] - - [290, 42.686] - - - [19584, 2048, 1, 128] - - [267, 43.263] - - - [19584, 11777, 1, 128] - - [264, 45.594] - - - [10496, 6785, 1, 128] - - [254, 44.614] - - - [23808, 128, 1, 128] - - [287, 32.621] - - - [9088, 512, 1, 128] - - [307, 31.659] - - - [8576, 4993, 1, 128] - - [252, 43.512] - - - [20608, 2048, 1, 128] - - [288, 43.039] - - - [23424, 15617, 1, 128] - - [272, 43.609] - - - [6656, 512, 1, 128] - - [287, 35.556] - - - [9984, 512, 1, 128] - - [263, 34.496] - - - [17920, 512, 1, 128] - - [266, 37.359] - - - [3968, 2433, 1, 128] - - [261, 38.283] - - - [22272, 1024, 1, 128] - - [264, 41.907] - - - [18560, 10753, 1, 128] - - [264, 45.476] - - - [12800, 1024, 1, 128] - - [256, 40.056] - - - [7296, 2048, 1, 128] - - [262, 39.711] - - - [11904, 512, 1, 128] - - [261, 35.236] - - - [27008, 1024, 1, 128] - - [262, 42.529] - - - [6272, 1024, 1, 128] - - [307, 36.799] - - - [26624, 10625, 1, 128] - - [256, 46.721] - - - [18304, 10625, 1, 128] - - [264, 45.509] - - - [21120, 128, 1, 128] - - [288, 29.43] - - - [25728, 9857, 1, 128] - - [256, 45.466] - - - [18560, 4096, 1, 128] - - [264, 45.201] - - - [7680, 3969, 1, 128] - - [262, 42.949] - - - [28160, 128, 1, 128] - - [305, 35.981] - - - [9856, 512, 1, 128] - - [276, 33.895] - - - [27392, 128, 1, 128] - - [285, 35.553] - - - [7808, 512, 1, 128] - - [254, 27.757] - - - [24576, 512, 1, 128] - - [252, 39.435] - - - [26496, 512, 1, 128] - - [287, 40.326] - - - [11648, 512, 1, 128] - - [287, 34.583] - - - [17920, 10241, 1, 128] - - [274, 45.751] - - - [6016, 2048, 1, 128] - - [254, 40.14] - - - [19712, 128, 1, 128] - - [261, 28.104] - - - [16768, 128, 1, 128] - - [266, 24.867] - - - [12800, 4993, 1, 128] - - [286, 44.444] - - - [26880, 10881, 1, 128] - - [265, 45.591] - - - [7040, 3457, 1, 128] - - [276, 42.449] - - - [20096, 4096, 1, 128] - - [282, 44.898] - - - [7936, 2048, 1, 128] - - [264, 40.79] - - - [21120, 2048, 1, 128] - - [296, 42.844] - - - [7168, 3457, 1, 128] - - [262, 41.781] - - - [26624, 1024, 1, 128] - - [262, 42.803] - - - [12544, 4865, 1, 128] - - [256, 44.396] - - - [20224, 128, 1, 128] - - [263, 28.834] - - - [16256, 8577, 1, 128] - - [256, 45.045] - - - [21248, 1024, 1, 128] - - [254, 41.174] - - - [5888, 2177, 1, 128] - - [256, 39.11] - - - [17536, 1024, 1, 128] - - [256, 42.672] - - - [13952, 6273, 1, 128] - - [256, 44.621] - - - [2304, 1793, 1, 128] - - [254, 28.752] - - - [22144, 128, 1, 128] - - [261, 30.683] - - - [27136, 4096, 1, 128] - - [276, 46.044] - - - [26240, 10241, 1, 128] - - [252, 41.87] - - - [21504, 128, 1, 128] - - [261, 30.135] - - - [18432, 1024, 1, 128] - - [264, 41.126] - - - [25344, 4096, 1, 128] - - [294, 45.027] - - - [21888, 2048, 1, 128] - - [314, 36.567] - - - [15104, 7425, 1, 128] - - [254, 45.481] - - - [13952, 4096, 1, 128] - - [256, 44.891] - - - [8064, 4481, 1, 128] - - [264, 42.758] - - - [20096, 12289, 1, 128] - - [252, 45.138] - - - [23168, 512, 1, 128] - - [267, 38.602] - - - [10624, 1024, 1, 128] - - [264, 36.621] - - - [22144, 14337, 1, 128] - - [264, 45.549] - - - [16768, 9089, 1, 128] - - [252, 45.453] - - - [20608, 512, 1, 128] - - [303, 40.424] - - - [7552, 1024, 1, 128] - - [254, 34.968] - - - [25216, 2048, 1, 128] - - [250, 43.233] - - - [10880, 2048, 1, 128] - - [276, 40.607] - - - [8320, 4609, 1, 128] - - [254, 42.483] - - - [22144, 2048, 1, 128] - - [287, 43.669] - - - [4096, 2561, 1, 128] - - [286, 36.95] - - - [26624, 4096, 1, 128] - - [252, 46.684] - - - [19584, 128, 1, 128] - - [280, 28.304] - - - [23552, 512, 1, 128] - - [261, 39.286] - - - [3840, 2305, 1, 128] - - [261, 36.194] - - - [16640, 8833, 1, 128] - - [256, 46.213] - - - [16896, 2048, 1, 128] - - [250, 43.271] - - - [25728, 8192, 1, 128] - - [256, 46.455] - - - [3840, 1024, 1, 128] - - [254, 27.77] - - - [12416, 1024, 1, 128] - - [294, 39.02] - - - [2944, 1024, 1, 128] - - [286, 31.974] - - - [24064, 128, 1, 128] - - [285, 32.609] - - - [2688, 1024, 1, 128] - - [274, 30.021] - - - [24320, 4096, 1, 128] - - [264, 45.703] - - - [13824, 1024, 1, 128] - - [252, 41.992] - - - [26240, 10369, 1, 128] - - [264, 41.875] - - - [9856, 6145, 1, 128] - - [252, 42.117] - - - [14976, 1024, 1, 128] - - [254, 40.146] - - - [12288, 512, 1, 128] - - [280, 35.624] - - - [28544, 4096, 1, 128] - - [282, 45.444] - - - [26112, 2048, 1, 128] - - [287, 44.634] - - - [2304, 1665, 1, 128] - - [261, 26.796] - - - [28672, 1024, 1, 128] - - [276, 42.181] - - - [24704, 1024, 1, 128] - - [290, 39.923] - - - [23424, 4096, 1, 128] - - [262, 43.442] - - - [23680, 2048, 1, 128] - - [264, 42.883] - - - [15360, 512, 1, 128] - - [287, 35.422] - - - [16128, 2048, 1, 128] - - [290, 43.058] - - - [27520, 4096, 1, 128] - - [254, 45.446] - - - [28672, 12673, 1, 128] - - [264, 46.81] - - - [22400, 4096, 1, 128] - - [265, 45.103] - - - [28416, 1024, 1, 128] - - [264, 42.268] - - - [13184, 1024, 1, 128] - - [254, 40.381] - - - [16000, 8321, 1, 128] - - [256, 45.088] - - - [25088, 1024, 1, 128] - - [252, 42.278] - - - [18432, 10625, 1, 128] - - [264, 46.727] - - - [26880, 11009, 1, 128] - - [256, 45.732] - - - [12288, 4096, 1, 128] - - [254, 45.625] - - - [13056, 2048, 1, 128] - - [264, 42.641] - - - [6784, 1024, 1, 128] - - [254, 39.104] - - - [27008, 11009, 1, 128] - - [277, 45.472] - - - [25600, 512, 1, 128] - - [255, 39.615] - - - [28160, 8192, 1, 128] - - [278, 46.763] - - - [18688, 4096, 1, 128] - - [252, 45.475] - - - [7296, 512, 1, 128] - - [263, 26.543] - - - [4608, 2945, 1, 128] - - [261, 41.341] - - - [9600, 1024, 1, 128] - - [307, 38.743] - - - [11904, 4097, 1, 128] - - [264, 43.378] - - - [11520, 7809, 1, 128] - - [276, 44.699] - - - [13696, 1024, 1, 128] - - [252, 41.476] - - - [25216, 512, 1, 128] - - [262, 39.351] - - - [17408, 4096, 1, 128] - - [252, 45.895] - - - [6912, 3329, 1, 128] - - [282, 41.631] - - - [22144, 4096, 1, 128] - - [282, 45.107] - - - [17920, 10113, 1, 128] - - [291, 46.032] - - - [26112, 128, 1, 128] - - [250, 34.81] - - - [20096, 128, 1, 128] - - [260, 28.324] - - - [25472, 512, 1, 128] - - [267, 39.729] - - - [20352, 12545, 1, 128] - - [256, 45.911] - - - [20480, 2048, 1, 128] - - [252, 43.517] - - - [22400, 128, 1, 128] - - [267, 30.98] - - - [24064, 8193, 1, 128] - - [257, 45.758] - - - [10880, 1024, 1, 128] - - [252, 36.012] - - - [15488, 1024, 1, 128] - - [276, 40.525] - - - [12416, 4609, 1, 128] - - [284, 43.37] - - - [25856, 8192, 1, 128] - - [265, 46.514] - - - [16256, 512, 1, 128] - - [267, 34.925] - - - [23168, 4096, 1, 128] - - [254, 45.243] - - - [24320, 512, 1, 128] - - [284, 39.869] - - - [12800, 5121, 1, 128] - - [274, 43.607] - - - [24576, 8192, 1, 128] - - [265, 43.71] - - - [12160, 4481, 1, 128] - - [256, 43.502] - - - [19072, 1024, 1, 128] - - [262, 41.458] - - - [9088, 2048, 1, 128] - - [264, 40.832] - - - [17408, 9729, 1, 128] - - [264, 46.24] - - - [4480, 512, 1, 128] - - [294, 25.95] - - - [8704, 1024, 1, 128] - - [280, 38.492] - - - [25856, 1024, 1, 128] - - [267, 42.188] - - - [23296, 128, 1, 128] - - [287, 31.743] - - - [26368, 10497, 1, 128] - - [254, 45.865] - - - [27776, 11905, 1, 128] - - [264, 45.554] - - - [21248, 4096, 1, 128] - - [262, 45.353] - - - [1920, 1409, 1, 128] - - [254, 28.905] - - - [13824, 6145, 1, 128] - - [254, 44.647] - - - [21376, 128, 1, 128] - - [250, 29.786] - - - [4352, 2817, 1, 128] - - [262, 38.359] - - - [27264, 11265, 1, 128] - - [252, 45.335] - - - [16256, 128, 1, 128] - - [299, 24.011] - - - [9600, 6017, 1, 128] - - [282, 44.038] - - - [20992, 13185, 1, 128] - - [256, 46.298] - - - [14720, 512, 1, 128] - - [263, 34.106] - - - [20352, 2048, 1, 128] - - [287, 43.425] - - - [8448, 2048, 1, 128] - - [256, 41.013] - - - [7808, 4097, 1, 128] - - [252, 42.053] - - - [12288, 4481, 1, 128] - - [254, 44.377] - - - [21760, 2048, 1, 128] - - [256, 43.441] - - - [13696, 2048, 1, 128] - - [256, 41.689] - - - [5632, 512, 1, 128] - - [286, 31.392] - - - [25216, 128, 1, 128] - - [287, 33.676] - - - [6144, 2048, 1, 128] - - [256, 39.869] - - - [15744, 128, 1, 128] - - [250, 23.537] - - - [16512, 1024, 1, 128] - - [302, 34.182] - - - [19456, 512, 1, 128] - - [261, 39.313] - - - [14080, 6401, 1, 128] - - [286, 43.172] - - - [4224, 2689, 1, 128] - - [258, 37.664] - - - [17280, 1024, 1, 128] - - [262, 42.398] - - - [7168, 3585, 1, 128] - - [254, 41.883] - - - [25728, 128, 1, 128] - - [285, 34.113] - - - [24832, 1024, 1, 128] - - [267, 41.954] - - - [17152, 1024, 1, 128] - - [264, 42.312] - - - [21120, 1024, 1, 128] - - [254, 42.256] - - - [21248, 13441, 1, 128] - - [256, 45.979] - - - [19712, 11905, 1, 128] - - [291, 44.366] - - - [24832, 8961, 1, 128] - - [278, 45.757] - - - [13696, 4096, 1, 128] - - [276, 43.197] - - - [21632, 13953, 1, 128] - - [249, 45.908] - - - [6272, 512, 1, 128] - - [290, 33.934] - - - [15232, 4096, 1, 128] - - [254, 44.986] - - - [2816, 1281, 1, 128] - - [252, 26.313] - - - [27648, 11777, 1, 128] - - [264, 46.371] - - - [13952, 6145, 1, 128] - - [261, 44.191] - - - [12544, 2048, 1, 128] - - [254, 42.423] - - - [24704, 2048, 1, 128] - - [255, 43.324] - - - [25088, 9217, 1, 128] - - [274, 45.765] - - - [12800, 4096, 1, 128] - - [294, 44.802] - - - [16512, 128, 1, 128] - - [290, 24.389] - - - [8576, 4865, 1, 128] - - [252, 43.84] - - - [28288, 2048, 1, 128] - - [250, 44.215] - - - [2560, 1921, 1, 128] - - [262, 33.5] - - - [6656, 2945, 1, 128] - - [284, 41.037] - - - [4608, 1024, 1, 128] - - [254, 32.411] - - - [23808, 2048, 1, 128] - - [287, 43.971] - - - [28416, 12417, 1, 128] - - [256, 45.628] - - - [21888, 128, 1, 128] - - [266, 30.215] - - - [22144, 1024, 1, 128] - - [276, 41.574] - - - [25600, 1024, 1, 256] - - [280, 69.238] - - - [15104, 512, 1, 256] - - [250, 55.153] - - - [34304, 8192, 1, 256] - - [269, 74.829] - - - [39424, 23552, 1, 256] - - [268, 75.106] - - - [28928, 13056, 1, 256] - - [278, 74.869] - - - [18432, 1024, 1, 256] - - [267, 66.708] - - - [33280, 17152, 1, 256] - - [251, 75.787] - - - [39680, 8192, 1, 256] - - [252, 74.356] - - - [25088, 8192, 1, 256] - - [251, 74.701] - - - [19712, 11776, 1, 256] - - [268, 74.182] - - - [44544, 4096, 1, 256] - - [268, 73.451] - - - [31744, 4096, 1, 256] - - [255, 73.158] - - - [32768, 8192, 1, 256] - - [315, 58.919] - - - [19968, 512, 1, 256] - - [285, 61.638] - - - [36864, 4096, 1, 256] - - [281, 72.634] - - - [22784, 14848, 1, 256] - - [255, 75.154] - - - [31488, 1024, 1, 256] - - [287, 69.495] - - - [33536, 512, 1, 256] - - [267, 65.285] - - - [15104, 7168, 1, 256] - - [254, 73.585] - - - [25344, 4096, 1, 256] - - [270, 72.024] - - - [21248, 13568, 1, 256] - - [264, 75.179] - - - [20992, 512, 1, 256] - - [254, 63.931] - - - [36864, 20736, 1, 256] - - [252, 75.147] - - - [37888, 512, 1, 256] - - [250, 66.052] - - - [35584, 19712, 1, 256] - - [249, 74.82] - - - [2816, 1024, 1, 256] - - [286, 46.28] - - - [39680, 4096, 1, 256] - - [264, 72.728] - - - [27648, 11776, 1, 256] - - [256, 75.73] - - - [4096, 2304, 1, 256] - - [261, 59.057] - - - [6400, 512, 1, 256] - - [267, 51.306] - - - [29440, 1024, 1, 256] - - [287, 68.859] - - - [19200, 4096, 1, 256] - - [264, 72.099] - - - [37888, 1024, 1, 256] - - [285, 70.228] - - - [26112, 1024, 1, 256] - - [296, 69.213] - - - [27136, 512, 1, 256] - - [254, 64.388] - - - [31232, 8192, 1, 256] - - [268, 74.803] - - - [27648, 512, 1, 256] - - [311, 65.08] - - - [44544, 2048, 1, 256] - - [268, 72.039] - - - [19456, 4096, 1, 256] - - [252, 72.834] - - - [35328, 4096, 1, 256] - - [270, 73.657] - - - [32768, 1024, 1, 256] - - [265, 59.341] - - - [15616, 7936, 1, 256] - - [264, 74.429] - - - [14592, 4096, 1, 256] - - [258, 70.479] - - - [37376, 21248, 1, 256] - - [264, 75.33] - - - [36608, 20480, 1, 256] - - [268, 74.602] - - - [44544, 28416, 1, 256] - - [265, 75.158] - - - [41216, 25344, 1, 256] - - [254, 74.919] - - - [30464, 14336, 1, 256] - - [268, 73.987] - - - [22528, 1024, 1, 256] - - [267, 68.222] - - - [27392, 1024, 1, 256] - - [278, 64.891] - - - [26368, 1024, 1, 256] - - [250, 68.193] - - - [23552, 15872, 1, 256] - - [252, 76.241] - - - [27392, 8192, 1, 256] - - [255, 74.062] - - - [16896, 9216, 1, 256] - - [254, 74.885] - - - [31488, 512, 1, 256] - - [252, 64.98] - - - [17152, 512, 1, 256] - - [290, 60.79] - - - [41984, 26112, 1, 256] - - [249, 75.545] - - - [16128, 8448, 1, 256] - - [261, 74.56] - - - [27904, 1024, 1, 256] - - [266, 68.722] - - - [20480, 12544, 1, 256] - - [249, 75.828] - - - [21504, 13824, 1, 256] - - [254, 76.158] - - - [32000, 512, 1, 256] - - [261, 63.572] - - - [34816, 8192, 1, 256] - - [264, 75.041] - - - [36608, 1024, 1, 256] - - [267, 69.423] - - - [30208, 14336, 1, 256] - - [269, 75.063] - - - [32256, 4096, 1, 256] - - [270, 73.355] - - - [10752, 6912, 1, 256] - - [254, 74.308] - - - [41472, 512, 1, 256] - - [252, 66.983] - - - [41472, 8192, 1, 256] - - [270, 74.446] - - - [44800, 512, 1, 256] - - [256, 66.471] - - - [19200, 1024, 1, 256] - - [299, 66.532] - - - [30976, 8192, 1, 256] - - [268, 73.668] - - - [22016, 14336, 1, 256] - - [264, 75.309] - - - [18432, 10752, 1, 256] - - [252, 75.858] - - - [21760, 13824, 1, 256] - - [256, 75.487] - - - [36096, 8192, 1, 256] - - [255, 73.872] - - - [35584, 19456, 1, 256] - - [268, 74.674] - - - [42240, 1024, 1, 256] - - [271, 69.828] - - - [28160, 12032, 1, 256] - - [254, 75.503] - - - [43520, 8192, 1, 256] - - [256, 74.662] - - - [6912, 3072, 1, 256] - - [261, 68.847] - - - [36352, 20224, 1, 256] - - [268, 75.358] - - - [37632, 1024, 1, 256] - - [250, 69.481] - - - [15616, 1024, 1, 256] - - [287, 65.832] - - - [24832, 8704, 1, 256] - - [278, 74.982] - - - [18176, 10496, 1, 256] - - [254, 75.104] - - - [16384, 1024, 1, 256] - - [281, 58.261] - - - [28672, 12544, 1, 256] - - [249, 75.492] - - - [22272, 14592, 1, 256] - - [274, 75.412] - - - [3584, 1792, 1, 256] - - [282, 56.36] - - - [32000, 4096, 1, 256] - - [281, 72.71] - - - [20480, 12800, 1, 256] - - [252, 75.76] - - - [44288, 8192, 1, 256] - - [268, 74.353] - - - [32512, 1024, 1, 256] - - [260, 68.481] - - - [40448, 1024, 1, 256] - - [287, 70.029] - - - [29952, 8192, 1, 256] - - [278, 74.456] - - - [29696, 4096, 1, 256] - - [252, 73.234] - - - [19968, 1024, 1, 256] - - [287, 66.898] - - - [33536, 17408, 1, 256] - - [255, 74.787] - - - [16384, 8704, 1, 256] - - [265, 63.818] - - - [14592, 6656, 1, 256] - - [294, 72.415] - - - [31744, 15872, 1, 256] - - [254, 75.897] - - - [23808, 1024, 1, 256] - - [267, 68.233] - - - [23296, 15360, 1, 256] - - [256, 75.187] - - - [22016, 1024, 1, 256] - - [290, 67.571] - - - [25600, 4096, 1, 256] - - [255, 73.152] - - - [30464, 8192, 1, 256] - - [268, 73.421] - - - [23808, 15872, 1, 256] - - [264, 75.524] - - - [44032, 4096, 1, 256] - - [264, 73.194] - - - [29184, 8192, 1, 256] - - [257, 74.597] - - - [36608, 8192, 1, 256] - - [252, 74.225] - - - [33024, 4096, 1, 256] - - [269, 73.199] - - - [23808, 16128, 1, 256] - - [254, 75.37] - - - [27392, 11520, 1, 256] - - [269, 74.289] - - - [20992, 13312, 1, 256] - - [255, 75.507] - - - [32512, 4096, 1, 256] - - [316, 72.937] - - - [9216, 1024, 1, 256] - - [263, 58.517] - - - [32256, 1024, 1, 256] - - [287, 69.566] - - - [29952, 1024, 1, 256] - - [287, 68.339] - - - [41728, 25600, 1, 256] - - [268, 73.877] - - - [38144, 22016, 1, 256] - - [249, 74.81] - - - [34048, 18176, 1, 256] - - [292, 74.655] - - - [19712, 12032, 1, 256] - - [268, 74.357] - - - [15872, 1024, 1, 256] - - [267, 64.076] - - - [9984, 1024, 1, 256] - - [267, 61.703] - - - [15360, 4096, 1, 256] - - [254, 72.39] - - - [44032, 8192, 1, 256] - - [264, 74.818] - - - [36352, 8192, 1, 256] - - [255, 74.844] - - - [20480, 1024, 1, 256] - - [254, 65.317] - - - [32000, 15872, 1, 256] - - [254, 75.055] - - - [15104, 7424, 1, 256] - - [256, 74.433] - - - [14848, 512, 1, 256] - - [252, 54.42] - - - [24832, 512, 1, 256] - - [250, 60.301] - - - [12544, 4608, 1, 256] - - [262, 72.008] - - - [29440, 512, 1, 256] - - [290, 63.112] - - - [18176, 4096, 1, 256] - - [256, 72.486] - - - [42240, 512, 1, 256] - - [263, 67.25] - - - [14848, 7168, 1, 256] - - [252, 73.661] - - - [15616, 7680, 1, 256] - - [256, 74.594] - - - [42752, 1024, 1, 256] - - [263, 69.863] - - - [36352, 1024, 1, 256] - - [305, 69.823] - - - [41472, 4096, 1, 256] - - [255, 73.656] - - - [22528, 14592, 1, 256] - - [252, 76.245] - - - [30720, 1024, 1, 256] - - [285, 69.647] - - - [22272, 1024, 1, 256] - - [267, 67.222] - - - [10240, 6400, 1, 256] - - [282, 74.368] - - - [41728, 25856, 1, 256] - - [268, 74.256] - - - [21760, 14080, 1, 256] - - [252, 75.486] - - - [24576, 8704, 1, 256] - - [265, 70.872] - - - [15360, 1024, 1, 256] - - [290, 65.031] - - - [27392, 512, 1, 256] - - [284, 62.874] - - - [40448, 4096, 1, 256] - - [288, 73.641] - - - [26368, 10496, 1, 256] - - [254, 74.947] - - - [33536, 1024, 1, 256] - - [250, 68.643] - - - [13824, 4096, 1, 256] - - [254, 71.984] - - - [9216, 5376, 1, 256] - - [276, 72.915] - - - [19456, 512, 1, 256] - - [254, 60.468] - - - [27392, 11264, 1, 256] - - [268, 74.537] - - - [44288, 4096, 1, 256] - - [268, 73.28] - - - [33280, 512, 1, 256] - - [287, 64.893] - - - [35584, 512, 1, 256] - - [250, 63.96] - - - [19968, 4096, 1, 256] - - [288, 72.937] - - - [41216, 8192, 1, 256] - - [268, 74.477] - - - [43520, 27392, 1, 256] - - [265, 75.44] - - - [38400, 22272, 1, 256] - - [265, 75.441] - - - [44800, 256, 1, 256] - - [285, 58.147] - - - [1792, 1024, 1, 256] - - [280, 30.733] - - - [30208, 1024, 1, 256] - - [290, 68.896] - - - [23040, 4096, 1, 256] - - [316, 73.309] - - - [38144, 512, 1, 256] - - [263, 65.417] - - - [14592, 6912, 1, 256] - - [295, 73.121] - - - [7168, 1024, 1, 256] - - [267, 53.439] - - - [19712, 1024, 1, 256] - - [254, 63.902] - - - [38656, 22784, 1, 256] - - [274, 74.655] - - - [4864, 1024, 1, 256] - - [276, 52.923] - - - [10240, 1024, 1, 256] - - [287, 62.923] - - - [24320, 4096, 1, 256] - - [274, 72.813] - - - [12032, 1024, 1, 256] - - [287, 63.985] - - - [16896, 4096, 1, 256] - - [254, 72.514] - - - [28416, 12544, 1, 256] - - [278, 74.959] - - - [19712, 4096, 1, 256] - - [268, 71.466] - - - [16384, 4096, 1, 256] - - [254, 61.887] - - - [38400, 8192, 1, 256] - - [264, 74.678] - - - [23296, 512, 1, 256] - - [262, 61.999] - - - [44544, 256, 1, 256] - - [263, 57.866] - - - [32512, 16640, 1, 256] - - [257, 75.357] - - - [43264, 512, 1, 256] - - [262, 65.863] - - - [26624, 8192, 1, 256] - - [264, 75.099] - - - [42752, 26880, 1, 256] - - [252, 74.838] - - - [34048, 8192, 1, 256] - - [270, 73.799] - - - [37632, 8192, 1, 256] - - [281, 74.221] - - - [8448, 4608, 1, 256] - - [276, 71.474] - - - [31488, 15616, 1, 256] - - [264, 74.939] - - - [37376, 8192, 1, 256] - - [268, 74.732] - - - [36608, 4096, 1, 256] - - [268, 72.898] - - - [5632, 512, 1, 256] - - [296, 46.409] - - - [17664, 9728, 1, 256] - - [284, 74.224] - - - [8448, 512, 1, 256] - - [262, 46.517] - - - [33536, 8192, 1, 256] - - [274, 74.444] - - - [43520, 4096, 1, 256] - - [268, 73.78] - - - [12544, 4096, 1, 256] - - [261, 71.348] - - - [14592, 512, 1, 256] - - [250, 53.515] - - - [18176, 1024, 1, 256] - - [267, 65.761] - - - [14848, 4096, 1, 256] - - [254, 72.5] - - - [16896, 8960, 1, 256] - - [264, 75.55] - - - [24064, 16128, 1, 256] - - [274, 75.907] - - - [25344, 9472, 1, 256] - - [255, 74.117] - - - [31488, 4096, 1, 256] - - [254, 72.728] - - - [28672, 12800, 1, 256] - - [252, 75.357] - - - [17664, 9984, 1, 256] - - [284, 74.504] - - - [11008, 1024, 1, 256] - - [286, 57.863] - - - [29952, 14080, 1, 256] - - [251, 75.218] - - - [9472, 512, 1, 256] - - [250, 51.769] - - - [5120, 1024, 1, 256] - - [267, 54.163] - - - [13568, 5632, 1, 256] - - [261, 73.703] - - - [16640, 8960, 1, 256] - - [278, 74.67] - - - [6912, 512, 1, 256] - - [280, 54.305] - - - [27648, 11520, 1, 256] - - [256, 75.978] - - - [34048, 4096, 1, 256] - - [269, 72.436] - - - [26368, 4096, 1, 256] - - [264, 72.565] - - - [15104, 1024, 1, 256] - - [267, 64.971] - - - [6144, 1024, 1, 256] - - [254, 54.464] - - - [25088, 512, 1, 256] - - [287, 61.479] - - - [27648, 1024, 1, 256] - - [267, 68.968] - - - [29440, 13312, 1, 256] - - [257, 75.126] - - - [23552, 512, 1, 256] - - [290, 62.539] - - - [44032, 28160, 1, 256] - - [264, 75.553] - - - [4352, 2560, 1, 256] - - [282, 60.33] - - - [12032, 4096, 1, 256] - - [252, 71.167] - - - [30208, 512, 1, 256] - - [262, 64.176] - - - [39424, 512, 1, 256] - - [284, 65.209] - - - [17152, 1024, 1, 256] - - [267, 66.735] - - - [27904, 11776, 1, 256] - - [269, 74.855] - - - [41728, 4096, 1, 256] - - [268, 72.671] - - - [44800, 28928, 1, 256] - - [252, 74.343] - - - [44032, 1024, 1, 256] - - [271, 70.885] - - - [15872, 8192, 1, 256] - - [252, 75.001] - - - [35584, 1024, 1, 256] - - [290, 69.271] - - - [36096, 20224, 1, 256] - - [255, 74.266] - - - [7936, 1024, 1, 256] - - [254, 56.833] - - - [38656, 8192, 1, 256] - - [268, 74.287] - - - [41984, 1024, 1, 256] - - [285, 70.499] - - - [32768, 16640, 1, 256] - - [315, 59.277] - - - [25856, 9984, 1, 256] - - [252, 74.885] - - - [37120, 4096, 1, 256] - - [268, 73.15] - - - [35072, 512, 1, 256] - - [263, 66.521] - - - [37376, 512, 1, 256] - - [250, 65.878] - - - [41984, 25856, 1, 256] - - [265, 75.626] - - - [31232, 15104, 1, 256] - - [278, 75.761] - - - [15360, 7424, 1, 256] - - [252, 75.274] - - - [26624, 4096, 1, 256] - - [254, 73.175] - - - [16128, 4096, 1, 256] - - [254, 72.085] - - - [15872, 512, 1, 256] - - [287, 57.075] - - - [10496, 6656, 1, 256] - - [262, 73.241] - - - [29184, 512, 1, 256] - - [263, 62.86] - - - [40192, 8192, 1, 256] - - [252, 74.473] - - - [25856, 8192, 1, 256] - - [252, 74.455] - - - [12288, 4352, 1, 256] - - [252, 73.324] - - - [28160, 1024, 1, 256] - - [250, 68.878] - - - [24832, 4096, 1, 256] - - [268, 72.684] - - - [22784, 4096, 1, 256] - - [288, 72.548] - - - [8704, 4864, 1, 256] - - [282, 72.098] - - - [39424, 23296, 1, 256] - - [249, 75.413] - - - [33792, 4096, 1, 256] - - [281, 72.911] - - - [25856, 512, 1, 256] - - [262, 62.098] - - - [40192, 4096, 1, 256] - - [301, 73.012] - - - [29184, 13312, 1, 256] - - [252, 75.113] - - - [37888, 21760, 1, 256] - - [254, 75.792] - - - [6912, 1024, 1, 256] - - [287, 60.997] - - - [42240, 26112, 1, 256] - - [256, 74.753] - - - [11776, 1024, 1, 256] - - [262, 62.652] - - - [34560, 18688, 1, 256] - - [256, 74.923] - - - [41728, 512, 1, 256] - - [256, 67.03] - - - [36864, 8192, 1, 256] - - [254, 74.544] - - - [30720, 14592, 1, 256] - - [252, 75.998] - - - [40192, 512, 1, 256] - - [254, 65.943] - - - [16640, 4096, 1, 256] - - [275, 71.565] - - - [27136, 8192, 1, 256] - - [257, 74.986] - - - [21504, 512, 1, 256] - - [254, 58.901] - - - [38912, 23040, 1, 256] - - [265, 75.609] - - - [38656, 22528, 1, 256] - - [255, 74.572] - - - [29184, 13056, 1, 256] - - [274, 75.458] - - - [38144, 8192, 1, 256] - - [268, 74.149] - - - [18432, 4096, 1, 256] - - [254, 72.431] - - - [29696, 1024, 1, 256] - - [280, 69.307] - - - [25600, 512, 1, 256] - - [264, 61.911] - - - [40448, 512, 1, 256] - - [271, 66.253] - - - [5376, 512, 1, 256] - - [260, 43.993] - - - [35840, 19968, 1, 256] - - [254, 75.549] - - - [34816, 18688, 1, 256] - - [254, 75.739] - - - [36352, 20480, 1, 256] - - [255, 75.176] - - - [28672, 1024, 1, 256] - - [256, 67.129] - - - [31488, 15360, 1, 256] - - [254, 74.816] - - - [12544, 512, 1, 256] - - [250, 55.681] - - - [26880, 8192, 1, 256] - - [254, 74.524] - - - [7680, 1024, 1, 256] - - [287, 55.641] - - - [39168, 8192, 1, 256] - - [255, 74.205] - - - [33536, 4096, 1, 256] - - [268, 73.191] - - - [34816, 18944, 1, 256] - - [252, 75.801] - - - [14848, 6912, 1, 256] - - [252, 74.685] - - - [32256, 16128, 1, 256] - - [257, 75.679] - - - [19968, 12288, 1, 256] - - [268, 75.413] - - - [15616, 4096, 1, 256] - - [256, 72.242] - - - [26112, 4096, 1, 256] - - [292, 73.758] - - - [15104, 4096, 1, 256] - - [254, 71.987] - - - [37888, 22016, 1, 256] - - [256, 75.62] - - - [41728, 1024, 1, 256] - - [254, 69.281] - - - [34048, 17920, 1, 256] - - [251, 74.538] - - - [11264, 1024, 1, 256] - - [267, 61.449] - - - [18432, 10496, 1, 256] - - [254, 75.833] - - - [41472, 25344, 1, 256] - - [264, 75.153] - - - [19456, 1024, 1, 256] - - [290, 65.93] - - - [8704, 512, 1, 256] - - [287, 48.105] - - - [8192, 1024, 1, 256] - - [254, 57.825] - - - [44544, 512, 1, 256] - - [258, 66.525] - - - [42752, 8192, 1, 256] - - [256, 74.408] - - - [42496, 26368, 1, 256] - - [254, 75.25] - - - [35840, 8192, 1, 256] - - [256, 74.949] - - - [21248, 4096, 1, 256] - - [264, 72.544] - - - [34304, 18176, 1, 256] - - [269, 75.541] - - - [25856, 9728, 1, 256] - - [254, 74.614] - - - [29440, 8192, 1, 256] - - [251, 74.577] - - - [21760, 512, 1, 256] - - [262, 59.218] - - - [28416, 12288, 1, 256] - - [257, 74.006] - - - [21504, 4096, 1, 256] - - [255, 72.907] - - - [13056, 5120, 1, 256] - - [276, 73.164] - - - [32768, 16896, 1, 256] - - [315, 59.383] - - - [33024, 8192, 1, 256] - - [270, 74.541] - - - [43008, 1024, 1, 256] - - [263, 70.968] - - - [11008, 512, 1, 256] - - [261, 49.571] - - - [43776, 512, 1, 256] - - [260, 63.657] - - - [12800, 4096, 1, 256] - - [301, 71.929] - - - [22528, 14848, 1, 256] - - [254, 75.994] - - - [40704, 1024, 1, 256] - - [299, 69.656] - - - [26112, 8192, 1, 256] - - [256, 74.852] - - - [15616, 512, 1, 256] - - [261, 56.189] - - - [31232, 4096, 1, 256] - - [268, 73.472] - - - [13312, 1024, 1, 256] - - [290, 64.256] - - - [4864, 512, 1, 256] - - [282, 40.82] - - - [43776, 1024, 1, 256] - - [303, 67.204] - - - [33792, 512, 1, 256] - - [261, 65.761] - - - [9216, 512, 1, 256] - - [252, 50.463] - - - [43008, 27136, 1, 256] - - [249, 75.512] - - - [40960, 25088, 1, 256] - - [265, 67.979] - - - [12288, 4096, 1, 256] - - [254, 71.721] - - - [40960, 24832, 1, 256] - - [265, 68.288] - - - [36096, 1281, 1, 256] - - [317, 63.047] - - - [35072, 2048, 1, 256] - - [301, 71.133] - - - [39168, 3328, 1, 256] - - [255, 72.947] - - - [39216, 5632, 1, 256] - - [285, 62.951] - - - [35328, 3072, 1, 256] - - [274, 73.434] - - - [36864, 3328, 1, 256] - - [256, 72.842] - - - [36352, 4352, 1, 256] - - [256, 74.346] - - - [38144, 256, 1, 256] - - [285, 57.666] - - - [35632, 1792, 1, 256] - - [285, 62.655] - - - [36144, 2816, 1, 256] - - [263, 62.834] - - - [35888, 2865, 1, 256] - - [280, 61.552] - - - [37120, 3072, 1, 256] - - [252, 73.092] - - - [39936, 3328, 1, 256] - - [254, 73.571] - - - [39680, 3329, 1, 256] - - [281, 69.552] - - - [38144, 5888, 1, 256] - - [264, 73.803] - - - [37376, 10240, 1, 256] - - [252, 74.986] - - - [39168, 5376, 1, 256] - - [255, 73.875] - - - [37376, 3584, 1, 256] - - [252, 74.017] - - - [39936, 6144, 1, 256] - - [264, 74.671] - - - [36352, 2304, 1, 256] - - [252, 73.025] - - - [36608, 1280, 1, 256] - - [252, 70.589] - - - [36608, 3329, 1, 256] - - [281, 69.48] - - - [37168, 3584, 1, 256] - - [271, 62.89] - - - [39424, 3329, 1, 256] - - [268, 70.143] - - - [39984, 2865, 1, 256] - - [263, 61.483] - - - [35584, 256, 1, 256] - - [271, 55.093] - - - [38960, 5376, 1, 256] - - [255, 62.631] - - - [39680, 1281, 1, 256] - - [254, 64.293] - - - [39936, 1280, 1, 256] - - [254, 71.595] - - - [35584, 10240, 1, 256] - - [251, 74.399] - - - [39424, 2816, 1, 256] - - [278, 73.142] - - - [38192, 2816, 1, 256] - - [263, 62.921] - - - [37936, 2865, 1, 256] - - [256, 61.41] - - - [38656, 10240, 1, 256] - - [268, 74.41] - - - [35120, 2816, 1, 256] - - [280, 63.338] - - - [37680, 2816, 1, 256] - - [263, 63.798] - - - [38144, 1281, 1, 256] - - [318, 64.27] - - - [39680, 256, 1, 256] - - [285, 59.321] - - - [39168, 2816, 1, 256] - - [268, 72.679] - - - [38912, 4608, 1, 256] - - [254, 74.158] - - - [35376, 1536, 1, 256] - - [263, 61.119] - - - [38192, 10240, 1, 256] - - [301, 62.37] - - - [39424, 1024, 1, 256] - - [303, 70.073] - - - [39984, 6144, 1, 256] - - [268, 63.37] - - - [35840, 2865, 1, 256] - - [254, 71.032] - - - [36864, 768, 1, 256] - - [263, 67.906] - - - [36912, 3328, 1, 256] - - [268, 61.472] - - - [38912, 1281, 1, 256] - - [254, 64.419] - - - [39936, 1281, 1, 256] - - [255, 64.808] - - - [37888, 2048, 1, 256] - - [255, 71.546] - - - [35376, 2816, 1, 256] - - [285, 63.641] - - - [36912, 256, 1, 256] - - [256, 51.073] - - - [37424, 256, 1, 256] - - [264, 51.135] - - - [35840, 3840, 1, 256] - - [252, 74.545] - - - [36096, 1280, 1, 256] - - [284, 68.901] - - - [39680, 3328, 1, 256] - - [256, 72.712] - - - [38912, 256, 1, 256] - - [256, 58.888] - - - [35888, 2816, 1, 256] - - [285, 63.218] - - - [38960, 10240, 1, 256] - - [268, 62.954] - - - [39424, 5120, 1, 256] - - [264, 74.236] - - - [37632, 4096, 1, 256] - - [254, 72.873] - - - [36144, 2560, 1, 256] - - [280, 63.37] - - - [39424, 5376, 1, 256] - - [264, 74.471] - - - [38656, 6144, 1, 256] - - [257, 73.826] - - - [36912, 2865, 1, 256] - - [254, 59.852] - - - [37632, 1536, 1, 256] - - [264, 70.658] - - - [35888, 2304, 1, 256] - - [263, 63.831] - - - [37680, 4096, 1, 256] - - [288, 62.597] - - - [37888, 3840, 1, 256] - - [254, 74.613] - - - [35328, 256, 1, 256] - - [311, 55.159] - - - [37888, 5888, 1, 256] - - [264, 74.629] - - - [35632, 256, 1, 256] - - [271, 50.1] - - - [37888, 2816, 1, 256] - - [256, 73.767] - - - [35328, 1536, 1, 256] - - [258, 70.794] - - - [35888, 10240, 1, 256] - - [255, 63.793] - - - [35072, 10240, 1, 256] - - [252, 74.629] - - - [35584, 3329, 1, 256] - - [300, 69.36] - - - [38144, 4096, 1, 256] - - [301, 73.033] - - - [36864, 3072, 1, 256] - - [252, 73.23] - - - [37632, 3584, 1, 256] - - [252, 73.496] - - - [38656, 512, 1, 256] - - [280, 66.314] - - - [36096, 2048, 1, 256] - - [255, 69.762] - - - [35840, 10240, 1, 256] - - [256, 75.284] - - - [38144, 4352, 1, 256] - - [256, 73.886] - - - [35840, 3329, 1, 256] - - [281, 70.438] - - - [38912, 10240, 1, 256] - - [252, 75.232] - - - [36096, 3840, 1, 256] - - [281, 72.331] - - - [39680, 1536, 1, 256] - - [259, 70.578] - - - [38912, 512, 1, 256] - - [254, 65.257] - - - [38704, 10240, 1, 256] - - [255, 62.954] - - - [35120, 256, 1, 256] - - [285, 53.951] - - - [37888, 1536, 1, 256] - - [264, 71.305] - - - [37376, 3329, 1, 256] - - [259, 70.049] - - - [39680, 6144, 1, 256] - - [256, 74.112] - - - [36608, 2865, 1, 256] - - [281, 69.97] - - - [35584, 1792, 1, 256] - - [261, 71.502] - - - [36352, 1281, 1, 256] - - [301, 64.906] - - - [38144, 3329, 1, 256] - - [281, 69.524] - - - [38656, 256, 1, 256] - - [252, 57.937] - - - [36352, 3329, 1, 256] - - [256, 70.011] - - - [39984, 6400, 1, 256] - - [268, 63.356] - - - [37888, 4096, 1, 256] - - [264, 73.073] - - - [36096, 256, 1, 256] - - [263, 55.392] - - - [36144, 2865, 1, 256] - - [280, 61.056] - - - [35072, 1024, 1, 256] - - [267, 69.616] - - - [39424, 1281, 1, 256] - - [318, 65.011] - - - [37632, 2865, 1, 256] - - [281, 70.226] - - - [37376, 5376, 1, 256] - - [264, 74.538] - - - [35584, 6144, 1, 256] - - [252, 73.895] - - - [36608, 3072, 1, 256] - - [264, 73.008] - - - [35328, 1281, 1, 256] - - [301, 64.856] - - - [39936, 1536, 1, 256] - - [259, 71.413] - - - [39168, 1281, 1, 256] - - [305, 64.281] - - - [38960, 2865, 1, 256] - - [254, 61.03] - - - [36864, 256, 1, 256] - - [285, 56.308] - - - [36096, 2304, 1, 256] - - [284, 70.752] - - - [38704, 2865, 1, 256] - - [285, 61.61] - - - [36608, 2560, 1, 256] - - [256, 72.736] - - - [35840, 2304, 1, 256] - - [264, 73.298] - - - [37376, 1280, 1, 256] - - [264, 71.316] - - - [35584, 2865, 1, 256] - - [259, 69.888] - - - [38912, 5120, 1, 256] - - [264, 74.698] - - - [37888, 1792, 1, 256] - - [254, 72.793] - - - [38192, 4352, 1, 256] - - [285, 62.726] - - - [35072, 1281, 1, 256] - - [305, 64.274] - - - [39728, 2865, 1, 256] - - [263, 61.192] - - - [38400, 4608, 1, 256] - - [269, 73.988] - - - [35072, 3072, 1, 256] - - [252, 72.885] - - - [39216, 2816, 1, 256] - - [285, 63.182] - - - [35632, 2048, 1, 256] - - [263, 62.969] - - - [37888, 3329, 1, 256] - - [256, 70.547] - - - [37376, 6144, 1, 256] - - [268, 74.51] - - - [36400, 256, 1, 256] - - [263, 50.53] - - - [38144, 10240, 1, 256] - - [264, 74.524] - - - [37168, 10240, 1, 256] - - [288, 62.52] - - - [37168, 2816, 1, 256] - - [280, 62.931] - - - [36352, 1280, 1, 256] - - [265, 70.957] - - - [38144, 3328, 1, 256] - - [292, 72.575] - - - [35328, 6144, 1, 256] - - [270, 74.535] - - - [36656, 10240, 1, 256] - - [268, 63.069] - - - [35120, 2865, 1, 256] - - [271, 61.404] - - - [37888, 2865, 1, 256] - - [264, 70.976] - - - [38704, 4864, 1, 256] - - [271, 63.525] - - - [38960, 5120, 1, 256] - - [255, 62.521] - - - [38656, 1280, 1, 256] - - [258, 70.603] - - - [39216, 2865, 1, 256] - - [263, 61.369] - - - [36096, 3328, 1, 256] - - [269, 71.742] - - - [37376, 3840, 1, 256] - - [256, 74.191] - - - [36864, 10240, 1, 256] - - [254, 74.71] - - - [37632, 5632, 1, 256] - - [264, 74.204] - - - [36864, 2865, 1, 256] - - [256, 70.301] - - - [37888, 1280, 1, 256] - - [261, 71.371] - - - [37632, 256, 1, 256] - - [261, 57.126] - - - [38656, 5120, 1, 256] - - [251, 73.766] - - - [38400, 3329, 1, 256] - - [281, 70.087] - - - [39680, 2865, 1, 256] - - [281, 70.163] - - - [37632, 10240, 1, 256] - - [264, 74.479] - - - [37424, 2865, 1, 256] - - [263, 61.628] - - - [36864, 6144, 1, 256] - - [264, 74.004] - - - [37936, 256, 1, 256] - - [285, 52.269] - - - [35840, 256, 1, 256] - - [305, 55.143] - - - [39728, 6144, 1, 256] - - [255, 62.67] - - - [38144, 2816, 1, 256] - - [254, 72.629] - - - [37376, 2816, 1, 256] - - [252, 73.31] - - - [39216, 256, 1, 256] - - [285, 52.562] - - - [39472, 10240, 1, 256] - - [255, 63.129] - - - [36096, 10240, 1, 256] - - [281, 73.83] - - - [39168, 5632, 1, 256] - - [274, 73.916] - - - [35072, 256, 1, 256] - - [263, 59.305] - - - [38400, 4864, 1, 256] - - [264, 74.544] - - - [35072, 3328, 1, 256] - - [268, 72.749] - - - [36352, 10240, 1, 256] - - [268, 75.006] - - - [35632, 2816, 1, 256] - - [263, 63.573] - - - [39936, 5888, 1, 256] - - [256, 74.727] - - - [38144, 3840, 1, 256] - - [254, 73.821] - - - [37376, 3328, 1, 256] - - [268, 73.555] - - - [36912, 3072, 1, 256] - - [255, 60.786] - - - [38144, 2865, 1, 256] - - [281, 70.012] - - - [38192, 4608, 1, 256] - - [285, 63.111] - - - [38400, 3328, 1, 256] - - [255, 73.36] - - - [38912, 4864, 1, 256] - - [254, 74.806] - - - [37120, 3329, 1, 256] - - [259, 69.606] - - - [38400, 1281, 1, 256] - - [269, 64.99] - - - [38144, 1280, 1, 256] - - [262, 70.558] - - - [39424, 6144, 1, 256] - - [268, 74.434] - - - [35328, 3329, 1, 256] - - [274, 70.11] - - - [36352, 4096, 1, 256] - - [255, 73.815] - - - [37168, 256, 1, 256] - - [285, 51.069] - - - [35840, 1536, 1, 256] - - [264, 71.264] - - - [37424, 3840, 1, 256] - - [263, 63.367] - - - [36864, 2560, 1, 256] - - [249, 72.877] - - - [36096, 2560, 1, 256] - - [284, 71.482] - - - [35328, 2048, 1, 256] - - [268, 71.615] - - - [38400, 4096, 1, 256] - - [268, 73.807] - - - [36144, 256, 1, 256] - - [254, 50.175] - - - [39168, 3329, 1, 256] - - [251, 69.417] - - - [37632, 3328, 1, 256] - - [264, 72.662] - - - [35120, 10240, 1, 256] - - [271, 62.831] - - - [36656, 2816, 1, 256] - - [285, 63.274] - - - [35328, 1792, 1, 256] - - [258, 72.387] - - - [39472, 2865, 1, 256] - - [263, 61.639] - - - [35888, 2048, 1, 256] - - [280, 62.81] - - - [37376, 2865, 1, 256] - - [264, 70.609] - - - [39472, 5632, 1, 256] - - [263, 63.752] - - - [37120, 2865, 1, 256] - - [281, 70.15] - - - [35584, 2816, 1, 256] - - [262, 72.373] - - - [36352, 2560, 1, 256] - - [254, 73.332] - - - [37632, 3840, 1, 256] - - [254, 73.728] - - - [37376, 3072, 1, 256] - - [265, 73.241] - - - [38912, 3328, 1, 256] - - [264, 73.576] - - - [35376, 256, 1, 256] - - [285, 49.669] - - - [44032, 5888, 1, 256] - - [264, 74.602] - - - [43312, 256, 1, 256] - - [263, 50.632] - - - [41216, 2816, 1, 256] - - [252, 72.943] - - - [43008, 2048, 1, 256] - - [255, 70.834] - - - [40704, 3328, 1, 256] - - [268, 72.847] - - - [40192, 1792, 1, 256] - - [254, 72.033] - - - [42032, 2865, 1, 256] - - [263, 61.445] - - - [41008, 7424, 1, 256] - - [255, 61.948] - - - [41264, 2865, 1, 256] - - [271, 62.119] - - - [40704, 2816, 1, 256] - - [254, 72.846] - - - [40960, 7168, 1, 256] - - [265, 66.483] - - - [41984, 6144, 1, 256] - - [254, 74.663] - - - [42752, 2865, 1, 256] - - [281, 70.315] - - - [40704, 7168, 1, 256] - - [252, 73.285] - - - [40960, 256, 1, 256] - - [263, 60.229] - - - [42752, 3329, 1, 256] - - [259, 69.627] - - - [40192, 6144, 1, 256] - - [252, 74.232] - - - [43520, 1281, 1, 256] - - [268, 65.119] - - - [43312, 10240, 1, 256] - - [271, 62.292] - - - [40192, 256, 1, 256] - - [280, 59.814] - - - [41216, 768, 1, 256] - - [263, 68.62] - - - [44288, 1280, 1, 256] - - [284, 70.78] - - - [42032, 8192, 1, 256] - - [268, 63.425] - - - [40448, 1280, 1, 256] - - [254, 71.25] - - - [42496, 1280, 1, 256] - - [252, 71.711] - - - [40240, 256, 1, 256] - - [252, 53.619] - - - [43264, 3328, 1, 256] - - [254, 72.954] - - - [42240, 1281, 1, 256] - - [268, 64.663] - - - [43008, 3328, 1, 256] - - [252, 73.593] - - - [43264, 8960, 1, 256] - - [256, 74.747] - - - [40448, 3328, 1, 256] - - [270, 73.139] - - - [40704, 6400, 1, 256] - - [252, 74.452] - - - [43312, 9472, 1, 256] - - [271, 62.742] - - - [43776, 5632, 1, 256] - - [268, 73.586] - - - [40448, 2816, 1, 256] - - [254, 73.248] - - - [42752, 2048, 1, 256] - - [301, 70.609] - - - [43008, 3329, 1, 256] - - [259, 70.532] - - - [40192, 2048, 1, 256] - - [301, 71.401] - - - [44288, 9984, 1, 256] - - [254, 74.699] - - - [41984, 2816, 1, 256] - - [252, 73.74] - - - [43776, 3329, 1, 256] - - [255, 68.689] - - - [43776, 9728, 1, 256] - - [255, 74.27] - - - [41472, 7424, 1, 256] - - [252, 74.806] - - - [43008, 6144, 1, 256] - - [254, 74.574] - - - [42496, 8192, 1, 256] - - [264, 74.717] - - - [44032, 3329, 1, 256] - - [281, 70.575] - - - [43264, 2048, 1, 256] - - [301, 71.216] - - - [40496, 2865, 1, 256] - - [271, 61.745] - - - [41008, 10240, 1, 256] - - [255, 62.391] - - - [43264, 1280, 1, 256] - - [256, 71.169] - - - [41216, 7680, 1, 256] - - [254, 74.657] - - - [44288, 2048, 1, 256] - - [255, 71.786] - - - [43264, 2865, 1, 256] - - [281, 70.362] - - - [41216, 6912, 1, 256] - - [264, 74.568] - - - [40192, 1281, 1, 256] - - [301, 64.585] - - - [41520, 10240, 1, 256] - - [301, 62.859] - - - [43264, 2816, 1, 256] - - [264, 72.884] - - - [43776, 1281, 1, 256] - - [288, 63.973] - - - [43264, 9216, 1, 256] - - [256, 74.059] - - - [42752, 3328, 1, 256] - - [264, 72.653] - - - [44288, 3329, 1, 256] - - [255, 69.201] - - - [40960, 10240, 1, 256] - - [265, 67.284] - - - [40496, 10240, 1, 256] - - [301, 62.619] - - - [43776, 6144, 1, 256] - - [255, 73.627] - - - [42288, 8704, 1, 256] - - [255, 62.698] - - - [41216, 2865, 1, 256] - - [281, 70.173] - - - [41984, 7936, 1, 256] - - [254, 75.243] - - - [41776, 256, 1, 256] - - [280, 55.02] - - - [41472, 7168, 1, 256] - - [252, 73.532] - - - [41728, 1280, 1, 256] - - [286, 69.927] - - - [40960, 2048, 1, 256] - - [254, 62.991] - - - [41472, 10240, 1, 256] - - [252, 74.705] - - - [42800, 8960, 1, 256] - - [280, 63.579] - - - [44032, 3328, 1, 256] - - [256, 73.632] - - - [41472, 2048, 1, 256] - - [275, 71.859] - - - [43776, 3328, 1, 256] - - [255, 72.192] - - - [40960, 3328, 1, 256] - - [265, 66.416] - - - [41984, 3329, 1, 256] - - [252, 70.588] - - - [42240, 1280, 1, 256] - - [249, 70.817] - - - [42752, 8448, 1, 256] - - [254, 74.38] - - - [44032, 1536, 1, 256] - - [252, 71.765] - - - [44288, 1792, 1, 256] - - [252, 71.649] - - - [43008, 8704, 1, 256] - - [252, 75.246] - - - [41728, 2048, 1, 256] - - [316, 69.934] - - - [42032, 256, 1, 256] - - [263, 55.382] - - - [43008, 9472, 1, 256] - - [252, 75.471] - - - [40704, 2865, 1, 256] - - [264, 70.238] - - - [40704, 2304, 1, 256] - - [254, 72.693] - - - [42240, 8704, 1, 256] - - [264, 74.597] - - - [43568, 9728, 1, 256] - - [268, 63.01] - - - [41472, 7936, 1, 256] - - [252, 74.77] - - - [41008, 2865, 1, 256] - - [255, 58.776] - - - [44032, 256, 1, 256] - - [263, 57.631] - - - [42032, 8448, 1, 256] - - [255, 63.624] - - - [44032, 1280, 1, 256] - - [254, 72.047] - - - [42800, 2865, 1, 256] - - [263, 61.011] - - - [41984, 8192, 1, 256] - - [252, 74.9] - - - [41728, 2865, 1, 256] - - [278, 68.836] - - - [41984, 3584, 1, 256] - - [264, 74.444] - - - [43520, 9472, 1, 256] - - [254, 75.262] - - - [41728, 1281, 1, 256] - - [318, 64.039] - - - [42288, 256, 1, 256] - - [252, 49.73] - - - [42752, 4352, 1, 256] - - [252, 74.11] - - - [41728, 3328, 1, 256] - - [268, 72.265] - - - [43008, 512, 1, 256] - - [271, 66.535] - - - [42800, 256, 1, 256] - - [280, 49.994] - - - [40192, 6400, 1, 256] - - [254, 74.557] - - - [42544, 8960, 1, 256] - - [271, 63.039] - - - [42288, 10240, 1, 256] - - [268, 62.646] - - - [42288, 8448, 1, 256] - - [271, 62.928] - - - [43008, 1281, 1, 256] - - [252, 64.667] - - - [41728, 256, 1, 256] - - [280, 61.757] - - - [42240, 3329, 1, 256] - - [259, 69.539] - - - [41216, 6144, 1, 256] - - [254, 74.096] - - - [42496, 8704, 1, 256] - - [256, 75.149] - - - [44032, 1281, 1, 256] - - [252, 64.951] - - - [41216, 7424, 1, 256] - - [252, 74.529] - - - [40960, 768, 1, 256] - - [256, 61.525] - - - [43264, 768, 1, 256] - - [285, 68.7] - - - [42800, 10240, 1, 256] - - [268, 62.897] - - - [40752, 6912, 1, 256] - - [263, 63.427] - - - [42288, 2816, 1, 256] - - [263, 62.835] - - - [42800, 2816, 1, 256] - - [285, 62.968] - - - [41728, 7680, 1, 256] - - [255, 73.673] - - - [41984, 1281, 1, 256] - - [301, 64.815] - - - [40240, 6400, 1, 256] - - [271, 63.269] - - - [42496, 4096, 1, 256] - - [268, 73.815] - - - [44288, 5888, 1, 256] - - [301, 73.788] - - - [43776, 9984, 1, 256] - - [268, 73.921] - - - [40960, 1281, 1, 256] - - [265, 57.818] - - - [40192, 2816, 1, 256] - - [254, 72.942] - - - [44032, 9984, 1, 256] - - [254, 75.513] - - - [42240, 3840, 1, 256] - - [252, 73.925] - - - [43520, 10240, 1, 256] - - [252, 74.931] - - - [40448, 2304, 1, 256] - - [256, 72.902] - - - [43520, 2816, 1, 256] - - [252, 73.456] - - - [41984, 256, 1, 256] - - [254, 61.95] - - - [43568, 2865, 1, 256] - - [280, 61.386] - - - [41472, 3328, 1, 256] - - [270, 73.35] - - - [40448, 2048, 1, 256] - - [255, 71.758] - - - [43520, 9984, 1, 256] - - [252, 75.254] - - - [42240, 2048, 1, 256] - - [268, 71.349] - - - [40192, 5888, 1, 256] - - [254, 73.973] - - - [41984, 10240, 1, 256] - - [252, 75.156] - - - [42752, 10240, 1, 256] - - [254, 74.606] - - - [41776, 10240, 1, 256] - - [268, 62.884] - - - [40960, 2560, 1, 256] - - [249, 66.523] - - - [41984, 1280, 1, 256] - - [254, 71.781] - - - [41776, 2816, 1, 256] - - [280, 63.494] - - - [43264, 4864, 1, 256] - - [256, 74.253] - - - [43520, 3328, 1, 256] - - [268, 73.577] - - - [40752, 7168, 1, 256] - - [255, 62.197] - - - [42032, 10240, 1, 256] - - [255, 63.661] - - - [40704, 10240, 1, 256] - - [254, 74.63] - - - [41520, 2865, 1, 256] - - [271, 61.695] - - - [40496, 6656, 1, 256] - - [271, 63.334] - - - [43568, 9984, 1, 256] - - [280, 63.212] - - - [43008, 768, 1, 256] - - [263, 69.577] - - - [43008, 2816, 1, 256] - - [249, 73.792] - - - [43520, 256, 1, 256] - - [280, 56.961] - - - [41984, 3840, 1, 256] - - [254, 74.662] - - - [43056, 2816, 1, 256] - - [268, 61.699] - - - [41776, 2865, 1, 256] - - [280, 61.177] - - - [41728, 6144, 1, 256] - - [255, 73.428] - - - [43568, 256, 1, 256] - - [285, 50.65] - - - [42544, 2816, 1, 256] - - [280, 63.276] - - - [41728, 2816, 1, 256] - - [318, 71.895] - - - [44032, 2048, 1, 256] - - [268, 71.938] - - - [40448, 256, 1, 256] - - [264, 60.316] - - - [43264, 6144, 1, 256] - - [252, 74.082] - - - [42752, 9216, 1, 256] - - [256, 74.194] - - - [40704, 512, 1, 256] - - [280, 66.176] - - - [44032, 9728, 1, 256] - - [254, 75.072] - - - [43264, 1024, 1, 256] - - [280, 69.973] - - - [40240, 2816, 1, 256] - - [285, 62.853] - - - [44288, 10240, 1, 256] - - [268, 74.537] - - - [41728, 7424, 1, 256] - - [255, 73.837] - - - [43264, 5120, 1, 256] - - [254, 74.085] - - - [11776, 6144, 1, 256] - - [264, 73.095] - - - [4352, 2865, 1, 256] - - [252, 59.825] - - - [4096, 2865, 1, 256] - - [286, 61.805] - - - [12544, 3328, 1, 256] - - [262, 70.717] - - - [16640, 3329, 1, 256] - - [252, 67.757] - - - [7424, 5888, 1, 256] - - [254, 71.721] - - - [15664, 2865, 1, 256] - - [252, 61.915] - - - [7680, 4096, 1, 256] - - [254, 70.334] - - - [11520, 256, 1, 256] - - [252, 47.463] - - - [10544, 2865, 1, 256] - - [254, 59.435] - - - [3072, 3072, 1, 256] - - [282, 59.056] - - - [8448, 3328, 1, 256] - - [261, 69.62] - - - [19200, 5376, 1, 256] - - [252, 73.642] - - - [12032, 3329, 1, 256] - - [262, 67.426] - - - [11520, 1281, 1, 256] - - [287, 60.407] - - - [15360, 1281, 1, 256] - - [252, 61.715] - - - [9216, 6400, 1, 256] - - [252, 73.86] - - - [5632, 2816, 1, 256] - - [282, 67.037] - - - [9984, 1280, 1, 256] - - [261, 61.779] - - - [14128, 256, 1, 256] - - [262, 38.254] - - - [18224, 5120, 1, 256] - - [285, 63.714] - - - [8192, 5376, 1, 256] - - [262, 71.577] - - - [11264, 3584, 1, 256] - - [276, 72.09] - - - [6144, 2560, 1, 256] - - [262, 67.197] - - - [8960, 5376, 1, 256] - - [284, 72.299] - - - [9984, 7168, 1, 256] - - [261, 72.618] - - - [2352, 2304, 1, 256] - - [263, 46.217] - - - [6656, 3840, 1, 256] - - [276, 69.201] - - - [8496, 5376, 1, 256] - - [256, 64.353] - - - [13312, 2865, 1, 256] - - [252, 69.024] - - - [11520, 6144, 1, 256] - - [286, 72.709] - - - [7936, 3328, 1, 256] - - [290, 69.124] - - - [19968, 6656, 1, 256] - - [264, 74.609] - - - [18432, 4608, 1, 256] - - [262, 73.713] - - - [19712, 1281, 1, 256] - - [313, 59.967] - - - [8192, 2865, 1, 256] - - [262, 65.783] - - - [14336, 2816, 1, 256] - - [254, 71.657] - - - [12544, 8960, 1, 256] - - [254, 74.392] - - - [19760, 2816, 1, 256] - - [285, 63.795] - - - [11824, 2865, 1, 256] - - [254, 60.194] - - - [12288, 8704, 1, 256] - - [264, 75.086] - - - [15920, 2816, 1, 256] - - [280, 64.786] - - - [14080, 2048, 1, 256] - - [319, 67.032] - - - [7216, 2865, 1, 256] - - [254, 60.212] - - - [5376, 3840, 1, 256] - - [261, 67.618] - - - [19200, 5632, 1, 256] - - [254, 73.869] - - - [12800, 3329, 1, 256] - - [294, 68.406] - - - [12800, 5376, 1, 256] - - [284, 73.266] - - - [19968, 1281, 1, 256] - - [285, 63.204] - - - [11264, 3329, 1, 256] - - [256, 68.433] - - - [6656, 5376, 1, 256] - - [286, 71.81] - - - [11568, 2865, 1, 256] - - [252, 58.964] - - - [19968, 2048, 1, 256] - - [285, 70.791] - - - [7728, 256, 1, 256] - - [280, 32.336] - - - [5376, 2048, 1, 256] - - [262, 58.901] - - - [7216, 2816, 1, 256] - - [252, 60.371] - - - [4656, 1792, 1, 256] - - [254, 54.518] - - - [19456, 1280, 1, 256] - - [282, 69.213] - - - [17920, 6144, 1, 256] - - [256, 74.033] - - - [19456, 6400, 1, 256] - - [264, 75.291] - - - [15664, 10240, 1, 256] - - [252, 63.248] - - - [18224, 2865, 1, 256] - - [264, 61.253] - - - [12288, 256, 1, 256] - - [276, 49.522] - - - [17712, 4352, 1, 256] - - [285, 63.205] - - - [8704, 5376, 1, 256] - - [276, 72.78] - - - [4352, 3329, 1, 256] - - [276, 61.736] - - - [15872, 2865, 1, 256] - - [252, 69.974] - - - [11520, 2048, 1, 256] - - [290, 66.369] - - - [5632, 1281, 1, 256] - - [261, 51.805] - - - [4912, 2865, 1, 256] - - [264, 54.733] - - - [7472, 2816, 1, 256] - - [280, 61.407] - - - [10752, 7936, 1, 256] - - [252, 74.542] - - - [11776, 1280, 1, 256] - - [294, 64.065] - - - [17408, 1536, 1, 256] - - [276, 68.826] - - - [16688, 3328, 1, 256] - - [271, 65.797] - - - [2352, 2097, 1, 256] - - [264, 49.148] - - - [10240, 6912, 1, 256] - - [276, 74.534] - - - [16640, 1024, 1, 256] - - [287, 63.0] - - - [7936, 1280, 1, 256] - - [261, 62.118] - - - [3328, 2048, 1, 256] - - [262, 58.869] - - - [16944, 3840, 1, 256] - - [280, 64.63] - - - [12288, 768, 1, 256] - - [263, 57.894] - - - [6192, 2865, 1, 256] - - [252, 56.627] - - - [11824, 256, 1, 256] - - [254, 45.599] - - - [11520, 7936, 1, 256] - - [286, 73.686] - - - [16944, 10240, 1, 256] - - [252, 64.198] - - - [9984, 6656, 1, 256] - - [262, 73.374] - - - [6912, 3329, 1, 256] - - [262, 65.222] - - - [7936, 4864, 1, 256] - - [258, 71.543] - - - [9728, 6144, 1, 256] - - [254, 72.968] - - - [13056, 5632, 1, 256] - - [264, 73.854] - - - [4096, 512, 1, 256] - - [286, 34.82] - - - [11264, 2048, 1, 256] - - [285, 67.528] - - - [8960, 3329, 1, 256] - - [282, 66.204] - - - [12848, 2816, 1, 256] - - [263, 61.496] - - - [2816, 2816, 1, 256] - - [282, 56.977] - - - [17408, 1792, 1, 256] - - [282, 70.156] - - - [3584, 2865, 1, 256] - - [262, 62.625] - - - [2096, 2048, 1, 256] - - [262, 44.187] - - - [11520, 1280, 1, 256] - - [276, 63.231] - - - [7984, 4864, 1, 256] - - [252, 64.026] - - - [18432, 6144, 1, 256] - - [252, 74.435] - - - [8448, 1024, 1, 256] - - [250, 59.313] - - - [1536, 1536, 1, 256] - - [261, 38.782] - - - [18944, 3328, 1, 256] - - [258, 72.475] - - - [8960, 6144, 1, 256] - - [262, 72.367] - - - [8704, 5120, 1, 256] - - [262, 72.429] - - - [19456, 1281, 1, 256] - - [288, 62.089] - - - [2352, 2353, 1, 256] - - [263, 46.546] - - - [13312, 2816, 1, 256] - - [261, 71.629] - - - [2048, 1792, 1, 256] - - [276, 41.156] - - - [16176, 10240, 1, 256] - - [254, 65.472] - - - [8704, 2816, 1, 256] - - [261, 69.527] - - - [6656, 1024, 1, 256] - - [254, 58.781] - - - [7680, 2816, 1, 256] - - [252, 69.69] - - - [10752, 1280, 1, 256] - - [261, 64.9] - - - [15872, 1281, 1, 256] - - [250, 62.2] - - - [6912, 3584, 1, 256] - - [261, 69.944] - - - [11264, 1281, 1, 256] - - [280, 60.24] - - - [17200, 2816, 1, 256] - - [285, 63.832] - - - [11520, 2865, 1, 256] - - [282, 68.192] - - - [16176, 256, 1, 256] - - [287, 43.839] - - - [4096, 768, 1, 256] - - [284, 49.726] - - - [13568, 256, 1, 256] - - [267, 53.158] - - - [9984, 6144, 1, 256] - - [262, 72.711] - - - [19200, 5888, 1, 256] - - [261, 73.564] - - - [15360, 1280, 1, 256] - - [276, 67.381] - - - [7168, 3328, 1, 256] - - [276, 68.647] - - - [18176, 3328, 1, 256] - - [276, 72.027] - - - [12544, 768, 1, 256] - - [280, 59.132] - - - [8704, 1280, 1, 256] - - [261, 60.386] - - - [2304, 1024, 1, 256] - - [284, 38.726] - - - [6912, 1280, 1, 256] - - [262, 61.687] - - - [17152, 6144, 1, 256] - - [254, 73.807] - - - [10240, 2560, 1, 256] - - [252, 70.12] - - - [7680, 6144, 1, 256] - - [261, 72.08] - - - [17456, 10240, 1, 256] - - [280, 63.79] - - - [13360, 10240, 1, 256] - - [254, 64.032] - - - [17408, 3584, 1, 256] - - [264, 73.531] - - - [12592, 9472, 1, 256] - - [263, 64.017] - - - [5376, 2865, 1, 256] - - [282, 63.726] - - - [14336, 2048, 1, 256] - - [288, 68.004] - - - [12288, 1280, 1, 256] - - [254, 65.855] - - - [15360, 1792, 1, 256] - - [276, 69.304] - - - [10240, 3328, 1, 256] - - [252, 71.06] - - - [11008, 6144, 1, 256] - - [302, 72.223] - - - [16688, 3584, 1, 256] - - [256, 65.642] - - - [18688, 1281, 1, 256] - - [267, 62.826] - - - [10496, 2816, 1, 256] - - [290, 69.463] - - - [9216, 2865, 1, 256] - - [254, 66.685] - - - [3632, 512, 1, 256] - - [262, 29.845] - - - [19456, 1792, 1, 256] - - [282, 70.657] - - - [10752, 3072, 1, 256] - - [254, 70.504] - - - [6192, 3072, 1, 256] - - [252, 60.601] - - - [2096, 2097, 1, 256] - - [252, 44.726] - - - [13056, 6144, 1, 256] - - [254, 73.792] - - - [12592, 9728, 1, 256] - - [285, 63.963] - - - [18176, 4864, 1, 256] - - [252, 73.781] - - - [12544, 256, 1, 256] - - [254, 50.485] - - - [17920, 2816, 1, 256] - - [291, 71.961] - - - [15872, 1280, 1, 256] - - [284, 66.943] - - - [18176, 4608, 1, 256] - - [276, 73.101] - - - [13616, 256, 1, 256] - - [263, 50.658] - - - [19712, 3328, 1, 256] - - [294, 70.314] - - - [11008, 3328, 1, 256] - - [302, 69.176] - - - [6656, 2865, 1, 256] - - [286, 65.529] - - - [6960, 4096, 1, 256] - - [271, 62.828] - - - [16128, 2560, 1, 256] - - [286, 71.226] - - - [6656, 1281, 1, 256] - - [319, 53.011] - - - [11520, 1792, 1, 256] - - [261, 67.255] - - - [10752, 2048, 1, 256] - - [267, 67.186] - - - [4352, 768, 1, 256] - - [250, 51.77] - - - [18992, 256, 1, 256] - - [267, 47.829] - - - [14640, 256, 1, 256] - - [276, 39.97] - - - [16896, 3329, 1, 256] - - [252, 69.544] - - - [14080, 10240, 1, 256] - - [251, 73.542] - - - [6912, 5632, 1, 256] - - [276, 71.919] - - - [6400, 3072, 1, 256] - - [262, 68.625] - - - [9264, 2816, 1, 256] - - [264, 61.26] - - - [11264, 1280, 1, 256] - - [262, 66.843] - - - [18736, 5376, 1, 256] - - [285, 62.844] - - - [6144, 4608, 1, 256] - - [252, 70.913] - - - [12032, 2865, 1, 256] - - [261, 68.265] - - - [16128, 3072, 1, 256] - - [286, 71.509] - - - [2816, 2561, 1, 256] - - [261, 52.938] - - - [2304, 1281, 1, 256] - - [254, 46.337] - - - [12800, 1024, 1, 256] - - [250, 62.114] - - - [18432, 1280, 1, 256] - - [254, 67.656] - - - [19712, 6400, 1, 256] - - [251, 73.172] - - - [6448, 2816, 1, 256] - - [263, 58.008] - - - [3072, 1792, 1, 256] - - [254, 49.282] - - - [13312, 1281, 1, 256] - - [252, 60.158] - - - [2048, 2048, 1, 256] - - [262, 46.271] - - - [11520, 2816, 1, 256] - - [294, 70.272] - - - [7472, 256, 1, 256] - - [271, 31.265] - - - [9520, 2865, 1, 256] - - [271, 60.006] - - - [18176, 256, 1, 256] - - [267, 49.441] - - - [13568, 5888, 1, 256] - - [276, 73.35] - - - [6656, 5120, 1, 256] - - [276, 71.447] - - - [1536, 1537, 1, 256] - - [261, 38.752] - - - [10032, 2816, 1, 256] - - [280, 62.235] - - - [13872, 10240, 1, 256] - - [271, 63.139] - - - [12544, 9728, 1, 256] - - [264, 74.151] - - - [4096, 1281, 1, 256] - - [276, 47.228] - - - [18432, 2816, 1, 256] - - [282, 72.467] - - - [19200, 256, 1, 256] - - [261, 51.609] - - - [5888, 3072, 1, 256] - - [262, 65.822] - - - [5888, 512, 1, 256] - - [254, 48.316] - - - [5120, 2865, 1, 256] - - [276, 62.345] - - - [13104, 2865, 1, 256] - - [271, 60.974] - - - [15408, 10240, 1, 256] - - [256, 63.988] - - - [17712, 2865, 1, 256] - - [254, 60.28] - - - [12848, 9728, 1, 256] - - [280, 64.542] - - - [10240, 6144, 1, 256] - - [256, 73.496] - - - [9008, 5888, 1, 256] - - [263, 63.331] - - - [12032, 1280, 1, 256] - - [276, 65.217] - - - [18432, 5120, 1, 256] - - [252, 74.354] - - - [9776, 2865, 1, 256] - - [263, 58.455] - - - [7680, 2048, 1, 256] - - [267, 65.594] - - - [13872, 2816, 1, 256] - - [263, 63.15] - - - [4144, 2865, 1, 256] - - [256, 58.002] - - - [16432, 2816, 1, 256] - - [263, 67.138] - - - [8240, 5120, 1, 256] - - [254, 65.285] - - - [12544, 2865, 1, 256] - - [254, 68.603] - - - [13616, 512, 1, 256] - - [267, 55.1] - - - [13312, 2048, 1, 256] - - [250, 67.86] - - - [7424, 1281, 1, 256] - - [262, 58.342] - - - [2048, 1281, 1, 256] - - [284, 42.578] - - - [15664, 2816, 1, 256] - - [263, 63.151] - - - [9776, 256, 1, 256] - - [280, 39.723] - - - [9728, 2304, 1, 256] - - [282, 68.567] - - - [13568, 6144, 1, 256] - - [256, 73.621] - - - [9264, 6400, 1, 256] - - [254, 63.431] - - - [11520, 4096, 1, 256] - - [267, 70.668] - - - [11776, 256, 1, 256] - - [263, 48.248] - - - [5120, 1280, 1, 256] - - [261, 57.075] - - - [10288, 2816, 1, 256] - - [256, 60.366] - - - [4864, 1281, 1, 256] - - [262, 54.304] - - - [5888, 3328, 1, 256] - - [262, 68.189] - - - [15360, 3328, 1, 256] - - [262, 72.12] - - - [9776, 2816, 1, 256] - - [280, 61.887] - - - [6656, 3072, 1, 256] - - [261, 66.992] - - - [13824, 2865, 1, 256] - - [262, 69.183] - - - [13568, 9984, 1, 256] - - [264, 74.874] - - - [11776, 1281, 1, 256] - - [271, 58.768] - - - [18992, 2816, 1, 256] - - [263, 63.728] - - - [5120, 2048, 1, 256] - - [254, 61.987] - - - [19200, 2865, 1, 256] - - [286, 69.04] - - - [2560, 2560, 1, 256] - - [261, 57.421] - - - [8704, 5632, 1, 256] - - [282, 72.624] - - - [18688, 3328, 1, 256] - - [262, 71.918] - - - [13104, 2816, 1, 256] - - [263, 63.036] - - - [14592, 2816, 1, 256] - - [286, 70.771] - - - [17920, 256, 1, 256] - - [290, 49.106] - - - [18688, 768, 1, 256] - - [280, 66.3] - - - [14896, 10240, 1, 256] - - [252, 63.546] - - - [14848, 1024, 1, 256] - - [287, 64.645] - - - [7424, 256, 1, 256] - - [258, 31.555] - - - [18176, 5120, 1, 256] - - [254, 73.841] - - - [5632, 4096, 1, 256] - - [267, 68.981] - - - [2864, 2609, 1, 256] - - [254, 49.44] - - - [3840, 1280, 1, 256] - - [276, 52.859] - - - [15104, 1281, 1, 256] - - [250, 62.75] - - - [12800, 5120, 1, 256] - - [286, 73.081] - - - [9728, 6912, 1, 256] - - [276, 74.008] - - - [2560, 2304, 1, 256] - - [282, 52.115] - - - [5888, 3329, 1, 256] - - [252, 64.312] - - - [18944, 10240, 1, 256] - - [274, 75.238] - - - [4144, 1280, 1, 256] - - [262, 53.132] - - - [14896, 1536, 1, 256] - - [252, 60.265] - - - [4912, 2816, 1, 256] - - [264, 59.825] - - - [7680, 4608, 1, 256] - - [282, 71.244] - - - [3584, 2048, 1, 256] - - [261, 53.777] - - - [10800, 7680, 1, 256] - - [271, 64.193] - - - [15104, 10240, 1, 256] - - [254, 74.662] - - - [16128, 1281, 1, 256] - - [296, 60.91] - - - [12544, 9472, 1, 256] - - [252, 74.527] - - - [9472, 3329, 1, 256] - - [264, 66.568] - - - [2048, 1841, 1, 256] - - [267, 41.594] - - - [12544, 5120, 1, 256] - - [276, 72.708] - - - [6912, 2048, 1, 256] - - [267, 65.866] - - - [9216, 1281, 1, 256] - - [263, 56.597] - - - [10752, 7424, 1, 256] - - [256, 74.342] - - - [12288, 9216, 1, 256] - - [256, 74.52] - - - [14336, 3328, 1, 256] - - [252, 72.042] - - - [19760, 2865, 1, 256] - - [263, 60.832] - - - [18688, 1024, 1, 256] - - [267, 66.573] - - - [5680, 2560, 1, 256] - - [264, 57.404] - - - [17408, 2048, 1, 256] - - [285, 69.586] - - - [14848, 1281, 1, 256] - - [280, 62.3] - - - [16384, 10240, 1, 256] - - [249, 63.0] - - - [8448, 5120, 1, 256] - - [252, 72.211] - - - [7984, 2816, 1, 256] - - [254, 61.507] - - - [16688, 2865, 1, 256] - - [276, 63.796] - - - [7424, 2865, 1, 256] - - [254, 65.069] - - - [17152, 2048, 1, 256] - - [290, 69.707] - - - [16384, 2048, 1, 256] - - [265, 59.349] - - - [18688, 256, 1, 256] - - [262, 50.462] - - - [19968, 6400, 1, 256] - - [261, 74.803] - - - [19456, 3328, 1, 256] - - [276, 72.717] - - - [15920, 2865, 1, 256] - - [254, 61.767] - - - [19200, 3328, 1, 256] - - [280, 71.858] - - - [7936, 3329, 1, 256] - - [284, 65.631] - - - [6144, 3329, 1, 256] - - [276, 66.13] - - - [12800, 9728, 1, 256] - - [301, 74.84] - - - [12544, 3329, 1, 256] - - [254, 67.735] - - - [7728, 2816, 1, 256] - - [250, 59.059] - - - [19968, 6144, 1, 256] - - [268, 74.294] - - - [1536, 1329, 1, 256] - - [296, 33.699] - - - [9472, 5888, 1, 256] - - [256, 72.702] - - - [8496, 256, 1, 256] - - [285, 35.102] - - - [11824, 2816, 1, 256] - - [271, 62.525] - - - [19968, 6912, 1, 256] - - [254, 74.983] - - - [3840, 2560, 1, 256] - - [276, 60.629] - - - [19248, 6144, 1, 256] - - [285, 63.548] - - - [16128, 3328, 1, 256] - - [276, 71.602] - - - [15664, 256, 1, 256] - - [282, 42.258] - - - [16384, 6144, 1, 256] - - [249, 63.174] - - - [14592, 1281, 1, 256] - - [261, 60.083] - - - [8960, 2048, 1, 256] - - [261, 64.636] - - - [2816, 2609, 1, 256] - - [261, 53.455] - - - [1792, 1536, 1, 256] - - [258, 44.237] - - - [5936, 2816, 1, 256] - - [264, 58.611] - - - [18944, 5888, 1, 256] - - [258, 74.241] - - - [19248, 2816, 1, 256] - - [280, 62.64] - - - [1584, 1585, 1, 256] - - [256, 38.322] - - - [13056, 9728, 1, 256] - - [252, 74.687] - - - [14336, 6144, 1, 256] - - [254, 74.178] - - - [18432, 1281, 1, 256] - - [252, 61.726] - - - [12544, 9216, 1, 256] - - [264, 73.968] - - - [8704, 5888, 1, 256] - - [252, 72.741] - - - [11312, 8448, 1, 256] - - [252, 63.968] - - - [16384, 3328, 1, 256] - - [265, 61.752] - - - [16432, 256, 1, 256] - - [250, 44.904] - - - [3584, 3328, 1, 256] - - [262, 64.076] - - - [19504, 256, 1, 256] - - [287, 48.745] - - - [2816, 1536, 1, 256] - - [262, 47.445] - - - [4656, 2816, 1, 256] - - [263, 57.374] - - - [9984, 2816, 1, 256] - - [262, 69.679] - - - [18432, 3328, 1, 256] - - [256, 72.629] - - - [4864, 3329, 1, 256] - - [262, 63.146] - - - [17152, 2816, 1, 256] - - [261, 71.888] - - - [1584, 1536, 1, 256] - - [261, 37.94] - - - [13824, 512, 1, 256] - - [254, 60.187] - - - [12592, 2865, 1, 256] - - [264, 60.455] - - - [6912, 1536, 1, 256] - - [282, 64.344] - - - [10496, 256, 1, 256] - - [267, 43.732] - - - [4608, 1536, 1, 256] - - [262, 60.86] - - - [3840, 1024, 1, 256] - - [262, 43.255] - - - [3840, 1281, 1, 256] - - [276, 52.315] - - - [13568, 1281, 1, 256] - - [267, 61.146] - - - [13056, 9472, 1, 256] - - [252, 74.981] - - - [13056, 2816, 1, 256] - - [262, 71.111] - - - [7936, 4352, 1, 256] - - [284, 70.426] - - - [5424, 2304, 1, 256] - - [280, 56.662] - - - [10752, 1024, 1, 256] - - [267, 58.928] - - - [4352, 3328, 1, 256] - - [262, 63.523] - - - [12032, 8704, 1, 256] - - [276, 74.067] - - - [6400, 3329, 1, 256] - - [262, 64.596] - - - [7472, 2865, 1, 256] - - [256, 57.811] - - - [18736, 10240, 1, 256] - - [285, 63.377] - - - [7936, 5120, 1, 256] - - [264, 71.824] - - - [3072, 1280, 1, 256] - - [262, 43.671] - - - [9472, 6144, 1, 256] - - [282, 72.458] - - - [17408, 2865, 1, 256] - - [252, 69.641] - - - [16384, 3329, 1, 256] - - [256, 58.92] - - - [18688, 2048, 1, 256] - - [290, 69.803] - - - [9472, 6400, 1, 256] - - [261, 73.356] - - - [4608, 1024, 1, 256] - - [282, 50.887] - - - [10752, 3329, 1, 256] - - [254, 68.157] - - - [9984, 2048, 1, 256] - - [290, 66.727] - - - [5120, 3584, 1, 256] - - [261, 66.629] - - - [18736, 256, 1, 256] - - [280, 47.713] - - - [17456, 4352, 1, 256] - - [252, 64.048] - - - [17664, 4608, 1, 256] - - [286, 72.479] - - - [6400, 3584, 1, 256] - - [276, 69.343] - - - [10752, 2816, 1, 256] - - [286, 70.312] - - - [8192, 1281, 1, 256] - - [254, 55.243] - - - [3632, 2816, 1, 256] - - [271, 57.042] - - - [5168, 2865, 1, 256] - - [254, 57.066] - - - [12544, 2816, 1, 256] - - [254, 70.656] - - - [11776, 2048, 1, 256] - - [266, 67.352] - - - [14336, 768, 1, 256] - - [285, 59.929] - - - [11776, 2816, 1, 256] - - [258, 70.084] - - - [5120, 1792, 1, 256] - - [254, 57.697] - - - [5376, 3329, 1, 256] - - [262, 63.193] - - - [9008, 2816, 1, 256] - - [252, 59.908] - - - [19968, 3329, 1, 256] - - [254, 69.474] - - - [13824, 256, 1, 256] - - [262, 54.449] - - - [5376, 2304, 1, 256] - - [261, 64.441] - - - [18432, 2048, 1, 256] - - [305, 68.684] - - - [8704, 3328, 1, 256] - - [261, 69.305] - - - [6192, 3328, 1, 256] - - [254, 60.553] - - - [5424, 2560, 1, 256] - - [250, 61.133] - - - [11776, 4352, 1, 256] - - [286, 72.393] - - - [13312, 3328, 1, 256] - - [254, 71.901] - - - [9216, 256, 1, 256] - - [267, 38.837] - - - [5936, 2865, 1, 256] - - [252, 59.11] - - - [19968, 2816, 1, 256] - - [276, 72.336] - - - [9728, 256, 1, 256] - - [252, 40.936] - - - [7680, 3329, 1, 256] - - [264, 67.112] - - - [11776, 2865, 1, 256] - - [258, 67.578] - - - [4352, 2048, 1, 256] - - [252, 61.808] - - - [12080, 2865, 1, 256] - - [254, 60.842] - - - [7936, 512, 1, 256] - - [267, 44.612] - - - [5632, 3329, 1, 256] - - [254, 65.402] - - - [6192, 2816, 1, 256] - - [254, 60.678] - - - [19712, 6656, 1, 256] - - [286, 73.027] - - - [6400, 1280, 1, 256] - - [261, 58.678] - - - [15616, 256, 1, 256] - - [250, 43.934] - - - [7936, 2865, 1, 256] - - [258, 66.845] - - - [5168, 2304, 1, 256] - - [252, 57.61] - - - [3840, 2816, 1, 256] - - [261, 65.639] - - - [3328, 3073, 1, 256] - - [254, 61.561] - - - [14592, 2048, 1, 256] - - [266, 67.195] - - - [13312, 5888, 1, 256] - - [282, 74.216] - - - [19712, 2816, 1, 256] - - [278, 70.242] - - - [7168, 4352, 1, 256] - - [254, 71.0] - - - [13568, 3328, 1, 256] - - [261, 70.67] - - - [11056, 2865, 1, 256] - - [256, 59.852] - - - [13824, 768, 1, 256] - - [267, 63.874] - - - [7216, 4352, 1, 256] - - [254, 62.523] - - - [11520, 3328, 1, 256] - - [263, 69.856] - - - [17408, 10240, 1, 256] - - [264, 75.574] - - - [16640, 1280, 1, 256] - - [287, 65.558] - - - [11776, 8960, 1, 256] - - [276, 74.797] - - - [11264, 3328, 1, 256] - - [282, 71.228] - - - [14848, 6144, 1, 256] - - [264, 73.826] - - - [2816, 2817, 1, 256] - - [262, 56.435] - - - [18176, 3329, 1, 256] - - [252, 69.35] - - - [6144, 3328, 1, 256] - - [262, 67.918] - - - [8960, 5888, 1, 256] - - [284, 71.764] - - - [7728, 4608, 1, 256] - - [280, 63.336] - - - [16944, 256, 1, 256] - - [261, 45.256] - - - [8448, 3329, 1, 256] - - [256, 66.474] - - - [5936, 3072, 1, 256] - - [254, 58.598] - - - [19456, 10240, 1, 256] - - [252, 75.579] - - - [18992, 5632, 1, 256] - - [271, 64.215] - - - [8704, 6144, 1, 256] - - [261, 72.779] - - - [3120, 2865, 1, 256] - - [254, 51.19] - - - [12080, 2816, 1, 256] - - [280, 62.716] - - - [15920, 10240, 1, 256] - - [280, 64.072] - - - [11568, 2816, 1, 256] - - [263, 61.211] - - - [13312, 5632, 1, 256] - - [264, 74.331] - - - [12288, 2865, 1, 256] - - [259, 68.287] - - - [19968, 256, 1, 256] - - [271, 53.43] - - - [11008, 3584, 1, 256] - - [302, 70.462] - - - [11264, 6144, 1, 256] - - [256, 73.603] - - - [3840, 2048, 1, 256] - - [254, 56.612] - - - [19200, 1536, 1, 256] - - [284, 68.062] - - - [12544, 1281, 1, 256] - - [267, 61.48] - - - [7680, 2865, 1, 256] - - [252, 66.362] - - - [15616, 3328, 1, 256] - - [252, 71.606] - - - [2560, 1024, 1, 256] - - [261, 42.786] - - - [6912, 3328, 1, 256] - - [282, 68.767] - - - [17408, 3328, 1, 256] - - [252, 72.343] - - - [15104, 1792, 1, 256] - - [262, 68.538] - - - [12032, 4608, 1, 256] - - [262, 72.088] - - - [9216, 1792, 1, 256] - - [254, 65.175] - - - [8448, 1280, 1, 256] - - [282, 64.286] - - - [2608, 2560, 1, 256] - - [254, 53.501] - - - [13056, 3328, 1, 256] - - [262, 70.928] - - - [14896, 2865, 1, 256] - - [254, 60.701] - - - [15872, 3329, 1, 256] - - [252, 69.435] - - - [4656, 2865, 1, 256] - - [264, 57.651] - - - [16432, 3072, 1, 256] - - [265, 65.493] - - - [4864, 1536, 1, 256] - - [262, 53.95] - - - [10496, 2048, 1, 256] - - [287, 68.578] - - - [18432, 256, 1, 256] - - [267, 50.137] - - - [10240, 512, 1, 256] - - [267, 54.602] - - - [10496, 3329, 1, 256] - - [264, 66.755] - - - [17920, 1281, 1, 256] - - [260, 63.538] - - - [10496, 768, 1, 256] - - [280, 56.926] - - - [12288, 2048, 1, 256] - - [264, 66.331] - - - [10496, 1024, 1, 256] - - [267, 64.161] - - - [5168, 2048, 1, 256] - - [252, 59.001] - - - [15152, 10240, 1, 256] - - [280, 63.573] - - - [12288, 512, 1, 256] - - [311, 54.504] - - - [6912, 4096, 1, 256] - - [252, 69.673] - - - [4096, 1024, 1, 256] - - [276, 45.746] - - - [16896, 1281, 1, 256] - - [280, 61.455] - - - [14592, 6144, 1, 256] - - [294, 72.431] - - - [8448, 2865, 1, 256] - - [262, 67.332] - - - [12032, 512, 1, 256] - - [250, 53.772] - - - [15360, 1536, 1, 256] - - [262, 67.238] - - - [13104, 256, 1, 256] - - [290, 49.375] - - - [19248, 2865, 1, 256] - - [285, 60.917] - - - [8192, 2048, 1, 256] - - [264, 64.148] - - - [7168, 2048, 1, 256] - - [250, 63.091] - - - [17920, 4352, 1, 256] - - [284, 73.831] - - - [19200, 1281, 1, 256] - - [287, 63.126] - - - [9984, 6400, 1, 256] - - [262, 73.374] - - - [3840, 2865, 1, 256] - - [262, 59.129] - - - [15616, 2816, 1, 256] - - [254, 71.289] - - - [14592, 768, 1, 256] - - [267, 59.758] - - - [11264, 1536, 1, 256] - - [276, 66.394] - - - [17664, 3328, 1, 256] - - [258, 71.465] - - - [15408, 2865, 1, 256] - - [252, 61.565] - - - [14128, 1024, 1, 256] - - [271, 55.462] - - - [17664, 2048, 1, 256] - - [280, 68.715] - - - [2304, 2097, 1, 256] - - [276, 51.622] - - - [9472, 1792, 1, 256] - - [252, 65.661] - - - [19200, 1280, 1, 256] - - [284, 68.368] - - - [16432, 3328, 1, 256] - - [255, 66.565] - - - [8752, 5632, 1, 256] - - [256, 63.332] - - - [9472, 1280, 1, 256] - - [254, 63.42] - - - [6656, 3329, 1, 256] - - [286, 65.83] - - - [5632, 2304, 1, 256] - - [276, 62.835] - - - [15872, 3328, 1, 256] - - [261, 72.279] - - - [13312, 9728, 1, 256] - - [252, 75.464] - - - [15664, 2560, 1, 256] - - [264, 62.896] - - - [16128, 2865, 1, 256] - - [254, 68.888] - - - [18224, 2816, 1, 256] - - [271, 62.778] - - - [1792, 1281, 1, 256] - - [276, 37.52] - - - [16640, 1281, 1, 256] - - [292, 60.184] - - - [17968, 256, 1, 256] - - [252, 46.195] - - - [15152, 2048, 1, 256] - - [285, 60.877] - - - [12800, 2816, 1, 256] - - [258, 70.881] - - - [17968, 4608, 1, 256] - - [271, 63.285] - - - [7472, 4352, 1, 256] - - [252, 61.464] - - - [4400, 2816, 1, 256] - - [254, 59.033] - - - [18944, 2816, 1, 256] - - [294, 72.175] - - - [8192, 768, 1, 256] - - [261, 52.953] - - - [12288, 1281, 1, 256] - - [254, 60.225] - - - [5888, 1280, 1, 256] - - [276, 54.731] - - - [19200, 3329, 1, 256] - - [294, 68.61] - - - [15360, 2865, 1, 256] - - [254, 69.192] - - - [17456, 2816, 1, 256] - - [256, 63.271] - - - [6144, 2816, 1, 256] - - [276, 67.471] - - - [16384, 1281, 1, 256] - - [254, 53.321] - - - [14080, 2816, 1, 256] - - [294, 70.205] - - - [15872, 10240, 1, 256] - - [252, 75.281] - - - [9984, 1281, 1, 256] - - [267, 60.139] - - - [11776, 4096, 1, 256] - - [301, 71.643] - - - [5376, 2816, 1, 256] - - [282, 64.272] - - - [4608, 2816, 1, 256] - - [282, 63.153] - - - [14080, 256, 1, 256] - - [280, 54.728] - - - [2816, 1280, 1, 256] - - [276, 54.728] - - - [12544, 4864, 1, 256] - - [282, 72.899] - - - [4912, 2048, 1, 256] - - [271, 56.406] - - - [19248, 5888, 1, 256] - - [263, 63.654] - - - [9728, 2816, 1, 256] - - [261, 69.216] - - - [16896, 3072, 1, 256] - - [261, 72.467] - - - [2608, 2609, 1, 256] - - [252, 46.653] - - - [17664, 2816, 1, 256] - - [258, 71.206] - - - [13312, 1536, 1, 256] - - [276, 66.916] - - - [4608, 3328, 1, 256] - - [282, 65.809] - - - [6912, 5376, 1, 256] - - [276, 71.656] - - - [4352, 3072, 1, 256] - - [261, 63.961] - - - [9216, 5632, 1, 256] - - [256, 73.616] - - - [13824, 10240, 1, 256] - - [278, 74.807] - - - [19712, 1792, 1, 256] - - [284, 68.68] - - - [13056, 2048, 1, 256] - - [287, 68.872] - - - [16176, 2816, 1, 256] - - [252, 65.223] - - - [10496, 2865, 1, 256] - - [252, 66.963] - - - [5888, 4608, 1, 256] - - [282, 68.523] - - - [19456, 5888, 1, 256] - - [254, 74.612] - - - [19456, 256, 1, 256] - - [267, 52.392] - - - [14384, 1024, 1, 256] - - [252, 56.205] - - - [3840, 3329, 1, 256] - - [262, 61.31] - - - [10240, 2816, 1, 256] - - [254, 71.023] - - - [11008, 2865, 1, 256] - - [312, 67.12] - - - [10240, 7168, 1, 256] - - [252, 73.574] - - - [8192, 4864, 1, 256] - - [252, 70.814] - - - [6448, 3584, 1, 256] - - [280, 61.758] - - - [9728, 2865, 1, 256] - - [254, 68.808] - - - [14848, 256, 1, 256] - - [287, 42.175] - - - [15360, 10240, 1, 256] - - [256, 75.504] - - - [3328, 3072, 1, 256] - - [261, 62.752] - - - [19760, 6656, 1, 256] - - [285, 63.827] - - - [14848, 2816, 1, 256] - - [261, 71.299] - - - [3888, 1024, 1, 256] - - [271, 41.614] - - - [4096, 2816, 1, 256] - - [286, 61.935] - - - [9472, 1281, 1, 256] - - [261, 57.761] - - - [8960, 2865, 1, 256] - - [276, 67.185] - - - [10288, 7168, 1, 256] - - [252, 62.857] - - - [10240, 2048, 1, 256] - - [263, 66.868] - - - [12592, 2816, 1, 256] - - [285, 62.814] - - - [11568, 256, 1, 256] - - [256, 44.265] - - - [5120, 3328, 1, 256] - - [276, 66.584] - - - [18944, 256, 1, 256] - - [262, 51.106] - - - [7984, 5120, 1, 256] - - [261, 64.029] - - - [6144, 1280, 1, 256] - - [276, 56.718] - - - [14592, 1024, 1, 256] - - [254, 62.882] - - - [13360, 256, 1, 256] - - [280, 50.34] - - - [16896, 2816, 1, 256] - - [282, 72.165] - - - [11056, 2816, 1, 256] - - [271, 62.578] - - - [8192, 5120, 1, 256] - - [254, 70.894] - - - [18432, 10240, 1, 256] - - [264, 75.394] - - - [14336, 1280, 1, 256] - - [276, 66.254] - - - [8192, 256, 1, 256] - - [262, 35.021] - - - [1840, 1841, 1, 256] - - [261, 35.183] - - - [17664, 256, 1, 256] - - [267, 48.45] - - - [13056, 3329, 1, 256] - - [264, 67.803] - - - [7424, 6144, 1, 256] - - [262, 71.699] - - - [6656, 3328, 1, 256] - - [280, 67.747] - - - [9984, 2865, 1, 256] - - [254, 67.115] - - - [17920, 3328, 1, 256] - - [303, 72.159] - - - [16640, 3584, 1, 256] - - [261, 71.857] - - - [12032, 8960, 1, 256] - - [261, 74.177] - - - [3072, 2816, 1, 256] - - [261, 60.572] - - - [11312, 8192, 1, 256] - - [256, 64.421] - - - [16640, 3072, 1, 256] - - [254, 70.804] - - - [6144, 512, 1, 256] - - [287, 49.726] - - - [17200, 256, 1, 256] - - [267, 45.57] - - - [18944, 1280, 1, 256] - - [276, 68.497] - - - [15104, 256, 1, 256] - - [267, 42.985] - - - [12336, 9216, 1, 256] - - [252, 63.141] - - - [7168, 5888, 1, 256] - - [254, 72.398] - - - [4864, 3584, 1, 256] - - [282, 67.73] - - - [18480, 10240, 1, 256] - - [252, 62.84] - - - [9472, 6656, 1, 256] - - [261, 73.381] - - - [5120, 3329, 1, 256] - - [261, 65.105] - - - [6960, 256, 1, 256] - - [261, 39.312] - - - [7424, 2048, 1, 256] - - [287, 64.503] - - - [19712, 10240, 1, 256] - - [268, 73.761] - - - [10752, 3328, 1, 256] - - [254, 71.314] - - - [9520, 256, 1, 256] - - [267, 38.314] - - - [11008, 3329, 1, 256] - - [302, 65.949] - - - [12336, 9472, 1, 256] - - [264, 63.293] - - - [7936, 6144, 1, 256] - - [286, 71.94] - - - [4352, 1281, 1, 256] - - [254, 49.182] - - - [16944, 3584, 1, 256] - - [263, 64.55] - - - [14336, 1281, 1, 256] - - [263, 59.694] - - - [9216, 6144, 1, 256] - - [256, 73.135] - - - [1792, 1537, 1, 256] - - [262, 43.657] - - - [8448, 4864, 1, 256] - - [261, 72.259] - - - [15104, 1280, 1, 256] - - [261, 67.135] - - - [15616, 2865, 1, 256] - - [262, 69.029] - - - [15104, 2816, 1, 256] - - [287, 71.358] - - - [13056, 1280, 1, 256] - - [261, 65.37] - - - [3840, 3328, 1, 256] - - [261, 62.248] - - - [8448, 1281, 1, 256] - - [261, 56.762] - - - [6144, 2865, 1, 256] - - [252, 63.396] - - - [3584, 1281, 1, 256] - - [261, 49.145] - - - [12288, 2816, 1, 256] - - [252, 70.398] - - - [9216, 2048, 1, 256] - - [285, 66.26] - - - [10496, 7424, 1, 256] - - [252, 73.594] - - - [14848, 10240, 1, 256] - - [254, 74.949] - - - [11008, 7936, 1, 256] - - [298, 73.234] - - - [7680, 4864, 1, 256] - - [261, 72.155] - - - [15616, 1792, 1, 256] - - [262, 69.504] - - - [8192, 3329, 1, 256] - - [254, 66.799] - - - [16896, 2865, 1, 256] - - [264, 69.675] - - - [6960, 2865, 1, 256] - - [254, 58.909] - - - [18480, 5376, 1, 256] - - [256, 62.704] - - - [9520, 6656, 1, 256] - - [263, 63.438] - - - [19712, 6144, 1, 256] - - [278, 72.546] - - - [1792, 1585, 1, 256] - - [252, 44.837] - - - [17408, 4096, 1, 256] - - [254, 72.71] - - - [15104, 2048, 1, 256] - - [287, 68.882] - - - [6144, 768, 1, 256] - - [287, 50.604] - - - [6912, 3840, 1, 256] - - [282, 70.36] - - - [13312, 1792, 1, 256] - - [262, 68.51] - - - [10496, 1281, 1, 256] - - [287, 58.196] - - - [2560, 2353, 1, 256] - - [261, 52.819] - - - [12336, 256, 1, 256] - - [263, 47.326] - - - [31744, 6144, 1, 256] - - [254, 74.524] - - - [22272, 512, 1, 256] - - [250, 60.583] - - - [23808, 4096, 1, 256] - - [301, 72.707] - - - [29440, 2865, 1, 256] - - [254, 70.006] - - - [24832, 10240, 1, 256] - - [254, 74.967] - - - [24368, 768, 1, 256] - - [285, 59.639] - - - [29184, 10240, 1, 256] - - [251, 74.869] - - - [23088, 10240, 1, 256] - - [280, 63.352] - - - [24320, 6144, 1, 256] - - [257, 74.167] - - - [29488, 2865, 1, 256] - - [264, 60.926] - - - [32000, 7936, 1, 256] - - [254, 74.425] - - - [31792, 10240, 1, 256] - - [255, 63.633] - - - [27136, 3072, 1, 256] - - [278, 73.24] - - - [34096, 2865, 1, 256] - - [271, 62.013] - - - [28928, 1280, 1, 256] - - [261, 69.965] - - - [27696, 4096, 1, 256] - - [252, 62.899] - - - [28928, 10240, 1, 256] - - [256, 74.419] - - - [26112, 3328, 1, 256] - - [316, 73.184] - - - [27904, 3328, 1, 256] - - [308, 72.375] - - - [29952, 1281, 1, 256] - - [260, 64.138] - - - [28160, 4096, 1, 256] - - [269, 73.284] - - - [34816, 768, 1, 256] - - [280, 68.672] - - - [27136, 2816, 1, 256] - - [284, 73.176] - - - [26624, 2865, 1, 256] - - [254, 70.664] - - - [29440, 5376, 1, 256] - - [284, 74.171] - - - [29232, 10240, 1, 256] - - [263, 62.969] - - - [27904, 6144, 1, 256] - - [255, 73.746] - - - [34816, 3329, 1, 256] - - [252, 70.685] - - - [21248, 7424, 1, 256] - - [264, 74.819] - - - [34560, 256, 1, 256] - - [285, 59.115] - - - [23600, 2865, 1, 256] - - [264, 61.367] - - - [30768, 10240, 1, 256] - - [268, 62.81] - - - [26624, 256, 1, 256] - - [252, 57.116] - - - [21248, 7936, 1, 256] - - [254, 74.601] - - - [29696, 3840, 1, 256] - - [254, 74.502] - - - [24064, 2048, 1, 256] - - [299, 70.68] - - - [32512, 8704, 1, 256] - - [274, 74.767] - - - [27648, 10240, 1, 256] - - [264, 75.468] - - - [28976, 5632, 1, 256] - - [280, 63.367] - - - [26112, 2816, 1, 256] - - [254, 73.015] - - - [29696, 1536, 1, 256] - - [282, 70.945] - - - [32000, 3328, 1, 256] - - [276, 72.394] - - - [33280, 9984, 1, 256] - - [278, 75.53] - - - [27904, 4096, 1, 256] - - [292, 72.935] - - - [32048, 10240, 1, 256] - - [254, 63.133] - - - [32816, 2816, 1, 256] - - [255, 65.244] - - - [20992, 3328, 1, 256] - - [252, 72.666] - - - [24576, 4864, 1, 256] - - [265, 70.216] - - - [28208, 256, 1, 256] - - [271, 46.063] - - - [20528, 7168, 1, 256] - - [252, 61.991] - - - [20736, 2816, 1, 256] - - [261, 72.019] - - - [31536, 8192, 1, 256] - - [285, 62.985] - - - [20224, 2865, 1, 256] - - [254, 69.451] - - - [24320, 2048, 1, 256] - - [296, 70.443] - - - [33072, 10240, 1, 256] - - [256, 64.214] - - - [21040, 7680, 1, 256] - - [263, 64.019] - - - [32304, 8704, 1, 256] - - [285, 63.345] - - - [27136, 1024, 1, 256] - - [250, 68.778] - - - [33584, 256, 1, 256] - - [254, 54.154] - - - [28928, 2865, 1, 256] - - [256, 69.869] - - - [29440, 3328, 1, 256] - - [288, 72.706] - - - [20480, 2816, 1, 256] - - [254, 72.352] - - - [20736, 10240, 1, 256] - - [254, 74.9] - - - [33792, 3328, 1, 256] - - [252, 73.479] - - - [25600, 1280, 1, 256] - - [256, 69.893] - - - [22784, 1281, 1, 256] - - [266, 63.786] - - - [33280, 3328, 1, 256] - - [256, 73.324] - - - [33024, 768, 1, 256] - - [250, 65.517] - - - [26368, 3329, 1, 256] - - [281, 69.349] - - - [23040, 6144, 1, 256] - - [274, 74.371] - - - [28672, 3328, 1, 256] - - [256, 72.77] - - - [32048, 2816, 1, 256] - - [263, 63.364] - - - [27392, 1536, 1, 256] - - [258, 66.59] - - - [28160, 256, 1, 256] - - [261, 59.165] - - - [25648, 10240, 1, 256] - - [268, 63.438] - - - [22528, 1281, 1, 256] - - [252, 63.117] - - - [27952, 2865, 1, 256] - - [285, 61.685] - - - [20480, 3329, 1, 256] - - [259, 69.603] - - - [26624, 2560, 1, 256] - - [252, 73.237] - - - [26672, 3328, 1, 256] - - [252, 61.969] - - - [26112, 2304, 1, 256] - - [264, 72.648] - - - [29744, 6144, 1, 256] - - [268, 62.89] - - - [22272, 2560, 1, 256] - - [258, 72.039] - - - [31792, 256, 1, 256] - - [271, 51.858] - - - [32816, 10240, 1, 256] - - [301, 65.69] - - - [34096, 10240, 1, 256] - - [268, 62.84] - - - [32256, 1281, 1, 256] - - [266, 64.32] - - - [20784, 7424, 1, 256] - - [263, 63.384] - - - [22272, 256, 1, 256] - - [263, 49.692] - - - [28720, 2816, 1, 256] - - [252, 61.361] - - - [31488, 1536, 1, 256] - - [256, 70.299] - - - [30512, 7168, 1, 256] - - [285, 62.559] - - - [25088, 3329, 1, 256] - - [278, 69.78] - - - [23040, 1280, 1, 256] - - [287, 69.267] - - - [20224, 3329, 1, 256] - - [264, 69.242] - - - [31232, 7424, 1, 256] - - [274, 75.246] - - - [25904, 2865, 1, 256] - - [263, 61.333] - - - [20736, 768, 1, 256] - - [250, 65.491] - - - [28672, 6144, 1, 256] - - [254, 74.135] - - - [28928, 768, 1, 256] - - [267, 66.936] - - - [32256, 6144, 1, 256] - - [274, 74.469] - - - [32560, 2865, 1, 256] - - [280, 63.06] - - - [22784, 9728, 1, 256] - - [255, 74.598] - - - [27392, 2865, 1, 256] - - [268, 67.581] - - - [30720, 6656, 1, 256] - - [254, 75.049] - - - [25392, 1792, 1, 256] - - [271, 61.999] - - - [20224, 256, 1, 256] - - [267, 53.677] - - - [30976, 256, 1, 256] - - [267, 54.826] - - - [26672, 2816, 1, 256] - - [264, 62.709] - - - [28160, 1281, 1, 256] - - [301, 64.386] - - - [21504, 7936, 1, 256] - - [252, 75.534] - - - [24880, 2816, 1, 256] - - [271, 64.271] - - - [34816, 2865, 1, 256] - - [254, 71.003] - - - [29488, 6144, 1, 256] - - [271, 62.997] - - - [25088, 256, 1, 256] - - [280, 54.654] - - - [23040, 3072, 1, 256] - - [254, 72.891] - - - [33792, 3329, 1, 256] - - [254, 70.407] - - - [30720, 4864, 1, 256] - - [264, 74.739] - - - [28160, 3329, 1, 256] - - [259, 69.566] - - - [27648, 1281, 1, 256] - - [255, 63.629] - - - [23344, 10240, 1, 256] - - [263, 62.842] - - - [33024, 9216, 1, 256] - - [269, 74.582] - - - [24064, 10240, 1, 256] - - [274, 75.204] - - - [28208, 10240, 1, 256] - - [301, 62.743] - - - [33792, 1792, 1, 256] - - [264, 72.448] - - - [24624, 2816, 1, 256] - - [268, 62.429] - - - [21760, 1792, 1, 256] - - [254, 70.593] - - - [34608, 10240, 1, 256] - - [271, 63.342] - - - [31536, 2816, 1, 256] - - [285, 63.566] - - - [23552, 2865, 1, 256] - - [256, 70.547] - - - [34304, 2048, 1, 256] - - [268, 71.565] - - - [29440, 5632, 1, 256] - - [258, 74.412] - - - [30464, 6656, 1, 256] - - [251, 73.443] - - - [24832, 1024, 1, 256] - - [296, 67.376] - - - [29184, 2048, 1, 256] - - [301, 71.095] - - - [25344, 1536, 1, 256] - - [284, 69.677] - - - [32560, 8960, 1, 256] - - [280, 64.641] - - - [26880, 3072, 1, 256] - - [264, 72.76] - - - [20224, 6400, 1, 256] - - [254, 74.537] - - - [24320, 512, 1, 256] - - [284, 63.719] - - - [30720, 768, 1, 256] - - [267, 67.188] - - - [22016, 3328, 1, 256] - - [261, 72.823] - - - [31232, 2816, 1, 256] - - [264, 72.918] - - - [27136, 256, 1, 256] - - [254, 57.836] - - - [30208, 4352, 1, 256] - - [278, 74.258] - - - [29184, 6144, 1, 256] - - [254, 74.248] - - - [22576, 2816, 1, 256] - - [264, 62.212] - - - [23856, 2816, 1, 256] - - [280, 63.075] - - - [21248, 6144, 1, 256] - - [254, 74.121] - - - [27952, 256, 1, 256] - - [256, 53.969] - - - [31744, 3328, 1, 256] - - [252, 73.375] - - - [34048, 3329, 1, 256] - - [300, 69.195] - - - [27904, 256, 1, 256] - - [262, 58.921] - - - [29952, 6144, 1, 256] - - [251, 74.05] - - - [20784, 256, 1, 256] - - [285, 51.079] - - - [31232, 5376, 1, 256] - - [275, 74.579] - - - [21248, 1280, 1, 256] - - [282, 68.285] - - - [30720, 3329, 1, 256] - - [264, 70.407] - - - [31024, 2816, 1, 256] - - [285, 63.748] - - - [20528, 256, 1, 256] - - [261, 50.618] - - - [21504, 7680, 1, 256] - - [252, 75.495] - - - [33280, 6144, 1, 256] - - [278, 74.485] - - - [30720, 1281, 1, 256] - - [254, 63.715] - - - [30976, 6912, 1, 256] - - [268, 73.297] - - - [20992, 3329, 1, 256] - - [256, 69.637] - - - [24832, 2816, 1, 256] - - [284, 72.466] - - - [26880, 1280, 1, 256] - - [276, 69.337] - - - [33280, 9472, 1, 256] - - [251, 75.514] - - - [28416, 2816, 1, 256] - - [286, 72.194] - - - [31232, 256, 1, 256] - - [282, 55.016] - - - [26416, 256, 1, 256] - - [254, 51.781] - - - [22784, 6144, 1, 256] - - [251, 73.854] - - - [28160, 4864, 1, 256] - - [252, 74.305] - - - [29696, 256, 1, 256] - - [254, 52.782] - - - [25088, 5376, 1, 256] - - [294, 74.491] - - - [21296, 10240, 1, 256] - - [285, 63.537] - - - [21760, 256, 1, 256] - - [264, 48.404] - - - [30976, 1281, 1, 256] - - [316, 62.501] - - - [34304, 6144, 1, 256] - - [270, 74.528] - - - [31744, 5888, 1, 256] - - [264, 74.707] - - - [23856, 10240, 1, 256] - - [280, 63.188] - - - [32000, 1792, 1, 256] - - [254, 71.451] - - - [28416, 3328, 1, 256] - - [286, 72.264] - - - [28416, 4608, 1, 256] - - [278, 72.828] - - - [27440, 2865, 1, 256] - - [280, 61.881] - - - [30976, 2816, 1, 256] - - [294, 70.984] - - - [34048, 10240, 1, 256] - - [274, 74.096] - - - [22064, 2865, 1, 256] - - [263, 61.128] - - - [22528, 1280, 1, 256] - - [254, 70.17] - - - [33072, 2865, 1, 256] - - [256, 62.888] - - - [34560, 2048, 1, 256] - - [305, 71.126] - - - [34560, 1280, 1, 256] - - [252, 70.838] - - - [32304, 2816, 1, 256] - - [285, 64.167] - - - [30208, 6400, 1, 256] - - [274, 74.958] - - - [33072, 2816, 1, 256] - - [280, 64.978] - - - [20016, 2816, 1, 256] - - [285, 63.492] - - - [25648, 2816, 1, 256] - - [285, 63.004] - - - [24576, 10240, 1, 256] - - [265, 70.414] - - - [33584, 10240, 1, 256] - - [252, 63.044] - - - [23296, 3329, 1, 256] - - [259, 69.072] - - - [29696, 5632, 1, 256] - - [264, 75.118] - - - [20528, 2865, 1, 256] - - [254, 60.85] - - - [28160, 2048, 1, 256] - - [301, 70.872] - - - [20736, 6144, 1, 256] - - [264, 73.931] - - - [21552, 8448, 1, 256] - - [263, 64.011] - - - [28464, 4864, 1, 256] - - [285, 63.713] - - - [34304, 512, 1, 256] - - [284, 66.06] - - - [25648, 2048, 1, 256] - - [271, 62.668] - - - [32256, 8960, 1, 256] - - [251, 75.392] - - - [20992, 7680, 1, 256] - - [264, 75.084] - - - [26112, 256, 1, 256] - - [254, 56.263] - - - [25904, 2560, 1, 256] - - [285, 63.309] - - - [25088, 6144, 1, 256] - - [278, 74.353] - - - [22016, 8192, 1, 256] - - [264, 74.983] - - - [31232, 2048, 1, 256] - - [270, 71.486] - - - [20992, 2816, 1, 256] - - [264, 72.698] - - - [20224, 1280, 1, 256] - - [261, 68.324] - - - [26624, 768, 1, 256] - - [250, 66.916] - - - [21760, 1281, 1, 256] - - [296, 63.106] - - - [32768, 256, 1, 256] - - [261, 57.515] - - - [23552, 3329, 1, 256] - - [254, 70.102] - - - [29696, 2816, 1, 256] - - [252, 73.527] - - - [23040, 2048, 1, 256] - - [296, 70.576] - - - [22016, 2304, 1, 256] - - [261, 71.872] - - - [27136, 1281, 1, 256] - - [260, 63.812] - - - [27184, 3840, 1, 256] - - [285, 63.75] - - - [28672, 768, 1, 256] - - [250, 65.658] - - - [26160, 256, 1, 256] - - [271, 51.484] - - - [30208, 3328, 1, 256] - - [270, 73.144] - - - [26624, 1280, 1, 256] - - [262, 70.172] - - - [32000, 8704, 1, 256] - - [264, 74.667] - - - [26416, 2865, 1, 256] - - [263, 60.973] - - - [28672, 3329, 1, 256] - - [254, 69.719] - - - [27184, 10240, 1, 256] - - [280, 63.078] - - - [33280, 2816, 1, 256] - - [278, 73.287] - - - [26112, 1280, 1, 256] - - [262, 70.457] - - - [32560, 9216, 1, 256] - - [264, 63.849] - - - [27904, 1280, 1, 256] - - [258, 70.022] - - - [33280, 1280, 1, 256] - - [254, 70.983] - - - [27184, 2865, 1, 256] - - [263, 61.678] - - - [20784, 7680, 1, 256] - - [280, 63.953] - - - [25856, 1280, 1, 256] - - [262, 69.588] - - - [20016, 256, 1, 256] - - [267, 49.813] - - - [22576, 9216, 1, 256] - - [268, 62.894] - - - [28928, 5632, 1, 256] - - [254, 74.019] - - - [21552, 2865, 1, 256] - - [254, 61.492] - - - [24320, 2865, 1, 256] - - [284, 70.104] - - - [33536, 10240, 1, 256] - - [278, 74.79] - - - [24832, 1280, 1, 256] - - [252, 69.452] - - - [27392, 6144, 1, 256] - - [255, 73.273] - - - [25856, 1281, 1, 256] - - [263, 62.913] - - - [31488, 7680, 1, 256] - - [252, 74.434] - - - [32512, 512, 1, 256] - - [284, 62.875] - - - [32512, 1280, 1, 256] - - [282, 70.012] - - - [27904, 10240, 1, 256] - - [256, 74.585] - - - [21248, 2048, 1, 256] - - [267, 70.141] - - - [24064, 512, 1, 256] - - [282, 63.612] - - - [33840, 256, 1, 256] - - [262, 53.89] - - - [21296, 7936, 1, 256] - - [271, 63.676] - - - [32256, 8704, 1, 256] - - [274, 75.201] - - - [28464, 2816, 1, 256] - - [285, 63.429] - - - [31792, 8448, 1, 256] - - [290, 64.158] - - - [25088, 5120, 1, 256] - - [258, 74.212] - - - [30720, 1280, 1, 256] - - [264, 70.942] - - - [28416, 512, 1, 256] - - [261, 61.95] - - - [32512, 3328, 1, 256] - - [292, 72.772] - - - [22784, 2048, 1, 256] - - [266, 70.375] - - - [25344, 2816, 1, 256] - - [284, 71.619] - - - [21552, 10240, 1, 256] - - [254, 63.385] - - - [24576, 6144, 1, 256] - - [259, 69.698] - - - [31488, 3328, 1, 256] - - [268, 72.631] - - - [21040, 256, 1, 256] - - [250, 51.366] - - - [27648, 4096, 1, 256] - - [268, 73.226] - - - [31744, 7936, 1, 256] - - [254, 75.389] - - - [28672, 2048, 1, 256] - - [264, 68.271] - - - [25344, 3329, 1, 256] - - [274, 68.833] - - - [22320, 8960, 1, 256] - - [280, 64.407] - - - [34304, 2304, 1, 256] - - [258, 72.816] - - - [27392, 1280, 1, 256] - - [250, 67.051] - - - [28720, 10240, 1, 256] - - [268, 62.448] - - - [31280, 256, 1, 256] - - [280, 50.995] - - - [30720, 2048, 1, 256] - - [301, 70.334] - - - [24576, 2865, 1, 256] - - [281, 65.98] - - - [33280, 1024, 1, 256] - - [299, 69.675] - - - [25648, 2304, 1, 256] - - [280, 63.621] - - - [23040, 1024, 1, 256] - - [287, 66.825] - - - [31792, 8192, 1, 256] - - [252, 63.352] - - - [27648, 1536, 1, 256] - - [256, 70.642] - - - [24624, 256, 1, 256] - - [280, 49.754] - - - [28720, 5376, 1, 256] - - [268, 62.018] - - - [33840, 2865, 1, 256] - - [252, 62.278] - - - [24064, 2816, 1, 256] - - [258, 72.665] - - - [33328, 256, 1, 256] - - [263, 53.917] - - - [34304, 256, 1, 256] - - [261, 58.916] - - - [31744, 1281, 1, 256] - - [255, 64.217] - - - [24064, 4352, 1, 256] - - [286, 74.349] - - - [21504, 1281, 1, 256] - - [303, 62.721] - - - [34816, 3328, 1, 256] - - [264, 73.475] - - - [22016, 3329, 1, 256] - - [256, 69.596] - - - [31744, 7680, 1, 256] - - [254, 75.385] - - - [23808, 2048, 1, 256] - - [287, 70.427] - - - [34816, 256, 1, 256] - - [252, 59.726] - - - [31232, 1280, 1, 256] - - [282, 70.796] - - - [22528, 8704, 1, 256] - - [252, 75.669] - - - [34816, 1280, 1, 256] - - [252, 71.284] - - - [32512, 2048, 1, 256] - - [269, 70.782] - - - [34048, 2048, 1, 256] - - [318, 70.237] - - - [22272, 3329, 1, 256] - - [278, 68.958] - - - [34048, 2816, 1, 256] - - [284, 72.312] - - - [25648, 256, 1, 256] - - [285, 50.986] - - - [27904, 2048, 1, 256] - - [250, 70.333] - - - [21504, 2865, 1, 256] - - [254, 70.242] - - - [22576, 9472, 1, 256] - - [252, 62.863] - - - [28672, 5120, 1, 256] - - [254, 74.274] - - - [32816, 9472, 1, 256] - - [255, 66.519] - - - [22272, 8704, 1, 256] - - [278, 74.842] - - - [26928, 2816, 1, 256] - - [285, 63.237] - - - [23552, 3328, 1, 256] - - [254, 73.147] - - - [28928, 6144, 1, 256] - - [252, 73.774] - - - [22272, 2865, 1, 256] - - [264, 69.352] - - - [21760, 10240, 1, 256] - - [254, 74.893] - - - [26672, 256, 1, 256] - - [280, 52.318] - - - [30720, 3328, 1, 256] - - [261, 73.371] - - - [25088, 3328, 1, 256] - - [294, 72.895] - - - [33536, 9472, 1, 256] - - [251, 74.945] - - - [30208, 6144, 1, 256] - - [257, 74.374] - - - [33280, 256, 1, 256] - - [262, 56.664] - - - [32560, 256, 1, 256] - - [290, 52.359] - - - [33536, 256, 1, 256] - - [267, 58.138] - - - [24832, 5120, 1, 256] - - [252, 73.988] - - - [30768, 256, 1, 256] - - [263, 50.495] - - - [24320, 1024, 1, 256] - - [296, 68.601] - - - [31488, 5376, 1, 256] - - [254, 73.859] - - - [27392, 3584, 1, 256] - - [268, 71.904] - - - [22528, 256, 1, 256] - - [262, 50.074] - - - [30256, 6912, 1, 256] - - [263, 63.907] - - - [21760, 8448, 1, 256] - - [252, 74.631] - - - [27696, 2816, 1, 256] - - [271, 63.638] - - - [22576, 10240, 1, 256] - - [268, 62.743] - - - [28976, 10240, 1, 256] - - [280, 62.845] - - - [26368, 2816, 1, 256] - - [261, 72.464] - - - [27648, 3329, 1, 256] - - [264, 70.276] - - - [32048, 8448, 1, 256] - - [280, 63.167] - - - [34048, 3328, 1, 256] - - [270, 72.364] - - - [25600, 2816, 1, 256] - - [252, 73.099] - - - [24576, 512, 1, 256] - - [285, 62.802] - - - [22576, 2865, 1, 256] - - [264, 60.853] - - - [24368, 2816, 1, 256] - - [285, 63.888] - - - [30512, 6912, 1, 256] - - [285, 64.4] - - - [23296, 256, 1, 256] - - [256, 51.395] - - - [29184, 1280, 1, 256] - - [284, 70.572] - - - [31280, 7680, 1, 256] - - [268, 63.157] - - - [26624, 6144, 1, 256] - - [256, 74.59] - - - [20528, 2816, 1, 256] - - [252, 61.315] - - - [22272, 6144, 1, 256] - - [257, 73.834] - - - [25344, 2865, 1, 256] - - [258, 69.458] - - - [23296, 3584, 1, 256] - - [261, 73.272] - - - [20272, 2865, 1, 256] - - [256, 60.656] - - - [21760, 7936, 1, 256] - - [256, 74.725] - - - [33072, 9472, 1, 256] - - [271, 64.768] - - - [34352, 512, 1, 256] - - [263, 60.082] - - - [27952, 4352, 1, 256] - - [280, 63.569] - - - [22320, 9216, 1, 256] - - [271, 63.672] - - - [21248, 7680, 1, 256] - - [264, 74.765] - - - [25648, 2865, 1, 256] - - [254, 60.967] - - - [28416, 2865, 1, 256] - - [252, 69.557] - - - [27648, 2048, 1, 256] - - [313, 71.036] - - - [26880, 2048, 1, 256] - - [271, 70.716] - - - [24064, 1280, 1, 256] - - [282, 69.177] - - - [22064, 10240, 1, 256] - - [254, 63.217] - - - [22016, 6144, 1, 256] - - [256, 74.548] - - - [25392, 2048, 1, 256] - - [271, 62.586] - - - [33024, 3329, 1, 256] - - [259, 69.227] - - - [28208, 4608, 1, 256] - - [263, 63.809] - - - [30464, 512, 1, 256] - - [299, 63.734] - - - [29696, 6144, 1, 256] - - [256, 74.778] - - - [33024, 9472, 1, 256] - - [292, 74.823] - - - [21248, 1281, 1, 256] - - [285, 63.451] - - - [22272, 2816, 1, 256] - - [286, 71.99] - - - [30464, 6400, 1, 256] - - [291, 73.497] - - - [23040, 9472, 1, 256] - - [278, 75.599] - - - [30976, 1024, 1, 256] - - [299, 68.274] - - - [32000, 2048, 1, 256] - - [285, 69.97] - - - [33328, 10240, 1, 256] - - [256, 63.577] - - - [30512, 256, 1, 256] - - [271, 49.716] - - - [34608, 2816, 1, 256] - - [271, 63.396] - - - [25600, 2865, 1, 256] - - [256, 70.481] - - - [27392, 2048, 1, 256] - - [305, 69.256] - - - [20016, 10240, 1, 256] - - [285, 63.271] - - - [28160, 3328, 1, 256] - - [268, 72.965] - - - [31232, 7936, 1, 256] - - [274, 75.186] - - - [24832, 1536, 1, 256] - - [282, 69.465] - - - [29440, 6144, 1, 256] - - [278, 74.249] - - - [33792, 6144, 1, 256] - - [264, 74.818] - - - [20480, 768, 1, 256] - - [287, 64.292] - - - [23808, 3328, 1, 256] - - [254, 72.471] - - - [27136, 2048, 1, 256] - - [275, 70.97] - - - [20480, 7168, 1, 256] - - [256, 73.728] - - - [21552, 8192, 1, 256] - - [254, 63.331] - - - [28160, 6144, 1, 256] - - [268, 74.237] - - - [23600, 10240, 1, 256] - - [255, 63.677] - - - [24576, 1280, 1, 256] - - [252, 65.76] - - - [23088, 2816, 1, 256] - - [280, 63.162] - - - [29952, 2816, 1, 256] - - [252, 72.497] - - - [30976, 6144, 1, 256] - - [255, 72.357] - - - [30208, 4096, 1, 256] - - [268, 73.684] - - - [31488, 1281, 1, 256] - - [267, 64.289] - - - [32512, 256, 1, 256] - - [285, 56.034] - - - [26112, 3329, 1, 256] - - [281, 70.013] - - - [28208, 2816, 1, 256] - - [263, 63.499] - - - [28160, 4352, 1, 256] - - [257, 74.11] - - - [23856, 256, 1, 256] - - [271, 48.368] - - - [20992, 1281, 1, 256] - - [260, 63.172] - - - [21040, 10240, 1, 256] - - [264, 62.806] - - - [30976, 7168, 1, 256] - - [268, 72.418] - - - [24368, 1024, 1, 256] - - [271, 63.2] - - - [23296, 1536, 1, 256] - - [261, 69.712] - - - [23552, 2816, 1, 256] - - [254, 73.152] - - - [22832, 2865, 1, 256] - - [263, 61.219] - - - [22272, 8448, 1, 256] - - [278, 74.457] - - - [25856, 5888, 1, 256] - - [254, 73.834] - - - [29184, 3329, 1, 256] - - [300, 69.841] - - - [29744, 2816, 1, 256] - - [280, 63.337] - - - [25600, 2048, 1, 256] - - [268, 71.033] - - - [25856, 3328, 1, 256] - - [268, 72.256] - - - [32256, 2865, 1, 256] - - [259, 70.537] - - - [21504, 3328, 1, 256] - - [252, 72.919] - - - [24112, 256, 1, 256] - - [254, 48.82] - - - [22064, 8960, 1, 256] - - [285, 64.15] - - - [21504, 1536, 1, 256] - - [276, 69.413] - - - [20272, 6912, 1, 256] - - [285, 63.868] - - - [25904, 10240, 1, 256] - - [255, 62.643] - - - [32512, 8448, 1, 256] - - [292, 74.783] - - - [23808, 9984, 1, 256] - - [256, 75.123] - - - [34816, 6144, 1, 256] - - [264, 74.652] - - - [26368, 3328, 1, 256] - - [254, 72.308] - - - [34608, 1024, 1, 256] - - [271, 62.214] - - - [22528, 6144, 1, 256] - - [254, 74.61] - - - [22528, 2865, 1, 256] - - [264, 70.368] - - - [22016, 8704, 1, 256] - - [274, 75.464] - - - [33024, 2048, 1, 256] - - [269, 70.653] - - - [33024, 3328, 1, 256] - - [270, 72.644] - - - [27136, 3584, 1, 256] - - [252, 73.896] - - - [34864, 1280, 1, 256] - - [252, 62.169] - - - [21760, 3329, 1, 256] - - [256, 69.145] - - - [28976, 2816, 1, 256] - - [280, 63.623] - - - [21248, 1536, 1, 256] - - [261, 68.786] - - - [32768, 9472, 1, 256] - - [249, 59.284] - - - [26880, 3584, 1, 256] - - [252, 73.397] - - - [27648, 3840, 1, 256] - - [264, 74.357] - - - [34816, 2560, 1, 256] - - [252, 73.621] - - - [34816, 10240, 1, 256] - - [254, 75.289] - - - [33792, 10240, 1, 256] - - [254, 75.354] - - - [31024, 7424, 1, 256] - - [263, 63.587] - - - [30976, 2048, 1, 256] - - [320, 68.691] - - - [33792, 2816, 1, 256] - - [264, 73.562] - - - [20736, 7424, 1, 256] - - [256, 74.802] - - - [31488, 256, 1, 256] - - [250, 55.599] - - - [21504, 1792, 1, 256] - - [282, 71.135] - - - [33024, 1024, 1, 256] - - [250, 67.629] - - - [23088, 256, 1, 256] - - [285, 47.102] - - - [27392, 256, 1, 256] - - [263, 57.922] - - - [32512, 9216, 1, 256] - - [270, 74.559] - - - [20016, 2865, 1, 256] - - [263, 61.493] - - - [31024, 2865, 1, 256] - - [280, 61.593] - - - [34608, 2865, 1, 256] - - [263, 61.135] - - - [28160, 2865, 1, 256] - - [281, 69.887] - - - [23552, 6144, 1, 256] - - [252, 74.767] - - - [28672, 4864, 1, 256] - - [252, 74.407] - - - [31488, 7424, 1, 256] - - [264, 74.459] - - - [30976, 3328, 1, 256] - - [258, 71.027] - - - [33536, 9984, 1, 256] - - [278, 74.907] - - - [29184, 3328, 1, 256] - - [288, 72.913] - - - [34304, 768, 1, 256] - - [266, 68.64] - - - [25088, 1281, 1, 256] - - [266, 63.796] - - - [25136, 2816, 1, 256] - - [285, 63.662] - - - [27392, 3329, 1, 256] - - [311, 67.992] - - - [30976, 10240, 1, 256] - - [255, 73.744] - - - [34304, 1280, 1, 256] - - [286, 71.143] - - - [26624, 10240, 1, 256] - - [256, 75.441] - - - [29696, 3328, 1, 256] - - [256, 73.404] - - - [24368, 2865, 1, 256] - - [271, 62.435] - - - [20224, 6656, 1, 256] - - [252, 74.229] - - - [21296, 2865, 1, 256] - - [280, 61.548] - - - [32000, 8448, 1, 256] - - [254, 74.291] - - - [30976, 7680, 1, 256] - - [302, 72.959] - - - [23344, 9984, 1, 256] - - [271, 63.465] - - - [29696, 6400, 1, 256] - - [254, 75.374] - - - [27184, 256, 1, 256] - - [285, 52.727] - - - [20736, 256, 1, 256] - - [254, 54.594] - - - [25856, 6144, 1, 256] - - [264, 74.009] - - - [23088, 2865, 1, 256] - - [263, 61.571] - - - [23040, 9728, 1, 256] - - [292, 75.267] - - - [31744, 1792, 1, 256] - - [262, 72.252] - - - [27696, 256, 1, 256] - - [285, 53.65] - - - [22784, 3072, 1, 256] - - [254, 72.397] - - - [28928, 2816, 1, 256] - - [254, 72.431] - - - [27440, 4096, 1, 256] - - [280, 63.269] - - - [32256, 3329, 1, 256] - - [257, 70.183] - - - [25600, 10240, 1, 256] - - [252, 75.488] - - - [21040, 7936, 1, 256] - - [267, 64.061] - - - [32512, 1281, 1, 256] - - [318, 63.536] - - - [34864, 2865, 1, 256] - - [254, 61.273] - - - [21040, 2865, 1, 256] - - [280, 61.177] - - - [23296, 10240, 1, 256] - - [256, 74.732] - - - [25904, 2816, 1, 256] - - [271, 63.446] - - - [27648, 3328, 1, 256] - - [252, 73.377] - - - [21248, 256, 1, 256] - - [254, 47.768] - - - [27184, 3584, 1, 256] - - [285, 63.863] - - - [34048, 512, 1, 256] - - [262, 65.439] - - - [28672, 2816, 1, 256] - - [252, 72.765] - - - [34560, 512, 1, 256] - - [262, 66.38] - - - [30208, 6912, 1, 256] - - [274, 75.246] - - - [24832, 3328, 1, 256] - - [284, 72.518] - - - [32768, 8704, 1, 256] - - [265, 59.402] - - - [32816, 9216, 1, 256] - - [255, 66.618] - - - [34608, 256, 1, 256] - - [252, 54.121] - - - [31232, 7680, 1, 256] - - [274, 75.166] - - - [30208, 2865, 1, 256] - - [264, 70.349] - - - [32512, 2816, 1, 256] - - [251, 72.51] - - - [28416, 2048, 1, 256] - - [275, 69.489] - - - [23040, 2816, 1, 256] - - [262, 72.336] - - - [24064, 1281, 1, 256] - - [266, 64.011] - - - [29440, 10240, 1, 256] - - [257, 74.925] - - - [32768, 3328, 1, 256] - - [249, 58.272] - - - [27904, 4608, 1, 256] - - [255, 73.253] - - - [29184, 1024, 1, 256] - - [266, 69.176] - - - [28416, 10240, 1, 256] - - [274, 74.509] - - - [27904, 3329, 1, 256] - - [257, 69.138] - - - [29952, 6400, 1, 256] - - [251, 74.528] - - - [28928, 1281, 1, 256] - - [260, 63.378] - - - [23088, 9984, 1, 256] - - [285, 64.288] - - - [22528, 2560, 1, 256] - - [254, 72.954] - - - [26368, 256, 1, 256] - - [271, 56.773] - - - [31792, 2816, 1, 256] - - [267, 63.597] - - - [24880, 256, 1, 256] - - [256, 49.526] - - - [33584, 9984, 1, 256] - - [271, 63.613] - - - [24832, 256, 1, 256] - - [284, 53.701] - - - [30720, 7424, 1, 256] - - [256, 75.416] - - - [30000, 6400, 1, 256] - - [285, 63.97] - - - [22832, 2816, 1, 256] - - [280, 63.64] - - - [24320, 4352, 1, 256] - - [284, 73.984] - - - [26880, 768, 1, 256] - - [267, 66.503] - - - [27904, 3840, 1, 256] - - [284, 73.388] - - - [23296, 1281, 1, 256] - - [267, 62.497] - - - [23344, 2816, 1, 256] - - [263, 63.552] - - - [21504, 3329, 1, 256] - - [256, 70.177] - - - [32768, 1280, 1, 256] - - [265, 56.265] - - - [27952, 2816, 1, 256] - - [285, 63.766] - - - [20992, 6144, 1, 256] - - [256, 74.373] - - - [33280, 10240, 1, 256] - - [254, 75.116] - - - [32768, 9216, 1, 256] - - [293, 58.928] - - - [22320, 10240, 1, 256] - - [271, 63.577] - - - [34816, 512, 1, 256] - - [256, 66.676] - - - [29232, 5888, 1, 256] - - [263, 63.527] - - - [33792, 9984, 1, 256] - - [256, 75.596] - - - [21504, 2048, 1, 256] - - [268, 70.357] - - - [20784, 10240, 1, 256] - - [264, 63.029] - - - [28976, 2865, 1, 256] - - [271, 61.422] - - - [23552, 256, 1, 256] - - [261, 51.689] - - - [23600, 2816, 1, 256] - - [263, 63.383] - - - [28928, 3329, 1, 256] - - [281, 69.267] - - - [25344, 256, 1, 256] - - [280, 55.05] - - - [28672, 1280, 1, 256] - - [276, 69.999] - - - [28720, 2865, 1, 256] - - [252, 59.965] - - - [25600, 1536, 1, 256] - - [261, 70.369] - - - [27136, 10240, 1, 256] - - [278, 75.178] - - - [24320, 10240, 1, 256] - - [278, 74.908] - - - [23296, 9984, 1, 256] - - [252, 75.04] - - - [32000, 2865, 1, 256] - - [281, 70.214] - - - [20272, 10240, 1, 256] - - [271, 63.816] - - - [21760, 3328, 1, 256] - - [252, 72.312] - - - [29952, 3840, 1, 256] - - [258, 73.676] - - - [23344, 2865, 1, 256] - - [263, 60.816] - - - [27648, 3584, 1, 256] - - [252, 74.08] - - - [29440, 3584, 1, 256] - - [254, 73.461] - - - [32304, 256, 1, 256] - - [280, 52.548] - - - [22784, 1024, 1, 256] - - [260, 67.558] - - - [25856, 2816, 1, 256] - - [276, 72.336] - - - [28160, 4608, 1, 256] - - [318, 73.703] - - - [30208, 256, 1, 256] - - [250, 53.789] - - - [26368, 2304, 1, 256] - - [262, 71.856] - - - [31280, 7936, 1, 256] - - [271, 63.951] - - - [22320, 2816, 1, 256] - - [285, 63.845] - - - [22272, 1280, 1, 256] - - [261, 69.431] - - - [29696, 2048, 1, 256] - - [268, 71.313] - - - [28416, 1280, 1, 256] - - [282, 69.731] - - - [33536, 2865, 1, 256] - - [264, 69.879] - - - [31744, 256, 1, 256] - - [254, 55.586] - - - [30768, 7168, 1, 256] - - [255, 62.311] - - - [32256, 3328, 1, 256] - - [251, 73.232] - - - [33024, 256, 1, 256] - - [283, 54.622] - - - [21504, 8448, 1, 256] - - [254, 75.477] - - - [34304, 10240, 1, 256] - - [270, 75.096] - - - [22528, 8960, 1, 256] - - [256, 75.783] - - - [21808, 2865, 1, 256] - - [280, 61.674] - - - [27392, 3840, 1, 256] - - [268, 72.145] - - - [34864, 10240, 1, 256] - - [255, 62.895] - - - [30256, 10240, 1, 256] - - [271, 62.69] - - - [28928, 2048, 1, 256] - - [266, 69.845] - - - [20736, 7168, 1, 256] - - [252, 73.681] - - - [26368, 1280, 1, 256] - - [261, 69.846] - - - [23040, 3329, 1, 256] - - [300, 69.455] - - - [28416, 3329, 1, 256] - - [281, 69.052] - - - [21808, 8448, 1, 256] - - [263, 64.072] - - - [22272, 8960, 1, 256] - - [258, 74.925] - - - [31232, 7168, 1, 256] - - [278, 73.728] - - - [21504, 256, 1, 256] - - [276, 48.344] - - - [29696, 1792, 1, 256] - - [254, 72.277] - - - [27136, 3329, 1, 256] - - [274, 70.101] - - - [26880, 1281, 1, 256] - - [285, 63.574] - - - [20480, 6144, 1, 256] - - [254, 74.369] - - - [29696, 5888, 1, 256] - - [264, 74.749] - - - [33792, 2865, 1, 256] - - [264, 71.066] - - - [30720, 256, 1, 256] - - [287, 54.635] - - - [33024, 9728, 1, 256] - - [270, 74.736] - - - [33024, 3840, 1, 384] - - [263, 89.14] - - - [44160, 8832, 1, 384] - - [280, 90.868] - - - [39168, 3072, 1, 384] - - [252, 89.312] - - - [36096, 13440, 1, 384] - - [255, 89.824] - - - [32256, 1153, 1, 384] - - [280, 78.124] - - - [31488, 7296, 1, 384] - - [256, 90.342] - - - [31872, 6144, 1, 384] - - [264, 89.989] - - - [33792, 1152, 1, 384] - - [254, 85.923] - - - [36480, 1153, 1, 384] - - [263, 78.52] - - - [37248, 8448, 1, 384] - - [254, 90.158] - - - [41856, 7296, 1, 384] - - [285, 90.634] - - - [35712, 1153, 1, 384] - - [263, 77.05] - - - [31488, 1153, 1, 384] - - [252, 76.43] - - - [36864, 9216, 1, 384] - - [281, 86.953] - - - [37632, 8832, 1, 384] - - [285, 90.755] - - - [36864, 3072, 1, 384] - - [256, 87.687] - - - [31872, 13440, 1, 384] - - [254, 90.709] - - - [39168, 10368, 1, 384] - - [280, 90.644] - - - [39552, 1536, 1, 384] - - [254, 87.432] - - - [40320, 1153, 1, 384] - - [263, 78.245] - - - [43776, 14976, 1, 384] - - [268, 89.905] - - - [33024, 4224, 1, 384] - - [280, 89.263] - - - [34560, 9216, 1, 384] - - [283, 90.365] - - - [34944, 7296, 1, 384] - - [254, 90.377] - - - [41856, 6912, 1, 384] - - [271, 90.762] - - - [41472, 12672, 1, 384] - - [252, 91.04] - - - [43008, 1920, 1, 384] - - [254, 88.654] - - - [31104, 13441, 1, 384] - - [256, 89.23] - - - [31488, 7297, 1, 384] - - [252, 87.888] - - - [40704, 5376, 1, 384] - - [285, 90.373] - - - [38784, 13440, 1, 384] - - [264, 90.948] - - - [41856, 15360, 1, 384] - - [263, 91.147] - - - [33792, 4608, 1, 384] - - [254, 88.828] - - - [31104, 3072, 1, 384] - - [252, 88.396] - - - [34176, 5376, 1, 384] - - [263, 89.995] - - - [33792, 1920, 1, 384] - - [256, 88.614] - - - [43392, 15360, 1, 384] - - [254, 90.778] - - - [43776, 1153, 1, 384] - - [278, 76.234] - - - [42240, 1153, 1, 384] - - [252, 78.784] - - - [31488, 5760, 1, 384] - - [256, 90.215] - - - [32256, 3072, 1, 384] - - [252, 88.739] - - - [43776, 8448, 1, 384] - - [268, 89.663] - - - [34944, 13440, 1, 384] - - [256, 90.894] - - - [43008, 14208, 1, 384] - - [252, 90.513] - - - [43392, 8448, 1, 384] - - [280, 90.947] - - - [31104, 2304, 1, 384] - - [280, 88.538] - - - [39552, 13440, 1, 384] - - [252, 90.967] - - - [31872, 1153, 1, 384] - - [271, 77.391] - - - [36864, 15360, 1, 384] - - [259, 88.677] - - - [43392, 7297, 1, 384] - - [254, 88.191] - - - [41472, 7296, 1, 384] - - [252, 90.726] - - - [38784, 1153, 1, 384] - - [252, 77.728] - - - [36864, 1152, 1, 384] - - [252, 85.952] - - - [33024, 768, 1, 384] - - [313, 79.528] - - - [33792, 7680, 1, 384] - - [256, 90.317] - - - [40320, 4992, 1, 384] - - [280, 90.42] - - - [39552, 10752, 1, 384] - - [263, 91.09] - - - [44160, 1152, 1, 384] - - [264, 86.378] - - - [37248, 15360, 1, 384] - - [264, 90.911] - - - [41088, 15360, 1, 384] - - [288, 89.764] - - - [35328, 6528, 1, 384] - - [256, 90.045] - - - [33024, 7297, 1, 384] - - [259, 86.942] - - - [40704, 1153, 1, 384] - - [263, 78.6] - - - [41472, 7297, 1, 384] - - [264, 88.199] - - - [40704, 13440, 1, 384] - - [252, 91.016] - - - [34176, 9216, 1, 384] - - [283, 90.031] - - - [38400, 1152, 1, 384] - - [280, 86.421] - - - [31488, 2304, 1, 384] - - [280, 87.961] - - - [39168, 768, 1, 384] - - [263, 85.104] - - - [39936, 13440, 1, 384] - - [252, 90.764] - - - [33024, 15360, 1, 384] - - [254, 90.317] - - - [40320, 15360, 1, 384] - - [256, 90.659] - - - [40320, 11136, 1, 384] - - [271, 90.855] - - - [44160, 14976, 1, 384] - - [271, 90.918] - - - [38016, 13441, 1, 384] - - [256, 88.959] - - - [41856, 6528, 1, 384] - - [263, 90.77] - - - [39936, 15360, 1, 384] - - [259, 89.281] - - - [38400, 9600, 1, 384] - - [256, 90.744] - - - [37248, 2304, 1, 384] - - [256, 88.9] - - - [42624, 1152, 1, 384] - - [256, 85.746] - - - [37632, 13440, 1, 384] - - [254, 90.935] - - - [32640, 1152, 1, 384] - - [264, 82.416] - - - [40320, 3072, 1, 384] - - [252, 89.67] - - - [35712, 9216, 1, 384] - - [285, 90.628] - - - [36096, 3072, 1, 384] - - [257, 87.825] - - - [33408, 13441, 1, 384] - - [254, 89.408] - - - [31488, 1152, 1, 384] - - [256, 84.3] - - - [35328, 1152, 1, 384] - - [264, 86.19] - - - [34944, 6144, 1, 384] - - [254, 89.899] - - - [39168, 15360, 1, 384] - - [264, 90.611] - - - [42624, 1536, 1, 384] - - [259, 86.032] - - - [33792, 9216, 1, 384] - - [288, 88.885] - - - [33408, 1536, 1, 384] - - [254, 86.603] - - - [43008, 13440, 1, 384] - - [264, 90.679] - - - [34560, 5376, 1, 384] - - [263, 89.95] - - - [33408, 1153, 1, 384] - - [254, 77.608] - - - [41856, 13440, 1, 384] - - [271, 91.068] - - - [34560, 8832, 1, 384] - - [256, 90.698] - - - [33408, 13440, 1, 384] - - [252, 90.855] - - - [43776, 3072, 1, 384] - - [257, 86.725] - - - [39936, 7296, 1, 384] - - [254, 90.369] - - - [43392, 9216, 1, 384] - - [285, 90.379] - - - [30720, 13441, 1, 384] - - [256, 88.74] - - - [41472, 1152, 1, 384] - - [271, 86.836] - - - [43776, 2304, 1, 384] - - [301, 87.775] - - - [36096, 7297, 1, 384] - - [257, 86.378] - - - [35712, 15360, 1, 384] - - [256, 90.899] - - - [36096, 9216, 1, 384] - - [301, 89.158] - - - [33408, 9216, 1, 384] - - [289, 90.217] - - - [38016, 2688, 1, 384] - - [263, 89.371] - - - [39168, 9216, 1, 384] - - [264, 89.958] - - - [43392, 2304, 1, 384] - - [271, 89.144] - - - [43008, 1152, 1, 384] - - [252, 86.658] - - - [43392, 14592, 1, 384] - - [254, 90.978] - - - [38784, 9984, 1, 384] - - [256, 90.687] - - - [31488, 5376, 1, 384] - - [271, 90.263] - - - [34176, 1152, 1, 384] - - [264, 86.48] - - - [34560, 15360, 1, 384] - - [256, 90.733] - - - [37632, 13441, 1, 384] - - [254, 89.069] - - - [41856, 12672, 1, 384] - - [271, 91.102] - - - [42624, 7297, 1, 384] - - [278, 84.477] - - - [41472, 9216, 1, 384] - - [301, 89.586] - - - [36864, 8064, 1, 384] - - [252, 90.015] - - - [41088, 7297, 1, 384] - - [281, 86.365] - - - [38784, 3072, 1, 384] - - [256, 89.002] - - - [36096, 1153, 1, 384] - - [289, 76.971] - - - [39936, 1152, 1, 384] - - [254, 86.321] - - - [39552, 1153, 1, 384] - - [263, 77.084] - - - [31104, 5376, 1, 384] - - [285, 90.199] - - - [33024, 9216, 1, 384] - - [288, 89.636] - - - [43008, 3072, 1, 384] - - [252, 87.535] - - - [43776, 1152, 1, 384] - - [301, 85.56] - - - [38016, 7297, 1, 384] - - [280, 88.208] - - - [38400, 7297, 1, 384] - - [264, 88.136] - - - [33792, 15360, 1, 384] - - [281, 89.277] - - - [38784, 3840, 1, 384] - - [280, 90.014] - - - [41088, 2688, 1, 384] - - [263, 88.083] - - - [42240, 7296, 1, 384] - - [263, 90.815] - - - [32640, 13440, 1, 384] - - [254, 89.719] - - - [36480, 13440, 1, 384] - - [254, 90.863] - - - [41088, 7296, 1, 384] - - [265, 89.193] - - - [43392, 1920, 1, 384] - - [254, 88.388] - - - [31488, 15360, 1, 384] - - [256, 90.903] - - - [43008, 7680, 1, 384] - - [256, 90.127] - - - [43776, 15360, 1, 384] - - [288, 89.613] - - - [37248, 1153, 1, 384] - - [263, 77.474] - - - [34560, 7297, 1, 384] - - [264, 88.375] - - - [33408, 7296, 1, 384] - - [252, 90.328] - - - [33024, 13440, 1, 384] - - [254, 90.417] - - - [42240, 9216, 1, 384] - - [280, 91.084] - - - [33792, 7297, 1, 384] - - [264, 88.085] - - - [34176, 7296, 1, 384] - - [264, 90.146] - - - [31872, 15360, 1, 384] - - [252, 90.709] - - - [31104, 4992, 1, 384] - - [264, 89.792] - - - [39552, 10368, 1, 384] - - [271, 90.976] - - - [36864, 7297, 1, 384] - - [254, 87.1] - - - [34560, 1152, 1, 384] - - [254, 84.888] - - - [31488, 9216, 1, 384] - - [256, 89.809] - - - [41088, 1153, 1, 384] - - [285, 76.435] - - - [42240, 6912, 1, 384] - - [263, 90.914] - - - [42240, 7297, 1, 384] - - [254, 88.344] - - - [35712, 7296, 1, 384] - - [264, 90.409] - - - [39936, 9216, 1, 384] - - [288, 89.072] - - - [42624, 7680, 1, 384] - - [252, 88.093] - - - [38784, 768, 1, 384] - - [271, 84.311] - - - [34560, 5760, 1, 384] - - [256, 90.265] - - - [38016, 8832, 1, 384] - - [264, 90.763] - - - [40320, 11520, 1, 384] - - [254, 91.004] - - - [33792, 13441, 1, 384] - - [252, 88.944] - - - [33024, 1153, 1, 384] - - [285, 76.254] - - - [38016, 1153, 1, 384] - - [254, 78.612] - - - [42624, 13441, 1, 384] - - [257, 84.193] - - - [39552, 9216, 1, 384] - - [271, 90.808] - - - [42624, 3072, 1, 384] - - [252, 88.928] - - - [42240, 13441, 1, 384] - - [256, 89.158] - - - [41472, 6144, 1, 384] - - [256, 89.786] - - - [33408, 7297, 1, 384] - - [254, 88.0] - - - [41472, 1153, 1, 384] - - [254, 78.0] - - - [38400, 3456, 1, 384] - - [263, 89.537] - - - [43392, 1152, 1, 384] - - [263, 87.397] - - - [40704, 15360, 1, 384] - - [256, 90.792] - - - [32640, 3456, 1, 384] - - [254, 86.766] - - - [31104, 15360, 1, 384] - - [252, 90.791] - - - [32640, 7296, 1, 384] - - [256, 88.652] - - - [32640, 13441, 1, 384] - - [281, 87.325] - - - [34176, 13440, 1, 384] - - [264, 90.629] - - - [44160, 7297, 1, 384] - - [271, 88.025] - - - [39936, 1536, 1, 384] - - [252, 86.0] - - - [35712, 7297, 1, 384] - - [252, 88.087] - - - [42624, 7296, 1, 384] - - [277, 88.397] - - - [41472, 13441, 1, 384] - - [264, 89.163] - - - [32256, 15360, 1, 384] - - [252, 90.612] - - - [31872, 9216, 1, 384] - - [253, 90.157] - - - [40320, 1152, 1, 384] - - [271, 87.104] - - - [38784, 3456, 1, 384] - - [280, 89.687] - - - [41856, 3072, 1, 384] - - [256, 89.414] - - - [39936, 1153, 1, 384] - - [254, 77.349] - - - [39552, 7296, 1, 384] - - [263, 90.452] - - - [34560, 2688, 1, 384] - - [280, 89.033] - - - [40320, 9216, 1, 384] - - [263, 90.595] - - - [43776, 13440, 1, 384] - - [255, 90.065] - - - [42624, 13440, 1, 384] - - [249, 87.56] - - - [33024, 3072, 1, 384] - - [256, 88.041] - - - [42624, 1153, 1, 384] - - [252, 77.163] - - - [36480, 3072, 1, 384] - - [264, 88.881] - - - [39168, 1152, 1, 384] - - [252, 85.477] - - - [35712, 13441, 1, 384] - - [252, 89.034] - - - [43392, 3072, 1, 384] - - [254, 89.924] - - - [32256, 13440, 1, 384] - - [254, 90.922] - - - [40704, 11904, 1, 384] - - [252, 90.964] - - - [38784, 15360, 1, 384] - - [252, 90.615] - - - [44160, 13440, 1, 384] - - [263, 91.059] - - - [32640, 768, 1, 384] - - [303, 78.277] - - - [32640, 6528, 1, 384] - - [256, 88.489] - - - [31488, 13441, 1, 384] - - [256, 89.238] - - - [41088, 5760, 1, 384] - - [281, 88.944] - - - [43392, 7296, 1, 384] - - [252, 90.595] - - - [43008, 1153, 1, 384] - - [252, 78.043] - - - [36096, 13441, 1, 384] - - [259, 87.233] - - - [37632, 7297, 1, 384] - - [254, 88.148] - - - [43008, 8064, 1, 384] - - [254, 90.492] - - - [37248, 8064, 1, 384] - - [264, 90.624] - - - [38016, 13440, 1, 384] - - [256, 90.967] - - - [37248, 13441, 1, 384] - - [256, 88.869] - - - [44160, 3072, 1, 384] - - [256, 89.316] - - - [34176, 2304, 1, 384] - - [256, 88.599] - - - [34176, 1153, 1, 384] - - [252, 76.738] - - - [38784, 13441, 1, 384] - - [252, 88.737] - - - [32256, 3456, 1, 384] - - [254, 89.567] - - - [36480, 15360, 1, 384] - - [285, 91.186] - - - [35328, 7296, 1, 384] - - [256, 90.268] - - - [38016, 1152, 1, 384] - - [285, 85.896] - - - [23040, 7296, 1, 384] - - [252, 89.767] - - - [4224, 4225, 1, 384] - - [254, 76.404] - - - [16128, 3072, 1, 384] - - [252, 86.309] - - - [24576, 13440, 1, 384] - - [265, 87.615] - - - [16512, 9216, 1, 384] - - [288, 87.679] - - - [18432, 13441, 1, 384] - - [256, 88.684] - - - [27264, 3072, 1, 384] - - [252, 88.364] - - - [22272, 5376, 1, 384] - - [271, 89.631] - - - [26880, 9216, 1, 384] - - [271, 90.633] - - - [16512, 2688, 1, 384] - - [311, 82.82] - - - [22656, 1153, 1, 384] - - [271, 73.989] - - - [23808, 9216, 1, 384] - - [280, 90.726] - - - [1920, 1152, 1, 384] - - [286, 41.978] - - - [25728, 1152, 1, 384] - - [254, 84.086] - - - [3840, 1153, 1, 384] - - [282, 58.679] - - - [13056, 12673, 1, 384] - - [256, 89.234] - - - [28416, 13440, 1, 384] - - [263, 91.161] - - - [24576, 10752, 1, 384] - - [281, 86.561] - - - [16512, 7296, 1, 384] - - [268, 87.442] - - - [24192, 10368, 1, 384] - - [271, 90.498] - - - [1920, 1536, 1, 384] - - [262, 55.125] - - - [13056, 3072, 1, 384] - - [254, 85.097] - - - [10368, 1152, 1, 384] - - [287, 76.853] - - - [13440, 7297, 1, 384] - - [254, 87.271] - - - [13440, 13441, 1, 384] - - [285, 88.895] - - - [27648, 14208, 1, 384] - - [252, 90.692] - - - [24960, 1153, 1, 384] - - [263, 76.623] - - - [30720, 3072, 1, 384] - - [256, 88.071] - - - [24576, 3072, 1, 384] - - [256, 84.927] - - - [10368, 10369, 1, 384] - - [252, 87.889] - - - [11520, 7296, 1, 384] - - [263, 89.153] - - - [28800, 7296, 1, 384] - - [271, 90.21] - - - [11904, 7296, 1, 384] - - [254, 88.549] - - - [21504, 13440, 1, 384] - - [264, 90.62] - - - [13824, 7296, 1, 384] - - [252, 89.863] - - - [20736, 7296, 1, 384] - - [264, 90.329] - - - [21888, 1153, 1, 384] - - [284, 73.218] - - - [7680, 7680, 1, 384] - - [252, 88.391] - - - [9600, 1920, 1, 384] - - [254, 79.204] - - - [29568, 3840, 1, 384] - - [251, 89.362] - - - [13056, 2688, 1, 384] - - [280, 84.981] - - - [19200, 1152, 1, 384] - - [264, 80.905] - - - [17664, 7296, 1, 384] - - [280, 89.781] - - - [21888, 7296, 1, 384] - - [291, 87.508] - - - [23040, 1153, 1, 384] - - [285, 74.924] - - - [21120, 4608, 1, 384] - - [283, 89.236] - - - [29184, 13441, 1, 384] - - [274, 89.21] - - - [30336, 3072, 1, 384] - - [254, 89.232] - - - [16128, 2304, 1, 384] - - [252, 85.472] - - - [24960, 7297, 1, 384] - - [252, 88.598] - - - [19200, 5760, 1, 384] - - [252, 89.282] - - - [6144, 1152, 1, 384] - - [261, 73.359] - - - [18816, 5376, 1, 384] - - [263, 88.84] - - - [20352, 3840, 1, 384] - - [254, 88.528] - - - [17280, 768, 1, 384] - - [280, 76.137] - - - [28416, 14592, 1, 384] - - [285, 90.909] - - - [25344, 13440, 1, 384] - - [256, 91.119] - - - [26880, 1152, 1, 384] - - [264, 83.551] - - - [24960, 8448, 1, 384] - - [280, 90.766] - - - [27648, 1920, 1, 384] - - [252, 87.035] - - - [23040, 6528, 1, 384] - - [263, 89.69] - - - [27648, 1153, 1, 384] - - [264, 76.656] - - - [19584, 13441, 1, 384] - - [254, 89.274] - - - [16512, 13441, 1, 384] - - [259, 86.662] - - - [28416, 14976, 1, 384] - - [285, 90.971] - - - [20352, 9216, 1, 384] - - [285, 90.003] - - - [20736, 768, 1, 384] - - [280, 81.601] - - - [27264, 13440, 1, 384] - - [264, 90.688] - - - [27648, 13440, 1, 384] - - [254, 90.679] - - - [6528, 1920, 1, 384] - - [276, 80.083] - - - [12288, 7297, 1, 384] - - [252, 87.098] - - - [13824, 13441, 1, 384] - - [254, 88.762] - - - [16512, 7297, 1, 384] - - [259, 84.44] - - - [24960, 1152, 1, 384] - - [252, 85.839] - - - [22272, 1153, 1, 384] - - [285, 76.45] - - - [11136, 7296, 1, 384] - - [254, 88.208] - - - [6912, 2688, 1, 384] - - [254, 79.957] - - - [15744, 13441, 1, 384] - - [254, 89.503] - - - [18816, 7297, 1, 384] - - [252, 87.989] - - - [29184, 3456, 1, 384] - - [252, 89.49] - - - [4992, 4609, 1, 384] - - [252, 78.626] - - - [21504, 4608, 1, 384] - - [256, 88.464] - - - [14592, 13441, 1, 384] - - [264, 89.353] - - - [14976, 3072, 1, 384] - - [256, 86.849] - - - [12672, 4992, 1, 384] - - [252, 87.364] - - - [19584, 1152, 1, 384] - - [254, 82.418] - - - [23040, 1152, 1, 384] - - [254, 84.389] - - - [25344, 3072, 1, 384] - - [264, 87.695] - - - [5760, 1536, 1, 384] - - [262, 76.188] - - - [21504, 8064, 1, 384] - - [264, 89.897] - - - [16128, 13441, 1, 384] - - [254, 89.135] - - - [23808, 7297, 1, 384] - - [285, 88.195] - - - [16896, 13441, 1, 384] - - [252, 89.152] - - - [21504, 1152, 1, 384] - - [256, 84.059] - - - [16512, 3072, 1, 384] - - [259, 82.448] - - - [21888, 8064, 1, 384] - - [277, 87.855] - - - [10752, 10753, 1, 384] - - [256, 88.712] - - - [28032, 7297, 1, 384] - - [256, 88.513] - - - [23040, 13440, 1, 384] - - [264, 90.83] - - - [4224, 4224, 1, 384] - - [262, 82.929] - - - [26112, 1152, 1, 384] - - [285, 85.034] - - - [28032, 9216, 1, 384] - - [280, 90.625] - - - [17664, 1153, 1, 384] - - [280, 73.888] - - - [21888, 1152, 1, 384] - - [256, 82.958] - - - [10752, 10369, 1, 384] - - [264, 88.454] - - - [13440, 13057, 1, 384] - - [271, 88.973] - - - [12672, 3072, 1, 384] - - [252, 86.363] - - - [29952, 3840, 1, 384] - - [280, 89.78] - - - [26112, 7297, 1, 384] - - [256, 88.499] - - - [17664, 4224, 1, 384] - - [254, 88.407] - - - [30720, 1920, 1, 384] - - [252, 87.964] - - - [17664, 13441, 1, 384] - - [280, 89.74] - - - [20736, 4224, 1, 384] - - [285, 89.19] - - - [18048, 13441, 1, 384] - - [263, 89.388] - - - [16896, 7296, 1, 384] - - [252, 89.528] - - - [9600, 1152, 1, 384] - - [280, 72.126] - - - [20736, 1153, 1, 384] - - [263, 75.779] - - - [29568, 7296, 1, 384] - - [280, 89.833] - - - [27264, 1152, 1, 384] - - [252, 83.817] - - - [5760, 5376, 1, 384] - - [264, 83.708] - - - [14208, 13441, 1, 384] - - [271, 89.455] - - - [7296, 6913, 1, 384] - - [254, 85.463] - - - [20352, 768, 1, 384] - - [280, 79.924] - - - [13056, 12672, 1, 384] - - [256, 90.482] - - - [11904, 1153, 1, 384] - - [280, 70.203] - - - [9984, 9985, 1, 384] - - [254, 87.492] - - - [30336, 15360, 1, 384] - - [252, 90.942] - - - [24192, 13441, 1, 384] - - [264, 89.274] - - - [24576, 7297, 1, 384] - - [264, 83.448] - - - [9600, 1153, 1, 384] - - [267, 70.514] - - - [6144, 1153, 1, 384] - - [282, 61.339] - - - [26880, 1153, 1, 384] - - [285, 75.08] - - - [6912, 6529, 1, 384] - - [252, 85.275] - - - [20736, 9216, 1, 384] - - [285, 90.051] - - - [11136, 1153, 1, 384] - - [267, 72.462] - - - [25344, 13441, 1, 384] - - [251, 89.36] - - - [14592, 13440, 1, 384] - - [252, 90.678] - - - [29568, 768, 1, 384] - - [299, 81.546] - - - [21888, 13441, 1, 384] - - [255, 84.654] - - - [25728, 11904, 1, 384] - - [271, 90.922] - - - [28416, 1153, 1, 384] - - [252, 75.623] - - - [23040, 7297, 1, 384] - - [254, 88.179] - - - [4608, 4225, 1, 384] - - [252, 76.866] - - - [3072, 1152, 1, 384] - - [294, 64.479] - - - [13440, 3072, 1, 384] - - [264, 87.428] - - - [18432, 13440, 1, 384] - - [264, 90.518] - - - [5376, 768, 1, 384] - - [262, 55.528] - - - [20352, 1152, 1, 384] - - [254, 80.551] - - - [23808, 9984, 1, 384] - - [252, 90.845] - - - [24576, 1152, 1, 384] - - [252, 82.609] - - - [4224, 3072, 1, 384] - - [254, 74.892] - - - [19200, 5376, 1, 384] - - [263, 88.93] - - - [18432, 4992, 1, 384] - - [252, 88.376] - - - [15744, 3072, 1, 384] - - [254, 87.77] - - - [20352, 6912, 1, 384] - - [254, 89.537] - - - [4224, 3840, 1, 384] - - [252, 82.671] - - - [16512, 13440, 1, 384] - - [259, 89.118] - - - [10368, 3072, 1, 384] - - [254, 85.551] - - - [14208, 6528, 1, 384] - - [263, 89.104] - - - [12672, 1152, 1, 384] - - [262, 75.651] - - - [19200, 2304, 1, 384] - - [280, 86.548] - - - [27264, 1153, 1, 384] - - [254, 75.153] - - - [11520, 11136, 1, 384] - - [280, 89.532] - - - [18048, 1153, 1, 384] - - [285, 75.368] - - - [22272, 7297, 1, 384] - - [271, 88.518] - - - [22656, 2688, 1, 384] - - [280, 88.542] - - - [8064, 7297, 1, 384] - - [254, 85.223] - - - [23808, 7296, 1, 384] - - [256, 90.104] - - - [14208, 6912, 1, 384] - - [254, 89.323] - - - [28416, 7296, 1, 384] - - [271, 90.064] - - - [17280, 7297, 1, 384] - - [256, 87.711] - - - [28032, 1153, 1, 384] - - [254, 77.624] - - - [8064, 3072, 1, 384] - - [254, 84.009] - - - [3840, 3456, 1, 384] - - [262, 77.031] - - - [11904, 11520, 1, 384] - - [252, 89.564] - - - [21120, 13440, 1, 384] - - [256, 90.747] - - - [6912, 6528, 1, 384] - - [263, 85.626] - - - [23808, 3072, 1, 384] - - [264, 88.181] - - - [29952, 9216, 1, 384] - - [285, 90.581] - - - [20352, 3456, 1, 384] - - [280, 87.447] - - - [3840, 3457, 1, 384] - - [282, 76.26] - - - [26112, 3072, 1, 384] - - [256, 88.29] - - - [21504, 1536, 1, 384] - - [264, 83.608] - - - [6528, 6144, 1, 384] - - [254, 85.453] - - - [3072, 1920, 1, 384] - - [282, 62.213] - - - [30336, 7297, 1, 384] - - [264, 87.864] - - - [26112, 9216, 1, 384] - - [285, 90.379] - - - [11520, 11137, 1, 384] - - [256, 89.105] - - - [29568, 13441, 1, 384] - - [278, 88.47] - - - [2688, 1152, 1, 384] - - [307, 57.57] - - - [25344, 11520, 1, 384] - - [254, 91.069] - - - [11904, 4608, 1, 384] - - [280, 87.354] - - - [11520, 3072, 1, 384] - - [252, 85.887] - - - [24960, 11136, 1, 384] - - [280, 90.638] - - - [19200, 2688, 1, 384] - - [280, 87.33] - - - [23040, 3072, 1, 384] - - [264, 87.388] - - - [5760, 5377, 1, 384] - - [256, 83.087] - - - [13824, 7297, 1, 384] - - [256, 88.168] - - - [16128, 8448, 1, 384] - - [263, 89.625] - - - [6144, 6144, 1, 384] - - [254, 86.096] - - - [10368, 7297, 1, 384] - - [263, 87.132] - - - [30720, 1153, 1, 384] - - [264, 77.37] - - - [9984, 9984, 1, 384] - - [254, 89.105] - - - [26112, 12672, 1, 384] - - [254, 90.97] - - - [19584, 13440, 1, 384] - - [264, 90.462] - - - [8832, 8449, 1, 384] - - [280, 86.262] - - - [7680, 1153, 1, 384] - - [250, 65.5] - - - [28800, 1152, 1, 384] - - [285, 84.104] - - - [12288, 1152, 1, 384] - - [254, 80.098] - - - [14976, 13441, 1, 384] - - [256, 89.511] - - - [18432, 4608, 1, 384] - - [252, 87.45] - - - [11520, 4224, 1, 384] - - [254, 88.351] - - - [28416, 2304, 1, 384] - - [280, 87.819] - - - [26112, 7296, 1, 384] - - [256, 90.453] - - - [4224, 1152, 1, 384] - - [287, 64.294] - - - [6144, 1536, 1, 384] - - [276, 70.378] - - - [8064, 8064, 1, 384] - - [254, 87.259] - - - [13056, 7297, 1, 384] - - [252, 87.919] - - - [25344, 7297, 1, 384] - - [256, 88.445] - - - [28416, 1152, 1, 384] - - [254, 83.942] - - - [28416, 7297, 1, 384] - - [264, 88.08] - - - [26496, 9216, 1, 384] - - [263, 90.68] - - - [14208, 1152, 1, 384] - - [254, 76.797] - - - [18048, 7296, 1, 384] - - [264, 89.317] - - - [28800, 14976, 1, 384] - - [271, 90.941] - - - [24576, 7680, 1, 384] - - [264, 86.674] - - - [20736, 3840, 1, 384] - - [271, 88.309] - - - [21888, 3072, 1, 384] - - [257, 84.822] - - - [14592, 1153, 1, 384] - - [263, 71.604] - - - [22272, 3072, 1, 384] - - [254, 88.543] - - - [8064, 7296, 1, 384] - - [254, 88.017] - - - [22272, 2688, 1, 384] - - [271, 87.261] - - - [8832, 8833, 1, 384] - - [263, 87.997] - - - [23424, 1153, 1, 384] - - [254, 75.606] - - - [10752, 10368, 1, 384] - - [263, 89.934] - - - [4608, 4608, 1, 384] - - [252, 83.853] - - - [12672, 2304, 1, 384] - - [252, 83.227] - - - [5376, 1153, 1, 384] - - [261, 63.522] - - - [27264, 7297, 1, 384] - - [256, 87.552] - - - [11136, 11136, 1, 384] - - [285, 90.048] - - - [26496, 13440, 1, 384] - - [256, 91.008] - - - [17280, 7296, 1, 384] - - [264, 90.133] - - - [23808, 13441, 1, 384] - - [263, 89.697] - - - [6144, 1920, 1, 384] - - [282, 76.159] - - - [30336, 13441, 1, 384] - - [252, 89.106] - - - [24576, 9216, 1, 384] - - [259, 84.679] - - - [18048, 4608, 1, 384] - - [280, 88.225] - - - [11136, 3840, 1, 384] - - [254, 87.16] - - - [19584, 7296, 1, 384] - - [254, 89.672] - - - [2304, 2305, 1, 384] - - [282, 56.221] - - - [15744, 1152, 1, 384] - - [252, 77.976] - - - [4992, 3072, 1, 384] - - [254, 79.07] - - - [2688, 2304, 1, 384] - - [261, 65.324] - - - [25344, 11904, 1, 384] - - [254, 90.852] - - - [8448, 8449, 1, 384] - - [256, 86.577] - - - [9216, 1153, 1, 384] - - [271, 68.145] - - - [25344, 2688, 1, 384] - - [285, 88.692] - - - [14208, 7297, 1, 384] - - [280, 87.622] - - - [4608, 4224, 1, 384] - - [262, 83.371] - - - [24576, 13441, 1, 384] - - [252, 84.434] - - - [21888, 4992, 1, 384] - - [277, 87.042] - - - [19968, 6144, 1, 384] - - [254, 88.894] - - - [9600, 7297, 1, 384] - - [254, 86.663] - - - [2304, 2304, 1, 384] - - [276, 69.873] - - - [18816, 1152, 1, 384] - - [254, 79.623] - - - [17280, 1153, 1, 384] - - [263, 72.434] - - - [29184, 7297, 1, 384] - - [256, 88.119] - - - [11520, 11520, 1, 384] - - [252, 90.049] - - - [18048, 13440, 1, 384] - - [285, 90.675] - - - [14976, 7680, 1, 384] - - [254, 90.015] - - - [15360, 7680, 1, 384] - - [252, 89.515] - - - [28032, 1920, 1, 384] - - [252, 88.185] - - - [6144, 6145, 1, 384] - - [252, 81.33] - - - [5760, 3072, 1, 384] - - [256, 82.465] - - - [14592, 768, 1, 384] - - [287, 72.813] - - - [6912, 1153, 1, 384] - - [267, 68.662] - - - [13056, 5760, 1, 384] - - [254, 89.082] - - - [13056, 13056, 1, 384] - - [254, 90.324] - - - [22272, 8448, 1, 384] - - [285, 90.311] - - - [17280, 3456, 1, 384] - - [263, 87.076] - - - [9216, 7297, 1, 384] - - [254, 87.23] - - - [16128, 7297, 1, 384] - - [254, 87.628] - - - [30336, 1153, 1, 384] - - [263, 76.888] - - - [27648, 1152, 1, 384] - - [256, 85.05] - - - [9216, 1920, 1, 384] - - [254, 81.957] - - - [21888, 5376, 1, 384] - - [257, 85.822] - - - [13440, 1153, 1, 384] - - [263, 71.698] - - - [29952, 4224, 1, 384] - - [254, 89.448] - - - [5760, 5760, 1, 384] - - [252, 85.038] - - - [24192, 3072, 1, 384] - - [252, 87.871] - - - [27264, 7296, 1, 384] - - [256, 90.021] - - - [18432, 1536, 1, 384] - - [252, 84.098] - - - [9600, 7296, 1, 384] - - [254, 89.09] - - - [29568, 9216, 1, 384] - - [263, 89.74] - - - [6528, 3072, 1, 384] - - [252, 79.503] - - - [1920, 1153, 1, 384] - - [261, 42.154] - - - [4992, 1153, 1, 384] - - [287, 60.197] - - - [9984, 1152, 1, 384] - - [261, 74.729] - - - [25344, 2304, 1, 384] - - [254, 87.575] - - - [9216, 8833, 1, 384] - - [256, 87.479] - - - [8448, 1153, 1, 384] - - [262, 71.078] - - - [11520, 7297, 1, 384] - - [256, 87.053] - - - [4224, 3841, 1, 384] - - [276, 76.561] - - - [2304, 1152, 1, 384] - - [284, 50.045] - - - [15360, 1153, 1, 384] - - [280, 74.403] - - - [19200, 1153, 1, 384] - - [254, 74.708] - - - [12672, 12288, 1, 384] - - [280, 90.041] - - - [26496, 1152, 1, 384] - - [263, 86.187] - - - [6528, 2304, 1, 384] - - [261, 77.781] - - - [19968, 1152, 1, 384] - - [285, 83.787] - - - [3840, 3840, 1, 384] - - [262, 76.876] - - - [24192, 7296, 1, 384] - - [252, 90.513] - - - [27648, 7297, 1, 384] - - [254, 87.86] - - - [3456, 3457, 1, 384] - - [261, 76.449] - - - [23808, 768, 1, 384] - - [263, 78.629] - - - [28416, 9216, 1, 384] - - [280, 90.634] - - - [25728, 12288, 1, 384] - - [285, 90.662] - - - [20736, 3072, 1, 384] - - [256, 87.451] - - - [29568, 7297, 1, 384] - - [274, 87.324] - - - [15744, 1153, 1, 384] - - [280, 71.28] - - - [17664, 7297, 1, 384] - - [256, 88.339] - - - [25344, 1153, 1, 384] - - [263, 77.647] - - - [25728, 9216, 1, 384] - - [280, 90.516] - - - [30336, 1152, 1, 384] - - [271, 84.968] - - - [18432, 9216, 1, 384] - - [251, 88.031] - - - [12288, 1536, 1, 384] - - [252, 80.308] - - - [19968, 7297, 1, 384] - - [264, 88.003] - - - [24960, 1920, 1, 384] - - [254, 87.186] - - - [30720, 7296, 1, 384] - - [254, 90.195] - - - [25344, 7296, 1, 384] - - [254, 90.358] - - - [26112, 13441, 1, 384] - - [256, 89.566] - - - [27648, 1536, 1, 384] - - [254, 85.661] - - - [5760, 1153, 1, 384] - - [287, 57.969] - - - [12672, 1153, 1, 384] - - [271, 74.098] - - - [18816, 13440, 1, 384] - - [271, 90.572] - - - [9216, 1152, 1, 384] - - [262, 77.636] - - - [8832, 7296, 1, 384] - - [254, 88.738] - - - [16512, 1153, 1, 384] - - [305, 70.083] - - - [11136, 7297, 1, 384] - - [254, 87.578] - - - [28800, 7297, 1, 384] - - [252, 87.828] - - - [23808, 10368, 1, 384] - - [280, 90.899] - - - [9216, 1536, 1, 384] - - [252, 80.124] - - - [15360, 3072, 1, 384] - - [254, 85.325] - - - [28416, 3072, 1, 384] - - [254, 88.406] - - - [21504, 3072, 1, 384] - - [264, 86.952] - - - [22656, 7297, 1, 384] - - [264, 88.163] - - - [12288, 11905, 1, 384] - - [254, 88.072] - - - [29184, 3072, 1, 384] - - [256, 88.836] - - - [16896, 7297, 1, 384] - - [264, 88.076] - - - [5376, 4993, 1, 384] - - [254, 80.619] - - - [8448, 8064, 1, 384] - - [252, 88.899] - - - [19200, 9216, 1, 384] - - [263, 89.79] - - - [13440, 5760, 1, 384] - - [252, 89.431] - - - [29952, 15360, 1, 384] - - [252, 90.909] - - - [10368, 9985, 1, 384] - - [254, 87.627] - - - [3456, 3073, 1, 384] - - [254, 68.565] - - - [8064, 8065, 1, 384] - - [264, 86.659] - - - [30336, 4608, 1, 384] - - [280, 89.816] - - - [11904, 11521, 1, 384] - - [280, 89.158] - - - [19200, 13440, 1, 384] - - [264, 90.448] - - - [23424, 3072, 1, 384] - - [254, 88.984] - - - [12672, 7297, 1, 384] - - [252, 86.974] - - - [18432, 1152, 1, 384] - - [254, 83.095] - - - [10752, 1152, 1, 384] - - [261, 79.228] - - - [11904, 3072, 1, 384] - - [264, 84.928] - - - [20736, 13441, 1, 384] - - [280, 89.398] - - - [23808, 13440, 1, 384] - - [252, 90.934] - - - [10752, 3072, 1, 384] - - [252, 84.298] - - - [7296, 7296, 1, 384] - - [264, 87.616] - - - [25728, 8832, 1, 384] - - [256, 90.353] - - - [4992, 1152, 1, 384] - - [282, 60.473] - - - [21504, 1153, 1, 384] - - [264, 74.005] - - - [10752, 7296, 1, 384] - - [264, 88.717] - - - [7296, 1152, 1, 384] - - [261, 72.559] - - - [19200, 7297, 1, 384] - - [254, 87.529] - - - [12672, 12673, 1, 384] - - [254, 89.458] - - - [23424, 6528, 1, 384] - - [252, 89.85] - - - [6528, 6145, 1, 384] - - [264, 82.886] - - - [21888, 8448, 1, 384] - - [255, 87.485] - - - [7680, 7296, 1, 384] - - [254, 86.882] - - - [24960, 11520, 1, 384] - - [254, 90.87] - - - [25728, 1153, 1, 384] - - [254, 75.264] - - - [13056, 13057, 1, 384] - - [252, 89.144] - - - [23424, 7297, 1, 384] - - [257, 87.659] - - - [11520, 1153, 1, 384] - - [261, 68.201] - - - [14976, 13440, 1, 384] - - [252, 90.603] - - - [24192, 10752, 1, 384] - - [254, 90.661] - - - [9984, 1153, 1, 384] - - [285, 66.306] - - - [12672, 12672, 1, 384] - - [254, 89.862] - - - [11520, 1152, 1, 384] - - [254, 76.213] - - - [20352, 1153, 1, 384] - - [280, 74.649] - - - [6912, 2304, 1, 384] - - [271, 81.625] - - - [17664, 9216, 1, 384] - - [280, 90.066] - - - [10752, 3456, 1, 384] - - [252, 86.158] - - - [9216, 9217, 1, 384] - - [255, 84.232] - - - [12288, 12289, 1, 384] - - [281, 83.752] - - - [11904, 7297, 1, 384] - - [254, 86.522] - - - [12288, 1153, 1, 384] - - [254, 71.726] - - - [13056, 1152, 1, 384] - - [254, 77.549] - - - [18816, 2304, 1, 384] - - [271, 85.066] - - - [16512, 1152, 1, 384] - - [272, 75.119] - - - [7296, 2688, 1, 384] - - [254, 83.159] - - - [4608, 4609, 1, 384] - - [254, 77.982] - - - [30720, 9216, 1, 384] - - [301, 88.322] - - - [42496, 10240, 1, 256] - - [256, 74.992] - - - [8960, 5632, 1, 256] - - [286, 72.47] - - - [2560, 1281, 1, 256] - - [321, 50.321] - - - [1536, 1153, 1, 384] - - [278, 33.947] - - - [14208, 128, 1, 384] - - [250, 34.859] - - - [14976, 128, 1, 384] - - [261, 36.622] - - - [44544, 2048, 1, 384] - - [280, 88.554] - - - [38016, 22145, 1, 384] - - [264, 89.555] - - - [39552, 23681, 1, 384] - - [256, 89.45] - - - [44544, 28673, 1, 384] - - [259, 87.03] - - - [43392, 1024, 1, 384] - - [263, 86.883] - - - [42624, 26369, 1, 384] - - [278, 83.329] - - - [36096, 384, 1, 384] - - [285, 77.693] - - - [39552, 384, 1, 384] - - [252, 77.319] - - - [42240, 25985, 1, 384] - - [252, 89.72] - - - [39168, 4096, 1, 384] - - [280, 89.71] - - - [43392, 2048, 1, 384] - - [263, 88.584] - - - [35712, 1024, 1, 384] - - [280, 84.814] - - - [40704, 512, 1, 384] - - [254, 81.464] - - - [37632, 21761, 1, 384] - - [264, 89.191] - - - [44160, 2048, 1, 384] - - [280, 88.431] - - - [43392, 384, 1, 384] - - [254, 77.508] - - - [38784, 1024, 1, 384] - - [271, 84.387] - - - [44544, 8192, 1, 384] - - [252, 89.941] - - - [39936, 512, 1, 384] - - [254, 80.271] - - - [42240, 2048, 1, 384] - - [285, 89.372] - - - [37248, 1024, 1, 384] - - [263, 84.781] - - - [43776, 384, 1, 384] - - [281, 76.267] - - - [39936, 1024, 1, 384] - - [285, 86.121] - - - [39936, 384, 1, 384] - - [254, 77.887] - - - [36864, 4096, 1, 384] - - [259, 84.148] - - - [36096, 512, 1, 384] - - [254, 78.538] - - - [40320, 512, 1, 384] - - [264, 80.915] - - - [43776, 27521, 1, 384] - - [281, 88.034] - - - [44160, 384, 1, 384] - - [252, 78.312] - - - [41088, 2048, 1, 384] - - [288, 86.645] - - - [41856, 1024, 1, 384] - - [263, 87.124] - - - [37632, 4096, 1, 384] - - [271, 90.013] - - - [44160, 28289, 1, 384] - - [254, 89.054] - - - [38400, 512, 1, 384] - - [263, 82.08] - - - [37632, 1024, 1, 384] - - [271, 85.403] - - - [42240, 512, 1, 384] - - [263, 83.781] - - - [36480, 20609, 1, 384] - - [280, 89.24] - - - [43008, 384, 1, 384] - - [271, 76.864] - - - [38400, 22145, 1, 384] - - [264, 89.347] - - - [36096, 19841, 1, 384] - - [281, 87.621] - - - [38400, 2048, 1, 384] - - [263, 88.327] - - - [42624, 2048, 1, 384] - - [264, 83.262] - - - [36480, 512, 1, 384] - - [254, 79.706] - - - [43776, 27905, 1, 384] - - [259, 88.04] - - - [35712, 19841, 1, 384] - - [256, 89.178] - - - [43008, 1024, 1, 384] - - [263, 86.036] - - - [41472, 384, 1, 384] - - [252, 80.457] - - - [36096, 2048, 1, 384] - - [289, 85.766] - - - [40320, 2048, 1, 384] - - [285, 89.135] - - - [35328, 8192, 1, 384] - - [256, 90.273] - - - [41856, 512, 1, 384] - - [285, 83.581] - - - [35712, 2048, 1, 384] - - [263, 88.452] - - - [36480, 20225, 1, 384] - - [256, 89.099] - - - [43008, 512, 1, 384] - - [252, 80.274] - - - [37248, 4096, 1, 384] - - [280, 89.751] - - - [41472, 512, 1, 384] - - [280, 82.889] - - - [41088, 384, 1, 384] - - [283, 76.69] - - - [37632, 2048, 1, 384] - - [280, 88.738] - - - [36864, 20993, 1, 384] - - [264, 87.812] - - - [40320, 384, 1, 384] - - [254, 78.52] - - - [42624, 8192, 1, 384] - - [252, 89.493] - - - [44160, 4096, 1, 384] - - [263, 90.077] - - - [36480, 1024, 1, 384] - - [263, 85.97] - - - [44544, 1024, 1, 384] - - [263, 85.806] - - - [36864, 384, 1, 384] - - [256, 79.013] - - - [39168, 8192, 1, 384] - - [254, 90.446] - - - [40320, 24065, 1, 384] - - [256, 89.226] - - - [38016, 384, 1, 384] - - [263, 75.268] - - - [37248, 512, 1, 384] - - [252, 80.906] - - - [35712, 8192, 1, 384] - - [252, 90.561] - - - [42240, 1024, 1, 384] - - [280, 87.472] - - - [41856, 2048, 1, 384] - - [285, 88.873] - - - [37632, 384, 1, 384] - - [254, 74.508] - - - [43008, 27137, 1, 384] - - [259, 88.817] - - - [36096, 20225, 1, 384] - - [259, 87.899] - - - [38784, 384, 1, 384] - - [254, 76.379] - - - [42624, 384, 1, 384] - - [263, 75.993] - - - [37632, 8192, 1, 384] - - [280, 90.488] - - - [36864, 1024, 1, 384] - - [271, 84.734] - - - [39168, 23297, 1, 384] - - [254, 89.091] - - - [41472, 2048, 1, 384] - - [285, 88.036] - - - [44160, 1024, 1, 384] - - [280, 85.41] - - - [42240, 384, 1, 384] - - [285, 80.885] - - - [37248, 21377, 1, 384] - - [252, 88.605] - - - [38784, 512, 1, 384] - - [252, 78.508] - - - [43776, 1024, 1, 384] - - [301, 84.107] - - - [40320, 8192, 1, 384] - - [252, 90.614] - - - [38400, 8192, 1, 384] - - [264, 90.376] - - - [42624, 1024, 1, 384] - - [252, 83.626] - - - [43008, 4096, 1, 384] - - [288, 87.372] - - - [41856, 8192, 1, 384] - - [263, 90.813] - - - [41472, 8192, 1, 384] - - [256, 90.544] - - - [37248, 20993, 1, 384] - - [254, 88.42] - - - [36096, 8192, 1, 384] - - [278, 88.991] - - - [41856, 25601, 1, 384] - - [259, 87.17] - - - [40320, 24449, 1, 384] - - [252, 89.215] - - - [35328, 4096, 1, 384] - - [283, 88.908] - - - [38016, 8192, 1, 384] - - [271, 90.868] - - - [39936, 2048, 1, 384] - - [301, 86.374] - - - [37248, 8192, 1, 384] - - [256, 90.594] - - - [26112, 1024, 1, 384] - - [285, 85.092] - - - [13440, 5761, 1, 384] - - [252, 87.024] - - - [21120, 1024, 1, 384] - - [280, 83.912] - - - [31488, 512, 1, 384] - - [263, 81.143] - - - [16896, 9217, 1, 384] - - [280, 86.173] - - - [29568, 1024, 1, 384] - - [283, 83.999] - - - [33408, 17537, 1, 384] - - [252, 89.304] - - - [5760, 1024, 1, 384] - - [280, 61.788] - - - [18816, 4096, 1, 384] - - [271, 88.971] - - - [17664, 9985, 1, 384] - - [263, 88.697] - - - [17664, 1024, 1, 384] - - [263, 78.042] - - - [31488, 384, 1, 384] - - [263, 77.016] - - - [5760, 3841, 1, 384] - - [252, 80.878] - - - [20352, 384, 1, 384] - - [280, 67.827] - - - [7680, 2048, 1, 384] - - [267, 80.651] - - - [22656, 14977, 1, 384] - - [280, 89.781] - - - [6912, 2048, 1, 384] - - [271, 80.824] - - - [31104, 384, 1, 384] - - [280, 75.937] - - - [17280, 2048, 1, 384] - - [285, 85.776] - - - [16896, 2048, 1, 384] - - [280, 84.012] - - - [21504, 512, 1, 384] - - [280, 71.604] - - - [15744, 1024, 1, 384] - - [280, 82.108] - - - [11904, 4096, 1, 384] - - [263, 85.832] - - - [34944, 4096, 1, 384] - - [285, 89.626] - - - [3840, 512, 1, 384] - - [280, 37.763] - - - [11136, 7553, 1, 384] - - [254, 87.11] - - - [17280, 9217, 1, 384] - - [320, 85.933] - - - [18048, 512, 1, 384] - - [287, 68.959] - - - [28416, 1024, 1, 384] - - [280, 82.714] - - - [34560, 18305, 1, 384] - - [264, 89.206] - - - [11520, 512, 1, 384] - - [261, 61.863] - - - [31872, 384, 1, 384] - - [252, 77.442] - - - [19968, 2048, 1, 384] - - [263, 86.677] - - - [11520, 384, 1, 384] - - [262, 59.07] - - - [25728, 9857, 1, 384] - - [256, 88.779] - - - [28032, 12161, 1, 384] - - [256, 89.361] - - - [28416, 12161, 1, 384] - - [280, 89.491] - - - [31488, 15617, 1, 384] - - [264, 89.332] - - - [21120, 2048, 1, 384] - - [253, 86.878] - - - [9600, 6017, 1, 384] - - [256, 86.235] - - - [34176, 512, 1, 384] - - [285, 80.817] - - - [21120, 384, 1, 384] - - [280, 69.931] - - - [31872, 2048, 1, 384] - - [280, 87.371] - - - [33408, 8192, 1, 384] - - [254, 90.609] - - - [13824, 384, 1, 384] - - [250, 68.66] - - - [28032, 1024, 1, 384] - - [263, 85.796] - - - [23040, 2048, 1, 384] - - [263, 85.758] - - - [22272, 14593, 1, 384] - - [285, 89.97] - - - [19584, 384, 1, 384] - - [262, 65.696] - - - [22656, 384, 1, 384] - - [256, 73.659] - - - [7296, 2048, 1, 384] - - [252, 77.098] - - - [6144, 2177, 1, 384] - - [261, 75.794] - - - [23424, 2048, 1, 384] - - [271, 85.898] - - - [6144, 2561, 1, 384] - - [256, 73.855] - - - [16128, 8065, 1, 384] - - [254, 87.841] - - - [18816, 2048, 1, 384] - - [271, 85.312] - - - [12672, 4993, 1, 384] - - [252, 86.778] - - - [13056, 4993, 1, 384] - - [264, 85.015] - - - [30336, 14465, 1, 384] - - [254, 89.258] - - - [17664, 9601, 1, 384] - - [271, 88.671] - - - [8832, 512, 1, 384] - - [261, 60.337] - - - [19200, 1024, 1, 384] - - [263, 82.967] - - - [26112, 10241, 1, 384] - - [259, 86.108] - - - [7296, 384, 1, 384] - - [250, 52.303] - - - [11904, 4225, 1, 384] - - [280, 85.165] - - - [6912, 2945, 1, 384] - - [254, 79.917] - - - [24576, 8705, 1, 384] - - [256, 83.163] - - - [13824, 6145, 1, 384] - - [256, 84.121] - - - [33792, 8192, 1, 384] - - [288, 88.394] - - - [21888, 384, 1, 384] - - [285, 71.162] - - - [23040, 14977, 1, 384] - - [252, 89.321] - - - [7680, 3713, 1, 384] - - [256, 80.768] - - - [34176, 384, 1, 384] - - [280, 74.807] - - - [29952, 4096, 1, 384] - - [285, 89.17] - - - [33408, 2048, 1, 384] - - [285, 88.753] - - - [34560, 4096, 1, 384] - - [263, 89.877] - - - [21120, 13057, 1, 384] - - [254, 89.01] - - - [28800, 12929, 1, 384] - - [256, 88.677] - - - [18816, 10753, 1, 384] - - [252, 89.03] - - - [23040, 512, 1, 384] - - [280, 75.666] - - - [17280, 512, 1, 384] - - [271, 75.461] - - - [12288, 4225, 1, 384] - - [254, 84.633] - - - [30720, 4096, 1, 384] - - [301, 86.995] - - - [15744, 512, 1, 384] - - [254, 69.785] - - - [6912, 3329, 1, 384] - - [252, 78.713] - - - [26496, 384, 1, 384] - - [254, 73.982] - - - [19584, 11905, 1, 384] - - [285, 89.153] - - - [18816, 384, 1, 384] - - [280, 63.634] - - - [16512, 384, 1, 384] - - [317, 62.882] - - - [34944, 1024, 1, 384] - - [285, 86.21] - - - [8832, 2048, 1, 384] - - [285, 77.616] - - - [22272, 1024, 1, 384] - - [285, 83.249] - - - [9216, 2048, 1, 384] - - [250, 78.405] - - - [26880, 8192, 1, 384] - - [280, 90.358] - - - [28800, 12545, 1, 384] - - [256, 88.809] - - - [27264, 11009, 1, 384] - - [254, 88.429] - - - [19200, 4096, 1, 384] - - [280, 88.577] - - - [9216, 5249, 1, 384] - - [264, 84.69] - - - [7680, 384, 1, 384] - - [261, 54.997] - - - [14976, 1024, 1, 384] - - [271, 78.752] - - - [3456, 1921, 1, 384] - - [262, 68.429] - - - [15360, 2048, 1, 384] - - [288, 82.111] - - - [31872, 8192, 1, 384] - - [256, 90.406] - - - [26496, 1024, 1, 384] - - [280, 81.816] - - - [31104, 4096, 1, 384] - - [280, 89.842] - - - [4224, 2689, 1, 384] - - [262, 73.263] - - - [8448, 512, 1, 384] - - [261, 58.027] - - - [27264, 384, 1, 384] - - [280, 75.65] - - - [14976, 6913, 1, 384] - - [254, 87.539] - - - [14208, 4096, 1, 384] - - [280, 87.288] - - - [6528, 2561, 1, 384] - - [254, 77.922] - - - [19968, 384, 1, 384] - - [254, 67.052] - - - [14592, 512, 1, 384] - - [267, 65.731] - - - [23424, 4096, 1, 384] - - [254, 87.192] - - - [21504, 1024, 1, 384] - - [280, 80.566] - - - [2688, 1024, 1, 384] - - [286, 51.497] - - - [20352, 2048, 1, 384] - - [280, 84.991] - - - [15360, 4096, 1, 384] - - [264, 86.111] - - - [14208, 384, 1, 384] - - [256, 57.537] - - - [23808, 512, 1, 384] - - [252, 78.073] - - - [4608, 512, 1, 384] - - [267, 45.066] - - - [11904, 1024, 1, 384] - - [280, 78.276] - - - [14208, 6529, 1, 384] - - [252, 87.106] - - - [26112, 9857, 1, 384] - - [256, 89.098] - - - [33024, 1024, 1, 384] - - [288, 83.484] - - - [13824, 512, 1, 384] - - [290, 73.084] - - - [15744, 384, 1, 384] - - [268, 62.389] - - - [24960, 9089, 1, 384] - - [285, 89.011] - - - [19968, 11905, 1, 384] - - [264, 89.202] - - - [20352, 12289, 1, 384] - - [320, 87.032] - - - [5376, 3457, 1, 384] - - [262, 79.024] - - - [11520, 1024, 1, 384] - - [285, 76.232] - - - [18432, 4096, 1, 384] - - [288, 85.855] - - - [28032, 4096, 1, 384] - - [285, 89.737] - - - [33792, 17537, 1, 384] - - [252, 89.18] - - - [27648, 1024, 1, 384] - - [280, 84.344] - - - [28032, 8192, 1, 384] - - [285, 90.546] - - - [16896, 512, 1, 384] - - [282, 74.187] - - - [3072, 1537, 1, 384] - - [262, 61.903] - - - [9600, 384, 1, 384] - - [262, 50.233] - - - [33024, 512, 1, 384] - - [282, 76.942] - - - [18048, 384, 1, 384] - - [254, 70.676] - - - [23808, 384, 1, 384] - - [264, 68.167] - - - [20736, 13057, 1, 384] - - [264, 89.631] - - - [14592, 6529, 1, 384] - - [264, 87.597] - - - [22656, 14593, 1, 384] - - [285, 89.381] - - - [25728, 9473, 1, 384] - - [252, 88.602] - - - [31488, 1024, 1, 384] - - [263, 86.085] - - - [4608, 2689, 1, 384] - - [282, 71.478] - - - [25728, 8192, 1, 384] - - [285, 90.683] - - - [3456, 1024, 1, 384] - - [307, 64.265] - - - [27264, 8192, 1, 384] - - [252, 90.393] - - - [20736, 12673, 1, 384] - - [264, 89.313] - - - [27648, 11777, 1, 384] - - [254, 88.558] - - - [24576, 2048, 1, 384] - - [256, 79.488] - - - [34944, 19073, 1, 384] - - [252, 89.106] - - - [15360, 7681, 1, 384] - - [254, 87.49] - - - [10368, 1024, 1, 384] - - [280, 77.886] - - - [34176, 2048, 1, 384] - - [320, 88.035] - - - [10752, 1024, 1, 384] - - [263, 71.496] - - - [5760, 2048, 1, 384] - - [250, 75.92] - - - [23424, 384, 1, 384] - - [280, 75.859] - - - [14592, 2048, 1, 384] - - [271, 84.533] - - - [6528, 2048, 1, 384] - - [264, 76.208] - - - [31872, 16001, 1, 384] - - [256, 89.19] - - - [15360, 7297, 1, 384] - - [264, 87.266] - - - [16896, 384, 1, 384] - - [262, 66.557] - - - [24576, 384, 1, 384] - - [264, 69.35] - - - [28416, 4096, 1, 384] - - [280, 89.71] - - - [22656, 512, 1, 384] - - [252, 74.849] - - - [14592, 1024, 1, 384] - - [263, 77.259] - - - [32640, 1024, 1, 384] - - [301, 79.399] - - - [24192, 384, 1, 384] - - [287, 68.969] - - - [25344, 8192, 1, 384] - - [263, 90.336] - - - [18432, 10753, 1, 384] - - [264, 88.005] - - - [28800, 4096, 1, 384] - - [271, 89.288] - - - [22272, 2048, 1, 384] - - [280, 86.157] - - - [29952, 8192, 1, 384] - - [285, 90.427] - - - [27264, 512, 1, 384] - - [280, 78.612] - - - [12288, 384, 1, 384] - - [267, 62.431] - - - [13056, 5377, 1, 384] - - [256, 86.897] - - - [30336, 512, 1, 384] - - [254, 79.244] - - - [14976, 384, 1, 384] - - [290, 59.917] - - - [16128, 512, 1, 384] - - [256, 70.223] - - - [18432, 2048, 1, 384] - - [288, 83.507] - - - [16896, 8833, 1, 384] - - [264, 89.018] - - - [8832, 4865, 1, 384] - - [254, 84.203] - - - [10752, 2048, 1, 384] - - [285, 80.532] - - - [20352, 1024, 1, 384] - - [263, 81.987] - - - [16512, 512, 1, 384] - - [260, 67.144] - - - [23424, 15745, 1, 384] - - [264, 89.763] - - - [32256, 2048, 1, 384] - - [263, 87.968] - - - [14976, 7297, 1, 384] - - [254, 87.868] - - - [16512, 4096, 1, 384] - - [303, 83.999] - - - [34560, 384, 1, 384] - - [263, 75.523] - - - [5376, 384, 1, 384] - - [294, 39.52] - - - [19200, 512, 1, 384] - - [280, 72.362] - - - [21504, 4096, 1, 384] - - [301, 86.954] - - - [4992, 512, 1, 384] - - [286, 48.394] - - - [31488, 15233, 1, 384] - - [251, 89.348] - - - [28800, 1024, 1, 384] - - [263, 83.054] - - - [13824, 4096, 1, 384] - - [285, 87.375] - - - [13056, 512, 1, 384] - - [252, 69.355] - - - [8064, 1024, 1, 384] - - [266, 70.051] - - - [12672, 4609, 1, 384] - - [252, 84.709] - - - [11520, 2048, 1, 384] - - [280, 80.803] - - - [19968, 512, 1, 384] - - [280, 75.065] - - - [30336, 8192, 1, 384] - - [285, 90.763] - - - [25344, 512, 1, 384] - - [271, 74.705] - - - [28416, 384, 1, 384] - - [263, 70.699] - - - [10368, 2048, 1, 384] - - [271, 83.477] - - - [4224, 512, 1, 384] - - [261, 41.356] - - - [23040, 384, 1, 384] - - [301, 74.434] - - - [28800, 384, 1, 384] - - [263, 71.175] - - - [28032, 384, 1, 384] - - [254, 77.554] - - - [23808, 16129, 1, 384] - - [263, 90.108] - - - [9600, 512, 1, 384] - - [261, 64.587] - - - [31488, 8192, 1, 384] - - [256, 90.292] - - - [34944, 8192, 1, 384] - - [264, 90.6] - - - [30720, 2048, 1, 384] - - [301, 85.262] - - - [27648, 8192, 1, 384] - - [301, 88.499] - - - [26112, 4096, 1, 384] - - [263, 88.996] - - - [33024, 17153, 1, 384] - - [259, 88.628] - - - [13440, 2048, 1, 384] - - [280, 82.8] - - - [26496, 10625, 1, 384] - - [252, 88.98] - - - [26112, 384, 1, 384] - - [263, 73.559] - - - [34944, 2048, 1, 384] - - [285, 88.407] - - - [23808, 15745, 1, 384] - - [280, 90.117] - - - [16128, 1024, 1, 384] - - [252, 76.093] - - - [29568, 13313, 1, 384] - - [281, 86.436] - - - [26496, 2048, 1, 384] - - [280, 86.456] - - - [12672, 512, 1, 384] - - [280, 67.275] - - - [17280, 1024, 1, 384] - - [271, 82.053] - - - [31104, 15233, 1, 384] - - [252, 89.393] - - - [33408, 1024, 1, 384] - - [263, 86.352] - - - [7296, 3329, 1, 384] - - [256, 82.434] - - - [8832, 1024, 1, 384] - - [254, 67.434] - - - [11520, 7553, 1, 384] - - [256, 86.704] - - - [20352, 512, 1, 384] - - [271, 76.379] - - - [31488, 2048, 1, 384] - - [285, 88.404] - - - [6144, 2048, 1, 384] - - [271, 78.784] - - - [21504, 13441, 1, 384] - - [252, 88.86] - - - [24960, 8705, 1, 384] - - [280, 88.641] - - - [26880, 11009, 1, 384] - - [252, 89.218] - - - [14976, 2048, 1, 384] - - [280, 82.793] - - - [31872, 512, 1, 384] - - [271, 76.429] - - - [3072, 1024, 1, 384] - - [254, 58.286] - - - [12672, 2048, 1, 384] - - [285, 83.079] - - - [14976, 512, 1, 384] - - [252, 66.951] - - - [31872, 15617, 1, 384] - - [264, 89.141] - - - [11136, 384, 1, 384] - - [250, 57.501] - - - [28800, 2048, 1, 384] - - [271, 87.385] - - - [17280, 4096, 1, 384] - - [280, 87.951] - - - [32640, 2048, 1, 384] - - [255, 83.394] - - - [20352, 4096, 1, 384] - - [285, 88.403] - - - [24192, 4096, 1, 384] - - [271, 89.825] - - - [6912, 512, 1, 384] - - [305, 64.604] - - - [26496, 512, 1, 384] - - [280, 77.278] - - - [33408, 512, 1, 384] - - [264, 79.233] - - - [21888, 14209, 1, 384] - - [268, 84.605] - - - [33792, 17921, 1, 384] - - [252, 88.783] - - - [18432, 384, 1, 384] - - [311, 72.436] - - - [22272, 384, 1, 384] - - [280, 73.09] - - - [6144, 1024, 1, 384] - - [280, 65.55] - - - [19584, 4096, 1, 384] - - [271, 88.664] - - - [29568, 384, 1, 384] - - [252, 72.775] - - - [13056, 384, 1, 384] - - [261, 65.235] - - - [26880, 2048, 1, 384] - - [263, 87.507] - - - [29952, 512, 1, 384] - - [264, 78.472] - - - [29568, 13697, 1, 384] - - [257, 88.429] - - - [27648, 384, 1, 384] - - [252, 76.844] - - - [26880, 1024, 1, 384] - - [280, 82.929] - - - [13824, 1024, 1, 384] - - [280, 80.611] - - - [34560, 2048, 1, 384] - - [280, 87.843] - - - [21504, 13825, 1, 384] - - [256, 88.505] - - - [12288, 4096, 1, 384] - - [252, 85.445] - - - [14208, 512, 1, 384] - - [303, 63.839] - - - [29568, 2048, 1, 384] - - [320, 86.694] - - - [29952, 384, 1, 384] - - [285, 73.967] - - - [30336, 14081, 1, 384] - - [256, 89.07] - - - [33792, 2048, 1, 384] - - [288, 85.928] - - - [8448, 1024, 1, 384] - - [267, 74.445] - - - [18432, 10369, 1, 384] - - [256, 88.481] - - - [23424, 512, 1, 384] - - [280, 76.757] - - - [29184, 12929, 1, 384] - - [278, 89.169] - - - [28416, 512, 1, 384] - - [285, 75.226] - - - [18432, 1024, 1, 384] - - [263, 80.413] - - - [32640, 4096, 1, 384] - - [318, 86.09] - - - [5376, 1024, 1, 384] - - [262, 57.879] - - - [8832, 5249, 1, 384] - - [254, 84.238] - - - [3840, 1024, 1, 384] - - [290, 53.498] - - - [19968, 12289, 1, 384] - - [271, 86.208] - - - [4992, 3073, 1, 384] - - [264, 77.085] - - - [34944, 384, 1, 384] - - [280, 75.839] - - - [29568, 128, 1, 384] - - [261, 47.913] - - - [25728, 128, 1, 384] - - [267, 57.278] - - - [7680, 7680, 1, 256] - - [284, 73.456] - - - [40448, 40448, 1, 256] - - [251, 74.937] - - - [15104, 15104, 1, 256] - - [254, 75.487] - - - [34688, 128, 1, 384] - - [250, 54.573] - - - [35328, 35328, 1, 256] - - [278, 75.386] - - - [31232, 31232, 1, 256] - - [275, 75.543] - - - [16128, 16128, 1, 256] - - [264, 75.578] - - - [21888, 21888, 1, 384] - - [255, 89.215] - - - [20224, 20224, 1, 256] - - [264, 75.749] - - - [29952, 29952, 1, 384] - - [254, 90.917] - - - [33408, 33408, 1, 384] - - [256, 90.847] - - - [23424, 23424, 1, 384] - - [254, 91.11] - - - [21504, 21504, 1, 384] - - [268, 89.214] - - - [40960, 40960, 1, 256] - - [281, 66.623] - - - [25088, 25088, 1, 256] - - [269, 75.905] - - - [35840, 35840, 1, 256] - - [264, 75.487] - - - [26368, 26368, 1, 256] - - [252, 75.32] - - - [28032, 28032, 1, 384] - - [254, 91.053] - - - [27008, 128, 1, 256] - - [260, 50.615] - - - [11008, 11008, 1, 256] - - [312, 73.892] - - - [37248, 37248, 1, 384] - - [252, 90.692] - - - [16640, 16640, 1, 256] - - [252, 75.73] - - - [18688, 18688, 1, 256] - - [254, 75.593] - - - [15488, 128, 1, 256] - - [286, 32.909] - - - [3840, 3840, 1, 256] - - [261, 64.173] - - - [43776, 43776, 1, 256] - - [255, 74.28] - - - [20736, 20736, 1, 256] - - [254, 75.611] - - - [18432, 18432, 1, 256] - - [254, 76.049] - - - [35328, 35328, 1, 384] - - [252, 90.526] - - - [26112, 26112, 1, 384] - - [256, 91.076] - - - [44928, 44928, 1, 384] - - [254, 90.696] - - - [42368, 128, 1, 256] - - [250, 44.691] - - - [17152, 17152, 1, 256] - - [274, 75.562] - - - [29184, 29184, 1, 256] - - [264, 75.558] - - - [27136, 27136, 1, 256] - - [269, 75.911] - - - [33792, 33792, 1, 384] - - [268, 89.232] - - - [14592, 14592, 1, 384] - - [264, 90.309] - - - [14080, 14080, 1, 256] - - [284, 74.446] - - - [16896, 16896, 1, 256] - - [264, 76.106] - - - [28416, 28416, 1, 256] - - [252, 74.869] - - - [27648, 27648, 1, 256] - - [252, 75.552] - - - [21888, 128, 1, 256] - - [262, 43.504] - - - [43776, 43776, 1, 384] - - [268, 89.916] - - - [24320, 24320, 1, 256] - - [251, 75.666] - - - [29184, 29184, 1, 384] - - [256, 90.781] - - - [36480, 36480, 1, 384] - - [252, 90.923] - - - [25728, 25728, 1, 384] - - [285, 90.933] - - - [9728, 9728, 1, 256] - - [256, 74.669] - - - [15488, 128, 1, 384] - - [284, 37.868] - - - [18176, 18176, 1, 256] - - [254, 75.705] - - - [16384, 16384, 1, 256] - - [281, 62.756] - - - [27904, 27904, 1, 256] - - [265, 75.163] - - - [35968, 128, 1, 384] - - [280, 55.639] - - - [14848, 14848, 1, 256] - - [255, 75.624] - - - [4608, 4608, 1, 50000] - - [280, 97.942] - - - [38528, 128, 1, 256] - - [285, 48.177] - - - [16768, 128, 1, 384] - - [284, 40.639] - - - [44160, 44160, 1, 384] - - [256, 90.838] - - - [14592, 14592, 1, 256] - - [295, 74.291] - - - [41216, 41216, 1, 256] - - [252, 74.854] - - - [25344, 25344, 1, 256] - - [268, 74.966] - - - [2560, 2048, 1, 256] - - [276, 56.06] - - - [19200, 19200, 1, 256] - - [264, 75.675] - - - [29952, 29952, 1, 256] - - [274, 75.275] - - - [17408, 17408, 1, 256] - - [252, 75.599] - - - [18816, 18816, 1, 384] - - [285, 90.869] - - - [18048, 18048, 1, 384] - - [285, 90.958] - - - [24576, 24576, 1, 256] - - [281, 70.192] - - - [8448, 8448, 1, 256] - - [284, 73.715] - - - [5376, 5376, 1, 256] - - [276, 69.366] - - - [3584, 3584, 1, 256] - - [262, 62.742] - - - [39936, 39936, 1, 384] - - [288, 89.019] - - - [35584, 35584, 1, 256] - - [264, 74.893] - - - [16896, 16896, 1, 384] - - [264, 90.639] - - - [33280, 33280, 1, 256] - - [251, 75.635] - - - [19456, 19456, 1, 256] - - [254, 75.886] - - - [14208, 128, 1, 256] - - [260, 30.277] - - - [30848, 128, 1, 384] - - [287, 49.625] - - - [39936, 39936, 1, 256] - - [252, 75.103] - - - [9984, 9984, 1, 256] - - [262, 74.421] - - - [23168, 128, 1, 256] - - [274, 45.196] - - - [44800, 44800, 1, 256] - - [252, 74.449] - - - [35712, 35712, 1, 384] - - [252, 90.953] - - - [41088, 41088, 1, 384] - - [268, 90.052] - - - [5120, 5120, 1, 256] - - [262, 70.401] - - - [12288, 12288, 1, 256] - - [254, 75.012] - - - [43264, 43264, 1, 256] - - [252, 74.81] - - - [34176, 34176, 1, 384] - - [254, 90.755] - - - [36608, 36608, 1, 256] - - [265, 74.827] - - - [13824, 13824, 1, 384] - - [256, 90.095] - - - [21120, 21120, 1, 384] - - [264, 91.055] - - - [42240, 42240, 1, 256] - - [254, 74.739] - - - [43648, 128, 1, 256] - - [267, 45.265] - - - [33792, 33792, 1, 256] - - [252, 75.516] - - - [768, 3072, 1, 384] - - [284, 44.769] - - - [20352, 20352, 1, 384] - - [254, 91.117] - - - [33536, 33536, 1, 256] - - [277, 75.178] - - - [1536, 3072, 1, 384] - - [262, 63.449] - - - [2688, 3072, 1, 384] - - [276, 72.101] - - - [6912, 6912, 1, 256] - - [261, 72.526] - - - [18944, 18944, 1, 256] - - [251, 76.049] - - - [31488, 31488, 1, 384] - - [256, 90.749] - - - [44928, 128, 1, 128] - - [267, 31.959] - - - [31872, 31872, 1, 384] - - [256, 90.846] - - - [15616, 15616, 1, 256] - - [252, 75.505] - - - [4352, 4352, 1, 256] - - [261, 67.359] - - - [32128, 128, 1, 256] - - [267, 42.898] - - - [8704, 8704, 1, 256] - - [256, 74.167] - - - [24576, 24576, 1, 384] - - [259, 83.07] - - - [34304, 34304, 1, 256] - - [270, 75.525] - - - [17664, 17664, 1, 384] - - [254, 90.703] - - - [26624, 26624, 1, 256] - - [264, 75.83] - - - [40704, 40704, 1, 256] - - [254, 74.86] - - - [33024, 33024, 1, 384] - - [252, 90.647] - - - [22656, 22656, 1, 384] - - [252, 91.179] - - - [39680, 39680, 1, 256] - - [256, 74.613] - - - [43392, 43392, 1, 384] - - [252, 90.699] - - - [38912, 38912, 1, 256] - - [256, 75.302] - - - [13312, 13312, 1, 256] - - [252, 75.812] - - - [25344, 25344, 1, 384] - - [254, 91.121] - - - [40192, 40192, 1, 256] - - [256, 74.915] - - - [44544, 44544, 1, 256] - - [249, 74.966] - - - [15360, 15360, 1, 384] - - [274, 89.355] - - - [26496, 26496, 1, 384] - - [256, 91.077] - - - [18048, 128, 1, 384] - - [286, 43.408] - - - [2304, 3072, 1, 384] - - [262, 73.922] - - - [43008, 43008, 1, 256] - - [252, 74.922] - - - [36352, 36352, 1, 256] - - [268, 75.406] - - - [23808, 23808, 1, 384] - - [280, 91.162] - - - [37888, 37888, 1, 256] - - [256, 75.212] - - - [13824, 13824, 1, 256] - - [278, 75.675] - - - [37632, 37632, 1, 384] - - [252, 90.859] - - - [14336, 14336, 1, 256] - - [256, 75.668] - - - [16512, 16512, 1, 384] - - [281, 89.482] - - - [21248, 21248, 1, 256] - - [264, 75.578] - - - [28672, 28672, 1, 256] - - [254, 75.173] - - - [9216, 9216, 1, 256] - - [254, 74.101] - - - [30336, 30336, 1, 384] - - [252, 90.946] - - - [7936, 7936, 1, 256] - - [262, 73.5] - - - [44288, 44288, 1, 256] - - [264, 74.601] - - - [7424, 7424, 1, 256] - - [276, 73.14] - - - [43648, 128, 1, 384] - - [263, 52.576] - - - [36096, 36096, 1, 384] - - [255, 89.977] - - - [22528, 22528, 1, 256] - - [256, 75.715] - - - [31744, 31744, 1, 256] - - [264, 75.52] - - - [26880, 26880, 1, 384] - - [264, 91.155] - - - [41088, 128, 1, 256] - - [286, 49.94] - - - [16768, 128, 1, 256] - - [286, 35.074] - - - [44928, 128, 1, 256] - - [254, 46.495] - - - [24832, 24832, 1, 256] - - [251, 75.662] - - - [40320, 40320, 1, 384] - - [256, 90.989] - - - [43008, 2304, 1, 384] - - [256, 89.185] - - - [33408, 1920, 1, 384] - - [256, 87.863] - - - [8064, 2688, 1, 384] - - [264, 79.14] - - - [31872, 1536, 1, 384] - - [252, 85.612] - - - [41088, 2304, 1, 384] - - [301, 87.329] - - - [16128, 1536, 1, 384] - - [256, 83.019] - - - [15360, 768, 1, 384] - - [287, 75.527] - - - [4992, 1536, 1, 384] - - [262, 67.772] - - - [10752, 2688, 1, 384] - - [263, 82.264] - - - [36096, 1536, 1, 384] - - [251, 85.555] - - - [42624, 2688, 1, 384] - - [254, 87.217] - - - [37248, 1536, 1, 384] - - [264, 87.151] - - - [22272, 1920, 1, 384] - - [252, 87.094] - - - [16896, 1920, 1, 384] - - [264, 86.53] - - - [5760, 2304, 1, 384] - - [276, 76.76] - - - [24576, 2304, 1, 384] - - [264, 85.039] - - - [36096, 2304, 1, 384] - - [251, 86.916] - - - [14592, 2304, 1, 384] - - [264, 85.738] - - - [16512, 1920, 1, 384] - - [268, 79.688] - - - [9216, 2688, 1, 384] - - [254, 83.651] - - - [18048, 768, 1, 384] - - [271, 78.673] - - - [14208, 1920, 1, 384] - - [256, 82.418] - - - [17280, 2304, 1, 384] - - [256, 84.863] - - - [18432, 768, 1, 384] - - [271, 79.512] - - - [34944, 1920, 1, 384] - - [264, 87.524] - - - [39936, 2304, 1, 384] - - [252, 88.199] - - - [19968, 2304, 1, 384] - - [254, 86.824] - - - [4224, 768, 1, 384] - - [261, 59.788] - - - [32640, 1920, 1, 384] - - [252, 84.928] - - - [35328, 1536, 1, 384] - - [254, 85.698] - - - [18048, 1920, 1, 384] - - [254, 84.439] - - - [21504, 2304, 1, 384] - - [252, 86.898] - - - [29184, 2688, 1, 384] - - [271, 88.494] - - - [21504, 2688, 1, 384] - - [254, 85.527] - - - [23808, 1536, 1, 384] - - [254, 84.573] - - - [13824, 768, 1, 384] - - [263, 77.721] - - - [20736, 2688, 1, 384] - - [285, 88.463] - - - [16512, 1536, 1, 384] - - [265, 75.676] - - - [5760, 2688, 1, 384] - - [282, 79.485] - - - [15360, 2688, 1, 384] - - [254, 86.326] - - - [43008, 2688, 1, 384] - - [256, 87.686] - - - [6528, 2688, 1, 384] - - [252, 81.588] - - - [40320, 1536, 1, 384] - - [254, 86.967] - - - [40320, 2688, 1, 384] - - [280, 88.955] - - - [24192, 2688, 1, 384] - - [271, 87.204] - - - [23040, 768, 1, 384] - - [285, 81.55] - - - [33024, 1920, 1, 384] - - [264, 86.647] - - - [42624, 1920, 1, 384] - - [264, 87.484] - - - [11520, 1536, 1, 384] - - [256, 82.185] - - - [39552, 1920, 1, 384] - - [254, 87.96] - - - [10752, 1536, 1, 384] - - [252, 77.35] - - - [12672, 2688, 1, 384] - - [280, 86.586] - - - [40704, 1536, 1, 384] - - [252, 87.483] - - - [18816, 2688, 1, 384] - - [271, 85.896] - - - [35712, 2688, 1, 384] - - [280, 88.852] - - - [26880, 1920, 1, 384] - - [252, 87.563] - - - [35328, 2688, 1, 384] - - [271, 88.976] - - - [19584, 1920, 1, 384] - - [254, 86.96] - - - [36864, 768, 1, 384] - - [271, 83.653] - - - [11136, 2304, 1, 384] - - [252, 82.333] - - - [9216, 2304, 1, 384] - - [256, 83.156] - - - [4992, 2688, 1, 384] - - [254, 77.175] - - - [37632, 1536, 1, 384] - - [254, 86.027] - - - [8448, 2688, 1, 384] - - [271, 82.716] - - - [40704, 1920, 1, 384] - - [256, 88.34] - - - [26112, 768, 1, 384] - - [271, 79.562] - - - [32640, 1536, 1, 384] - - [256, 83.298] - - - [13056, 768, 1, 384] - - [280, 73.998] - - - [16896, 768, 1, 384] - - [263, 74.655] - - - [27264, 2304, 1, 384] - - [256, 88.195] - - - [30720, 2688, 1, 384] - - [254, 87.806] - - - [5760, 1920, 1, 384] - - [262, 72.496] - - - [8448, 1920, 1, 384] - - [261, 82.584] - - - [44160, 2304, 1, 384] - - [285, 89.108] - - - [6144, 2688, 1, 384] - - [262, 77.707] - - - [11904, 1920, 1, 384] - - [285, 83.359] - - - [21120, 1920, 1, 384] - - [254, 86.0] - - - [19968, 2688, 1, 384] - - [271, 87.953] - - - [18432, 2304, 1, 384] - - [256, 86.351] - - - [33024, 2688, 1, 384] - - [285, 87.96] - - - [14208, 2688, 1, 384] - - [271, 84.865] - - - [16896, 2688, 1, 384] - - [263, 85.692] - - - [41472, 2304, 1, 384] - - [252, 88.447] - - - [26496, 2688, 1, 384] - - [263, 88.317] - - - [11520, 1920, 1, 384] - - [252, 81.321] - - - [6144, 2304, 1, 384] - - [261, 80.665] - - - [19584, 2304, 1, 384] - - [256, 85.331] - - - [15744, 2688, 1, 384] - - [280, 86.233] - - - [8448, 2304, 1, 384] - - [252, 82.987] - - - [39936, 768, 1, 384] - - [263, 82.561] - - - [38400, 768, 1, 384] - - [280, 83.34] - - - [13824, 1536, 1, 384] - - [252, 82.987] - - - [10368, 768, 1, 384] - - [290, 69.246] - - - [28800, 768, 1, 384] - - [263, 80.331] - - - [21888, 1536, 1, 384] - - [254, 82.833] - - - [27648, 2304, 1, 384] - - [256, 87.437] - - - [37248, 768, 1, 384] - - [271, 85.091] - - - [37632, 1920, 1, 384] - - [252, 87.723] - - - [38016, 1920, 1, 384] - - [252, 88.34] - - - [37248, 2688, 1, 384] - - [271, 88.92] - - - [38400, 2688, 1, 384] - - [280, 88.725] - - - [8832, 1920, 1, 384] - - [280, 79.336] - - - [15744, 768, 1, 384] - - [280, 77.357] - - - [33792, 2304, 1, 384] - - [254, 87.915] - - - [39168, 2688, 1, 384] - - [263, 89.136] - - - [35328, 2304, 1, 384] - - [254, 88.11] - - - [18048, 2688, 1, 384] - - [285, 87.839] - - - [18816, 1536, 1, 384] - - [254, 81.82] - - - [34944, 1536, 1, 384] - - [254, 87.087] - - - [15360, 2304, 1, 384] - - [252, 85.764] - - - [11136, 1536, 1, 384] - - [254, 79.702] - - - [24960, 2688, 1, 384] - - [285, 87.696] - - - [17280, 2688, 1, 384] - - [271, 87.383] - - - [11904, 2688, 1, 384] - - [263, 85.826] - - - [19200, 1536, 1, 384] - - [256, 83.176] - - - [33792, 768, 1, 384] - - [280, 82.575] - - - [2688, 1920, 1, 384] - - [262, 67.714] - - - [12288, 2688, 1, 384] - - [254, 83.987] - - - [28416, 768, 1, 384] - - [271, 80.118] - - - [39936, 2688, 1, 384] - - [254, 88.07] - - - [26880, 1536, 1, 384] - - [264, 86.713] - - - [41472, 2688, 1, 384] - - [285, 89.01] - - - [31104, 2688, 1, 384] - - [280, 88.587] - - - [41856, 2688, 1, 384] - - [263, 89.376] - - - [3456, 2688, 1, 384] - - [262, 70.02] - - - [36480, 1920, 1, 384] - - [271, 88.451] - - - [36864, 2688, 1, 384] - - [264, 87.725] - - - [24064, 3072, 1, 256] - - [258, 72.89] - - - [31744, 3072, 1, 256] - - [256, 73.477] - - - [35328, 22785, 1, 256] - - [300, 73.723] - - - [39168, 1792, 1, 256] - - [286, 71.707] - - - [16640, 4353, 1, 256] - - [300, 69.432] - - - [6912, 3585, 1, 256] - - [256, 65.964] - - - [30976, 18688, 1, 256] - - [255, 74.426] - - - [6400, 1792, 1, 256] - - [262, 62.034] - - - [36864, 24577, 1, 256] - - [259, 71.805] - - - [26368, 1536, 1, 256] - - [276, 69.857] - - - [14336, 1793, 1, 256] - - [252, 64.85] - - - [2560, 3072, 1, 256] - - [262, 56.708] - - - [27136, 1792, 1, 256] - - [261, 71.697] - - - [23296, 1792, 1, 256] - - [261, 70.456] - - - [768, 3072, 1, 256] - - [284, 38.337] - - - [40960, 769, 1, 256] - - [256, 55.161] - - - [7680, 4353, 1, 256] - - [252, 67.882] - - - [7936, 4609, 1, 256] - - [284, 69.169] - - - [38656, 3072, 1, 256] - - [257, 72.725] - - - [13824, 3072, 1, 256] - - [254, 71.6] - - - [35584, 23041, 1, 256] - - [255, 73.114] - - - [37888, 768, 1, 256] - - [271, 68.261] - - - [15872, 9216, 1, 256] - - [264, 74.823] - - - [26368, 14081, 1, 256] - - [281, 72.876] - - - [27648, 15105, 1, 256] - - [252, 74.161] - - - [23296, 9216, 1, 256] - - [256, 74.259] - - - [12032, 1792, 1, 256] - - [276, 69.388] - - - [16128, 768, 1, 256] - - [287, 62.842] - - - [38912, 26369, 1, 256] - - [281, 73.922] - - - [39168, 26880, 1, 256] - - [265, 74.625] - - - [10496, 1792, 1, 256] - - [262, 66.402] - - - [27392, 768, 1, 256] - - [266, 63.787] - - - [11520, 8193, 1, 256] - - [281, 70.412] - - - [14080, 1537, 1, 256] - - [254, 63.305] - - - [19200, 6913, 1, 256] - - [258, 71.548] - - - [5632, 3072, 1, 256] - - [282, 66.804] - - - [29440, 9216, 1, 256] - - [274, 74.429] - - - [10240, 3072, 1, 256] - - [261, 70.816] - - - [42240, 9216, 1, 256] - - [255, 74.271] - - - [27392, 1792, 1, 256] - - [313, 68.339] - - - [29440, 16897, 1, 256] - - [255, 73.235] - - - [11264, 8193, 1, 256] - - [256, 70.854] - - - [28928, 16641, 1, 256] - - [300, 73.123] - - - [44288, 1536, 1, 256] - - [256, 70.471] - - - [18176, 5633, 1, 256] - - [252, 71.129] - - - [17920, 5632, 1, 256] - - [251, 74.331] - - - [39936, 3072, 1, 256] - - [264, 73.824] - - - [24832, 1792, 1, 256] - - [294, 71.196] - - - [7168, 4097, 1, 256] - - [252, 67.186] - - - [13312, 3072, 1, 256] - - [254, 71.791] - - - [12032, 1536, 1, 256] - - [261, 65.445] - - - [44288, 27648, 1, 256] - - [255, 74.594] - - - [23808, 11520, 1, 256] - - [252, 75.256] - - - [40448, 1792, 1, 256] - - [256, 72.271] - - - [22784, 10496, 1, 256] - - [294, 74.874] - - - [9728, 3072, 1, 256] - - [254, 70.812] - - - [7680, 3072, 1, 256] - - [262, 67.931] - - - [12800, 3072, 1, 256] - - [258, 71.387] - - - [12288, 3072, 1, 256] - - [254, 71.431] - - - [21504, 8961, 1, 256] - - [252, 73.17] - - - [3584, 513, 1, 256] - - [260, 30.521] - - - [13056, 9216, 1, 256] - - [264, 74.371] - - - [33024, 1536, 1, 256] - - [282, 69.126] - - - [44032, 27648, 1, 256] - - [254, 74.877] - - - [32000, 19712, 1, 256] - - [264, 75.0] - - - [24064, 9216, 1, 256] - - [270, 74.935] - - - [27392, 3072, 1, 256] - - [270, 70.269] - - - [8960, 5633, 1, 256] - - [256, 70.165] - - - [23808, 9216, 1, 256] - - [252, 74.572] - - - [11776, 1536, 1, 256] - - [261, 64.33] - - - [30208, 9216, 1, 256] - - [269, 75.008] - - - [12800, 513, 1, 256] - - [264, 48.009] - - - [42496, 2305, 1, 256] - - [264, 68.926] - - - [20224, 1792, 1, 256] - - [261, 70.026] - - - [26368, 768, 1, 256] - - [267, 66.092] - - - [24320, 12033, 1, 256] - - [300, 73.218] - - - [36352, 1792, 1, 256] - - [256, 72.37] - - - [19712, 7424, 1, 256] - - [286, 73.379] - - - [10752, 1536, 1, 256] - - [256, 64.306] - - - [29696, 17409, 1, 256] - - [281, 72.74] - - - [18944, 1536, 1, 256] - - [276, 68.276] - - - [37376, 25089, 1, 256] - - [268, 73.77] - - - [8448, 1792, 1, 256] - - [282, 64.45] - - - [24320, 11777, 1, 256] - - [300, 73.035] - - - [8192, 4865, 1, 256] - - [264, 69.019] - - - [32000, 19713, 1, 256] - - [281, 73.259] - - - [31232, 3072, 1, 256] - - [294, 73.089] - - - [43776, 3585, 1, 256] - - [255, 68.553] - - - [29696, 17408, 1, 256] - - [254, 75.352] - - - [32768, 20481, 1, 256] - - [293, 55.844] - - - [19456, 7168, 1, 256] - - [254, 74.12] - - - [22272, 768, 1, 256] - - [287, 65.541] - - - [16896, 1792, 1, 256] - - [282, 70.328] - - - [37888, 3072, 1, 256] - - [252, 73.795] - - - [36608, 24065, 1, 256] - - [255, 72.88] - - - [28928, 16640, 1, 256] - - [274, 75.132] - - - [17152, 3072, 1, 256] - - [276, 71.874] - - - [42496, 9216, 1, 256] - - [255, 74.809] - - - [4352, 1792, 1, 256] - - [261, 55.99] - - - [40704, 513, 1, 256] - - [285, 56.078] - - - [33536, 20993, 1, 256] - - [300, 73.277] - - - [3584, 3072, 1, 256] - - [252, 59.974] - - - [44032, 9216, 1, 256] - - [252, 74.553] - - - [32512, 20225, 1, 256] - - [300, 73.399] - - - [17664, 5377, 1, 256] - - [257, 70.624] - - - [8448, 5121, 1, 256] - - [256, 68.835] - - - [40448, 513, 1, 256] - - [263, 56.249] - - - [17920, 3072, 1, 256] - - [284, 72.264] - - - [34816, 22273, 1, 256] - - [281, 74.005] - - - [39168, 9216, 1, 256] - - [255, 74.256] - - - [11776, 3072, 1, 256] - - [252, 70.075] - - - [24576, 12033, 1, 256] - - [259, 68.22] - - - [32512, 3072, 1, 256] - - [284, 72.723] - - - [22016, 9473, 1, 256] - - [251, 73.241] - - - [29440, 3072, 1, 256] - - [257, 73.013] - - - [41472, 768, 1, 256] - - [280, 68.878] - - - [35840, 3072, 1, 256] - - [254, 73.788] - - - [9728, 1792, 1, 256] - - [276, 67.491] - - - [32768, 3072, 1, 256] - - [249, 59.221] - - - [36864, 24576, 1, 256] - - [252, 74.304] - - - [12032, 8961, 1, 256] - - [274, 71.392] - - - [20224, 7937, 1, 256] - - [281, 72.001] - - - [30720, 18432, 1, 256] - - [254, 75.598] - - - [27136, 14848, 1, 256] - - [274, 75.662] - - - [18688, 6145, 1, 256] - - [259, 70.227] - - - [36608, 9216, 1, 256] - - [255, 74.237] - - - [29952, 17664, 1, 256] - - [251, 75.301] - - - [24064, 11776, 1, 256] - - [251, 75.569] - - - [18688, 6401, 1, 256] - - [281, 70.993] - - - [18432, 3072, 1, 256] - - [252, 72.829] - - - [25344, 768, 1, 256] - - [296, 66.478] - - - [34816, 9216, 1, 256] - - [252, 74.783] - - - [30464, 9216, 1, 256] - - [268, 73.654] - - - [20480, 1792, 1, 256] - - [276, 70.652] - - - [18688, 1792, 1, 256] - - [276, 70.3] - - - [38144, 3072, 1, 256] - - [264, 73.144] - - - [35584, 23296, 1, 256] - - [277, 74.79] - - - [40704, 1536, 1, 256] - - [256, 71.038] - - - [32256, 19969, 1, 256] - - [300, 73.669] - - - [37632, 9216, 1, 256] - - [256, 73.983] - - - [17920, 5377, 1, 256] - - [251, 71.454] - - - [34304, 22016, 1, 256] - - [269, 75.492] - - - [22272, 1536, 1, 256] - - [258, 69.158] - - - [19712, 3072, 1, 256] - - [278, 70.534] - - - [42496, 27648, 1, 256] - - [255, 75.01] - - - [42496, 2561, 1, 256] - - [252, 68.824] - - - [43008, 3073, 1, 256] - - [281, 69.489] - - - [24832, 3072, 1, 256] - - [284, 72.622] - - - [1024, 2048, 1, 256] - - [299, 34.714] - - - [9984, 1792, 1, 256] - - [282, 67.945] - - - [11008, 768, 1, 256] - - [299, 56.84] - - - [17920, 9216, 1, 256] - - [301, 74.862] - - - [19712, 9216, 1, 256] - - [268, 74.04] - - - [36864, 1792, 1, 256] - - [254, 72.242] - - - [39936, 9216, 1, 256] - - [252, 74.655] - - - [17664, 5121, 1, 256] - - [278, 69.096] - - - [27136, 14849, 1, 256] - - [270, 73.926] - - - [33024, 3072, 1, 256] - - [251, 72.339] - - - [6912, 1792, 1, 256] - - [262, 64.897] - - - [34048, 3072, 1, 256] - - [258, 72.477] - - - [14080, 9216, 1, 256] - - [278, 73.043] - - - [43776, 9216, 1, 256] - - [268, 74.086] - - - [37120, 24577, 1, 256] - - [268, 72.438] - - - [37632, 3072, 1, 256] - - [252, 72.923] - - - [5120, 3072, 1, 256] - - [262, 66.228] - - - [24064, 11521, 1, 256] - - [257, 73.393] - - - [21760, 9472, 1, 256] - - [256, 75.121] - - - [8192, 1536, 1, 256] - - [254, 63.981] - - - [21248, 9216, 1, 256] - - [254, 74.493] - - - [35072, 22529, 1, 256] - - [281, 72.351] - - - [35840, 9216, 1, 256] - - [252, 74.674] - - - [26368, 9216, 1, 256] - - [256, 74.3] - - - [26112, 1792, 1, 256] - - [286, 71.895] - - - [43264, 27648, 1, 256] - - [254, 74.283] - - - [38912, 3072, 1, 256] - - [254, 73.845] - - - [37376, 9216, 1, 256] - - [268, 74.853] - - - [35840, 23552, 1, 256] - - [256, 75.421] - - - [34816, 3072, 1, 256] - - [252, 73.722] - - - [36864, 1536, 1, 256] - - [254, 70.739] - - - [31232, 18945, 1, 256] - - [268, 73.789] - - - [19968, 7681, 1, 256] - - [268, 72.249] - - - [33280, 1792, 1, 256] - - [258, 72.275] - - - [30976, 18689, 1, 256] - - [268, 72.645] - - - [31488, 3072, 1, 256] - - [264, 72.752] - - - [11520, 8449, 1, 256] - - [257, 71.344] - - - [18176, 5889, 1, 256] - - [252, 70.944] - - - [39936, 27393, 1, 256] - - [281, 73.828] - - - [39168, 26625, 1, 256] - - [255, 72.658] - - - [13824, 1536, 1, 256] - - [254, 68.41] - - - [37632, 25345, 1, 256] - - [281, 73.074] - - - [23552, 3072, 1, 256] - - [252, 73.223] - - - [5888, 2561, 1, 256] - - [261, 63.656] - - - [22272, 1792, 1, 256] - - [261, 70.322] - - - [3840, 769, 1, 256] - - [261, 46.289] - - - [24576, 3072, 1, 256] - - [249, 68.65] - - - [25344, 12801, 1, 256] - - [255, 72.186] - - - [17664, 3072, 1, 256] - - [284, 71.451] - - - [40960, 3072, 1, 256] - - [265, 66.819] - - - [22784, 10497, 1, 256] - - [268, 72.571] - - - [39680, 27137, 1, 256] - - [259, 72.405] - - - [24320, 1536, 1, 256] - - [284, 69.964] - - - [17664, 5376, 1, 256] - - [284, 73.382] - - - [23296, 10753, 1, 256] - - [259, 72.528] - - - [27648, 15361, 1, 256] - - [259, 72.759] - - - [15104, 2817, 1, 256] - - [256, 68.216] - - - [24064, 11777, 1, 256] - - [270, 73.397] - - - [38912, 9216, 1, 256] - - [256, 74.717] - - - [29440, 1792, 1, 256] - - [282, 71.532] - - - [14080, 1793, 1, 256] - - [254, 65.232] - - - [17920, 768, 1, 256] - - [267, 64.388] - - - [40704, 27648, 1, 256] - - [254, 74.406] - - - [38144, 9216, 1, 256] - - [268, 74.19] - - - [2304, 3072, 1, 256] - - [276, 60.894] - - - [39424, 1792, 1, 256] - - [286, 72.285] - - - [11264, 3072, 1, 256] - - [256, 70.634] - - - [24576, 1792, 1, 256] - - [252, 67.901] - - - [14848, 2561, 1, 256] - - [254, 66.512] - - - [23552, 11009, 1, 256] - - [281, 73.699] - - - [16640, 1792, 1, 256] - - [282, 67.869] - - - [38656, 26369, 1, 256] - - [255, 72.925] - - - [28928, 1792, 1, 256] - - [254, 71.293] - - - [3840, 1792, 1, 256] - - [261, 60.099] - - - [27904, 1536, 1, 256] - - [294, 70.041] - - - [14592, 2305, 1, 256] - - [258, 66.345] - - - [26368, 13825, 1, 256] - - [300, 72.918] - - - [30720, 18177, 1, 256] - - [281, 74.079] - - - [24320, 3072, 1, 256] - - [252, 72.794] - - - [44032, 4097, 1, 256] - - [281, 70.115] - - - [34048, 21761, 1, 256] - - [255, 72.608] - - - [24832, 12545, 1, 256] - - [300, 73.308] - - - [40448, 257, 1, 256] - - [263, 45.308] - - - [8192, 5121, 1, 256] - - [252, 68.016] - - - [5888, 2817, 1, 256] - - [282, 63.618] - - - [16384, 1792, 1, 256] - - [249, 60.792] - - - [36352, 24064, 1, 256] - - [255, 75.214] - - - [8704, 1536, 1, 256] - - [282, 63.632] - - - [12032, 3072, 1, 256] - - [254, 70.991] - - - [28160, 1536, 1, 256] - - [281, 70.098] - - - [28416, 1792, 1, 256] - - [258, 70.953] - - - [28928, 16385, 1, 256] - - [281, 72.397] - - - [26368, 1792, 1, 256] - - [282, 71.252] - - - [25600, 13312, 1, 256] - - [254, 75.771] - - - [20736, 9216, 1, 256] - - [264, 74.492] - - - [43264, 3073, 1, 256] - - [281, 69.053] - - - [4864, 1793, 1, 256] - - [261, 53.98] - - - [20992, 3072, 1, 256] - - [252, 72.977] - - - [16640, 4097, 1, 256] - - [281, 67.999] - - - [38656, 1536, 1, 256] - - [258, 70.607] - - - [12544, 1792, 1, 256] - - [261, 68.266] - - - [17152, 4609, 1, 256] - - [254, 70.191] - - - [34560, 22017, 1, 256] - - [259, 73.029] - - - [14848, 768, 1, 256] - - [267, 60.74] - - - [8448, 5377, 1, 256] - - [258, 70.346] - - - [33792, 21504, 1, 256] - - [254, 75.406] - - - [37376, 1792, 1, 256] - - [252, 72.422] - - - [4096, 1025, 1, 256] - - [262, 45.482] - - - [19200, 3072, 1, 256] - - [276, 71.676] - - - [36864, 9216, 1, 256] - - [249, 74.017] - - - [44288, 9216, 1, 256] - - [268, 74.399] - - - [26112, 3072, 1, 256] - - [254, 73.247] - - - [4864, 1537, 1, 256] - - [282, 53.44] - - - [20224, 7681, 1, 256] - - [255, 71.553] - - - [21504, 3072, 1, 256] - - [252, 73.129] - - - [32256, 19713, 1, 256] - - [297, 73.799] - - - [18176, 5888, 1, 256] - - [261, 73.909] - - - [26624, 14336, 1, 256] - - [264, 75.385] - - - [7936, 1792, 1, 256] - - [261, 66.373] - - - [18944, 3072, 1, 256] - - [286, 72.445] - - - [37120, 24832, 1, 256] - - [254, 74.866] - - - [16896, 4609, 1, 256] - - [257, 70.699] - - - [39936, 768, 1, 256] - - [267, 68.67] - - - [26624, 9216, 1, 256] - - [264, 74.827] - - - [36352, 9216, 1, 256] - - [255, 74.823] - - - [37888, 25600, 1, 256] - - [254, 75.356] - - - [44544, 9216, 1, 256] - - [255, 74.614] - - - [33536, 21249, 1, 256] - - [297, 73.389] - - - [9984, 768, 1, 256] - - [250, 54.574] - - - [41728, 3072, 1, 256] - - [257, 71.64] - - - [32512, 768, 1, 256] - - [266, 67.213] - - - [14848, 9216, 1, 256] - - [268, 74.684] - - - [13568, 9216, 1, 256] - - [256, 74.107] - - - [42752, 3072, 1, 256] - - [252, 73.224] - - - [14592, 1792, 1, 256] - - [284, 68.208] - - - [26624, 1792, 1, 256] - - [254, 72.142] - - - [36096, 23808, 1, 256] - - [268, 74.247] - - - [43520, 3072, 1, 256] - - [254, 73.788] - - - [43008, 1792, 1, 256] - - [252, 72.967] - - - [35328, 23040, 1, 256] - - [278, 75.432] - - - [16128, 3585, 1, 256] - - [252, 69.134] - - - [17408, 5121, 1, 256] - - [281, 70.042] - - - [13056, 768, 1, 256] - - [280, 61.047] - - - [43520, 3585, 1, 256] - - [281, 70.545] - - - [27904, 15617, 1, 256] - - [269, 73.165] - - - [38912, 1536, 1, 256] - - [256, 71.532] - - - [18432, 1792, 1, 256] - - [252, 70.635] - - - [4352, 1025, 1, 256] - - [250, 47.919] - - - [28160, 15617, 1, 256] - - [281, 73.595] - - - [7424, 4353, 1, 256] - - [261, 68.112] - - - [27648, 15360, 1, 256] - - [256, 75.75] - - - [21248, 8961, 1, 256] - - [297, 72.223] - - - [26112, 13569, 1, 256] - - [270, 73.707] - - - [6656, 3585, 1, 256] - - [286, 66.739] - - - [9472, 768, 1, 256] - - [280, 52.718] - - - [15616, 3072, 1, 256] - - [264, 71.795] - - - [11520, 3072, 1, 256] - - [294, 70.214] - - - [28416, 768, 1, 256] - - [267, 66.259] - - - [20736, 1536, 1, 256] - - [282, 69.294] - - - [36608, 24321, 1, 256] - - [268, 72.971] - - - [10752, 768, 1, 256] - - [267, 58.055] - - - [9216, 5889, 1, 256] - - [282, 70.705] - - - [38144, 25856, 1, 256] - - [264, 74.767] - - - [29184, 16896, 1, 256] - - [252, 75.487] - - - [32768, 20225, 1, 256] - - [322, 57.927] - - - [15872, 3585, 1, 256] - - [252, 69.965] - - - [15360, 9216, 1, 256] - - [264, 74.876] - - - [30208, 1792, 1, 256] - - [256, 71.803] - - - [32256, 19968, 1, 256] - - [278, 75.446] - - - [41728, 1793, 1, 256] - - [252, 66.538] - - - [40192, 257, 1, 256] - - [252, 44.954] - - - [40704, 768, 1, 256] - - [260, 68.398] - - - [24576, 9216, 1, 256] - - [265, 70.049] - - - [29696, 9216, 1, 256] - - [254, 74.863] - - - [30208, 17920, 1, 256] - - [251, 75.773] - - - [22016, 9728, 1, 256] - - [264, 75.247] - - - [19200, 6657, 1, 256] - - [297, 71.173] - - - [18176, 768, 1, 256] - - [250, 64.992] - - - [12288, 1792, 1, 256] - - [262, 67.603] - - - [14848, 2305, 1, 256] - - [252, 66.318] - - - [8192, 3072, 1, 256] - - [254, 69.015] - - - [21760, 3072, 1, 256] - - [262, 72.375] - - - [29184, 1536, 1, 256] - - [258, 70.393] - - - [30720, 9216, 1, 256] - - [254, 74.894] - - - [9728, 1536, 1, 256] - - [262, 63.991] - - - [12032, 8705, 1, 256] - - [264, 71.355] - - - [18688, 1536, 1, 256] - - [282, 69.385] - - - [38656, 768, 1, 256] - - [260, 68.285] - - - [36352, 3072, 1, 256] - - [254, 73.493] - - - [3840, 513, 1, 256] - - [261, 32.748] - - - [5376, 768, 1, 256] - - [276, 44.727] - - - [17408, 5120, 1, 256] - - [254, 74.321] - - - [35072, 1792, 1, 256] - - [256, 71.863] - - - [11008, 7937, 1, 256] - - [298, 70.797] - - - [13568, 3072, 1, 256] - - [261, 71.084] - - - [34048, 768, 1, 256] - - [267, 67.567] - - - [28416, 16129, 1, 256] - - [281, 73.072] - - - [22528, 3072, 1, 256] - - [254, 73.326] - - - [35584, 3072, 1, 256] - - [256, 72.72] - - - [30464, 1792, 1, 256] - - [258, 70.013] - - - [20992, 9216, 1, 256] - - [268, 75.069] - - - [31488, 9216, 1, 256] - - [254, 74.117] - - - [41728, 1537, 1, 256] - - [254, 65.171] - - - [30464, 768, 1, 256] - - [260, 66.682] - - - [41472, 1537, 1, 256] - - [264, 66.289] - - - [39424, 27137, 1, 256] - - [268, 73.628] - - - [32000, 3072, 1, 256] - - [252, 72.768] - - - [15872, 3072, 1, 256] - - [252, 72.001] - - - [10496, 7425, 1, 256] - - [281, 71.106] - - - [9984, 6913, 1, 256] - - [252, 71.091] - - - [14336, 1792, 1, 256] - - [276, 69.041] - - - [30464, 18176, 1, 256] - - [255, 74.481] - - - [1536, 2048, 1, 256] - - [284, 49.311] - - - [42240, 3072, 1, 256] - - [256, 73.139] - - - [41984, 2049, 1, 256] - - [301, 66.57] - - - [13824, 1537, 1, 256] - - [254, 63.771] - - - [36096, 23809, 1, 256] - - [268, 72.71] - - - [25600, 768, 1, 256] - - [287, 67.051] - - - [25600, 9216, 1, 256] - - [256, 75.002] - - - [15616, 9216, 1, 256] - - [256, 74.411] - - - [7168, 3841, 1, 256] - - [256, 67.57] - - - [6144, 3073, 1, 256] - - [264, 64.657] - - - [18688, 6400, 1, 256] - - [264, 74.512] - - - [21760, 768, 1, 256] - - [287, 64.205] - - - [21760, 9216, 1, 256] - - [252, 74.502] - - - [42496, 3072, 1, 256] - - [264, 73.716] - - - [27392, 15105, 1, 256] - - [255, 72.88] - - - [44032, 3072, 1, 256] - - [264, 73.868] - - - [8448, 1536, 1, 256] - - [276, 61.889] - - - [23552, 768, 1, 256] - - [285, 64.869] - - - [44288, 4353, 1, 256] - - [281, 70.362] - - - [15616, 3073, 1, 256] - - [254, 68.049] - - - [38144, 25857, 1, 256] - - [255, 72.949] - - - [6144, 2817, 1, 256] - - [264, 62.577] - - - [39168, 26881, 1, 256] - - [268, 72.928] - - - [32512, 19969, 1, 256] - - [270, 73.225] - - - [26112, 13825, 1, 256] - - [270, 73.821] - - - [9728, 6657, 1, 256] - - [254, 71.388] - - - [24832, 9216, 1, 256] - - [256, 74.486] - - - [33280, 20992, 1, 256] - - [278, 75.582] - - - [28672, 16384, 1, 256] - - [254, 74.944] - - - [7936, 768, 1, 256] - - [254, 52.95] - - - [25088, 3072, 1, 256] - - [256, 72.818] - - - [44800, 3072, 1, 256] - - [252, 72.719] - - - [30464, 18177, 1, 256] - - [255, 72.706] - - - [7168, 768, 1, 256] - - [254, 49.311] - - - [33280, 20737, 1, 256] - - [297, 73.951] - - - [27648, 9216, 1, 256] - - [264, 74.837] - - - [28160, 9216, 1, 256] - - [275, 74.804] - - - [22016, 1792, 1, 256] - - [256, 71.088] - - - [21504, 9217, 1, 256] - - [259, 72.143] - - - [33536, 21248, 1, 256] - - [278, 75.086] - - - [28672, 9216, 1, 256] - - [264, 74.277] - - - [44544, 3072, 1, 256] - - [252, 73.355] - - - [35840, 23297, 1, 256] - - [259, 74.027] - - - [13312, 1025, 1, 256] - - [263, 58.916] - - - [17920, 5633, 1, 256] - - [286, 71.536] - - - [1024, 3072, 1, 256] - - [299, 49.177] - - - [21760, 9473, 1, 256] - - [259, 72.475] - - - [16896, 1536, 1, 256] - - [282, 68.548] - - - [23552, 11264, 1, 256] - - [264, 75.562] - - - [19712, 7425, 1, 256] - - [300, 71.126] - - - [24320, 9216, 1, 256] - - [274, 74.48] - - - [17408, 3072, 1, 256] - - [264, 72.646] - - - [14848, 2560, 1, 256] - - [261, 71.053] - - - [15104, 9216, 1, 256] - - [252, 74.312] - - - [38912, 26625, 1, 256] - - [281, 72.727] - - - [39424, 3072, 1, 256] - - [256, 73.417] - - - [18944, 6657, 1, 256] - - [251, 72.217] - - - [18176, 1792, 1, 256] - - [282, 69.577] - - - [42752, 2561, 1, 256] - - [259, 68.13] - - - [43776, 1792, 1, 256] - - [268, 69.792] - - - [33792, 21505, 1, 256] - - [259, 72.863] - - - [29952, 3072, 1, 256] - - [257, 72.875] - - - [20480, 9216, 1, 256] - - [264, 74.621] - - - [23040, 10753, 1, 256] - - [300, 73.355] - - - [16640, 4352, 1, 256] - - [261, 72.785] - - - [16128, 9216, 1, 256] - - [252, 74.406] - - - [39424, 9216, 1, 256] - - [268, 74.936] - - - [28416, 3072, 1, 256] - - [274, 72.357] - - - [43264, 1792, 1, 256] - - [252, 72.181] - - - [30976, 9216, 1, 256] - - [268, 73.289] - - - [31232, 1792, 1, 256] - - [262, 71.89] - - - [9984, 3072, 1, 256] - - [252, 69.262] - - - [37120, 1536, 1, 256] - - [276, 70.444] - - - [7424, 768, 1, 256] - - [287, 50.368] - - - [27392, 9216, 1, 256] - - [268, 74.083] - - - [19200, 9216, 1, 256] - - [255, 74.336] - - - [16128, 3841, 1, 256] - - [254, 69.422] - - - [43776, 3841, 1, 256] - - [255, 69.142] - - - [20224, 3072, 1, 256] - - [256, 72.17] - - - [20736, 8449, 1, 256] - - [259, 72.238] - - - [36864, 24321, 1, 256] - - [259, 73.196] - - - [7424, 4097, 1, 256] - - [252, 66.537] - - - [12032, 768, 1, 256] - - [290, 57.752] - - - [6400, 3073, 1, 256] - - [252, 63.363] - - - [8704, 1792, 1, 256] - - [254, 65.844] - - - [16384, 3841, 1, 256] - - [265, 58.468] - - - [23040, 1792, 1, 256] - - [276, 71.001] - - - [26368, 14080, 1, 256] - - [252, 75.251] - - - [33024, 20736, 1, 256] - - [270, 75.079] - - - [9472, 6145, 1, 256] - - [254, 69.301] - - - [41984, 27648, 1, 256] - - [264, 74.925] - - - [33280, 1536, 1, 256] - - [276, 71.098] - - - [22528, 1792, 1, 256] - - [261, 71.567] - - - [30976, 1024, 1, 128] - - [254, 42.513] - - - [33024, 17025, 1, 128] - - [254, 46.355] - - - [30848, 1024, 1, 128] - - [287, 42.967] - - - [39552, 23553, 1, 128] - - [249, 45.133] - - - [29952, 14081, 1, 128] - - [252, 45.934] - - - [40320, 128, 1, 128] - - [261, 32.824] - - - [36096, 1024, 1, 128] - - [286, 40.444] - - - [36096, 20097, 1, 128] - - [291, 44.703] - - - [39552, 23681, 1, 128] - - [249, 45.361] - - - [40320, 4096, 1, 128] - - [254, 45.636] - - - [29824, 2048, 1, 128] - - [294, 43.127] - - - [42752, 26753, 1, 128] - - [249, 45.655] - - - [36480, 20481, 1, 128] - - [277, 45.313] - - - [38784, 4096, 1, 128] - - [261, 45.248] - - - [34560, 128, 1, 128] - - [261, 29.604] - - - [29056, 2048, 1, 128] - - [262, 43.507] - - - [38272, 2048, 1, 128] - - [323, 36.638] - - - [30848, 512, 1, 128] - - [261, 40.458] - - - [40448, 24577, 1, 128] - - [291, 45.618] - - - [30208, 14209, 1, 128] - - [257, 46.282] - - - [31360, 512, 1, 128] - - [256, 40.429] - - - [39680, 1024, 1, 128] - - [261, 42.937] - - - [39552, 4096, 1, 128] - - [265, 45.488] - - - [40832, 24833, 1, 128] - - [265, 45.035] - - - [44672, 1024, 1, 128] - - [287, 43.451] - - - [40704, 4096, 1, 128] - - [254, 45.772] - - - [30720, 14849, 1, 128] - - [265, 46.757] - - - [30208, 1024, 1, 128] - - [262, 42.874] - - - [38528, 1024, 1, 128] - - [263, 42.892] - - - [38400, 22529, 1, 128] - - [291, 45.897] - - - [32256, 128, 1, 128] - - [285, 28.12] - - - [34560, 8192, 1, 128] - - [254, 46.641] - - - [33536, 128, 1, 128] - - [276, 28.894] - - - [32896, 4096, 1, 128] - - [301, 43.557] - - - [29440, 4096, 1, 128] - - [265, 45.779] - - - [36992, 512, 1, 128] - - [299, 39.86] - - - [41728, 128, 1, 128] - - [286, 33.563] - - - [31360, 1024, 1, 128] - - [287, 42.98] - - - [38016, 2048, 1, 128] - - [290, 44.223] - - - [30464, 512, 1, 128] - - [287, 39.585] - - - [40448, 4096, 1, 128] - - [278, 46.116] - - - [41088, 8192, 1, 128] - - [288, 45.778] - - - [40832, 4096, 1, 128] - - [252, 45.606] - - - [36352, 1024, 1, 128] - - [267, 43.102] - - - [30976, 512, 1, 128] - - [267, 40.197] - - - [29056, 1024, 1, 128] - - [267, 42.348] - - - [28928, 512, 1, 128] - - [261, 39.211] - - - [34816, 4096, 1, 128] - - [254, 46.696] - - - [29696, 8192, 1, 128] - - [252, 47.304] - - - [35584, 2048, 1, 128] - - [290, 44.122] - - - [33280, 17281, 1, 128] - - [274, 46.514] - - - [42112, 8192, 1, 128] - - [252, 46.462] - - - [41600, 4096, 1, 128] - - [265, 45.603] - - - [29952, 1024, 1, 128] - - [262, 42.84] - - - [31744, 512, 1, 128] - - [250, 40.693] - - - [41216, 128, 1, 128] - - [271, 33.152] - - - [32512, 2048, 1, 128] - - [271, 44.312] - - - [32640, 4096, 1, 128] - - [254, 45.254] - - - [32256, 512, 1, 128] - - [284, 40.905] - - - [35968, 128, 1, 128] - - [267, 30.387] - - - [42880, 512, 1, 128] - - [282, 40.788] - - - [43904, 1024, 1, 128] - - [267, 42.98] - - - [41856, 512, 1, 128] - - [280, 42.279] - - - [33408, 17537, 1, 128] - - [265, 45.966] - - - [41216, 25345, 1, 128] - - [291, 45.806] - - - [40192, 24193, 1, 128] - - [265, 45.525] - - - [38272, 1024, 1, 128] - - [257, 38.221] - - - [31104, 8192, 1, 128] - - [254, 46.434] - - - [29312, 2048, 1, 128] - - [285, 43.896] - - - [42240, 4096, 1, 128] - - [252, 45.827] - - - [29056, 512, 1, 128] - - [287, 39.186] - - - [30464, 128, 1, 128] - - [250, 26.971] - - - [43904, 28033, 1, 128] - - [254, 45.375] - - - [41088, 4096, 1, 128] - - [263, 45.105] - - - [36736, 128, 1, 128] - - [262, 30.789] - - - [43008, 128, 1, 128] - - [261, 31.442] - - - [31104, 2048, 1, 128] - - [287, 44.22] - - - [43392, 27521, 1, 128] - - [265, 45.358] - - - [33920, 17921, 1, 128] - - [265, 45.594] - - - [43264, 512, 1, 128] - - [282, 40.916] - - - [31616, 2048, 1, 128] - - [265, 43.609] - - - [43136, 27265, 1, 128] - - [249, 45.393] - - - [40320, 24449, 1, 128] - - [265, 45.352] - - - [35072, 1024, 1, 128] - - [254, 43.152] - - - [44160, 8192, 1, 128] - - [252, 46.21] - - - [37632, 1024, 1, 128] - - [261, 43.181] - - - [38400, 8192, 1, 128] - - [257, 46.925] - - - [32512, 16641, 1, 128] - - [278, 46.246] - - - [30976, 8192, 1, 128] - - [303, 45.168] - - - [38656, 512, 1, 128] - - [285, 40.966] - - - [33664, 17793, 1, 128] - - [249, 45.754] - - - [34944, 1024, 1, 128] - - [261, 43.102] - - - [43392, 8192, 1, 128] - - [264, 46.144] - - - [29440, 8192, 1, 128] - - [249, 46.686] - - - [34304, 18433, 1, 128] - - [277, 46.162] - - - [29184, 4096, 1, 128] - - [284, 45.73] - - - [39040, 128, 1, 128] - - [266, 31.994] - - - [38144, 4096, 1, 128] - - [252, 45.883] - - - [42368, 4096, 1, 128] - - [294, 45.365] - - - [41984, 512, 1, 128] - - [271, 42.307] - - - [33920, 2048, 1, 128] - - [263, 44.095] - - - [34432, 8192, 1, 128] - - [257, 44.91] - - - [31744, 15873, 1, 128] - - [256, 46.488] - - - [42240, 26241, 1, 128] - - [265, 45.662] - - - [42624, 26625, 1, 128] - - [254, 40.471] - - - [40576, 4096, 1, 128] - - [252, 45.622] - - - [42624, 512, 1, 128] - - [288, 40.003] - - - [36864, 1024, 1, 128] - - [262, 42.756] - - - [41344, 2048, 1, 128] - - [267, 44.187] - - - [40064, 2048, 1, 128] - - [249, 42.994] - - - [39808, 1024, 1, 128] - - [254, 42.27] - - - [42496, 1024, 1, 128] - - [290, 43.847] - - - [31232, 512, 1, 128] - - [276, 40.907] - - - [37376, 2048, 1, 128] - - [285, 44.486] - - - [42368, 8192, 1, 128] - - [257, 46.148] - - - [41984, 1024, 1, 128] - - [285, 43.616] - - - [29952, 128, 1, 128] - - [250, 26.55] - - - [32512, 16513, 1, 128] - - [291, 46.298] - - - [34304, 1024, 1, 128] - - [284, 43.493] - - - [30464, 4096, 1, 128] - - [303, 43.433] - - - [36480, 20609, 1, 128] - - [265, 45.594] - - - [36864, 512, 1, 128] - - [250, 40.566] - - - [43648, 4096, 1, 128] - - [252, 45.327] - - - [41600, 512, 1, 128] - - [301, 42.222] - - - [37632, 512, 1, 128] - - [262, 40.728] - - - [37248, 8192, 1, 128] - - [264, 46.313] - - - [38528, 22529, 1, 128] - - [265, 45.314] - - - [40832, 128, 1, 128] - - [286, 32.914] - - - [43776, 27777, 1, 128] - - [288, 43.983] - - - [33792, 4096, 1, 128] - - [252, 46.375] - - - [37888, 512, 1, 128] - - [261, 41.05] - - - [30592, 1024, 1, 128] - - [280, 42.842] - - - [43264, 1024, 1, 128] - - [252, 43.316] - - - [28928, 1024, 1, 128] - - [262, 42.433] - - - [41344, 1024, 1, 128] - - [267, 43.141] - - - [41344, 4096, 1, 128] - - [256, 45.428] - - - [38272, 128, 1, 128] - - [311, 31.539] - - - [44416, 4096, 1, 128] - - [252, 45.725] - - - [37760, 1024, 1, 128] - - [267, 43.235] - - - [44544, 28673, 1, 128] - - [291, 45.502] - - - [38144, 512, 1, 128] - - [254, 40.804] - - - [38144, 2048, 1, 128] - - [264, 44.289] - - - [42240, 8192, 1, 128] - - [252, 46.544] - - - [41088, 512, 1, 128] - - [284, 39.121] - - - [31744, 4096, 1, 128] - - [254, 46.397] - - - [32000, 1024, 1, 128] - - [261, 42.737] - - - [42496, 2048, 1, 128] - - [266, 44.786] - - - [42752, 26881, 1, 128] - - [265, 45.625] - - - [44800, 512, 1, 128] - - [264, 41.246] - - - [38272, 8192, 1, 128] - - [288, 42.399] - - - [37120, 8192, 1, 128] - - [249, 46.547] - - - [43008, 1024, 1, 128] - - [262, 43.999] - - - [36736, 4096, 1, 128] - - [254, 45.56] - - - [32768, 512, 1, 128] - - [263, 36.086] - - - [43008, 27137, 1, 128] - - [249, 46.575] - - - [43136, 2048, 1, 128] - - [264, 43.944] - - - [40064, 24193, 1, 128] - - [291, 44.519] - - - [29184, 2048, 1, 128] - - [266, 44.175] - - - [35328, 128, 1, 128] - - [261, 30.192] - - - [29952, 4096, 1, 128] - - [284, 45.58] - - - [34176, 4096, 1, 128] - - [258, 45.382] - - - [44160, 4096, 1, 128] - - [252, 45.544] - - - [33792, 512, 1, 128] - - [287, 41.448] - - - [31872, 1024, 1, 128] - - [252, 42.624] - - - [44672, 4096, 1, 128] - - [256, 45.698] - - - [39168, 8192, 1, 128] - - [291, 46.6] - - - [38016, 22145, 1, 128] - - [265, 45.578] - - - [30208, 4096, 1, 128] - - [286, 45.872] - - - [30336, 128, 1, 128] - - [267, 26.666] - - - [38912, 4096, 1, 128] - - [264, 46.69] - - - [42752, 8192, 1, 128] - - [254, 46.612] - - - [36608, 1024, 1, 128] - - [287, 42.897] - - - [29312, 13441, 1, 128] - - [252, 45.652] - - - [31616, 8192, 1, 128] - - [252, 45.923] - - - [35328, 2048, 1, 128] - - [271, 44.595] - - - [32000, 2048, 1, 128] - - [254, 44.031] - - - [37504, 128, 1, 128] - - [267, 31.255] - - - [31872, 8192, 1, 128] - - [252, 46.308] - - - [37632, 128, 1, 128] - - [261, 31.221] - - - [44544, 8192, 1, 128] - - [249, 46.449] - - - [39296, 512, 1, 128] - - [261, 41.084] - - - [39168, 1024, 1, 128] - - [276, 43.49] - - - [29568, 13569, 1, 128] - - [249, 45.271] - - - [32768, 2048, 1, 128] - - [256, 34.701] - - - [29568, 4096, 1, 128] - - [288, 44.664] - - - [37376, 1024, 1, 128] - - [261, 43.367] - - - [31488, 8192, 1, 128] - - [254, 46.506] - - - [41472, 4096, 1, 128] - - [257, 45.992] - - - [34560, 18561, 1, 128] - - [256, 45.867] - - - [32256, 16257, 1, 128] - - [251, 46.612] - - - [29312, 4096, 1, 128] - - [262, 45.362] - - - [44288, 1024, 1, 128] - - [287, 43.32] - - - [44160, 28289, 1, 128] - - [264, 45.17] - - - [37888, 22017, 1, 128] - - [265, 46.388] - - - [35840, 2048, 1, 128] - - [271, 44.646] - - - [41728, 4096, 1, 128] - - [294, 45.048] - - - [29824, 4096, 1, 128] - - [288, 44.859] - - - [30592, 4096, 1, 128] - - [252, 45.582] - - - [39552, 8192, 1, 128] - - [265, 46.1] - - - [29312, 128, 1, 128] - - [250, 25.614] - - - [33664, 2048, 1, 128] - - [280, 44.161] - - - [36224, 128, 1, 128] - - [261, 30.533] - - - [29440, 1024, 1, 128] - - [262, 42.53] - - - [42880, 4096, 1, 128] - - [284, 44.933] - - - [39936, 24065, 1, 128] - - [265, 46.243] - - - [32256, 2048, 1, 128] - - [263, 44.555] - - - [40192, 4096, 1, 128] - - [264, 45.816] - - - [44032, 512, 1, 128] - - [250, 41.391] - - - [35584, 8192, 1, 128] - - [256, 46.454] - - - [37504, 1024, 1, 128] - - [286, 42.416] - - - [32384, 1024, 1, 128] - - [321, 42.482] - - - [35200, 128, 1, 128] - - [290, 29.841] - - - [36608, 128, 1, 128] - - [250, 30.717] - - - [28928, 4096, 1, 128] - - [256, 45.55] - - - [41216, 8192, 1, 128] - - [252, 46.646] - - - [40064, 128, 1, 128] - - [284, 32.473] - - - [34688, 128, 1, 128] - - [261, 29.748] - - - [37888, 8192, 1, 128] - - [264, 47.162] - - - [38272, 512, 1, 128] - - [254, 37.579] - - - [42496, 26625, 1, 128] - - [277, 45.874] - - - [44800, 8192, 1, 128] - - [256, 46.255] - - - [37504, 4096, 1, 128] - - [290, 44.983] - - - [31104, 512, 1, 128] - - [250, 40.757] - - - [31616, 512, 1, 128] - - [290, 40.723] - - - [43136, 8192, 1, 128] - - [264, 46.264] - - - [38656, 4096, 1, 128] - - [278, 45.653] - - - [39424, 1024, 1, 128] - - [266, 43.586] - - - [35840, 4096, 1, 128] - - [256, 46.364] - - - [37632, 8192, 1, 128] - - [264, 46.625] - - - [38528, 512, 1, 128] - - [252, 40.582] - - - [31104, 4096, 1, 128] - - [265, 45.51] - - - [43904, 8192, 1, 128] - - [264, 46.33] - - - [38016, 1024, 1, 128] - - [282, 43.185] - - - [42496, 128, 1, 128] - - [269, 31.099] - - - [41728, 1024, 1, 128] - - [294, 42.868] - - - [34560, 2048, 1, 128] - - [290, 44.545] - - - [33664, 4096, 1, 128] - - [265, 45.445] - - - [39808, 4096, 1, 128] - - [264, 43.349] - - - [31360, 4096, 1, 128] - - [254, 45.616] - - - [32000, 512, 1, 128] - - [276, 40.442] - - - [30080, 14081, 1, 128] - - [274, 43.41] - - - [44672, 2048, 1, 128] - - [290, 44.597] - - - [42496, 4096, 1, 128] - - [249, 46.1] - - - [40704, 2048, 1, 128] - - [271, 44.366] - - - [44544, 2048, 1, 128] - - [296, 44.537] - - - [43776, 27905, 1, 128] - - [288, 44.016] - - - [36480, 512, 1, 128] - - [266, 40.173] - - - [41216, 4096, 1, 128] - - [264, 45.814] - - - [33152, 2048, 1, 128] - - [290, 44.512] - - - [39552, 512, 1, 128] - - [262, 41.308] - - - [40960, 4096, 1, 128] - - [249, 41.231] - - - [35968, 4096, 1, 128] - - [262, 45.611] - - - [30976, 15105, 1, 128] - - [303, 44.642] - - - [39424, 23425, 1, 128] - - [291, 45.929] - - - [44288, 28417, 1, 128] - - [277, 45.373] - - - [35712, 512, 1, 128] - - [254, 40.09] - - - [34048, 8192, 1, 128] - - [291, 46.442] - - - [39168, 23297, 1, 128] - - [277, 45.842] - - - [36864, 20993, 1, 128] - - [265, 46.547] - - - [38784, 8192, 1, 128] - - [264, 46.238] - - - [34560, 4096, 1, 128] - - [254, 45.799] - - - [36480, 1024, 1, 128] - - [287, 42.697] - - - [39808, 23809, 1, 128] - - [291, 43.154] - - - [32768, 16897, 1, 128] - - [264, 34.596] - - - [32640, 16769, 1, 128] - - [256, 45.395] - - - [29184, 128, 1, 128] - - [267, 25.776] - - - [30848, 2048, 1, 128] - - [267, 43.898] - - - [35456, 8192, 1, 128] - - [254, 46.552] - - - [36736, 2048, 1, 128] - - [280, 44.039] - - - [39808, 2048, 1, 128] - - [254, 42.312] - - - [34688, 4096, 1, 128] - - [276, 45.486] - - - [40960, 25089, 1, 128] - - [249, 40.357] - - - [41600, 25601, 1, 128] - - [249, 45.352] - - - [32128, 16129, 1, 128] - - [256, 45.863] - - - [35840, 512, 1, 128] - - [276, 40.542] - - - [32512, 8192, 1, 128] - - [252, 46.813] - - - [42880, 1024, 1, 128] - - [266, 43.034] - - - [36224, 1024, 1, 128] - - [267, 42.95] - - - [32384, 512, 1, 128] - - [262, 39.622] - - - [37120, 2048, 1, 128] - - [250, 44.324] - - - [36096, 4096, 1, 128] - - [303, 44.072] - - - [36864, 4096, 1, 128] - - [252, 46.428] - - - [37760, 128, 1, 128] - - [261, 31.327] - - - [36224, 4096, 1, 128] - - [264, 45.663] - - - [34688, 1024, 1, 128] - - [290, 43.274] - - - [44672, 512, 1, 128] - - [254, 41.103] - - - [39424, 512, 1, 128] - - [284, 41.0] - - - [33408, 2048, 1, 128] - - [267, 44.337] - - - [43520, 27649, 1, 128] - - [291, 45.797] - - - [33152, 17153, 1, 128] - - [264, 46.124] - - - [39680, 128, 1, 128] - - [284, 32.482] - - - [42112, 2048, 1, 128] - - [254, 44.11] - - - [30464, 8192, 1, 128] - - [305, 44.92] - - - [32896, 16897, 1, 128] - - [301, 44.774] - - - [34304, 512, 1, 128] - - [296, 41.954] - - - [41216, 1024, 1, 128] - - [299, 43.046] - - - [34048, 4096, 1, 128] - - [262, 45.503] - - - [41728, 512, 1, 128] - - [256, 42.02] - - - [41088, 1024, 1, 128] - - [311, 42.006] - - - [41088, 25089, 1, 128] - - [301, 44.401] - - - [29184, 1024, 1, 128] - - [286, 42.608] - - - [41600, 8192, 1, 128] - - [249, 46.389] - - - [34048, 512, 1, 128] - - [261, 41.814] - - - [44288, 4096, 1, 128] - - [256, 45.46] - - - [41856, 25857, 1, 128] - - [249, 45.643] - - - [34688, 512, 1, 128] - - [290, 41.886] - - - [38656, 1024, 1, 128] - - [264, 42.922] - - - [30336, 512, 1, 128] - - [262, 40.215] - - - [44544, 4096, 1, 128] - - [277, 45.793] - - - [39936, 128, 1, 128] - - [287, 32.476] - - - [35200, 4096, 1, 128] - - [252, 45.526] - - - [31104, 15105, 1, 128] - - [256, 45.786] - - - [32384, 16513, 1, 128] - - [252, 46.085] - - - [32768, 4096, 1, 128] - - [252, 35.663] - - - [32128, 512, 1, 128] - - [261, 40.604] - - - [37376, 21505, 1, 128] - - [264, 45.849] - - - [33536, 8192, 1, 128] - - [291, 46.748] - - - [44032, 28033, 1, 128] - - [265, 46.336] - - - [39296, 4096, 1, 128] - - [252, 45.716] - - - [37504, 8192, 1, 128] - - [288, 45.639] - - - [29952, 2048, 1, 128] - - [250, 44.232] - - - [44672, 28673, 1, 128] - - [249, 45.372] - - - [37888, 2048, 1, 128] - - [288, 44.739] - - - [29824, 8192, 1, 128] - - [301, 45.814] - - - [37376, 4096, 1, 128] - - [262, 45.946] - - - [34304, 8192, 1, 128] - - [274, 47.014] - - - [36992, 2048, 1, 128] - - [266, 43.993] - - - [31360, 15489, 1, 128] - - [264, 45.835] - - - [30080, 2048, 1, 128] - - [300, 41.788] - - - [34176, 128, 1, 128] - - [290, 29.275] - - - [35072, 128, 1, 128] - - [262, 29.938] - - - [36480, 4096, 1, 128] - - [264, 45.435] - - - [38400, 128, 1, 128] - - [252, 31.61] - - - [41344, 512, 1, 128] - - [276, 41.676] - - - [39936, 8192, 1, 128] - - [256, 47.064] - - - [43008, 512, 1, 128] - - [287, 41.471] - - - [36608, 512, 1, 128] - - [276, 40.617] - - - [29824, 13825, 1, 128] - - [277, 45.018] - - - [30336, 1024, 1, 128] - - [262, 42.605] - - - [38144, 22273, 1, 128] - - [265, 45.787] - - - [29568, 13697, 1, 128] - - [265, 45.321] - - - [42112, 4096, 1, 128] - - [264, 45.686] - - - [35584, 1024, 1, 128] - - [290, 43.056] - - - [40960, 8192, 1, 128] - - [252, 41.571] - - - [29312, 512, 1, 128] - - [267, 39.334] - - - [42240, 512, 1, 128] - - [280, 41.921] - - - [35072, 19201, 1, 128] - - [264, 45.897] - - - [33792, 2048, 1, 128] - - [280, 44.646] - - - [40576, 128, 1, 128] - - [290, 32.708] - - - [29824, 1024, 1, 128] - - [258, 42.016] - - - [43008, 27009, 1, 128] - - [249, 46.557] - - - [37120, 21249, 1, 128] - - [256, 45.742] - - - [33024, 1024, 1, 128] - - [261, 42.574] - - - [30592, 14721, 1, 128] - - [265, 45.745] - - - [31232, 128, 1, 128] - - [267, 27.421] - - - [32896, 1024, 1, 128] - - [324, 37.593] - - - [35968, 8192, 1, 128] - - [264, 46.444] - - - [36864, 2048, 1, 128] - - [254, 44.373] - - - [34688, 18817, 1, 128] - - [249, 45.573] - - - [29312, 1024, 1, 128] - - [261, 42.221] - - - [39808, 512, 1, 128] - - [271, 40.895] - - - [32128, 4096, 1, 128] - - [264, 45.8] - - - [41472, 512, 1, 128] - - [252, 42.267] - - - [30464, 1024, 1, 128] - - [254, 42.124] - - - [31872, 15873, 1, 128] - - [265, 45.676] - - - [30720, 2048, 1, 128] - - [264, 44.652] - - - [36096, 8192, 1, 128] - - [286, 45.087] - - - [30208, 2048, 1, 128] - - [266, 44.35] - - - [42880, 128, 1, 128] - - [258, 30.95] - - - [40960, 2048, 1, 128] - - [249, 39.906] - - - [41728, 25857, 1, 128] - - [291, 45.145] - - - [41728, 2048, 1, 128] - - [286, 43.251] - - - [30720, 14721, 1, 128] - - [256, 46.809] - - - [35456, 19457, 1, 128] - - [254, 45.548] - - - [34944, 128, 1, 128] - - [299, 29.692] - - - [35584, 19585, 1, 128] - - [265, 45.685] - - - [39936, 1024, 1, 128] - - [290, 43.517] - - - [32512, 128, 1, 128] - - [290, 28.309] - - - [30464, 2048, 1, 128] - - [278, 42.134] - - - [39040, 512, 1, 128] - - [261, 40.947] - - - [29824, 128, 1, 128] - - [287, 26.31] - - - [41728, 25729, 1, 128] - - [277, 45.1] - - - [42624, 128, 1, 128] - - [306, 30.795] - - - [38784, 2048, 1, 128] - - [287, 44.143] - - - [30976, 128, 1, 128] - - [262, 26.972] - - - [41984, 8192, 1, 128] - - [254, 47.088] - - - [35584, 512, 1, 128] - - [296, 39.6] - - - [34048, 128, 1, 128] - - [267, 29.335] - - - [43136, 128, 1, 128] - - [260, 31.165] - - - [34816, 18945, 1, 128] - - [249, 46.764] - - - [34304, 128, 1, 128] - - [271, 29.487] - - - [36096, 2048, 1, 128] - - [305, 42.932] - - - [37760, 4096, 1, 128] - - [264, 45.537] - - - [31744, 8192, 1, 128] - - [254, 47.209] - - - [39808, 128, 1, 128] - - [260, 32.336] - - - [30464, 14593, 1, 128] - - [305, 42.867] - - - [41472, 1024, 1, 128] - - [254, 43.428] - - - [41600, 128, 1, 128] - - [286, 33.352] - - - [37120, 4096, 1, 128] - - [265, 45.686] - - - [41088, 25217, 1, 128] - - [265, 44.43] - - - [38912, 1024, 1, 128] - - [261, 43.79] - - - [42496, 26497, 1, 128] - - [291, 46.082] - - - [43520, 512, 1, 128] - - [262, 41.357] - - - [29568, 2048, 1, 128] - - [267, 43.692] - - - [29824, 512, 1, 128] - - [287, 39.696] - - - [33408, 4096, 1, 128] - - [252, 45.857] - - - [40832, 2048, 1, 128] - - [254, 44.139] - - - [34048, 2048, 1, 128] - - [249, 43.336] - - - [39296, 23297, 1, 128] - - [249, 45.524] - - - [35840, 1024, 1, 128] - - [250, 43.597] - - - [42752, 2048, 1, 128] - - [265, 44.089] - - - [29440, 13569, 1, 128] - - [252, 45.99] - - - [35840, 19841, 1, 128] - - [264, 46.395] - - - [40192, 2048, 1, 128] - - [287, 44.606] - - - [43648, 512, 1, 128] - - [252, 40.198] - - - [39680, 23809, 1, 128] - - [249, 45.422] - - - [34816, 2048, 1, 128] - - [252, 44.568] - - - [39040, 4096, 1, 128] - - [254, 45.454] - - - [44416, 2048, 1, 128] - - [263, 44.36] - - - [31616, 1024, 1, 128] - - [254, 43.131] - - - [39424, 23553, 1, 128] - - [265, 45.76] - - - [42752, 128, 1, 128] - - [250, 31.193] - - - [36096, 20225, 1, 128] - - [274, 44.722] - - - [39936, 2048, 1, 128] - - [285, 44.806] - - - [29440, 128, 1, 128] - - [261, 26.096] - - - [36608, 4096, 1, 128] - - [264, 45.643] - - - [44160, 28161, 1, 128] - - [256, 45.034] - - - [29056, 13185, 1, 128] - - [256, 45.782] - - - [29696, 128, 1, 128] - - [262, 26.197] - - - [33024, 2048, 1, 128] - - [267, 44.345] - - - [44032, 128, 1, 128] - - [264, 31.413] - - - [29056, 8192, 1, 128] - - [264, 46.411] - - - [34816, 1024, 1, 128] - - [262, 43.646] - - - [39552, 1024, 1, 128] - - [250, 43.178] - - - [44544, 1024, 1, 128] - - [267, 43.492] - - - [29440, 512, 1, 128] - - [262, 39.452] - - - [41472, 2048, 1, 128] - - [299, 44.493] - - - [33408, 512, 1, 128] - - [287, 40.96] - - - [31744, 15745, 1, 128] - - [264, 46.536] - - - [30080, 8192, 1, 128] - - [301, 44.644] - - - [40704, 24833, 1, 128] - - [277, 45.754] - - - [33920, 512, 1, 128] - - [262, 41.639] - - - [36224, 20225, 1, 128] - - [265, 45.701] - - - [31488, 15617, 1, 128] - - [254, 45.941] - - - [30208, 512, 1, 128] - - [252, 40.117] - - - [37504, 21633, 1, 128] - - [257, 44.759] - - - [39168, 2048, 1, 128] - - [250, 44.037] - - - [29440, 13441, 1, 128] - - [254, 45.952] - - - [35072, 19073, 1, 128] - - [265, 45.943] - - - [40576, 8192, 1, 128] - - [256, 46.372] - - - [38656, 128, 1, 128] - - [316, 31.82] - - - [29952, 512, 1, 128] - - [261, 39.867] - - - [31872, 4096, 1, 128] - - [265, 45.435] - - - [38912, 23041, 1, 128] - - [249, 46.582] - - - [40448, 8192, 1, 128] - - [274, 46.84] - - - [43776, 4096, 1, 128] - - [288, 44.709] - - - [37760, 8192, 1, 128] - - [252, 46.459] - - - [44544, 128, 1, 128] - - [284, 32.088] - - - [32640, 512, 1, 128] - - [252, 39.477] - - - [41216, 512, 1, 128] - - [286, 41.238] - - - [34432, 512, 1, 128] - - [254, 41.713] - - - [42240, 128, 1, 128] - - [299, 33.682] - - - [30336, 14337, 1, 128] - - [274, 45.421] - - - [38400, 22401, 1, 128] - - [291, 46.128] - - - [28928, 8192, 1, 128] - - [254, 46.561] - - - [31104, 15233, 1, 128] - - [252, 45.866] - - - [34048, 18049, 1, 128] - - [291, 45.789] - - - [42880, 26881, 1, 128] - - [291, 44.892] - - - [43264, 8192, 1, 128] - - [264, 46.531] - - - [32256, 4096, 1, 128] - - [284, 46.088] - - - [33280, 17409, 1, 128] - - [291, 46.304] - - - [37120, 21121, 1, 128] - - [254, 45.711] - - - [36352, 20481, 1, 128] - - [265, 45.734] - - - [36352, 512, 1, 128] - - [287, 40.948] - - - [29056, 4096, 1, 128] - - [282, 45.545] - - - [40320, 512, 1, 128] - - [261, 41.566] - - - [41472, 8192, 1, 128] - - [278, 46.812] - - - [34176, 18305, 1, 128] - - [278, 45.623] - - - [29696, 13825, 1, 128] - - [264, 46.514] - - - [30208, 8192, 1, 128] - - [278, 46.932] - - - [31360, 2048, 1, 128] - - [287, 43.773] - - - [38528, 2048, 1, 128] - - [250, 43.961] - - - [42240, 1024, 1, 128] - - [287, 43.243] - - - [37248, 1024, 1, 128] - - [254, 42.966] - - - [39296, 23425, 1, 128] - - [249, 45.506] - - - [31104, 128, 1, 128] - - [287, 27.18] - - - [44288, 128, 1, 128] - - [278, 31.718] - - - [39552, 128, 1, 128] - - [250, 32.199] - - - [28928, 13057, 1, 128] - - [254, 45.96] - - - [35328, 4096, 1, 128] - - [291, 46.012] - - - [44800, 28801, 1, 128] - - [265, 45.329] - - - [40960, 24961, 1, 128] - - [249, 40.336] - - - [31744, 1024, 1, 128] - - [271, 43.136] - - - [42112, 26113, 1, 128] - - [249, 45.455] - - - [42112, 1024, 1, 128] - - [250, 43.21] - - - [38144, 128, 1, 128] - - [267, 31.54] - - - [36992, 128, 1, 128] - - [262, 30.794] - - - [40064, 24065, 1, 128] - - [277, 44.533] - - - [36352, 8192, 1, 128] - - [252, 46.787] - - - [32384, 4096, 1, 128] - - [254, 45.721] - - - [16512, 512, 1, 128] - - [325, 30.672] - - - [20992, 512, 1, 128] - - [254, 41.196] - - - [18048, 2048, 1, 128] - - [252, 42.348] - - - [28160, 512, 1, 128] - - [294, 41.669] - - - [19328, 1024, 1, 128] - - [261, 41.669] - - - [10624, 512, 1, 128] - - [276, 32.382] - - - [10240, 6529, 1, 128] - - [264, 45.357] - - - [13184, 2048, 1, 128] - - [267, 41.458] - - - [21248, 512, 1, 128] - - [282, 36.614] - - - [18688, 512, 1, 128] - - [261, 37.935] - - - [15232, 1024, 1, 128] - - [264, 40.493] - - - [6016, 2305, 1, 128] - - [256, 37.786] - - - [11264, 7553, 1, 128] - - [264, 45.632] - - - [15872, 8193, 1, 128] - - [264, 45.629] - - - [22784, 2048, 1, 128] - - [280, 43.471] - - - [14720, 2048, 1, 128] - - [250, 42.96] - - - [23040, 15233, 1, 128] - - [264, 46.577] - - - [18816, 1024, 1, 128] - - [258, 40.938] - - - [18176, 1024, 1, 128] - - [256, 40.995] - - - [16000, 2048, 1, 128] - - [262, 42.288] - - - [14336, 1024, 1, 128] - - [256, 39.739] - - - [23040, 128, 1, 128] - - [250, 31.621] - - - [19072, 128, 1, 128] - - [284, 27.345] - - - [7808, 1024, 1, 128] - - [256, 35.397] - - - [21504, 4096, 1, 128] - - [256, 46.162] - - - [12928, 5121, 1, 128] - - [280, 43.389] - - - [18304, 2048, 1, 128] - - [290, 43.223] - - - [24576, 128, 1, 128] - - [261, 33.175] - - - [24448, 4096, 1, 128] - - [262, 45.432] - - - [24064, 512, 1, 128] - - [287, 39.773] - - - [14592, 1024, 1, 128] - - [256, 39.377] - - - [17280, 128, 1, 128] - - [290, 25.266] - - - [4608, 512, 1, 128] - - [266, 26.791] - - - [11008, 7425, 1, 128] - - [312, 44.734] - - - [18048, 512, 1, 128] - - [254, 37.199] - - - [22656, 512, 1, 128] - - [267, 38.172] - - - [4864, 1024, 1, 128] - - [254, 33.684] - - - [18176, 2048, 1, 128] - - [267, 43.601] - - - [22528, 128, 1, 128] - - [250, 30.976] - - - [3328, 1665, 1, 128] - - [254, 32.65] - - - [18560, 10881, 1, 128] - - [256, 45.473] - - - [19456, 4096, 1, 128] - - [254, 45.916] - - - [1664, 1025, 1, 128] - - [261, 19.797] - - - [13312, 2048, 1, 128] - - [252, 42.713] - - - [19200, 2048, 1, 128] - - [250, 43.454] - - - [18688, 128, 1, 128] - - [271, 27.163] - - - [17792, 9985, 1, 128] - - [277, 45.306] - - - [11648, 8065, 1, 128] - - [264, 44.991] - - - [17024, 1024, 1, 128] - - [252, 41.744] - - - [20736, 512, 1, 128] - - [280, 40.399] - - - [20736, 13057, 1, 128] - - [264, 46.071] - - - [27776, 1024, 1, 128] - - [276, 42.75] - - - [6400, 2048, 1, 128] - - [254, 40.695] - - - [7040, 2048, 1, 128] - - [254, 42.133] - - - [2688, 1153, 1, 128] - - [307, 32.625] - - - [11648, 7937, 1, 128] - - [252, 44.948] - - - [9984, 6401, 1, 128] - - [254, 44.572] - - - [4224, 512, 1, 128] - - [284, 24.753] - - - [5632, 2049, 1, 128] - - [286, 37.024] - - - [5632, 3969, 1, 128] - - [286, 41.798] - - - [25472, 2048, 1, 128] - - [250, 43.855] - - - [14592, 128, 1, 128] - - [260, 22.261] - - - [15360, 7681, 1, 128] - - [264, 45.649] - - - [6400, 512, 1, 128] - - [262, 34.368] - - - [16768, 1024, 1, 128] - - [280, 40.445] - - - [26624, 2048, 1, 128] - - [265, 44.32] - - - [24448, 512, 1, 128] - - [262, 38.266] - - - [17024, 9345, 1, 128] - - [254, 45.64] - - - [5888, 2048, 1, 128] - - [263, 39.591] - - - [14336, 6529, 1, 128] - - [256, 45.676] - - - [25600, 8192, 1, 128] - - [256, 47.314] - - - [7680, 1024, 1, 128] - - [252, 35.722] - - - [6528, 2048, 1, 128] - - [256, 40.762] - - - [13568, 5889, 1, 128] - - [254, 44.448] - - - [26112, 10113, 1, 128] - - [277, 46.101] - - - [16384, 512, 1, 128] - - [256, 33.327] - - - [26368, 1024, 1, 128] - - [287, 42.303] - - - [3968, 2305, 1, 128] - - [261, 36.793] - - - [10368, 1024, 1, 128] - - [261, 40.775] - - - [11776, 4097, 1, 128] - - [258, 43.237] - - - [17792, 4096, 1, 128] - - [284, 45.005] - - - [12416, 512, 1, 128] - - [307, 36.201] - - - [21888, 14209, 1, 128] - - [303, 41.982] - - - [19712, 2048, 1, 128] - - [268, 41.116] - - - [3584, 1921, 1, 128] - - [276, 33.069] - - - [26880, 128, 1, 128] - - [290, 35.57] - - - [15488, 4096, 1, 128] - - [252, 44.919] - - - [22016, 14337, 1, 128] - - [257, 46.472] - - - [25600, 128, 1, 128] - - [287, 33.998] - - - [5376, 3713, 1, 128] - - [261, 41.376] - - - [12928, 4096, 1, 128] - - [261, 44.633] - - - [12032, 1024, 1, 128] - - [261, 40.109] - - - [19840, 12161, 1, 128] - - [264, 45.78] - - - [19840, 4096, 1, 128] - - [261, 45.206] - - - [27136, 11137, 1, 128] - - [257, 46.29] - - - [15744, 512, 1, 128] - - [280, 35.825] - - - [9344, 1024, 1, 128] - - [254, 38.357] - - - [13440, 5633, 1, 128] - - [277, 43.699] - - - [24064, 4096, 1, 128] - - [257, 45.78] - - - [22912, 15233, 1, 128] - - [254, 46.056] - - - [8064, 4353, 1, 128] - - [254, 42.586] - - - [19840, 1024, 1, 128] - - [252, 41.559] - - - [21632, 512, 1, 128] - - [262, 36.974] - - - [23936, 128, 1, 128] - - [261, 32.37] - - - [27008, 512, 1, 128] - - [261, 40.97] - - - [4096, 1024, 1, 128] - - [284, 29.651] - - - [12416, 4737, 1, 128] - - [258, 43.615] - - - [11136, 2048, 1, 128] - - [267, 41.832] - - - [14976, 4096, 1, 128] - - [252, 44.963] - - - [8064, 1024, 1, 128] - - [252, 34.668] - - - [5248, 3585, 1, 128] - - [256, 40.924] - - - [19328, 11521, 1, 128] - - [264, 45.479] - - - [3584, 2049, 1, 128] - - [254, 33.674] - - - [27392, 11393, 1, 128] - - [252, 44.508] - - - [18176, 4096, 1, 128] - - [264, 45.274] - - - [15232, 128, 1, 128] - - [299, 22.86] - - - [23936, 512, 1, 128] - - [261, 39.341] - - - [28032, 12161, 1, 128] - - [252, 45.603] - - - [14336, 6657, 1, 128] - - [249, 45.789] - - - [3328, 1793, 1, 128] - - [252, 34.56] - - - [16384, 8705, 1, 128] - - [256, 37.424] - - - [26368, 128, 1, 128] - - [290, 35.08] - - - [25216, 9345, 1, 128] - - [254, 45.479] - - - [24832, 512, 1, 128] - - [296, 38.5] - - - [20480, 4096, 1, 128] - - [261, 46.324] - - - [13312, 5505, 1, 128] - - [256, 45.025] - - - [20608, 12929, 1, 128] - - [256, 45.702] - - - [27776, 2048, 1, 128] - - [290, 43.876] - - - [15744, 2048, 1, 128] - - [287, 43.428] - - - [7808, 2048, 1, 128] - - [254, 40.69] - - - [26752, 512, 1, 128] - - [261, 40.792] - - - [25856, 128, 1, 128] - - [286, 34.214] - - - [28160, 4096, 1, 128] - - [284, 45.418] - - - [16384, 4096, 1, 128] - - [264, 38.021] - - - [27520, 11521, 1, 128] - - [254, 45.627] - - - [16768, 512, 1, 128] - - [254, 35.966] - - - [9344, 5633, 1, 128] - - [261, 44.287] - - - [2176, 1665, 1, 128] - - [252, 25.457] - - - [11776, 4096, 1, 128] - - [286, 44.562] - - - [5504, 3841, 1, 128] - - [326, 36.733] - - - [26880, 8192, 1, 128] - - [264, 46.482] - - - [26368, 10369, 1, 128] - - [264, 45.792] - - - [15872, 2048, 1, 128] - - [271, 43.266] - - - [24320, 8449, 1, 128] - - [264, 45.716] - - - [17152, 9473, 1, 128] - - [254, 45.713] - - - [28160, 12289, 1, 128] - - [278, 45.864] - - - [19200, 1024, 1, 128] - - [254, 41.453] - - - [12928, 512, 1, 128] - - [271, 37.398] - - - [27008, 2048, 1, 128] - - [267, 43.353] - - - [24960, 8192, 1, 128] - - [254, 46.265] - - - [23552, 15873, 1, 128] - - [252, 46.69] - - - [24320, 128, 1, 128] - - [284, 32.949] - - - [3200, 1665, 1, 128] - - [254, 31.363] - - - [6144, 512, 1, 128] - - [261, 33.481] - - - [23424, 128, 1, 128] - - [299, 31.852] - - - [19072, 11265, 1, 128] - - [252, 45.403] - - - [21888, 4096, 1, 128] - - [301, 40.382] - - - [23680, 4096, 1, 128] - - [277, 44.576] - - - [19968, 4096, 1, 128] - - [282, 45.484] - - - [8320, 2048, 1, 128] - - [263, 40.584] - - - [19200, 11393, 1, 128] - - [264, 45.778] - - - [14080, 2048, 1, 128] - - [254, 42.359] - - - [17024, 128, 1, 128] - - [274, 24.991] - - - [28800, 1024, 1, 128] - - [294, 41.985] - - - [23040, 2048, 1, 128] - - [263, 43.938] - - - [23680, 512, 1, 128] - - [261, 39.204] - - - [24192, 2048, 1, 128] - - [263, 44.024] - - - [5760, 2177, 1, 128] - - [262, 38.6] - - - [17024, 4096, 1, 128] - - [256, 45.364] - - - [5888, 1024, 1, 128] - - [267, 34.886] - - - [11520, 1024, 1, 128] - - [254, 38.952] - - - [10368, 6657, 1, 128] - - [254, 44.392] - - - [26880, 2048, 1, 128] - - [280, 43.869] - - - [26624, 8192, 1, 128] - - [252, 47.587] - - - [24192, 8321, 1, 128] - - [264, 45.195] - - - [27648, 8192, 1, 128] - - [254, 47.306] - - - [27776, 4096, 1, 128] - - [262, 45.284] - - - [13056, 512, 1, 128] - - [290, 37.842] - - - [17408, 9601, 1, 128] - - [264, 46.144] - - - [13824, 512, 1, 128] - - [261, 39.335] - - - [16768, 4096, 1, 128] - - [262, 45.208] - - - [27392, 512, 1, 128] - - [296, 38.691] - - - [22272, 14465, 1, 128] - - [257, 46.223] - - - [18176, 10497, 1, 128] - - [252, 45.819] - - - [14976, 7169, 1, 128] - - [256, 44.407] - - - [24704, 128, 1, 128] - - [254, 33.286] - - - [10880, 7169, 1, 128] - - [254, 44.091] - - - [22528, 1024, 1, 128] - - [254, 42.125] - - - [4480, 2945, 1, 128] - - [261, 40.291] - - - [28672, 128, 1, 128] - - [267, 25.507] - - - [11392, 512, 1, 128] - - [261, 33.918] - - - [14464, 1024, 1, 128] - - [252, 39.175] - - - [6528, 2945, 1, 128] - - [264, 40.713] - - - [26624, 128, 1, 128] - - [250, 35.168] - - - [4480, 2817, 1, 128] - - [262, 39.015] - - - [16000, 128, 1, 128] - - [284, 23.867] - - - [12416, 2048, 1, 128] - - [254, 41.554] - - - [14464, 512, 1, 128] - - [261, 33.638] - - - [28032, 4096, 1, 128] - - [276, 45.397] - - - [20224, 12545, 1, 128] - - [254, 45.854] - - - [3456, 1793, 1, 128] - - [261, 35.322] - - - [21248, 2048, 1, 128] - - [287, 43.649] - - - [15232, 7425, 1, 128] - - [249, 44.631] - - - [22912, 15105, 1, 128] - - [256, 45.856] - - - [3072, 1024, 1, 128] - - [284, 32.874] - - - [14464, 6657, 1, 128] - - [256, 44.75] - - - [26880, 4096, 1, 128] - - [252, 45.686] - - - [17152, 512, 1, 128] - - [252, 38.061] - - - [10368, 2048, 1, 128] - - [267, 42.664] - - - [17664, 128, 1, 128] - - [286, 26.138] - - - [24064, 2048, 1, 128] - - [299, 43.879] - - - [23168, 15489, 1, 128] - - [256, 45.702] - - - [19712, 12033, 1, 128] - - [277, 44.208] - - - [17152, 9345, 1, 128] - - [257, 45.62] - - - [25600, 9729, 1, 128] - - [254, 46.199] - - - [25088, 8192, 1, 128] - - [291, 46.92] - - - [4736, 3201, 1, 128] - - [284, 39.823] - - - [8192, 1024, 1, 128] - - [261, 36.996] - - - [11136, 7425, 1, 128] - - [264, 45.079] - - - [27904, 2048, 1, 128] - - [285, 43.926] - - - [24960, 2048, 1, 128] - - [285, 44.025] - - - [20480, 128, 1, 128] - - [250, 29.193] - - - [26496, 128, 1, 128] - - [261, 34.688] - - - [23296, 512, 1, 128] - - [262, 38.895] - - - [19968, 128, 1, 128] - - [284, 28.741] - - - [27648, 1024, 1, 128] - - [262, 43.172] - - - [3072, 1537, 1, 128] - - [262, 31.708] - - - [20992, 2048, 1, 128] - - [290, 44.058] - - - [20864, 128, 1, 128] - - [299, 29.177] - - - [24320, 8321, 1, 128] - - [254, 45.636] - - - [3840, 2177, 1, 128] - - [264, 36.982] - - - [13184, 512, 1, 128] - - [267, 38.064] - - - [11904, 2048, 1, 128] - - [271, 42.903] - - - [8448, 4737, 1, 128] - - [254, 43.198] - - - [21760, 14081, 1, 128] - - [254, 46.293] - - - [2432, 1793, 1, 128] - - [276, 29.769] - - - [28544, 512, 1, 128] - - [286, 38.922] - - - [4864, 3329, 1, 128] - - [256, 40.718] - - - [4736, 512, 1, 128] - - [260, 27.214] - - - [19840, 12033, 1, 128] - - [254, 45.797] - - - [21376, 4096, 1, 128] - - [264, 45.323] - - - [6400, 2689, 1, 128] - - [262, 41.995] - - - [22528, 14849, 1, 128] - - [254, 46.927] - - - [6656, 1024, 1, 128] - - [261, 38.433] - - - [6016, 512, 1, 128] - - [311, 32.844] - - - [17408, 512, 1, 128] - - [287, 38.114] - - - [20608, 12801, 1, 128] - - [264, 45.592] - - - [17792, 512, 1, 128] - - [267, 36.774] - - - [19968, 12161, 1, 128] - - [252, 46.15] - - - [12032, 4353, 1, 128] - - [254, 43.908] - - - [27648, 4096, 1, 128] - - [264, 46.345] - - - [28544, 2048, 1, 128] - - [271, 43.938] - - - [13184, 5377, 1, 128] - - [258, 43.357] - - - [16384, 2048, 1, 128] - - [252, 36.708] - - - [12800, 512, 1, 128] - - [267, 37.247] - - - [15360, 1024, 1, 128] - - [254, 40.76] - - - [11392, 1024, 1, 128] - - [254, 38.453] - - - [9856, 2048, 1, 128] - - [252, 41.072] - - - [10496, 512, 1, 128] - - [311, 35.51] - - - [3456, 1024, 1, 128] - - [262, 36.005] - - - [11904, 4096, 1, 128] - - [254, 44.407] - - - [12800, 2048, 1, 128] - - [296, 42.43] - - - [5760, 2049, 1, 128] - - [250, 38.035] - - - [11648, 2048, 1, 128] - - [280, 42.355] - - - [22912, 1024, 1, 128] - - [254, 42.084] - - - [23808, 1024, 1, 128] - - [282, 42.655] - - - [22656, 2048, 1, 128] - - [285, 43.771] - - - [28032, 512, 1, 128] - - [256, 41.815] - - - [22144, 512, 1, 128] - - [250, 37.331] - - - [10240, 2048, 1, 128] - - [264, 42.314] - - - [13696, 5889, 1, 128] - - [276, 42.592] - - - [17152, 128, 1, 128] - - [267, 25.431] - - - [16512, 8705, 1, 128] - - [254, 44.284] - - - [22400, 14593, 1, 128] - - [249, 45.652] - - - [9472, 5889, 1, 128] - - [264, 44.292] - - - [27136, 1024, 1, 128] - - [258, 43.113] - - - [25472, 1024, 1, 128] - - [267, 42.1] - - - [5504, 512, 1, 128] - - [261, 29.719] - - - [8832, 5249, 1, 128] - - [252, 43.648] - - - [5632, 2048, 1, 128] - - [251, 38.174] - - - [12672, 2048, 1, 128] - - [256, 42.183] - - - [15360, 7553, 1, 128] - - [254, 45.889] - - - [27136, 128, 1, 128] - - [303, 35.338] - - - [11648, 1024, 1, 128] - - [264, 39.161] - - - [11264, 7681, 1, 128] - - [254, 45.447] - - - [14080, 4096, 1, 128] - - [264, 43.534] - - - [27392, 4096, 1, 128] - - [303, 44.661] - - - [19584, 1024, 1, 128] - - [263, 41.301] - - - [20480, 12801, 1, 128] - - [256, 47.065] - - - [9728, 1024, 1, 128] - - [261, 39.467] - - - [21120, 13313, 1, 128] - - [278, 44.673] - - - [16640, 4096, 1, 128] - - [252, 45.585] - - - [28672, 512, 1, 128] - - [264, 39.061] - - - [25344, 9473, 1, 128] - - [284, 44.899] - - - [25600, 4096, 1, 128] - - [264, 46.394] - - - [21504, 13825, 1, 128] - - [265, 46.611] - - - [7040, 512, 1, 128] - - [262, 36.099] - - - [9216, 512, 1, 128] - - [267, 32.251] - - - [18944, 11137, 1, 128] - - [277, 46.171] - - - [21888, 1024, 1, 128] - - [261, 37.4] - - - [27776, 8192, 1, 128] - - [252, 46.27] - - - [18432, 10753, 1, 128] - - [264, 46.711] - - - [16640, 512, 1, 128] - - [311, 35.332] - - - [28672, 2048, 1, 128] - - [256, 43.883] - - - [14848, 1024, 1, 128] - - [254, 40.215] - - - [7424, 1024, 1, 128] - - [252, 34.422] - - - [15488, 7809, 1, 128] - - [256, 45.064] - - - [5760, 512, 1, 128] - - [262, 31.739] - - - [28160, 2048, 1, 128] - - [250, 44.07] - - - [11776, 1024, 1, 128] - - [254, 39.659] - - - [27904, 12033, 1, 128] - - [277, 45.586] - - - [18944, 11265, 1, 128] - - [291, 46.055] - - - [12032, 4225, 1, 128] - - [252, 43.658] - - - [26752, 10881, 1, 128] - - [254, 45.393] - - - [14464, 128, 1, 128] - - [284, 21.619] - - - [28544, 12673, 1, 128] - - [254, 45.789] - - - [16896, 1024, 1, 128] - - [284, 41.586] - - - [11392, 2048, 1, 128] - - [254, 41.914] - - - [26880, 1024, 1, 128] - - [285, 42.478] - - - [26368, 2048, 1, 128] - - [252, 43.697] - - - [23168, 128, 1, 128] - - [250, 31.797] - - - [16896, 9089, 1, 128] - - [251, 46.11] - - - [22912, 512, 1, 128] - - [263, 38.232] - - - [19584, 512, 1, 128] - - [285, 39.162] - - - [4992, 512, 1, 128] - - [260, 28.409] - - - [16640, 2048, 1, 128] - - [263, 42.998] - - - [2944, 1281, 1, 128] - - [262, 26.563] - - - [10496, 6913, 1, 128] - - [256, 44.723] - - - [11008, 7297, 1, 128] - - [327, 44.623] - - - [5888, 512, 1, 128] - - [261, 32.264] - - - [14208, 6529, 1, 128] - - [254, 44.534] - - - [14336, 2048, 1, 128] - - [264, 43.096] - - - [22656, 14849, 1, 128] - - [249, 45.629] - - - [5248, 1024, 1, 128] - - [252, 35.108] - - - [24448, 8449, 1, 128] - - [254, 45.255] - - - [25472, 4096, 1, 128] - - [252, 45.651] - - - [23168, 1024, 1, 128] - - [262, 42.044] - - - [24192, 8192, 1, 128] - - [252, 46.474] - - - [28288, 1024, 1, 128] - - [264, 42.197] - - - [18432, 2048, 1, 128] - - [252, 43.921] - - - [18944, 4096, 1, 128] - - [258, 45.601] - - - [16128, 512, 1, 128] - - [284, 34.593] - - - [23552, 128, 1, 128] - - [280, 31.909] - - - [10240, 1024, 1, 128] - - [252, 40.434] - - - [17664, 9985, 1, 128] - - [256, 45.395] - - - [7552, 2048, 1, 128] - - [287, 40.591] - - - [7168, 1024, 1, 128] - - [254, 33.982] - - - [24192, 1024, 1, 128] - - [254, 42.824] - - - [10368, 512, 1, 128] - - [263, 35.36] - - - [26752, 4096, 1, 128] - - [265, 45.179] - - - [8832, 1024, 1, 128] - - [307, 36.792] - - - [5120, 1024, 1, 128] - - [252, 34.644] - - - [24960, 512, 1, 128] - - [252, 38.841] - - - [22272, 14593, 1, 128] - - [277, 45.954] - - - [10496, 1024, 1, 128] - - [252, 41.279] - - - [6144, 1024, 1, 128] - - [252, 35.581] - - - [17280, 4096, 1, 128] - - [261, 45.135] - - - [18944, 128, 1, 128] - - [283, 27.644] - - - [19456, 1024, 1, 128] - - [264, 41.595] - - - [25216, 9217, 1, 128] - - [249, 45.153] - - - [13952, 2048, 1, 128] - - [290, 43.274] - - - [12672, 1024, 1, 128] - - [254, 39.816] - - - [11008, 2048, 1, 128] - - [302, 40.313] - - - [10880, 512, 1, 128] - - [261, 32.199] - - - [7424, 512, 1, 128] - - [254, 26.773] - - - [23296, 15489, 1, 128] - - [252, 46.155] - - - [11392, 7681, 1, 128] - - [254, 44.48] - - - [15616, 7809, 1, 128] - - [256, 45.364] - - - [7680, 4097, 1, 128] - - [254, 42.92] - - - [18944, 1024, 1, 128] - - [276, 41.631] - - - [22272, 2048, 1, 128] - - [280, 43.542] - - - [26240, 2048, 1, 128] - - [256, 41.678] - - - [20736, 1024, 1, 128] - - [264, 42.428] - - - [8064, 512, 1, 128] - - [262, 27.566] - - - [25984, 10113, 1, 128] - - [257, 45.563] - - - [27520, 8192, 1, 128] - - [264, 46.441] - - - [27264, 4096, 1, 128] - - [276, 45.21] - - - [17536, 9729, 1, 128] - - [264, 45.307] - - - [25344, 512, 1, 128] - - [252, 39.459] - - - [8704, 512, 1, 128] - - [267, 30.861] - - - [19328, 4096, 1, 128] - - [264, 45.15] - - - [10112, 1024, 1, 128] - - [254, 40.387] - - - [15616, 7937, 1, 128] - - [256, 45.393] - - - [25984, 2048, 1, 128] - - [266, 43.611] - - - [23424, 1024, 1, 128] - - [252, 41.891] - - - [27136, 2048, 1, 128] - - [271, 44.323] - - - [15488, 128, 1, 128] - - [262, 23.15] - - - [21888, 14081, 1, 128] - - [303, 41.962] - - - [21632, 1024, 1, 128] - - [267, 41.233] - - - [4352, 2689, 1, 128] - - [262, 38.62] - - - [4992, 1024, 1, 128] - - [252, 34.368] - - - [26752, 128, 1, 128] - - [287, 35.086] - - - [21504, 512, 1, 128] - - [250, 37.142] - - - [25088, 9089, 1, 128] - - [274, 45.869] - - - [4224, 2561, 1, 128] - - [258, 36.881] - - - [21632, 128, 1, 128] - - [284, 30.024] - - - [20352, 512, 1, 128] - - [290, 40.397] - - - [27904, 512, 1, 128] - - [254, 41.498] - - - [5760, 1024, 1, 128] - - [262, 34.613] - - - [16512, 8833, 1, 128] - - [254, 44.373] - - - [19968, 12289, 1, 128] - - [254, 45.801] - - - [28800, 128, 1, 128] - - [267, 25.256] - - - [20736, 128, 1, 128] - - [285, 29.22] - - - [26496, 8192, 1, 128] - - [284, 45.86] - - - [27264, 128, 1, 128] - - [267, 35.195] - - - [2432, 1921, 1, 128] - - [262, 31.632] - - - [6784, 2048, 1, 128] - - [264, 41.505] - - - [19328, 11649, 1, 128] - - [254, 45.711] - - - [23808, 16129, 1, 128] - - [254, 46.207] - - - [27904, 1024, 1, 128] - - [282, 42.724] - - - [26368, 4096, 1, 128] - - [252, 45.563] - - - [26112, 10241, 1, 128] - - [251, 45.807] - - - [4352, 1024, 1, 128] - - [254, 30.751] - - - [23808, 16001, 1, 128] - - [252, 46.117] - - - [23936, 4096, 1, 128] - - [256, 45.442] - - - [8704, 4993, 1, 128] - - [254, 44.036] - - - [27264, 8192, 1, 128] - - [256, 46.344] - - - [14080, 6273, 1, 128] - - [262, 43.219] - - - [27648, 512, 1, 128] - - [267, 41.305] - - - [15104, 128, 1, 128] - - [284, 22.807] - - - [15232, 512, 1, 128] - - [252, 34.983] - - - [11392, 7809, 1, 128] - - [281, 44.52] - - - [28800, 2048, 1, 128] - - [263, 43.549] - - - [19456, 11649, 1, 128] - - [256, 46.451] - - - [17792, 2048, 1, 128] - - [280, 43.056] - - - [25728, 4096, 1, 128] - - [276, 45.444] - - - [1792, 1281, 1, 128] - - [284, 25.763] - - - [16128, 1024, 1, 128] - - [250, 39.928] - - - [14720, 4096, 1, 128] - - [256, 44.948] - - - [22528, 512, 1, 128] - - [271, 38.262] - - - [25344, 1024, 1, 128] - - [261, 42.112] - - - [9088, 5377, 1, 128] - - [256, 44.163] - - - [15616, 512, 1, 128] - - [303, 35.235] - - - [20096, 512, 1, 128] - - [267, 39.969] - - - [28800, 4096, 1, 128] - - [261, 45.168] - - - [3968, 1024, 1, 128] - - [261, 28.48] - - - [24576, 4096, 1, 128] - - [256, 42.846] - - - [22528, 14721, 1, 128] - - [264, 46.947] - - - [8832, 512, 1, 128] - - [263, 30.98] - - - [3200, 1024, 1, 128] - - [307, 33.876] - - - [28032, 2048, 1, 128] - - [267, 44.151] - - - [13440, 1024, 1, 128] - - [254, 41.263] - - - [6912, 2048, 1, 128] - - [254, 42.332] - - - [17920, 128, 1, 128] - - [284, 26.677] - - - [23936, 16129, 1, 128] - - [256, 46.1] - - - [27648, 2048, 1, 128] - - [287, 44.374] - - - [20864, 2048, 1, 128] - - [256, 43.381] - - - [3584, 512, 1, 128] - - [266, 22.053] - - - [3328, 1024, 1, 128] - - [306, 34.549] - - - [23552, 15745, 1, 128] - - [264, 46.73] - - - [21120, 512, 1, 128] - - [252, 41.201] - - - [24448, 8577, 1, 128] - - [264, 45.437] - - - [18304, 4096, 1, 128] - - [262, 45.075] - - - [12160, 4096, 1, 128] - - [262, 45.148] - - - [27008, 4096, 1, 128] - - [291, 45.118] - - - [14336, 512, 1, 128] - - [256, 33.551] - - - [6272, 2048, 1, 128] - - [261, 39.859] - - - [9600, 5889, 1, 128] - - [256, 43.887] - - - [21376, 512, 1, 128] - - [267, 36.706] - - - [6400, 1024, 1, 128] - - [254, 37.431] - - - [22400, 14721, 1, 128] - - [249, 45.543] - - - [17408, 128, 1, 128] - - [267, 25.811] - - - [22400, 1024, 1, 128] - - [276, 41.793] - - - [15872, 1024, 1, 128] - - [284, 40.799] - - - [22784, 4096, 1, 128] - - [284, 45.163] - - - [16256, 1024, 1, 128] - - [252, 39.961] - - - [7936, 4225, 1, 128] - - [256, 43.039] - - - [9472, 2048, 1, 128] - - [267, 41.446] - - - [18816, 512, 1, 128] - - [263, 38.09] - - - [17664, 9857, 1, 128] - - [265, 45.257] - - - [24448, 2048, 1, 128] - - [254, 43.225] - - - [13952, 512, 1, 128] - - [252, 39.585] - - - [27520, 128, 1, 128] - - [286, 35.463] - - - [8576, 512, 1, 128] - - [267, 30.226] - - - [17408, 2048, 1, 128] - - [287, 43.526] - - - [16384, 1024, 1, 128] - - [254, 37.324] - - - [9728, 512, 1, 128] - - [276, 33.763] - - - [26112, 1024, 1, 128] - - [287, 42.492] - - - [8960, 1024, 1, 128] - - [252, 37.168] - - - [6144, 2561, 1, 128] - - [262, 39.951] - - - [12928, 1024, 1, 128] - - [254, 40.278] - - - [15744, 4096, 1, 128] - - [256, 45.477] - - - [21760, 128, 1, 128] - - [284, 30.145] - - - [24448, 8192, 1, 128] - - [256, 46.582] - - - [12160, 1024, 1, 128] - - [256, 40.129] - - - [15232, 7553, 1, 128] - - [265, 44.833] - - - [25344, 2048, 1, 128] - - [254, 43.168] - - - [21120, 4096, 1, 128] - - [286, 44.469] - - - [14592, 4096, 1, 128] - - [258, 43.855] - - - [24192, 4096, 1, 128] - - [256, 45.486] - - - [25216, 1024, 1, 128] - - [252, 42.089] - - - [21504, 1024, 1, 128] - - [254, 41.566] - - - [9728, 2048, 1, 128] - - [267, 41.96] - - - [28800, 512, 1, 128] - - [299, 38.717] - - - [20224, 1024, 1, 128] - - [262, 42.303] - - - [22784, 1024, 1, 128] - - [256, 42.032] - - - [27264, 512, 1, 128] - - [254, 39.551] - - - [27904, 4096, 1, 128] - - [291, 45.283] - - - [3712, 2177, 1, 128] - - [262, 35.942] - - - [28672, 4096, 1, 128] - - [254, 46.523] - - - [26880, 512, 1, 128] - - [290, 40.713] - - - [6016, 1024, 1, 128] - - [276, 35.431] - - - [22016, 2048, 1, 128] - - [299, 43.988] - - - [20992, 4096, 1, 128] - - [276, 45.563] - - - [12672, 4096, 1, 128] - - [254, 44.756] - - - [14848, 2048, 1, 128] - - [252, 43.185] - - - [13312, 1024, 1, 128] - - [264, 40.722] - - - [3840, 512, 1, 128] - - [254, 22.912] - - - [25472, 9473, 1, 128] - - [264, 45.643] - - - [12288, 2048, 1, 128] - - [254, 42.3] - - - [2816, 1024, 1, 128] - - [284, 31.092] - - - [5120, 3457, 1, 128] - - [256, 40.119] - - - [12672, 4865, 1, 128] - - [254, 43.916] - - - [18176, 512, 1, 128] - - [256, 37.436] - - - [16000, 4096, 1, 128] - - [262, 44.862] - - - [22784, 512, 1, 128] - - [282, 38.3] - - - [24576, 2048, 1, 128] - - [254, 41.044] - - - [18560, 2048, 1, 128] - - [267, 43.454] - - - [23936, 2048, 1, 128] - - [267, 43.931] - - - [12032, 4096, 1, 128] - - [276, 44.644] - - - [25088, 128, 1, 128] - - [250, 33.681] - - - [17280, 2048, 1, 128] - - [263, 43.519] - - - [22656, 128, 1, 128] - - [250, 31.094] - - - [16128, 128, 1, 128] - - [299, 24.353] - - - [24064, 1024, 1, 128] - - [252, 42.816] - - - [20608, 128, 1, 128] - - [267, 28.874] - - - [6656, 2048, 1, 128] - - [254, 40.68] - - - [15744, 7937, 1, 128] - - [252, 45.436] - - - [24192, 8193, 1, 128] - - [256, 44.945] - - - [24576, 8705, 1, 128] - - [249, 42.522] - - - [27776, 512, 1, 128] - - [267, 41.391] - - - [16512, 4096, 1, 128] - - [264, 42.737] - - - [27392, 1024, 1, 128] - - [267, 40.737] - - - [6528, 2817, 1, 128] - - [254, 40.738] - - - [5248, 512, 1, 128] - - [305, 29.412] - - - [21248, 13569, 1, 128] - - [278, 45.898] - - - [14208, 6401, 1, 128] - - [254, 44.543] - - - [6912, 1024, 1, 128] - - [256, 39.297] - - - [14592, 6913, 1, 128] - - [284, 43.881] - - - [22400, 512, 1, 128] - - [254, 37.762] - - - [24832, 128, 1, 128] - - [267, 33.52] - - - [25984, 9985, 1, 128] - - [257, 45.49] - - - [21376, 2048, 1, 128] - - [263, 43.564] - - - [15872, 512, 1, 128] - - [267, 36.255] - - - [25216, 4096, 1, 128] - - [262, 45.281] - - - [28288, 512, 1, 128] - - [261, 38.591] - - - [25344, 9345, 1, 128] - - [274, 45.119] - - - [27008, 8192, 1, 128] - - [277, 46.31] - - - [10240, 512, 1, 128] - - [276, 34.804] - - - [27520, 1024, 1, 128] - - [285, 42.634] - - - [6784, 512, 1, 128] - - [287, 35.653] - - - [16128, 4096, 1, 128] - - [254, 45.42] - - - [28800, 12801, 1, 128] - - [256, 45.526] - - - [20096, 12417, 1, 128] - - [256, 45.622] - - - [9088, 1024, 1, 128] - - [250, 37.203] - - - [21760, 4096, 1, 128] - - [261, 45.548] - - - [14208, 1024, 1, 128] - - [252, 39.018] - - - [23680, 1024, 1, 128] - - [252, 42.082] - - - [22016, 4096, 1, 128] - - [257, 45.788] - - - [10752, 1024, 1, 128] - - [254, 37.471] - - - [19200, 4096, 1, 128] - - [254, 45.06] - - - [18816, 11009, 1, 128] - - [254, 45.335] - - - [7168, 512, 1, 128] - - [287, 26.137] - - - [17920, 1024, 1, 128] - - [252, 40.839] - - - [22784, 14977, 1, 128] - - [274, 46.074] - - - [15104, 1024, 1, 128] - - [264, 40.371] - - - [5376, 512, 1, 128] - - [284, 29.296] - - - [25728, 512, 1, 128] - - [263, 39.372] - - - [22016, 512, 1, 128] - - [284, 37.306] - - - [23296, 2048, 1, 128] - - [290, 43.486] - - - [14848, 4096, 1, 128] - - [264, 45.374] - - - [28032, 8192, 1, 128] - - [264, 46.536] - - - [15360, 4096, 1, 128] - - [252, 45.83] - - - [18304, 128, 1, 128] - - [287, 26.658] - - - [26624, 10753, 1, 128] - - [264, 46.719] - - - [26496, 10497, 1, 128] - - [274, 45.036] - - - [9216, 5505, 1, 128] - - [249, 44.431] - - - [6656, 3073, 1, 128] - - [264, 41.59] - - - [23808, 512, 1, 128] - - [287, 39.24] - - - [24704, 8705, 1, 128] - - [252, 44.615] - - - [3712, 1024, 1, 128] - - [262, 26.643] - - - [8320, 1024, 1, 128] - - [286, 35.307] - - - [18944, 2048, 1, 128] - - [285, 43.781] - - - [15616, 2048, 1, 128] - - [264, 43.257] - - - [2688, 1025, 1, 128] - - [284, 29.542] - - - [4352, 512, 1, 128] - - [261, 25.253] - - - [8320, 512, 1, 128] - - [262, 27.956] - - - [2560, 1025, 1, 128] - - [286, 28.887] - - - [2816, 1153, 1, 128] - - [264, 33.809] - - - [26368, 8192, 1, 128] - - [264, 46.613] - - - [18816, 4096, 1, 128] - - [264, 45.154] - - - [11136, 1024, 1, 128] - - [287, 37.354] - - - [19072, 2048, 1, 128] - - [250, 43.381] - - - [7552, 3841, 1, 128] - - [256, 42.352] - - - [23424, 512, 1, 128] - - [280, 38.672] - - - [18560, 1024, 1, 128] - - [254, 40.909] - - - [18688, 1024, 1, 128] - - [287, 41.268] - - - [16256, 8449, 1, 128] - - [256, 44.905] - - - [24704, 4096, 1, 128] - - [288, 44.774] - - - [6528, 512, 1, 128] - - [262, 34.74] - - - [11264, 1024, 1, 128] - - [252, 38.262] - - - [22272, 4096, 1, 128] - - [249, 45.315] - - - [28160, 1024, 1, 128] - - [258, 43.105] - - - [18688, 2048, 1, 128] - - [267, 43.547] - - - [23552, 1024, 1, 128] - - [252, 42.519] - - - [20992, 1024, 1, 128] - - [264, 42.981] - - - [28032, 1024, 1, 128] - - [252, 43.043] - - - [15360, 128, 1, 128] - - [267, 23.052] - - - [6272, 2689, 1, 128] - - [262, 41.19] - - - [25472, 128, 1, 128] - - [287, 33.95] - - - [28416, 4096, 1, 128] - - [261, 45.681] - - - [21376, 1024, 1, 128] - - [254, 40.864] - - - [28288, 12417, 1, 128] - - [252, 45.754] - - - [22016, 1024, 1, 128] - - [286, 41.391] - - - [25216, 8192, 1, 128] - - [249, 46.249] - - - [23680, 128, 1, 128] - - [254, 32.26] - - - [3968, 512, 1, 128] - - [286, 23.346] - - - [25984, 4096, 1, 128] - - [278, 45.164] - - - [9984, 1024, 1, 128] - - [254, 39.768] - - - [20736, 2048, 1, 128] - - [290, 43.746] - - - [13696, 512, 1, 128] - - [262, 38.896] - - - [25728, 2048, 1, 128] - - [264, 43.634] - - - [14720, 1024, 1, 128] - - [254, 39.831] - - - [14976, 128, 1, 128] - - [287, 22.567] - - - [14080, 512, 1, 128] - - [261, 39.158] - - - [19072, 512, 1, 128] - - [294, 38.216] - - - [7936, 512, 1, 128] - - [254, 28.619] - - - [12160, 2048, 1, 128] - - [290, 42.781] - - - [23296, 1024, 1, 128] - - [262, 42.225] - - - [13568, 2048, 1, 128] - - [290, 43.078] - - - [10112, 2048, 1, 128] - - [264, 42.288] - - - [19328, 512, 1, 128] - - [267, 38.887] - - - [4864, 3201, 1, 128] - - [276, 41.106] - - - [23040, 4096, 1, 128] - - [262, 45.59] - - - [9472, 1024, 1, 128] - - [276, 38.481] - - - [28416, 2048, 1, 128] - - [256, 43.52] - - - [13952, 1024, 1, 128] - - [254, 41.942] - - - [10624, 7041, 1, 128] - - [256, 44.381] - - - [17280, 9473, 1, 128] - - [265, 45.407] - - - [10240, 6657, 1, 128] - - [256, 45.524] - - - [17792, 1024, 1, 128] - - [256, 40.407] - - - [25856, 512, 1, 128] - - [262, 39.671] - - - [25984, 1024, 1, 128] - - [260, 41.903] - - - [17920, 4096, 1, 128] - - [284, 45.539] - - - [24704, 8833, 1, 128] - - [256, 44.724] - - - [12672, 4993, 1, 128] - - [254, 43.969] - - - [25856, 9985, 1, 128] - - [249, 45.586] - - - [9728, 6145, 1, 128] - - [282, 44.553] - - - [14464, 4096, 1, 128] - - [256, 44.862] - - - [23680, 15873, 1, 128] - - [277, 45.366] - - - [17664, 512, 1, 256] - - [287, 56.371] - - - [28928, 512, 1, 256] - - [290, 62.502] - - - [38912, 1024, 1, 256] - - [250, 70.324] - - - [23552, 1024, 1, 256] - - [290, 68.047] - - - [9472, 1024, 1, 256] - - [267, 59.523] - - - [42496, 1024, 1, 256] - - [296, 70.355] - - - [40192, 24320, 1, 256] - - [254, 74.909] - - - [27904, 512, 1, 256] - - [254, 65.331] - - - [28160, 8192, 1, 256] - - [256, 74.637] - - - [23040, 15360, 1, 256] - - [252, 75.671] - - - [43520, 27648, 1, 256] - - [268, 74.947] - - - [20224, 4096, 1, 256] - - [254, 72.535] - - - [33024, 16896, 1, 256] - - [269, 75.252] - - - [42752, 4096, 1, 256] - - [254, 72.964] - - - [10496, 512, 1, 256] - - [262, 55.906] - - - [12288, 1024, 1, 256] - - [264, 62.574] - - - [17152, 9472, 1, 256] - - [258, 74.859] - - - [25344, 1024, 1, 256] - - [296, 67.149] - - - [28672, 8192, 1, 256] - - [256, 74.718] - - - [38144, 22272, 1, 256] - - [265, 74.833] - - - [6400, 2560, 1, 256] - - [261, 64.98] - - - [2304, 1536, 1, 256] - - [276, 53.512] - - - [3072, 1024, 1, 256] - - [262, 49.378] - - - [39936, 24064, 1, 256] - - [264, 75.501] - - - [39680, 1024, 1, 256] - - [264, 69.234] - - - [25600, 9728, 1, 256] - - [252, 75.44] - - - [13056, 1024, 1, 256] - - [287, 62.574] - - - [4864, 3072, 1, 256] - - [282, 64.205] - - - [13056, 4096, 1, 256] - - [264, 72.06] - - - [34048, 1024, 1, 256] - - [266, 68.468] - - - [15872, 4096, 1, 256] - - [264, 72.281] - - - [7424, 512, 1, 256] - - [290, 41.766] - - - [37376, 21504, 1, 256] - - [268, 75.215] - - - [40704, 24832, 1, 256] - - [254, 74.84] - - - [25088, 8960, 1, 256] - - [286, 75.437] - - - [4608, 512, 1, 256] - - [254, 38.941] - - - [33024, 512, 1, 256] - - [262, 60.923] - - - [28928, 4096, 1, 256] - - [254, 72.464] - - - [43776, 27648, 1, 256] - - [268, 74.437] - - - [33536, 17664, 1, 256] - - [251, 75.192] - - - [13312, 5376, 1, 256] - - [261, 74.004] - - - [39168, 4096, 1, 256] - - [268, 73.18] - - - [41472, 25600, 1, 256] - - [265, 74.75] - - - [37632, 21760, 1, 256] - - [252, 74.813] - - - [14336, 6656, 1, 256] - - [256, 74.692] - - - [32256, 16384, 1, 256] - - [270, 75.223] - - - [26112, 512, 1, 256] - - [262, 62.957] - - - [16640, 512, 1, 256] - - [328, 56.008] - - - [13312, 4096, 1, 256] - - [256, 72.101] - - - [5632, 1024, 1, 256] - - [253, 50.714] - - - [27648, 8192, 1, 256] - - [252, 75.078] - - - [43776, 4096, 1, 256] - - [255, 73.189] - - - [26624, 10496, 1, 256] - - [254, 75.778] - - - [17408, 9728, 1, 256] - - [252, 75.456] - - - [26880, 11008, 1, 256] - - [254, 74.818] - - - [14336, 6400, 1, 256] - - [276, 74.846] - - - [7168, 512, 1, 256] - - [276, 41.029] - - - [11776, 512, 1, 256] - - [267, 52.738] - - - [11520, 1024, 1, 256] - - [252, 61.695] - - - [37888, 8192, 1, 256] - - [256, 74.991] - - - [23552, 15616, 1, 256] - - [256, 76.268] - - - [27136, 11008, 1, 256] - - [251, 75.587] - - - [26624, 1024, 1, 256] - - [267, 68.378] - - - [23040, 512, 1, 256] - - [296, 61.25] - - - [43264, 8192, 1, 256] - - [256, 74.358] - - - [43520, 512, 1, 256] - - [286, 66.566] - - - [33792, 17664, 1, 256] - - [254, 75.868] - - - [29696, 8192, 1, 256] - - [256, 74.986] - - - [44544, 8192, 1, 256] - - [254, 74.517] - - - [38656, 4096, 1, 256] - - [268, 73.31] - - - [37376, 4096, 1, 256] - - [268, 73.81] - - - [39680, 23552, 1, 256] - - [264, 74.227] - - - [25856, 1024, 1, 256] - - [267, 68.429] - - - [17152, 9216, 1, 256] - - [252, 74.262] - - - [5376, 1024, 1, 256] - - [267, 48.52] - - - [15360, 512, 1, 256] - - [261, 55.596] - - - [42752, 26624, 1, 256] - - [254, 74.446] - - - [6144, 2304, 1, 256] - - [282, 66.961] - - - [26880, 4096, 1, 256] - - [252, 72.673] - - - [18688, 10752, 1, 256] - - [254, 75.22] - - - [8192, 4352, 1, 256] - - [264, 70.832] - - - [31744, 15616, 1, 256] - - [264, 75.832] - - - [41216, 25088, 1, 256] - - [249, 74.747] - - - [18944, 512, 1, 256] - - [252, 59.145] - - - [44288, 512, 1, 256] - - [258, 66.064] - - - [21504, 13568, 1, 256] - - [264, 76.086] - - - [21248, 1024, 1, 256] - - [287, 66.339] - - - [25600, 8192, 1, 256] - - [264, 75.08] - - - [16128, 1024, 1, 256] - - [262, 62.905] - - - [36864, 20992, 1, 256] - - [264, 75.008] - - - [35072, 19200, 1, 256] - - [252, 74.991] - - - [28416, 1024, 1, 256] - - [290, 67.534] - - - [21760, 4096, 1, 256] - - [318, 72.584] - - - [33280, 4096, 1, 256] - - [255, 73.587] - - - [38400, 22528, 1, 256] - - [255, 75.187] - - - [5120, 512, 1, 256] - - [262, 42.717] - - - [40448, 24576, 1, 256] - - [270, 74.509] - - - [21504, 1024, 1, 256] - - [285, 67.191] - - - [16896, 512, 1, 256] - - [287, 59.514] - - - [19712, 512, 1, 256] - - [254, 60.709] - - - [29952, 13824, 1, 256] - - [257, 75.112] - - - [36864, 1024, 1, 256] - - [264, 68.423] - - - [24832, 8960, 1, 256] - - [258, 75.119] - - - [19200, 11264, 1, 256] - - [255, 74.727] - - - [34304, 18432, 1, 256] - - [269, 75.048] - - - [40704, 4096, 1, 256] - - [255, 73.066] - - - [3328, 1024, 1, 256] - - [262, 52.353] - - - [39424, 8192, 1, 256] - - [264, 74.587] - - - [8960, 1024, 1, 256] - - [280, 56.85] - - - [35584, 8192, 1, 256] - - [269, 74.268] - - - [13824, 1024, 1, 256] - - [271, 65.854] - - - [30976, 512, 1, 256] - - [250, 64.406] - - - [11776, 7936, 1, 256] - - [254, 74.374] - - - [22272, 14336, 1, 256] - - [268, 74.944] - - - [30464, 4096, 1, 256] - - [268, 71.791] - - - [44800, 1024, 1, 256] - - [256, 69.404] - - - [23296, 15616, 1, 256] - - [254, 75.408] - - - [20224, 12544, 1, 256] - - [252, 75.301] - - - [38400, 512, 1, 256] - - [299, 65.71] - - - [36352, 512, 1, 256] - - [276, 65.269] - - - [32000, 1024, 1, 256] - - [250, 68.824] - - - [27136, 11264, 1, 256] - - [274, 75.389] - - - [22016, 14080, 1, 256] - - [286, 76.113] - - - [23552, 4096, 1, 256] - - [268, 72.963] - - - [37120, 512, 1, 256] - - [276, 65.473] - - - [32512, 8192, 1, 256] - - [251, 74.44] - - - [13056, 512, 1, 256] - - [261, 57.597] - - - [24576, 8192, 1, 256] - - [281, 70.1] - - - [26368, 8192, 1, 256] - - [264, 74.449] - - - [35072, 8192, 1, 256] - - [264, 74.34] - - - [40704, 24576, 1, 256] - - [255, 74.2] - - - [25856, 4096, 1, 256] - - [255, 72.553] - - - [39424, 4096, 1, 256] - - [292, 73.507] - - - [44288, 28416, 1, 256] - - [249, 74.674] - - - [12800, 4864, 1, 256] - - [286, 73.386] - - - [40192, 24064, 1, 256] - - [264, 74.898] - - - [40960, 8192, 1, 256] - - [281, 67.262] - - - [9728, 512, 1, 256] - - [276, 52.767] - - - [44800, 28672, 1, 256] - - [281, 74.049] - - - [33280, 17408, 1, 256] - - [269, 75.374] - - - [35328, 8192, 1, 256] - - [274, 74.689] - - - [35840, 1024, 1, 256] - - [301, 70.114] - - - [3584, 1024, 1, 256] - - [287, 41.069] - - - [44032, 512, 1, 256] - - [271, 66.675] - - - [39680, 512, 1, 256] - - [285, 65.345] - - - [42240, 26368, 1, 256] - - [256, 74.63] - - - [30976, 14848, 1, 256] - - [255, 74.145] - - - [28928, 12800, 1, 256] - - [274, 74.82] - - - [5888, 1024, 1, 256] - - [287, 52.899] - - - [38912, 4096, 1, 256] - - [252, 73.118] - - - [18688, 512, 1, 256] - - [261, 58.254] - - - [22784, 512, 1, 256] - - [267, 61.037] - - - [19200, 512, 1, 256] - - [262, 59.411] - - - [25088, 9216, 1, 256] - - [269, 74.874] - - - [25600, 9472, 1, 256] - - [254, 75.887] - - - [5632, 3840, 1, 256] - - [261, 70.207] - - - [8960, 5120, 1, 256] - - [258, 71.62] - - - [24576, 4096, 1, 256] - - [259, 68.369] - - - [35072, 4096, 1, 256] - - [268, 72.993] - - - [40448, 8192, 1, 256] - - [278, 74.465] - - - [22272, 4096, 1, 256] - - [268, 72.461] - - - [10752, 512, 1, 256] - - [262, 48.931] - - - [34304, 1024, 1, 256] - - [266, 69.634] - - - [35840, 19712, 1, 256] - - [254, 75.823] - - - [30720, 14848, 1, 256] - - [254, 75.633] - - - [40192, 1024, 1, 256] - - [288, 69.757] - - - [34560, 18432, 1, 256] - - [252, 74.758] - - - [5376, 3584, 1, 256] - - [262, 67.587] - - - [31232, 512, 1, 256] - - [261, 65.399] - - - [35840, 4096, 1, 256] - - [254, 73.139] - - - [31744, 512, 1, 256] - - [303, 63.607] - - - [19456, 11520, 1, 256] - - [254, 76.138] - - - [18944, 11264, 1, 256] - - [269, 75.342] - - - [26112, 9984, 1, 256] - - [264, 75.563] - - - [24320, 8192, 1, 256] - - [274, 74.704] - - - [20480, 4096, 1, 256] - - [264, 72.506] - - - [31232, 15360, 1, 256] - - [251, 75.415] - - - [33792, 1024, 1, 256] - - [285, 69.904] - - - [39168, 512, 1, 256] - - [262, 64.9] - - - [36096, 1024, 1, 256] - - [303, 66.757] - - - [30720, 4096, 1, 256] - - [252, 73.127] - - - [37120, 21248, 1, 256] - - [264, 74.681] - - - [18688, 11008, 1, 256] - - [256, 75.091] - - - [11264, 7424, 1, 256] - - [254, 74.924] - - - [25344, 512, 1, 256] - - [262, 61.71] - - - [43264, 4096, 1, 256] - - [281, 72.828] - - - [39936, 8192, 1, 256] - - [256, 74.974] - - - [16640, 8704, 1, 256] - - [264, 74.811] - - - [35328, 19456, 1, 256] - - [269, 75.247] - - - [25344, 9216, 1, 256] - - [270, 73.445] - - - [18688, 4096, 1, 256] - - [256, 72.394] - - - [19968, 12032, 1, 256] - - [254, 75.67] - - - [2048, 1536, 1, 256] - - [284, 49.785] - - - [35328, 512, 1, 256] - - [310, 64.709] - - - [33792, 8192, 1, 256] - - [264, 74.959] - - - [28416, 8192, 1, 256] - - [274, 74.104] - - - [13568, 4096, 1, 256] - - [254, 71.849] - - - [13568, 1024, 1, 256] - - [250, 64.791] - - - [22528, 4096, 1, 256] - - [256, 72.967] - - - [23296, 1024, 1, 256] - - [250, 67.242] - - - [39936, 23808, 1, 256] - - [264, 75.482] - - - [22016, 4096, 1, 256] - - [301, 73.077] - - - [39936, 4096, 1, 256] - - [254, 73.242] - - - [44800, 2048, 1, 256] - - [270, 70.237] - - - [20224, 12288, 1, 256] - - [268, 74.421] - - - [43264, 27392, 1, 256] - - [264, 74.802] - - - [24576, 8448, 1, 256] - - [249, 70.765] - - - [3328, 1536, 1, 256] - - [262, 54.557] - - - [43008, 8192, 1, 256] - - [264, 74.904] - - - [20736, 13056, 1, 256] - - [254, 75.359] - - - [17920, 1024, 1, 256] - - [266, 65.342] - - - [44800, 4096, 1, 256] - - [259, 72.511] - - - [39168, 23296, 1, 256] - - [268, 74.594] - - - [27904, 8192, 1, 256] - - [255, 74.273] - - - [15872, 7936, 1, 256] - - [252, 75.157] - - - [17920, 9984, 1, 256] - - [264, 75.469] - - - [41984, 4096, 1, 256] - - [254, 73.1] - - - [28416, 4096, 1, 256] - - [259, 72.453] - - - [7680, 3840, 1, 256] - - [276, 70.372] - - - [11264, 512, 1, 256] - - [261, 50.753] - - - [21248, 13312, 1, 256] - - [254, 75.106] - - - [35072, 18944, 1, 256] - - [254, 74.936] - - - [44800, 8192, 1, 256] - - [264, 74.033] - - - [27904, 12032, 1, 256] - - [270, 74.855] - - - [28928, 8192, 1, 256] - - [274, 74.128] - - - [43264, 27136, 1, 256] - - [265, 74.753] - - - [26880, 10752, 1, 256] - - [256, 75.05] - - - [36096, 512, 1, 256] - - [286, 63.214] - - - [38912, 8192, 1, 256] - - [252, 74.954] - - - [37632, 512, 1, 256] - - [267, 65.67] - - - [24320, 8448, 1, 256] - - [284, 74.743] - - - [37632, 21504, 1, 256] - - [265, 74.577] - - - [22784, 15104, 1, 256] - - [278, 75.335] - - - [43008, 26880, 1, 256] - - [265, 75.604] - - - [25344, 8192, 1, 256] - - [268, 73.596] - - - [38912, 22784, 1, 256] - - [265, 75.639] - - - [30976, 4096, 1, 256] - - [255, 71.622] - - - [40960, 4096, 1, 256] - - [259, 66.076] - - - [18944, 4096, 1, 256] - - [303, 72.836] - - - [11008, 7168, 1, 256] - - [312, 71.456] - - - [44288, 1024, 1, 256] - - [266, 69.726] - - - [38144, 1024, 1, 256] - - [256, 69.01] - - - [20736, 512, 1, 256] - - [287, 63.106] - - - [40960, 1024, 1, 256] - - [252, 64.607] - - - [29696, 13824, 1, 256] - - [256, 75.882] - - - [37120, 20992, 1, 256] - - [252, 74.854] - - - [38656, 1024, 1, 256] - - [263, 69.426] - - - [32000, 16128, 1, 256] - - [256, 74.91] - - - [23040, 15104, 1, 256] - - [257, 76.021] - - - [25088, 4096, 1, 256] - - [288, 73.174] - - - [16384, 8448, 1, 256] - - [249, 63.356] - - - [14080, 6400, 1, 256] - - [312, 73.029] - - - [43776, 27904, 1, 256] - - [268, 74.364] - - - [43776, 8192, 1, 256] - - [268, 74.087] - - - [13312, 512, 1, 256] - - [254, 58.378] - - - [15360, 7680, 1, 256] - - [252, 75.288] - - - [34560, 8192, 1, 256] - - [256, 74.307] - - - [32512, 16384, 1, 256] - - [269, 75.174] - - - [29440, 13568, 1, 256] - - [278, 75.133] - - - [32768, 4096, 1, 256] - - [293, 58.306] - - - [39680, 23808, 1, 256] - - [264, 74.405] - - - [7936, 4096, 1, 256] - - [267, 69.283] - - - [29696, 512, 1, 256] - - [290, 63.904] - - - [17408, 9472, 1, 256] - - [252, 75.849] - - - [17920, 512, 1, 256] - - [280, 56.82] - - - [44288, 28160, 1, 256] - - [249, 74.632] - - - [21248, 512, 1, 256] - - [287, 58.27] - - - [44544, 1024, 1, 256] - - [287, 70.345] - - - [33792, 17920, 1, 256] - - [254, 75.738] - - - [41216, 512, 1, 256] - - [266, 66.051] - - - [30720, 8192, 1, 256] - - [264, 75.059] - - - [43008, 4096, 1, 256] - - [265, 73.053] - - - [41984, 512, 1, 256] - - [271, 67.355] - - - [2560, 1792, 1, 256] - - [262, 49.744] - - - [40704, 8192, 1, 256] - - [254, 74.341] - - - [32256, 512, 1, 256] - - [262, 64.09] - - - [8960, 512, 1, 256] - - [276, 49.372] - - - [30976, 15104, 1, 256] - - [268, 74.383] - - - [30208, 8192, 1, 256] - - [274, 74.616] - - - [9728, 5888, 1, 256] - - [261, 73.206] - - - [7424, 1024, 1, 256] - - [261, 54.579] - - - [26880, 512, 1, 256] - - [262, 63.666] - - - [44032, 27904, 1, 256] - - [252, 75.577] - - - [29952, 512, 1, 256] - - [262, 63.804] - - - [28160, 12288, 1, 256] - - [255, 74.924] - - - [39936, 1024, 1, 256] - - [285, 70.178] - - - [40448, 24320, 1, 256] - - [251, 75.043] - - - [7424, 3584, 1, 256] - - [282, 69.863] - - - [27136, 4096, 1, 256] - - [270, 73.399] - - - [18944, 11008, 1, 256] - - [286, 75.428] - - - [38400, 1024, 1, 256] - - [285, 70.053] - - - [36608, 20736, 1, 256] - - [249, 74.849] - - - [42496, 26624, 1, 256] - - [268, 75.087] - - - [20992, 4096, 1, 256] - - [255, 73.044] - - - [24064, 8192, 1, 256] - - [274, 74.93] - - - [4352, 512, 1, 256] - - [299, 36.778] - - - [16128, 8192, 1, 256] - - [252, 74.553] - - - [29440, 4096, 1, 256] - - [274, 72.927] - - - [33024, 17152, 1, 256] - - [274, 75.207] - - - [22016, 512, 1, 256] - - [261, 59.438] - - - [20736, 12800, 1, 256] - - [252, 75.324] - - - [7680, 512, 1, 256] - - [267, 43.371] - - - [29184, 4096, 1, 256] - - [301, 73.343] - - - [44544, 28672, 1, 256] - - [249, 74.533] - - - [26624, 10752, 1, 256] - - [256, 75.829] - - - [23296, 4096, 1, 256] - - [256, 72.489] - - - [11520, 7680, 1, 256] - - [286, 73.86] - - - [2304, 1792, 1, 256] - - [262, 45.323] - - - [33280, 8192, 1, 256] - - [251, 74.883] - - - [21760, 1024, 1, 256] - - [271, 66.154] - - - [41216, 4096, 1, 256] - - [270, 73.138] - - - [17408, 1024, 1, 256] - - [280, 66.969] - - - [36096, 19968, 1, 256] - - [268, 74.202] - - - [30464, 1024, 1, 256] - - [286, 66.822] - - - [11520, 512, 1, 256] - - [252, 51.67] - - - [20736, 4096, 1, 256] - - [252, 72.521] - - - [39936, 512, 1, 256] - - [267, 65.84] - - - [28672, 4096, 1, 256] - - [264, 72.526] - - - [17664, 1024, 1, 256] - - [263, 63.994] - - - [28160, 512, 1, 256] - - [285, 65.774] - - - [19200, 11520, 1, 256] - - [254, 75.066] - - - [35328, 19200, 1, 256] - - [278, 75.548] - - - [20992, 13056, 1, 256] - - [252, 75.803] - - - [12800, 512, 1, 256] - - [267, 56.426] - - - [29696, 13568, 1, 256] - - [254, 75.794] - - - [30464, 14592, 1, 256] - - [274, 74.281] - - - [34560, 4096, 1, 256] - - [256, 72.865] - - - [13824, 5888, 1, 256] - - [249, 73.544] - - - [14080, 4096, 1, 256] - - [274, 70.417] - - - [31744, 1024, 1, 256] - - [267, 69.565] - - - [24832, 8192, 1, 256] - - [258, 74.582] - - - [24064, 1024, 1, 256] - - [299, 68.455] - - - [19456, 11776, 1, 256] - - [252, 75.91] - - - [30208, 14080, 1, 256] - - [251, 75.797] - - - [35584, 4096, 1, 256] - - [270, 73.157] - - - [9472, 5632, 1, 256] - - [261, 72.733] - - - [34816, 4096, 1, 256] - - [256, 73.248] - - - [9728, 1024, 1, 256] - - [276, 60.648] - - - [20224, 1024, 1, 256] - - [287, 66.946] - - - [34304, 4096, 1, 256] - - [269, 73.894] - - - [37120, 8192, 1, 256] - - [252, 74.285] - - - [42496, 512, 1, 256] - - [252, 66.06] - - - [14336, 4096, 1, 256] - - [254, 72.202] - - - [35840, 512, 1, 256] - - [276, 64.842] - - - [17408, 512, 1, 256] - - [276, 60.954] - - - [6656, 512, 1, 256] - - [252, 52.846] - - - [39168, 23040, 1, 256] - - [255, 74.625] - - - [38656, 2816, 1, 256] - - [301, 72.503] - - - [39424, 2865, 1, 256] - - [281, 70.508] - - - [36096, 2865, 1, 256] - - [278, 68.618] - - - [38144, 6144, 1, 256] - - [254, 74.027] - - - [39936, 256, 1, 256] - - [268, 59.512] - - - [39168, 6144, 1, 256] - - [268, 73.921] - - - [37680, 10240, 1, 256] - - [268, 62.962] - - - [37632, 1281, 1, 256] - - [301, 63.847] - - - [36096, 4096, 1, 256] - - [268, 72.242] - - - [36352, 256, 1, 256] - - [252, 55.862] - - - [38912, 1280, 1, 256] - - [256, 71.695] - - - [38448, 10240, 1, 256] - - [288, 62.59] - - - [39168, 10240, 1, 256] - - [255, 74.409] - - - [37168, 2865, 1, 256] - - [280, 61.603] - - - [37120, 1281, 1, 256] - - [301, 64.097] - - - [38704, 5120, 1, 256] - - [271, 62.977] - - - [38656, 2865, 1, 256] - - [281, 69.906] - - - [35072, 6144, 1, 256] - - [252, 74.189] - - - [37632, 5376, 1, 256] - - [252, 73.89] - - - [35840, 2048, 1, 256] - - [313, 71.487] - - - [39936, 1792, 1, 256] - - [254, 73.01] - - - [35072, 3329, 1, 256] - - [281, 69.553] - - - [36096, 1792, 1, 256] - - [258, 70.106] - - - [39424, 2048, 1, 256] - - [268, 71.915] - - - [38448, 256, 1, 256] - - [329, 51.782] - - - [36608, 10240, 1, 256] - - [256, 74.583] - - - [36352, 2048, 1, 256] - - [268, 71.757] - - - [36608, 2304, 1, 256] - - [256, 72.296] - - - [39680, 5376, 1, 256] - - [264, 74.025] - - - [36864, 512, 1, 256] - - [276, 65.314] - - - [35840, 2816, 1, 256] - - [252, 73.659] - - - [35584, 2048, 1, 256] - - [268, 70.786] - - - [39936, 2865, 1, 256] - - [256, 70.99] - - - [36608, 2048, 1, 256] - - [288, 70.999] - - - [39424, 5888, 1, 256] - - [252, 74.33] - - - [36096, 6144, 1, 256] - - [268, 73.136] - - - [37632, 2048, 1, 256] - - [288, 70.581] - - - [37680, 3840, 1, 256] - - [280, 63.791] - - - [38192, 2865, 1, 256] - - [280, 60.856] - - - [37120, 2048, 1, 256] - - [268, 71.342] - - - [38448, 4864, 1, 256] - - [263, 63.358] - - - [37632, 2816, 1, 256] - - [256, 72.628] - - - [39168, 256, 1, 256] - - [285, 58.604] - - - [38144, 4608, 1, 256] - - [264, 73.45] - - - [36352, 6144, 1, 256] - - [268, 74.483] - - - [37888, 3328, 1, 256] - - [252, 73.568] - - - [37120, 3584, 1, 256] - - [252, 73.471] - - - [39472, 256, 1, 256] - - [256, 52.657] - - - [37376, 5120, 1, 256] - - [264, 74.278] - - - [37632, 1280, 1, 256] - - [256, 70.731] - - - [38656, 3329, 1, 256] - - [257, 69.446] - - - [39168, 768, 1, 256] - - [296, 68.628] - - - [38448, 2816, 1, 256] - - [280, 63.43] - - - [38912, 2865, 1, 256] - - [249, 71.049] - - - [35328, 1024, 1, 256] - - [303, 69.48] - - - [36352, 2816, 1, 256] - - [252, 73.066] - - - [39424, 3328, 1, 256] - - [255, 73.378] - - - [36096, 2816, 1, 256] - - [291, 71.255] - - - [35840, 3584, 1, 256] - - [254, 74.213] - - - [37376, 1024, 1, 256] - - [299, 69.945] - - - [36400, 2865, 1, 256] - - [280, 61.927] - - - [36864, 1281, 1, 256] - - [254, 63.521] - - - [36400, 2816, 1, 256] - - [285, 63.396] - - - [37888, 5632, 1, 256] - - [254, 75.071] - - - [38656, 4352, 1, 256] - - [257, 73.699] - - - [35072, 2865, 1, 256] - - [256, 70.225] - - - [38912, 3329, 1, 256] - - [281, 70.513] - - - [38400, 6144, 1, 256] - - [255, 74.375] - - - [36608, 3328, 1, 256] - - [255, 72.821] - - - [36096, 3329, 1, 256] - - [255, 68.384] - - - [36864, 3329, 1, 256] - - [281, 69.763] - - - [38656, 4864, 1, 256] - - [274, 73.807] - - - [37120, 3328, 1, 256] - - [255, 72.974] - - - [35328, 1280, 1, 256] - - [284, 70.894] - - - [36400, 10240, 1, 256] - - [268, 62.71] - - - [39680, 2816, 1, 256] - - [264, 72.727] - - - [36656, 256, 1, 256] - - [285, 50.805] - - - [39936, 2816, 1, 256] - - [252, 73.724] - - - [37936, 2816, 1, 256] - - [280, 63.534] - - - [39984, 10240, 1, 256] - - [268, 63.649] - - - [35328, 10240, 1, 256] - - [278, 74.983] - - - [37120, 4864, 1, 256] - - [252, 73.988] - - - [39680, 5632, 1, 256] - - [264, 74.307] - - - [36400, 2560, 1, 256] - - [263, 63.434] - - - [37424, 10240, 1, 256] - - [268, 62.738] - - - [35072, 768, 1, 256] - - [271, 68.0] - - - [36656, 3072, 1, 256] - - [254, 62.236] - - - [36656, 2865, 1, 256] - - [263, 61.222] - - - [38400, 256, 1, 256] - - [252, 57.601] - - - [38912, 5376, 1, 256] - - [252, 74.867] - - - [38656, 3328, 1, 256] - - [255, 72.917] - - - [39216, 5376, 1, 256] - - [263, 63.737] - - - [37680, 256, 1, 256] - - [280, 51.618] - - - [38400, 2865, 1, 256] - - [281, 70.515] - - - [38912, 768, 1, 256] - - [267, 69.012] - - - [39424, 256, 1, 256] - - [252, 58.69] - - - [38400, 2048, 1, 256] - - [255, 71.748] - - - [39680, 2048, 1, 256] - - [280, 69.64] - - - [39728, 2816, 1, 256] - - [285, 63.085] - - - [38400, 10240, 1, 256] - - [252, 74.925] - - - [37888, 4352, 1, 256] - - [264, 74.699] - - - [35584, 1536, 1, 256] - - [262, 70.284] - - - [36864, 2048, 1, 256] - - [254, 68.862] - - - [35584, 3584, 1, 256] - - [264, 73.116] - - - [36352, 3328, 1, 256] - - [268, 73.516] - - - [39728, 10240, 1, 256] - - [268, 62.94] - - - [37168, 3328, 1, 256] - - [263, 63.032] - - - [35328, 2865, 1, 256] - - [297, 70.435] - - - [38704, 2816, 1, 256] - - [280, 63.386] - - - [39680, 1280, 1, 256] - - [252, 70.822] - - - [36608, 512, 1, 256] - - [267, 65.036] - - - [39728, 5888, 1, 256] - - [285, 63.022] - - - [37120, 256, 1, 256] - - [285, 56.339] - - - [39168, 5120, 1, 256] - - [251, 73.651] - - - [36912, 2816, 1, 256] - - [255, 61.151] - - - [35840, 3328, 1, 256] - - [254, 73.553] - - - [38144, 1792, 1, 256] - - [264, 71.784] - - - [35376, 10240, 1, 256] - - [268, 63.016] - - - [35840, 1280, 1, 256] - - [264, 71.379] - - - [35584, 3328, 1, 256] - - [269, 72.682] - - - [39936, 10240, 1, 256] - - [254, 75.215] - - - [37632, 3329, 1, 256] - - [259, 69.618] - - - [36608, 1281, 1, 256] - - [255, 64.424] - - - [39936, 2048, 1, 256] - - [268, 72.048] - - - [39472, 2816, 1, 256] - - [280, 63.357] - - - [38400, 2816, 1, 256] - - [252, 73.171] - - - [39216, 10240, 1, 256] - - [288, 62.289] - - - [38400, 4352, 1, 256] - - [256, 74.322] - - - [35840, 1792, 1, 256] - - [256, 72.46] - - - [38704, 256, 1, 256] - - [263, 52.174] - - - [37888, 6144, 1, 256] - - [254, 74.7] - - - [37376, 256, 1, 256] - - [258, 56.495] - - - [37936, 4096, 1, 256] - - [268, 62.768] - - - [35376, 1792, 1, 256] - - [263, 62.315] - - - [39984, 2816, 1, 256] - - [271, 63.532] - - - [39424, 5632, 1, 256] - - [254, 74.685] - - - [39680, 10240, 1, 256] - - [256, 74.481] - - - [39168, 2048, 1, 256] - - [268, 71.015] - - - [35632, 2865, 1, 256] - - [280, 61.238] - - - [36608, 4352, 1, 256] - - [264, 73.899] - - - [36608, 6144, 1, 256] - - [254, 74.021] - - - [39168, 1024, 1, 256] - - [296, 69.668] - - - [39168, 2865, 1, 256] - - [256, 69.967] - - - [36864, 2816, 1, 256] - - [252, 73.037] - - - [35584, 1281, 1, 256] - - [266, 64.143] - - - [35120, 1280, 1, 256] - - [285, 62.736] - - - [38912, 6144, 1, 256] - - [254, 74.651] - - - [37424, 2816, 1, 256] - - [280, 63.705] - - - [35328, 3328, 1, 256] - - [270, 73.449] - - - [35072, 1536, 1, 256] - - [264, 70.539] - - - [36864, 4608, 1, 256] - - [256, 73.415] - - - [36864, 4864, 1, 256] - - [256, 74.336] - - - [38912, 2816, 1, 256] - - [254, 73.697] - - - [35072, 1280, 1, 256] - - [276, 70.409] - - - [35840, 1281, 1, 256] - - [311, 64.22] - - - [38192, 256, 1, 256] - - [280, 51.414] - - - [39984, 256, 1, 256] - - [279, 53.316] - - - [37888, 256, 1, 256] - - [271, 57.593] - - - [38656, 1281, 1, 256] - - [316, 64.467] - - - [39168, 4864, 1, 256] - - [257, 73.766] - - - [37888, 10240, 1, 256] - - [256, 75.274] - - - [39680, 5888, 1, 256] - - [264, 73.797] - - - [39728, 256, 1, 256] - - [254, 53.166] - - - [36352, 2865, 1, 256] - - [254, 70.628] - - - [37888, 1281, 1, 256] - - [301, 64.407] - - - [37376, 2048, 1, 256] - - [270, 71.636] - - - [35584, 1280, 1, 256] - - [276, 70.281] - - - [39936, 5632, 1, 256] - - [254, 74.951] - - - [35376, 2865, 1, 256] - - [280, 61.863] - - - [35888, 256, 1, 256] - - [271, 50.379] - - - [37936, 10240, 1, 256] - - [268, 63.616] - - - [38656, 2048, 1, 256] - - [268, 71.141] - - - [36912, 10240, 1, 256] - - [255, 62.221] - - - [38448, 2865, 1, 256] - - [271, 61.645] - - - [37632, 6144, 1, 256] - - [264, 74.0] - - - [36608, 2816, 1, 256] - - [254, 72.815] - - - [37120, 2816, 1, 256] - - [255, 72.663] - - - [38912, 2048, 1, 256] - - [268, 70.594] - - - [37120, 5120, 1, 256] - - [254, 74.028] - - - [35632, 10240, 1, 256] - - [255, 62.964] - - - [37120, 10240, 1, 256] - - [254, 74.59] - - - [35840, 6144, 1, 256] - - [254, 74.653] - - - [36144, 10240, 1, 256] - - [268, 62.723] - - - [39424, 10240, 1, 256] - - [254, 74.889] - - - [39424, 1280, 1, 256] - - [261, 70.931] - - - [39472, 5888, 1, 256] - - [280, 63.366] - - - [38656, 4608, 1, 256] - - [255, 73.336] - - - [38144, 2048, 1, 256] - - [268, 71.056] - - - [37120, 1280, 1, 256] - - [262, 70.842] - - - [37424, 3584, 1, 256] - - [280, 63.462] - - - [39936, 6400, 1, 256] - - [256, 75.294] - - - [37120, 768, 1, 256] - - [271, 68.579] - - - [36608, 4608, 1, 256] - - [255, 73.42] - - - [35328, 2816, 1, 256] - - [278, 73.251] - - - [36608, 256, 1, 256] - - [282, 55.964] - - - [38448, 4608, 1, 256] - - [285, 62.818] - - - [37680, 2865, 1, 256] - - [280, 61.765] - - - [37120, 6144, 1, 256] - - [268, 73.953] - - - [39936, 3329, 1, 256] - - [281, 70.508] - - - [35120, 1536, 1, 256] - - [263, 61.566] - - - [37936, 4352, 1, 256] - - [252, 63.579] - - - [37888, 3584, 1, 256] - - [254, 74.221] - - - [36864, 1280, 1, 256] - - [282, 70.756] - - - [37120, 1024, 1, 256] - - [287, 69.346] - - - [38400, 1280, 1, 256] - - [261, 71.196] - - - [37376, 1281, 1, 256] - - [303, 64.736] - - - [36144, 2304, 1, 256] - - [271, 62.853] - - - [38960, 2816, 1, 256] - - [256, 62.02] - - - [35072, 2816, 1, 256] - - [254, 72.753] - - - [38960, 256, 1, 256] - - [285, 52.833] - - - [39168, 1280, 1, 256] - - [254, 70.66] - - - [40192, 2865, 1, 256] - - [252, 70.132] - - - [43520, 1280, 1, 256] - - [252, 71.545] - - - [41520, 7936, 1, 256] - - [263, 63.037] - - - [42496, 2048, 1, 256] - - [268, 72.118] - - - [41776, 7936, 1, 256] - - [271, 63.282] - - - [43520, 6144, 1, 256] - - [264, 74.517] - - - [41472, 3329, 1, 256] - - [281, 70.049] - - - [40448, 2865, 1, 256] - - [254, 70.49] - - - [43312, 9728, 1, 256] - - [285, 62.838] - - - [42544, 8704, 1, 256] - - [280, 62.483] - - - [41216, 3329, 1, 256] - - [300, 69.564] - - - [42240, 10240, 1, 256] - - [254, 74.534] - - - [41216, 1280, 1, 256] - - [258, 71.322] - - - [41216, 10240, 1, 256] - - [255, 74.599] - - - [40704, 2560, 1, 256] - - [252, 73.113] - - - [43264, 3329, 1, 256] - - [281, 69.794] - - - [43008, 10240, 1, 256] - - [252, 75.153] - - - [42496, 8960, 1, 256] - - [256, 75.285] - - - [44032, 6144, 1, 256] - - [252, 74.572] - - - [41984, 1536, 1, 256] - - [254, 71.539] - - - [40752, 256, 1, 256] - - [252, 54.047] - - - [43520, 9216, 1, 256] - - [268, 74.767] - - - [41728, 3584, 1, 256] - - [277, 71.854] - - - [41216, 7168, 1, 256] - - [254, 73.415] - - - [40448, 6656, 1, 256] - - [251, 74.479] - - - [41264, 2816, 1, 256] - - [263, 63.512] - - - [43008, 9216, 1, 256] - - [252, 74.511] - - - [42288, 2865, 1, 256] - - [280, 61.45] - - - [40496, 256, 1, 256] - - [252, 53.635] - - - [43056, 9472, 1, 256] - - [268, 62.988] - - - [41776, 8192, 1, 256] - - [268, 62.693] - - - [41984, 7680, 1, 256] - - [264, 75.275] - - - [40192, 1280, 1, 256] - - [264, 71.066] - - - [41984, 2865, 1, 256] - - [259, 71.09] - - - [42240, 3328, 1, 256] - - [268, 73.03] - - - [42240, 256, 1, 256] - - [285, 61.761] - - - [44032, 5632, 1, 256] - - [252, 75.031] - - - [41216, 256, 1, 256] - - [253, 60.115] - - - [43008, 1280, 1, 256] - - [256, 71.961] - - - [42752, 6144, 1, 256] - - [264, 74.06] - - - [43008, 2865, 1, 256] - - [249, 71.08] - - - [42240, 7936, 1, 256] - - [254, 74.473] - - - [43776, 5376, 1, 256] - - [255, 73.395] - - - [41216, 3072, 1, 256] - - [264, 73.182] - - - [40704, 6144, 1, 256] - - [252, 74.159] - - - [43520, 2048, 1, 256] - - [288, 71.975] - - - [40448, 1281, 1, 256] - - [288, 64.998] - - - [40448, 6144, 1, 256] - - [252, 74.257] - - - [43056, 2865, 1, 256] - - [254, 60.62] - - - [40192, 10240, 1, 256] - - [254, 74.615] - - - [41008, 7168, 1, 256] - - [268, 62.179] - - - [41472, 6144, 1, 256] - - [270, 74.195] - - - [40448, 6912, 1, 256] - - [257, 74.755] - - - [41984, 1792, 1, 256] - - [254, 72.966] - - - [40960, 3329, 1, 256] - - [281, 62.664] - - - [44032, 10240, 1, 256] - - [252, 75.084] - - - [43008, 4608, 1, 256] - - [264, 74.106] - - - [40240, 6656, 1, 256] - - [271, 63.007] - - - [40704, 3329, 1, 256] - - [259, 69.77] - - - [41984, 2048, 1, 256] - - [268, 71.786] - - - [40192, 3328, 1, 256] - - [268, 72.906] - - - [42496, 256, 1, 256] - - [290, 56.259] - - - [40496, 6912, 1, 256] - - [280, 63.78] - - - [40960, 1280, 1, 256] - - [252, 64.486] - - - [42496, 3328, 1, 256] - - [255, 73.671] - - - [42496, 3329, 1, 256] - - [259, 70.164] - - - [41472, 1281, 1, 256] - - [288, 65.032] - - - [42752, 8960, 1, 256] - - [249, 74.694] - - - [40240, 10240, 1, 256] - - [268, 62.723] - - - [40752, 2865, 1, 256] - - [263, 61.872] - - - [41472, 7680, 1, 256] - - [264, 74.81] - - - [41264, 7680, 1, 256] - - [280, 63.108] - - - [41728, 10240, 1, 256] - - [268, 74.303] - - - [40704, 6912, 1, 256] - - [254, 74.49] - - - [40960, 6144, 1, 256] - - [265, 67.029] - - - [42496, 2865, 1, 256] - - [254, 70.925] - - - [41728, 7936, 1, 256] - - [268, 73.453] - - - [43008, 256, 1, 256] - - [254, 57.193] - - - [43776, 10240, 1, 256] - - [255, 74.271] - - - [42496, 1281, 1, 256] - - [268, 65.226] - - - [40960, 2816, 1, 256] - - [249, 66.747] - - - [43264, 1281, 1, 256] - - [301, 64.133] - - - [41728, 1536, 1, 256] - - [284, 69.755] - - - [43520, 9728, 1, 256] - - [254, 74.909] - - - [43776, 256, 1, 256] - - [285, 56.609] - - - [44032, 1792, 1, 256] - - [264, 72.939] - - - [42240, 1792, 1, 256] - - [254, 71.959] - - - [42800, 9216, 1, 256] - - [255, 62.893] - - - [42496, 6144, 1, 256] - - [252, 74.501] - - - [40704, 2048, 1, 256] - - [255, 71.158] - - - [42752, 2816, 1, 256] - - [254, 72.973] - - - [40960, 6912, 1, 256] - - [265, 68.041] - - - [42496, 4352, 1, 256] - - [252, 74.491] - - - [42752, 4608, 1, 256] - - [256, 73.474] - - - [44288, 6144, 1, 256] - - [268, 74.003] - - - [41008, 2816, 1, 256] - - [268, 61.755] - - - [43264, 256, 1, 256] - - [287, 56.868] - - - [43520, 5120, 1, 256] - - [252, 74.521] - - - [41216, 3328, 1, 256] - - [255, 73.012] - - - [43264, 9728, 1, 256] - - [252, 74.48] - - - [40704, 1281, 1, 256] - - [268, 64.284] - - - [40960, 512, 1, 256] - - [254, 64.162] - - - [40752, 10240, 1, 256] - - [255, 62.91] - - - [43568, 2816, 1, 256] - - [263, 62.54] - - - [41216, 2048, 1, 256] - - [268, 71.246] - - - [43312, 2816, 1, 256] - - [263, 63.107] - - - [41264, 7424, 1, 256] - - [263, 63.342] - - - [41472, 256, 1, 256] - - [264, 61.092] - - - [43520, 1024, 1, 256] - - [280, 70.335] - - - [43776, 9472, 1, 256] - - [301, 74.186] - - - [43008, 8960, 1, 256] - - [265, 75.419] - - - [41520, 7680, 1, 256] - - [255, 62.77] - - - [41472, 2816, 1, 256] - - [274, 72.969] - - - [40752, 2816, 1, 256] - - [271, 62.707] - - - [40448, 6400, 1, 256] - - [264, 74.798] - - - [41472, 1024, 1, 256] - - [290, 70.004] - - - [43312, 2865, 1, 256] - - [263, 61.386] - - - [40240, 2865, 1, 256] - - [271, 61.619] - - - [40192, 3329, 1, 256] - - [264, 69.591] - - - [42752, 512, 1, 256] - - [282, 65.629] - - - [42240, 8192, 1, 256] - - [255, 74.303] - - - [40960, 2865, 1, 256] - - [281, 63.581] - - - [42496, 2816, 1, 256] - - [256, 73.572] - - - [42240, 8448, 1, 256] - - [268, 74.494] - - - [41984, 3328, 1, 256] - - [254, 73.564] - - - [44288, 256, 1, 256] - - [250, 57.649] - - - [43520, 3329, 1, 256] - - [281, 70.231] - - - [42752, 1281, 1, 256] - - [264, 64.042] - - - [41008, 256, 1, 256] - - [252, 53.831] - - - [40704, 6656, 1, 256] - - [256, 74.314] - - - [43264, 10240, 1, 256] - - [254, 74.641] - - - [43520, 2865, 1, 256] - - [254, 70.933] - - - [43056, 9216, 1, 256] - - [255, 62.923] - - - [42544, 10240, 1, 256] - - [268, 62.521] - - - [40704, 1280, 1, 256] - - [284, 70.86] - - - [43520, 5376, 1, 256] - - [254, 74.646] - - - [43776, 1280, 1, 256] - - [288, 68.773] - - - [41520, 256, 1, 256] - - [271, 55.016] - - - [40192, 6656, 1, 256] - - [252, 74.325] - - - [41520, 2816, 1, 256] - - [285, 63.286] - - - [42544, 2865, 1, 256] - - [285, 61.581] - - - [41472, 1280, 1, 256] - - [294, 71.256] - - - [42240, 2865, 1, 256] - - [254, 70.024] - - - [41264, 10240, 1, 256] - - [254, 62.503] - - - [41216, 1024, 1, 256] - - [299, 69.891] - - - [40960, 7424, 1, 256] - - [265, 67.903] - - - [41216, 1281, 1, 256] - - [316, 64.548] - - - [40448, 3329, 1, 256] - - [259, 69.924] - - - [40704, 256, 1, 256] - - [263, 60.412] - - - [42240, 6144, 1, 256] - - [254, 74.187] - - - [42752, 256, 1, 256] - - [271, 56.12] - - - [43776, 1536, 1, 256] - - [318, 67.546] - - - [43008, 4864, 1, 256] - - [252, 74.853] - - - [44288, 1281, 1, 256] - - [301, 64.844] - - - [41264, 256, 1, 256] - - [330, 53.76] - - - [42240, 4096, 1, 256] - - [268, 73.206] - - - [43264, 9472, 1, 256] - - [254, 74.68] - - - [43776, 2816, 1, 256] - - [268, 72.397] - - - [41984, 8448, 1, 256] - - [256, 75.147] - - - [43776, 2865, 1, 256] - - [255, 68.682] - - - [42544, 256, 1, 256] - - [264, 49.843] - - - [41472, 3072, 1, 256] - - [264, 73.12] - - - [43056, 256, 1, 256] - - [264, 50.543] - - - [42496, 8448, 1, 256] - - [255, 74.955] - - - [42752, 8704, 1, 256] - - [254, 74.639] - - - [40448, 10240, 1, 256] - - [274, 74.672] - - - [43568, 10240, 1, 256] - - [268, 62.816] - - - [41472, 2865, 1, 256] - - [252, 70.643] - - - [41728, 3329, 1, 256] - - [255, 68.274] - - - [40960, 6656, 1, 256] - - [249, 67.684] - - - [40496, 2816, 1, 256] - - [271, 63.636] - - - [43056, 10240, 1, 256] - - [268, 63.048] - - - [41728, 8192, 1, 256] - - [268, 74.078] - - - [43776, 2048, 1, 256] - - [255, 70.652] - - - [42752, 1280, 1, 256] - - [254, 71.138] - - - [42240, 2816, 1, 256] - - [254, 72.834] - - - [44288, 3328, 1, 256] - - [255, 72.991] - - - [42032, 2816, 1, 256] - - [280, 63.145] - - - [11264, 1792, 1, 256] - - [252, 67.19] - - - [14640, 1536, 1, 256] - - [254, 59.341] - - - [19968, 3328, 1, 256] - - [254, 72.493] - - - [15408, 2816, 1, 256] - - [252, 62.707] - - - [3840, 512, 1, 256] - - [286, 32.591] - - - [16384, 768, 1, 256] - - [252, 54.969] - - - [12544, 2048, 1, 256] - - [250, 67.949] - - - [8240, 5376, 1, 256] - - [264, 65.489] - - - [12800, 256, 1, 256] - - [282, 51.019] - - - [10032, 6912, 1, 256] - - [263, 63.949] - - - [5888, 2865, 1, 256] - - [276, 64.896] - - - [17920, 4096, 1, 256] - - [301, 72.746] - - - [16432, 2865, 1, 256] - - [265, 64.164] - - - [11776, 8704, 1, 256] - - [294, 74.604] - - - [19760, 10240, 1, 256] - - [271, 62.731] - - - [19712, 2865, 1, 256] - - [257, 67.478] - - - [18944, 3329, 1, 256] - - [257, 69.341] - - - [13872, 256, 1, 256] - - [262, 51.278] - - - [19248, 10240, 1, 256] - - [263, 63.303] - - - [12080, 9216, 1, 256] - - [280, 63.728] - - - [17712, 4608, 1, 256] - - [285, 63.487] - - - [8752, 5888, 1, 256] - - [263, 63.141] - - - [5376, 3328, 1, 256] - - [252, 68.034] - - - [9264, 256, 1, 256] - - [254, 38.214] - - - [2608, 2353, 1, 256] - - [262, 50.213] - - - [7424, 3329, 1, 256] - - [254, 65.354] - - - [9984, 512, 1, 256] - - [261, 53.859] - - - [17408, 3329, 1, 256] - - [264, 69.421] - - - [11264, 3840, 1, 256] - - [261, 72.724] - - - [3584, 768, 1, 256] - - [250, 44.664] - - - [15360, 2048, 1, 256] - - [288, 69.024] - - - [6144, 1281, 1, 256] - - [262, 55.503] - - - [15152, 256, 1, 256] - - [276, 41.36] - - - [6400, 4864, 1, 256] - - [282, 70.408] - - - [1792, 1280, 1, 256] - - [286, 37.644] - - - [10544, 2816, 1, 256] - - [285, 61.779] - - - [16384, 1280, 1, 256] - - [249, 58.499] - - - [13824, 1281, 1, 256] - - [280, 61.987] - - - [8448, 5376, 1, 256] - - [276, 71.936] - - - [6656, 3584, 1, 256] - - [262, 68.057] - - - [11312, 256, 1, 256] - - [263, 44.077] - - - [12032, 8448, 1, 256] - - [261, 73.765] - - - [6400, 5120, 1, 256] - - [262, 70.545] - - - [4400, 1280, 1, 256] - - [252, 47.188] - - - [7168, 2816, 1, 256] - - [261, 67.248] - - - [4144, 1024, 1, 256] - - [254, 44.266] - - - [6400, 2865, 1, 256] - - [264, 64.555] - - - [7168, 1536, 1, 256] - - [276, 59.413] - - - [17664, 1281, 1, 256] - - [266, 62.694] - - - [17712, 256, 1, 256] - - [250, 46.132] - - - [13056, 5376, 1, 256] - - [261, 73.45] - - - [3328, 1281, 1, 256] - - [276, 46.053] - - - [2304, 2048, 1, 256] - - [262, 50.783] - - - [7424, 4352, 1, 256] - - [282, 71.451] - - - [19456, 2816, 1, 256] - - [252, 72.622] - - - [18688, 5376, 1, 256] - - [252, 73.843] - - - [10240, 768, 1, 256] - - [271, 56.426] - - - [18432, 3329, 1, 256] - - [256, 69.63] - - - [1536, 1280, 1, 256] - - [274, 32.732] - - - [15360, 6144, 1, 256] - - [252, 74.341] - - - [4864, 3328, 1, 256] - - [282, 67.47] - - - [13056, 9984, 1, 256] - - [252, 75.033] - - - [7168, 3840, 1, 256] - - [262, 70.01] - - - [14592, 10240, 1, 256] - - [295, 73.199] - - - [16128, 2816, 1, 256] - - [254, 71.212] - - - [13568, 512, 1, 256] - - [262, 59.368] - - - [12032, 1281, 1, 256] - - [261, 59.533] - - - [12032, 2048, 1, 256] - - [285, 67.9] - - - [13312, 9984, 1, 256] - - [256, 75.719] - - - [15408, 2304, 1, 256] - - [254, 63.5] - - - [18688, 10240, 1, 256] - - [256, 74.915] - - - [2048, 1793, 1, 256] - - [262, 40.502] - - - [10032, 256, 1, 256] - - [261, 39.665] - - - [14080, 2865, 1, 256] - - [294, 68.381] - - - [18944, 1024, 1, 256] - - [267, 67.129] - - - [11008, 2048, 1, 256] - - [331, 64.317] - - - [8448, 768, 1, 256] - - [299, 55.654] - - - [11824, 8960, 1, 256] - - [280, 64.102] - - - [6960, 3840, 1, 256] - - [256, 62.457] - - - [16944, 2865, 1, 256] - - [264, 63.601] - - - [3376, 2816, 1, 256] - - [256, 53.634] - - - [17152, 3329, 1, 256] - - [282, 69.147] - - - [5888, 1281, 1, 256] - - [262, 53.849] - - - [18688, 1280, 1, 256] - - [261, 67.925] - - - [15616, 1281, 1, 256] - - [263, 61.056] - - - [12032, 4352, 1, 256] - - [262, 71.886] - - - [17408, 1281, 1, 256] - - [288, 62.295] - - - [16176, 2865, 1, 256] - - [256, 63.843] - - - [7168, 256, 1, 256] - - [278, 30.817] - - - [10752, 7680, 1, 256] - - [252, 74.394] - - - [17152, 2865, 1, 256] - - [254, 68.888] - - - [3584, 2304, 1, 256] - - [261, 59.247] - - - [18944, 5376, 1, 256] - - [294, 74.399] - - - [7936, 1281, 1, 256] - - [267, 54.833] - - - [9472, 3328, 1, 256] - - [267, 70.075] - - - [15616, 10240, 1, 256] - - [256, 74.822] - - - [10032, 2865, 1, 256] - - [254, 59.522] - - - [9472, 2816, 1, 256] - - [262, 69.497] - - - [9216, 1536, 1, 256] - - [276, 65.881] - - - [16896, 3328, 1, 256] - - [264, 72.404] - - - [11008, 1280, 1, 256] - - [258, 63.427] - - - [2096, 1841, 1, 256] - - [261, 40.144] - - - [17968, 10240, 1, 256] - - [263, 63.221] - - - [7728, 4864, 1, 256] - - [252, 61.971] - - - [4608, 1792, 1, 256] - - [276, 58.808] - - - [18736, 2816, 1, 256] - - [271, 62.4] - - - [19456, 6144, 1, 256] - - [252, 74.371] - - - [17200, 3840, 1, 256] - - [254, 64.105] - - - [17408, 2816, 1, 256] - - [282, 72.23] - - - [8960, 2816, 1, 256] - - [276, 70.027] - - - [17152, 10240, 1, 256] - - [284, 74.66] - - - [10288, 2865, 1, 256] - - [256, 60.769] - - - [7424, 4096, 1, 256] - - [254, 70.152] - - - [16384, 2865, 1, 256] - - [249, 59.821] - - - [11520, 3329, 1, 256] - - [258, 66.836] - - - [3120, 2816, 1, 256] - - [261, 56.127] - - - [13824, 6144, 1, 256] - - [256, 73.534] - - - [12032, 3328, 1, 256] - - [282, 70.445] - - - [19456, 5632, 1, 256] - - [252, 74.96] - - - [11520, 3840, 1, 256] - - [294, 71.601] - - - [14848, 1280, 1, 256] - - [262, 66.953] - - - [16384, 256, 1, 256] - - [267, 45.395] - - - [4864, 2816, 1, 256] - - [261, 65.058] - - - [4608, 2048, 1, 256] - - [276, 58.79] - - - [3632, 768, 1, 256] - - [271, 42.625] - - - [5632, 2865, 1, 256] - - [262, 63.299] - - - [12800, 1280, 1, 256] - - [261, 64.295] - - - [11264, 8448, 1, 256] - - [264, 74.919] - - - [4096, 2560, 1, 256] - - [282, 63.739] - - - [11568, 8448, 1, 256] - - [271, 63.978] - - - [13056, 1536, 1, 256] - - [252, 65.618] - - - [3072, 1281, 1, 256] - - [262, 43.281] - - - [9216, 2816, 1, 256] - - [262, 69.713] - - - [18992, 2865, 1, 256] - - [252, 60.345] - - - [2304, 1280, 1, 256] - - [262, 46.866] - - - [15920, 2560, 1, 256] - - [264, 64.2] - - - [14080, 3328, 1, 256] - - [284, 70.137] - - - [8448, 5632, 1, 256] - - [276, 72.533] - - - [6704, 2865, 1, 256] - - [252, 57.23] - - - [18992, 10240, 1, 256] - - [271, 63.555] - - - [18480, 5120, 1, 256] - - [264, 62.77] - - - [7424, 3328, 1, 256] - - [267, 69.576] - - - [13568, 3329, 1, 256] - - [252, 67.986] - - - [7168, 3584, 1, 256] - - [254, 69.608] - - - [6400, 1281, 1, 256] - - [252, 57.395] - - - [19248, 256, 1, 256] - - [290, 48.424] - - - [17920, 4864, 1, 256] - - [258, 74.051] - - - [18224, 4864, 1, 256] - - [263, 63.328] - - - [12288, 4864, 1, 256] - - [264, 73.366] - - - [17664, 3329, 1, 256] - - [254, 68.974] - - - [11776, 3329, 1, 256] - - [284, 68.167] - - - [12592, 256, 1, 256] - - [332, 47.498] - - - [10800, 2816, 1, 256] - - [287, 63.108] - - - [17152, 1536, 1, 256] - - [282, 68.551] - - - [8192, 4608, 1, 256] - - [252, 70.244] - - - [16640, 10240, 1, 256] - - [252, 74.759] - - - [18176, 2816, 1, 256] - - [254, 71.857] - - - [18944, 6144, 1, 256] - - [251, 74.1] - - - [13056, 1281, 1, 256] - - [290, 59.592] - - - [7728, 2865, 1, 256] - - [254, 59.378] - - - [17456, 256, 1, 256] - - [261, 45.992] - - - [7168, 5632, 1, 256] - - [276, 72.334] - - - [15104, 3329, 1, 256] - - [264, 68.888] - - - [10032, 7168, 1, 256] - - [280, 63.028] - - - [10496, 6144, 1, 256] - - [262, 72.906] - - - [17664, 3840, 1, 256] - - [286, 72.723] - - - [10496, 6912, 1, 256] - - [261, 73.557] - - - [15872, 256, 1, 256] - - [261, 44.267] - - - [3584, 512, 1, 256] - - [286, 31.134] - - - [11264, 8192, 1, 256] - - [264, 74.546] - - - [18688, 3329, 1, 256] - - [254, 69.295] - - - [18432, 512, 1, 256] - - [254, 58.069] - - - [13568, 2865, 1, 256] - - [252, 68.418] - - - [11264, 256, 1, 256] - - [261, 46.464] - - - [18944, 5120, 1, 256] - - [274, 74.02] - - - [15872, 2560, 1, 256] - - [276, 71.856] - - - [13360, 2865, 1, 256] - - [252, 61.513] - - - [15616, 3329, 1, 256] - - [254, 69.058] - - - [18176, 512, 1, 256] - - [267, 57.477] - - - [1328, 1329, 1, 256] - - [276, 29.006] - - - [5376, 1792, 1, 256] - - [262, 59.533] - - - [6400, 3328, 1, 256] - - [261, 69.436] - - - [12288, 4608, 1, 256] - - [254, 72.894] - - - [5888, 2304, 1, 256] - - [262, 64.782] - - - [3072, 1536, 1, 256] - - [262, 50.641] - - - [15616, 2048, 1, 256] - - [290, 69.79] - - - [8960, 3328, 1, 256] - - [262, 69.542] - - - [8448, 2048, 1, 256] - - [267, 65.944] - - - [14384, 1280, 1, 256] - - [256, 58.663] - - - [18176, 2048, 1, 256] - - [250, 70.358] - - - [12080, 8960, 1, 256] - - [263, 63.832] - - - [16688, 2816, 1, 256] - - [254, 64.939] - - - [14336, 256, 1, 256] - - [305, 41.029] - - - [14384, 10240, 1, 256] - - [254, 63.156] - - - [8704, 1024, 1, 256] - - [250, 61.245] - - - [18480, 256, 1, 256] - - [287, 47.093] - - - [15872, 6144, 1, 256] - - [254, 74.514] - - - [3328, 3328, 1, 256] - - [261, 59.627] - - - [9520, 2816, 1, 256] - - [285, 61.723] - - - [3072, 2865, 1, 256] - - [307, 55.246] - - - [12032, 9216, 1, 256] - - [281, 73.508] - - - [13824, 2048, 1, 256] - - [271, 69.586] - - - [17664, 4352, 1, 256] - - [284, 73.069] - - - [17408, 1280, 1, 256] - - [276, 68.046] - - - [18736, 2865, 1, 256] - - [285, 60.889] - - - [11008, 7680, 1, 256] - - [298, 73.369] - - - [11776, 8192, 1, 256] - - [254, 74.076] - - - [7936, 2816, 1, 256] - - [282, 67.886] - - - [12032, 2816, 1, 256] - - [276, 70.599] - - - [5680, 2865, 1, 256] - - [264, 57.715] - - - [18432, 4864, 1, 256] - - [254, 74.451] - - - [9728, 3328, 1, 256] - - [256, 71.105] - - - [2304, 2305, 1, 256] - - [284, 46.736] - - - [13312, 1280, 1, 256] - - [276, 66.37] - - - [7424, 3840, 1, 256] - - [276, 70.87] - - - [18480, 2816, 1, 256] - - [256, 62.353] - - - [7424, 4608, 1, 256] - - [250, 70.809] - - - [5632, 2560, 1, 256] - - [261, 67.482] - - - [6704, 2816, 1, 256] - - [285, 59.919] - - - [17968, 2816, 1, 256] - - [271, 62.628] - - - [14640, 1280, 1, 256] - - [254, 58.845] - - - [16128, 1280, 1, 256] - - [261, 66.94] - - - [4352, 2816, 1, 256] - - [262, 64.474] - - - [6144, 2048, 1, 256] - - [290, 63.394] - - - [5632, 2048, 1, 256] - - [254, 60.737] - - - [9728, 1281, 1, 256] - - [254, 59.385] - - - [10800, 256, 1, 256] - - [263, 41.971] - - - [5376, 1280, 1, 256] - - [261, 59.027] - - - [2048, 1280, 1, 256] - - [284, 42.717] - - - [4096, 3329, 1, 256] - - [286, 63.485] - - - [19968, 2865, 1, 256] - - [252, 70.059] - - - [15360, 2304, 1, 256] - - [282, 71.231] - - - [19200, 2048, 1, 256] - - [250, 69.891] - - - [18688, 5632, 1, 256] - - [252, 74.26] - - - [17152, 256, 1, 256] - - [267, 47.478] - - - [14896, 1792, 1, 256] - - [271, 61.582] - - - [7984, 2865, 1, 256] - - [264, 58.999] - - - [7216, 256, 1, 256] - - [267, 30.145] - - - [7424, 2816, 1, 256] - - [261, 68.231] - - - [10544, 256, 1, 256] - - [261, 41.412] - - - [8240, 2816, 1, 256] - - [254, 63.984] - - - [18224, 10240, 1, 256] - - [280, 63.963] - - - [5376, 1281, 1, 256] - - [262, 50.065] - - - [18176, 4352, 1, 256] - - [264, 73.61] - - - [15616, 6144, 1, 256] - - [264, 73.791] - - - [14640, 10240, 1, 256] - - [285, 63.516] - - - [19968, 10240, 1, 256] - - [254, 75.102] - - - [15152, 2816, 1, 256] - - [280, 63.055] - - - [8448, 2816, 1, 256] - - [282, 67.887] - - - [17200, 4096, 1, 256] - - [271, 63.61] - - - [13056, 256, 1, 256] - - [254, 51.76] - - - [7424, 1792, 1, 256] - - [261, 63.557] - - - [12336, 2865, 1, 256] - - [264, 59.727] - - - [12800, 2048, 1, 256] - - [263, 68.659] - - - [16384, 2816, 1, 256] - - [249, 62.061] - - - [9008, 2865, 1, 256] - - [254, 59.73] - - - [16640, 256, 1, 256] - - [266, 45.255] - - - [10240, 3329, 1, 256] - - [256, 68.28] - - - [6912, 2865, 1, 256] - - [252, 64.594] - - - [12544, 1024, 1, 256] - - [290, 61.314] - - - [16896, 2048, 1, 256] - - [287, 69.794] - - - [11008, 1281, 1, 256] - - [258, 56.944] - - - [7168, 3329, 1, 256] - - [256, 66.789] - - - [12288, 8960, 1, 256] - - [254, 75.243] - - - [9728, 6656, 1, 256] - - [261, 73.905] - - - [10544, 7424, 1, 256] - - [254, 63.368] - - - [3888, 2816, 1, 256] - - [254, 53.57] - - - [17200, 10240, 1, 256] - - [256, 63.374] - - - [9728, 1280, 1, 256] - - [261, 65.329] - - - [5888, 4352, 1, 256] - - [261, 69.203] - - - [19456, 3329, 1, 256] - - [254, 69.935] - - - [8496, 2865, 1, 256] - - [252, 59.372] - - - [9984, 3329, 1, 256] - - [282, 66.875] - - - [13616, 2816, 1, 256] - - [285, 62.517] - - - [10240, 6656, 1, 256] - - [256, 74.233] - - - [16896, 1280, 1, 256] - - [282, 68.869] - - - [11008, 7424, 1, 256] - - [312, 73.158] - - - [6912, 2816, 1, 256] - - [262, 67.987] - - - [17152, 4096, 1, 256] - - [281, 72.113] - - - [13824, 3328, 1, 256] - - [276, 71.534] - - - [9472, 256, 1, 256] - - [284, 39.795] - - - [7168, 1281, 1, 256] - - [276, 55.992] - - - [16176, 3072, 1, 256] - - [264, 65.657] - - - [14336, 512, 1, 256] - - [254, 52.927] - - - [4864, 1280, 1, 256] - - [261, 54.748] - - - [8240, 256, 1, 256] - - [261, 34.619] - - - [16128, 10240, 1, 256] - - [254, 74.965] - - - [7680, 1281, 1, 256] - - [254, 59.333] - - - [10752, 2865, 1, 256] - - [252, 68.123] - - - [19968, 1280, 1, 256] - - [262, 68.465] - - - [19200, 6144, 1, 256] - - [256, 73.718] - - - [13824, 3329, 1, 256] - - [256, 68.838] - - - [17920, 3329, 1, 256] - - [278, 69.263] - - - [12032, 256, 1, 256] - - [252, 49.22] - - - [14336, 10240, 1, 256] - - [252, 75.425] - - - [14896, 2816, 1, 256] - - [271, 62.421] - - - [8192, 2816, 1, 256] - - [262, 68.313] - - - [8960, 1281, 1, 256] - - [276, 59.542] - - - [8752, 2816, 1, 256] - - [256, 62.219] - - - [2560, 2561, 1, 256] - - [261, 57.087] - - - [12544, 6144, 1, 256] - - [252, 73.28] - - - [3328, 2865, 1, 256] - - [261, 58.513] - - - [11264, 2865, 1, 256] - - [256, 68.379] - - - [16128, 3329, 1, 256] - - [252, 68.765] - - - [11264, 2816, 1, 256] - - [282, 70.812] - - - [14848, 3329, 1, 256] - - [254, 69.124] - - - [10288, 7424, 1, 256] - - [252, 63.45] - - - [13616, 10240, 1, 256] - - [263, 63.511] - - - [9984, 3328, 1, 256] - - [282, 70.3] - - - [11264, 7936, 1, 256] - - [282, 74.907] - - - [17664, 1792, 1, 256] - - [261, 69.801] - - - [16944, 2816, 1, 256] - - [256, 64.078] - - - [3376, 2865, 1, 256] - - [254, 54.171] - - - [11776, 8448, 1, 256] - - [254, 74.414] - - - [14080, 512, 1, 256] - - [254, 61.021] - - - [6656, 2048, 1, 256] - - [250, 63.698] - - - [17664, 1280, 1, 256] - - [262, 67.773] - - - [9984, 256, 1, 256] - - [278, 41.767] - - - [4608, 3329, 1, 256] - - [276, 64.498] - - - [2864, 2865, 1, 256] - - [254, 53.819] - - - [14080, 1281, 1, 256] - - [252, 60.65] - - - [10240, 7424, 1, 256] - - [264, 74.721] - - - [19712, 256, 1, 256] - - [267, 52.927] - - - [9728, 3329, 1, 256] - - [254, 68.007] - - - [13568, 1792, 1, 256] - - [276, 68.552] - - - [8704, 2048, 1, 256] - - [267, 67.585] - - - [12800, 2865, 1, 256] - - [258, 68.74] - - - [5120, 1536, 1, 256] - - [262, 56.426] - - - [10240, 256, 1, 256] - - [261, 42.778] - - - [12800, 1281, 1, 256] - - [260, 61.721] - - - [15360, 3329, 1, 256] - - [254, 69.087] - - - [17152, 3328, 1, 256] - - [252, 71.81] - - - [13824, 2816, 1, 256] - - [276, 71.392] - - - [3632, 2865, 1, 256] - - [264, 51.217] - - - [3840, 768, 1, 256] - - [285, 46.673] - - - [17712, 10240, 1, 256] - - [280, 63.117] - - - [12800, 6144, 1, 256] - - [286, 73.468] - - - [17920, 10240, 1, 256] - - [264, 75.114] - - - [11312, 2865, 1, 256] - - [252, 60.631] - - - [4096, 1280, 1, 256] - - [333, 54.347] - - - [14640, 2865, 1, 256] - - [264, 59.354] - - - [7680, 1280, 1, 256] - - [262, 60.456] - - - [14080, 1280, 1, 256] - - [262, 67.641] - - - [7936, 256, 1, 256] - - [253, 33.726] - - - [2816, 2560, 1, 256] - - [276, 62.208] - - - [12800, 9984, 1, 256] - - [252, 75.144] - - - [18176, 2865, 1, 256] - - [254, 69.649] - - - [3328, 2816, 1, 256] - - [276, 58.508] - - - [13824, 1280, 1, 256] - - [261, 67.666] - - - [17152, 1281, 1, 256] - - [267, 62.238] - - - [14848, 1792, 1, 256] - - [276, 70.07] - - - [18176, 1281, 1, 256] - - [271, 62.024] - - - [11824, 8704, 1, 256] - - [285, 64.185] - - - [8704, 256, 1, 256] - - [286, 37.042] - - - [14128, 768, 1, 256] - - [285, 52.887] - - - [9264, 6144, 1, 256] - - [254, 63.829] - - - [18944, 5632, 1, 256] - - [294, 74.526] - - - [2048, 2049, 1, 256] - - [262, 45.631] - - - [15616, 2560, 1, 256] - - [262, 71.001] - - - [14848, 1536, 1, 256] - - [262, 68.167] - - - [6400, 1024, 1, 256] - - [262, 57.281] - - - [4656, 1536, 1, 256] - - [262, 47.999] - - - [2864, 2816, 1, 256] - - [262, 53.201] - - - [3840, 2304, 1, 256] - - [252, 62.014] - - - [16384, 2560, 1, 256] - - [265, 60.625] - - - [16688, 10240, 1, 256] - - [252, 65.624] - - - [7936, 4608, 1, 256] - - [252, 70.813] - - - [14336, 3329, 1, 256] - - [256, 69.103] - - - [16896, 6144, 1, 256] - - [254, 74.464] - - - [17408, 6144, 1, 256] - - [252, 74.582] - - - [17968, 4864, 1, 256] - - [263, 63.918] - - - [14336, 2865, 1, 256] - - [264, 69.431] - - - [13568, 2048, 1, 256] - - [267, 68.59] - - - [5424, 2816, 1, 256] - - [280, 59.939] - - - [7984, 256, 1, 256] - - [271, 33.496] - - - [19712, 3329, 1, 256] - - [274, 67.597] - - - [17152, 1280, 1, 256] - - [262, 67.098] - - - [4864, 1792, 1, 256] - - [261, 61.43] - - - [11520, 8448, 1, 256] - - [256, 73.729] - - - [13568, 10240, 1, 256] - - [256, 74.74] - - - [19504, 6144, 1, 256] - - [256, 63.873] - - - [18688, 2865, 1, 256] - - [261, 69.555] - - - [18688, 6144, 1, 256] - - [256, 74.125] - - - [12288, 9472, 1, 256] - - [264, 75.314] - - - [16128, 512, 1, 256] - - [276, 56.775] - - - [9216, 3329, 1, 256] - - [261, 67.666] - - - [6448, 3328, 1, 256] - - [254, 58.765] - - - [11520, 8704, 1, 256] - - [294, 74.005] - - - [10496, 7168, 1, 256] - - [264, 72.807] - - - [10752, 7168, 1, 256] - - [254, 73.268] - - - [12080, 256, 1, 256] - - [263, 46.156] - - - [4608, 1280, 1, 256] - - [261, 51.827] - - - [8704, 3329, 1, 256] - - [282, 67.535] - - - [11008, 2816, 1, 256] - - [312, 67.657] - - - [11008, 1536, 1, 256] - - [284, 62.887] - - - [7168, 4096, 1, 256] - - [254, 69.656] - - - [19504, 2865, 1, 256] - - [264, 61.604] - - - [1840, 1792, 1, 256] - - [252, 47.791] - - - [9008, 256, 1, 256] - - [285, 36.85] - - - [11056, 7936, 1, 256] - - [254, 64.522] - - - [17664, 10240, 1, 256] - - [278, 74.207] - - - [16640, 768, 1, 256] - - [250, 59.572] - - - [19504, 2816, 1, 256] - - [263, 63.411] - - - [14080, 768, 1, 256] - - [287, 64.176] - - - [5376, 2560, 1, 256] - - [282, 64.677] - - - [5120, 3840, 1, 256] - - [276, 68.426] - - - [4608, 3072, 1, 256] - - [261, 67.182] - - - [14896, 256, 1, 256] - - [261, 40.437] - - - [12800, 9216, 1, 256] - - [301, 74.395] - - - [17920, 4608, 1, 256] - - [288, 73.278] - - - [13312, 3329, 1, 256] - - [254, 68.915] - - - [14128, 10240, 1, 256] - - [285, 63.75] - - - [8960, 1536, 1, 256] - - [261, 64.968] - - - [4096, 3328, 1, 256] - - [282, 64.747] - - - [9728, 2048, 1, 256] - - [250, 65.56] - - - [15616, 2304, 1, 256] - - [254, 71.067] - - - [9216, 5888, 1, 256] - - [262, 73.129] - - - [9520, 6400, 1, 256] - - [263, 62.601] - - - [17408, 4352, 1, 256] - - [261, 74.236] - - - [18480, 2865, 1, 256] - - [256, 60.402] - - - [17664, 6144, 1, 256] - - [294, 73.429] - - - [14128, 2865, 1, 256] - - [264, 60.047] - - - [18992, 5888, 1, 256] - - [271, 63.717] - - - [15104, 6144, 1, 256] - - [254, 73.579] - - - [13056, 2865, 1, 256] - - [254, 68.468] - - - [11312, 2816, 1, 256] - - [254, 62.941] - - - [8192, 512, 1, 256] - - [305, 44.765] - - - [10496, 7680, 1, 256] - - [264, 73.763] - - - [15104, 2865, 1, 256] - - [254, 69.417] - - - [3328, 1792, 1, 256] - - [254, 52.483] - - - [12288, 6144, 1, 256] - - [264, 73.86] - - - [3584, 1280, 1, 256] - - [261, 50.073] - - - [18176, 6144, 1, 256] - - [256, 74.114] - - - [3888, 768, 1, 256] - - [267, 45.092] - - - [3072, 3073, 1, 256] - - [282, 57.596] - - - [7168, 2865, 1, 256] - - [254, 66.573] - - - [1840, 1585, 1, 256] - - [287, 43.249] - - - [4912, 1792, 1, 256] - - [276, 56.232] - - - [18432, 2865, 1, 256] - - [252, 69.917] - - - [6448, 2865, 1, 256] - - [262, 57.933] - - - [10496, 3328, 1, 256] - - [287, 70.271] - - - [15872, 2048, 1, 256] - - [285, 68.977] - - - [19200, 10240, 1, 256] - - [264, 74.775] - - - [15360, 2816, 1, 256] - - [256, 71.927] - - - [19760, 6400, 1, 256] - - [271, 64.027] - - - [7680, 4352, 1, 256] - - [282, 71.265] - - - [3072, 2817, 1, 256] - - [310, 54.528] - - - [13568, 1280, 1, 256] - - [252, 66.879] - - - [6144, 3072, 1, 256] - - [252, 67.753] - - - [10240, 1280, 1, 256] - - [276, 62.93] - - - [9776, 6656, 1, 256] - - [271, 63.826] - - - [10544, 7680, 1, 256] - - [280, 63.447] - - - [3328, 1280, 1, 256] - - [261, 46.673] - - - [17456, 4096, 1, 256] - - [285, 64.105] - - - [3584, 2816, 1, 256] - - [261, 62.134] - - - [19200, 2816, 1, 256] - - [282, 71.241] - - - [8192, 6144, 1, 256] - - [254, 72.019] - - - [13312, 256, 1, 256] - - [261, 52.493] - - - [14592, 3329, 1, 256] - - [284, 68.572] - - - [11520, 8192, 1, 256] - - [284, 73.806] - - - [5120, 2304, 1, 256] - - [276, 62.892] - - - [11008, 8192, 1, 256] - - [302, 72.75] - - - [6704, 3840, 1, 256] - - [254, 61.112] - - - [7168, 1792, 1, 256] - - [282, 62.425] - - - [18736, 5632, 1, 256] - - [285, 63.345] - - - [14848, 2865, 1, 256] - - [254, 68.873] - - - [9472, 2865, 1, 256] - - [261, 66.928] - - - [14336, 1024, 1, 256] - - [267, 63.079] - - - [16896, 3840, 1, 256] - - [294, 73.437] - - - [15664, 2304, 1, 256] - - [285, 62.137] - - - [19456, 1536, 1, 256] - - [262, 69.248] - - - [8960, 256, 1, 256] - - [261, 37.859] - - - [5120, 2816, 1, 256] - - [276, 66.912] - - - [11568, 8704, 1, 256] - - [263, 63.361] - - - [10288, 256, 1, 256] - - [261, 41.229] - - - [11776, 3328, 1, 256] - - [284, 71.261] - - - [14384, 2865, 1, 256] - - [254, 60.698] - - - [16688, 256, 1, 256] - - [261, 44.404] - - - [4864, 2048, 1, 256] - - [262, 60.744] - - - [12848, 9984, 1, 256] - - [280, 64.505] - - - [16896, 3584, 1, 256] - - [254, 73.171] - - - [17664, 4096, 1, 256] - - [274, 71.584] - - - [18432, 768, 1, 256] - - [280, 65.406] - - - [11008, 256, 1, 256] - - [306, 44.969] - - - [5888, 2048, 1, 256] - - [267, 63.123] - - - [11264, 7680, 1, 256] - - [256, 74.867] - - - [17200, 2865, 1, 256] - - [254, 61.583] - - - [7472, 4608, 1, 256] - - [280, 61.625] - - - [13360, 2816, 1, 256] - - [285, 62.643] - - - [15408, 256, 1, 256] - - [267, 41.789] - - - [15360, 256, 1, 256] - - [282, 43.165] - - - [18944, 2865, 1, 256] - - [264, 69.777] - - - [8192, 3328, 1, 256] - - [254, 68.008] - - - [17664, 2865, 1, 256] - - [286, 69.269] - - - [18944, 1281, 1, 256] - - [299, 63.675] - - - [8496, 5632, 1, 256] - - [261, 64.342] - - - [14848, 2048, 1, 256] - - [267, 69.541] - - - [4096, 2048, 1, 256] - - [286, 58.762] - - - [17152, 3840, 1, 256] - - [252, 72.897] - - - [16896, 256, 1, 256] - - [261, 46.769] - - - [5888, 2560, 1, 256] - - [276, 64.841] - - - [1328, 1280, 1, 256] - - [261, 36.117] - - - [12544, 1280, 1, 256] - - [252, 66.492] - - - [17920, 1280, 1, 256] - - [284, 68.79] - - - [7936, 2048, 1, 256] - - [271, 63.392] - - - [16896, 1024, 1, 256] - - [267, 66.25] - - - [17456, 2865, 1, 256] - - [256, 62.344] - - - [14080, 1024, 1, 256] - - [260, 65.233] - - - [5168, 2816, 1, 256] - - [252, 56.571] - - - [1584, 1329, 1, 256] - - [254, 33.18] - - - [17712, 2816, 1, 256] - - [280, 63.326] - - - [15872, 2816, 1, 256] - - [262, 72.252] - - - [5632, 4352, 1, 256] - - [262, 69.961] - - - [6656, 2816, 1, 256] - - [282, 66.455] - - - [13056, 10240, 1, 256] - - [256, 74.986] - - - [5376, 4096, 1, 256] - - [254, 67.138] - - - [13568, 2816, 1, 256] - - [261, 70.507] - - - [6960, 2816, 1, 256] - - [256, 61.413] - - - [15104, 3328, 1, 256] - - [250, 71.526] - - - [15872, 2304, 1, 256] - - [262, 71.22] - - - [16128, 6144, 1, 256] - - [254, 73.909] - - - [5424, 2865, 1, 256] - - [271, 60.264] - - - [12848, 256, 1, 256] - - [267, 48.34] - - - [13872, 512, 1, 256] - - [252, 55.706] - - - [2816, 1281, 1, 256] - - [262, 41.044] - - - [12800, 3328, 1, 256] - - [284, 71.351] - - - [17152, 3584, 1, 256] - - [282, 72.518] - - - [18176, 10240, 1, 256] - - [252, 74.987] - - - [18176, 1280, 1, 256] - - [261, 68.702] - - - [9216, 3328, 1, 256] - - [252, 69.932] - - - [18688, 2816, 1, 256] - - [262, 71.742] - - - [10800, 7936, 1, 256] - - [271, 64.944] - - - [16640, 6144, 1, 256] - - [254, 73.468] - - - [11056, 8192, 1, 256] - - [256, 64.782] - - - [19712, 1280, 1, 256] - - [252, 67.903] - - - [19456, 2865, 1, 256] - - [254, 70.001] - - - [7216, 4096, 1, 256] - - [256, 61.81] - - - [10496, 3072, 1, 256] - - [261, 70.522] - - - [9984, 6912, 1, 256] - - [261, 73.374] - - - [9984, 2304, 1, 256] - - [262, 68.987] - - - [19712, 2048, 1, 256] - - [311, 67.452] - - - [19504, 6400, 1, 256] - - [271, 64.123] - - - [15104, 1536, 1, 256] - - [262, 67.917] - - - [14384, 256, 1, 256] - - [267, 39.048] - - - [9776, 6912, 1, 256] - - [271, 64.424] - - - [9472, 2048, 1, 256] - - [267, 67.028] - - - [9984, 2560, 1, 256] - - [262, 68.769] - - - [16640, 2865, 1, 256] - - [261, 68.013] - - - [5632, 3328, 1, 256] - - [262, 67.157] - - - [3328, 3329, 1, 256] - - [261, 58.687] - - - [12288, 3329, 1, 256] - - [254, 68.787] - - - [6656, 1280, 1, 256] - - [276, 59.572] - - - [6912, 1281, 1, 256] - - [276, 55.428] - - - [14592, 3328, 1, 256] - - [284, 70.537] - - - [9728, 6400, 1, 256] - - [261, 73.801] - - - [16432, 10240, 1, 256] - - [281, 68.845] - - - [12336, 2816, 1, 256] - - [254, 61.954] - - - [14384, 2816, 1, 256] - - [256, 62.004] - - - [16640, 2816, 1, 256] - - [261, 70.493] - - - [2304, 2304, 1, 256] - - [262, 56.085] - - - [13104, 10240, 1, 256] - - [263, 64.67] - - - [13104, 9984, 1, 256] - - [271, 64.266] - - - [3888, 2865, 1, 256] - - [252, 53.93] - - - [15920, 256, 1, 256] - - [250, 43.417] - - - [16128, 2304, 1, 256] - - [262, 70.651] - - - [6144, 4864, 1, 256] - - [264, 71.369] - - - [4352, 1024, 1, 256] - - [261, 48.505] - - - [19504, 10240, 1, 256] - - [255, 63.61] - - - [1536, 1281, 1, 256] - - [258, 32.663] - - - [3584, 3329, 1, 256] - - [261, 62.683] - - - [4400, 2865, 1, 256] - - [254, 54.738] - - - [16640, 2048, 1, 256] - - [267, 68.188] - - - [14640, 2816, 1, 256] - - [271, 62.79] - - - [8192, 1280, 1, 256] - - [262, 62.359] - - - [4144, 2816, 1, 256] - - [252, 57.389] - - - [14592, 256, 1, 256] - - [267, 41.321] - - - [17920, 2865, 1, 256] - - [274, 69.514] - - - [14592, 1536, 1, 256] - - [262, 66.733] - - - [6704, 3584, 1, 256] - - [256, 61.656] - - - [4400, 1536, 1, 256] - - [262, 53.849] - - - [19712, 5888, 1, 256] - - [268, 72.495] - - - [8240, 2865, 1, 256] - - [262, 61.677] - - - [8496, 2816, 1, 256] - - [264, 61.931] - - - [14592, 2865, 1, 256] - - [258, 68.747] - - - [8448, 256, 1, 256] - - [286, 36.005] - - - [15152, 2865, 1, 256] - - [264, 61.182] - - - [15616, 1280, 1, 256] - - [276, 66.097] - - - [17408, 256, 1, 256] - - [267, 48.142] - - - [14592, 1280, 1, 256] - - [262, 65.738] - - - [13616, 2865, 1, 256] - - [263, 60.21] - - - [10240, 1281, 1, 256] - - [276, 60.343] - - - [12800, 9472, 1, 256] - - [256, 75.03] - - - [19456, 2048, 1, 256] - - [263, 69.833] - - - [11056, 256, 1, 256] - - [254, 43.423] - - - [14848, 3328, 1, 256] - - [256, 71.856] - - - [10752, 1281, 1, 256] - - [267, 59.249] - - - [14080, 3329, 1, 256] - - [284, 68.263] - - - [13312, 10240, 1, 256] - - [254, 75.296] - - - [17408, 3840, 1, 256] - - [264, 73.865] - - - [10800, 2865, 1, 256] - - [280, 61.201] - - - [14128, 2816, 1, 256] - - [285, 61.885] - - - [2560, 2305, 1, 256] - - [262, 51.575] - - - [16128, 2048, 1, 256] - - [267, 68.171] - - - [18688, 4864, 1, 256] - - [262, 73.833] - - - [8752, 2865, 1, 256] - - [256, 58.991] - - - [16640, 3328, 1, 256] - - [275, 70.911] - - - [7680, 3328, 1, 256] - - [262, 68.466] - - - [14080, 6144, 1, 256] - - [294, 72.26] - - - [7424, 1280, 1, 256] - - [261, 59.134] - - - [10240, 2865, 1, 256] - - [254, 68.642] - - - [7680, 256, 1, 256] - - [274, 32.732] - - - [12032, 6144, 1, 256] - - [256, 72.815] - - - [10752, 256, 1, 256] - - [254, 44.79] - - - [9008, 6144, 1, 256] - - [285, 63.58] - - - [5632, 1280, 1, 256] - - [262, 61.021] - - - [6400, 768, 1, 256] - - [290, 52.653] - - - [18432, 5376, 1, 256] - - [254, 74.701] - - - [5680, 2816, 1, 256] - - [252, 60.823] - - - [4608, 2865, 1, 256] - - [262, 62.836] - - - [1280, 1281, 1, 256] - - [290, 36.52] - - - [15152, 1792, 1, 256] - - [264, 60.278] - - - [12848, 2865, 1, 256] - - [264, 59.68] - - - [8752, 256, 1, 256] - - [261, 36.102] - - - [2304, 2049, 1, 256] - - [262, 50.153] - - - [10752, 6144, 1, 256] - - [254, 73.602] - - - [17968, 2865, 1, 256] - - [252, 61.071] - - - [15408, 2048, 1, 256] - - [280, 62.18] - - - [9264, 2865, 1, 256] - - [252, 59.195] - - - [16896, 10240, 1, 256] - - [252, 75.376] - - - [13872, 2865, 1, 256] - - [252, 60.894] - - - [12288, 3328, 1, 256] - - [254, 71.562] - - - [9216, 1280, 1, 256] - - [261, 62.921] - - - [4608, 1281, 1, 256] - - [262, 51.593] - - - [5888, 2816, 1, 256] - - [262, 65.221] - - - [18688, 5120, 1, 256] - - [252, 73.865] - - - [4864, 2865, 1, 256] - - [261, 65.382] - - - [13872, 768, 1, 256] - - [263, 58.587] - - - [16384, 512, 1, 256] - - [261, 54.525] - - - [17920, 2048, 1, 256] - - [266, 69.882] - - - [1792, 1792, 1, 256] - - [254, 49.797] - - - [19760, 256, 1, 256] - - [285, 49.25] - - - [1792, 1793, 1, 256] - - [252, 49.099] - - - [2560, 1280, 1, 256] - - [261, 51.366] - - - [8960, 1280, 1, 256] - - [276, 61.258] - - - [6400, 2048, 1, 256] - - [282, 62.695] - - - [18944, 2048, 1, 256] - - [267, 70.368] - - - [4352, 1536, 1, 256] - - [262, 57.987] - - - [6400, 2816, 1, 256] - - [256, 68.737] - - - [13312, 6144, 1, 256] - - [252, 74.102] - - - [16128, 256, 1, 256] - - [296, 44.519] - - - [8704, 2865, 1, 256] - - [256, 66.521] - - - [10496, 1280, 1, 256] - - [262, 63.682] - - - [8704, 1281, 1, 256] - - [250, 58.499] - - - [5120, 1281, 1, 256] - - [282, 55.721] - - - [8448, 6144, 1, 256] - - [261, 72.701] - - - [4352, 1280, 1, 256] - - [261, 49.744] - - - [21504, 10240, 1, 256] - - [256, 75.583] - - - [27648, 1280, 1, 256] - - [261, 70.518] - - - [29184, 256, 1, 256] - - [261, 52.208] - - - [30720, 7168, 1, 256] - - [264, 73.869] - - - [25600, 5632, 1, 256] - - [254, 75.075] - - - [22784, 2865, 1, 256] - - [256, 69.758] - - - [21760, 8192, 1, 256] - - [264, 74.528] - - - [26368, 6144, 1, 256] - - [254, 74.073] - - - [29952, 4096, 1, 256] - - [256, 72.803] - - - [32256, 2048, 1, 256] - - [292, 71.243] - - - [21808, 2816, 1, 256] - - [285, 63.226] - - - [23040, 10240, 1, 256] - - [252, 75.226] - - - [24320, 1281, 1, 256] - - [296, 64.001] - - - [31488, 6144, 1, 256] - - [254, 74.077] - - - [33024, 8960, 1, 256] - - [257, 74.681] - - - [31488, 5632, 1, 256] - - [254, 74.077] - - - [31488, 10240, 1, 256] - - [264, 74.538] - - - [26160, 10240, 1, 256] - - [263, 62.983] - - - [28928, 4864, 1, 256] - - [256, 73.672] - - - [29184, 5376, 1, 256] - - [257, 74.353] - - - [24832, 6144, 1, 256] - - [251, 74.183] - - - [24320, 1280, 1, 256] - - [284, 69.334] - - - [34816, 1281, 1, 256] - - [254, 63.991] - - - [32256, 8192, 1, 256] - - [257, 74.855] - - - [23808, 3840, 1, 256] - - [254, 73.539] - - - [30464, 10240, 1, 256] - - [251, 73.683] - - - [27136, 1280, 1, 256] - - [284, 70.23] - - - [33024, 2816, 1, 256] - - [269, 72.269] - - - [34048, 1792, 1, 256] - - [258, 71.4] - - - [29952, 256, 1, 256] - - [250, 53.356] - - - [26368, 3072, 1, 256] - - [264, 72.767] - - - [30720, 512, 1, 256] - - [262, 64.809] - - - [28928, 1024, 1, 256] - - [290, 68.461] - - - [26928, 10240, 1, 256] - - [280, 63.381] - - - [34304, 2816, 1, 256] - - [286, 73.087] - - - [27696, 10240, 1, 256] - - [254, 63.529] - - - [33536, 6144, 1, 256] - - [257, 74.074] - - - [21552, 2816, 1, 256] - - [280, 63.568] - - - [22272, 2048, 1, 256] - - [260, 70.036] - - - [30720, 10240, 1, 256] - - [252, 75.334] - - - [20528, 10240, 1, 256] - - [255, 62.326] - - - [31536, 2865, 1, 256] - - [280, 61.567] - - - [20784, 2865, 1, 256] - - [252, 60.364] - - - [25600, 3329, 1, 256] - - [259, 70.116] - - - [29488, 256, 1, 256] - - [285, 48.225] - - - [31280, 2865, 1, 256] - - [280, 61.474] - - - [34096, 2816, 1, 256] - - [271, 63.755] - - - [32768, 1281, 1, 256] - - [322, 51.657] - - - [30464, 3328, 1, 256] - - [286, 71.563] - - - [23552, 1280, 1, 256] - - [276, 70.066] - - - [34560, 2865, 1, 256] - - [264, 70.251] - - - [26880, 3328, 1, 256] - - [252, 72.345] - - - [31744, 8448, 1, 256] - - [252, 75.166] - - - [22528, 2048, 1, 256] - - [288, 69.864] - - - [32512, 8960, 1, 256] - - [274, 74.928] - - - [24880, 10240, 1, 256] - - [256, 63.243] - - - [26368, 10240, 1, 256] - - [252, 74.681] - - - [33536, 1281, 1, 256] - - [301, 63.723] - - - [33792, 1281, 1, 256] - - [268, 64.358] - - - [20528, 7424, 1, 256] - - [254, 62.655] - - - [22016, 2048, 1, 256] - - [275, 70.212] - - - [30208, 2048, 1, 256] - - [271, 71.22] - - - [25904, 256, 1, 256] - - [261, 51.108] - - - [32816, 256, 1, 256] - - [285, 53.638] - - - [22272, 1281, 1, 256] - - [296, 63.01] - - - [33280, 3329, 1, 256] - - [251, 70.393] - - - [25392, 2816, 1, 256] - - [280, 63.703] - - - [21760, 1280, 1, 256] - - [262, 68.663] - - - [25088, 1792, 1, 256] - - [282, 71.376] - - - [34560, 3328, 1, 256] - - [288, 72.635] - - - [30464, 2048, 1, 256] - - [316, 68.048] - - - [30464, 6912, 1, 256] - - [257, 73.595] - - - [33792, 9728, 1, 256] - - [252, 75.271] - - - [24112, 512, 1, 256] - - [256, 57.899] - - - [34816, 2048, 1, 256] - - [255, 70.698] - - - [20992, 10240, 1, 256] - - [252, 75.383] - - - [25904, 2304, 1, 256] - - [280, 62.878] - - - [30976, 768, 1, 256] - - [266, 66.489] - - - [33072, 256, 1, 256] - - [252, 52.096] - - - [28208, 2865, 1, 256] - - [285, 61.173] - - - [26880, 2865, 1, 256] - - [281, 69.839] - - - [29184, 5120, 1, 256] - - [264, 74.128] - - - [30464, 3329, 1, 256] - - [297, 68.208] - - - [25344, 2048, 1, 256] - - [266, 69.446] - - - [21248, 2816, 1, 256] - - [262, 71.886] - - - [27136, 3840, 1, 256] - - [286, 74.215] - - - [24576, 4608, 1, 256] - - [249, 69.32] - - - [26928, 2865, 1, 256] - - [280, 60.727] - - - [24112, 2816, 1, 256] - - [285, 64.168] - - - [30208, 1281, 1, 256] - - [296, 64.377] - - - [20992, 1280, 1, 256] - - [261, 69.476] - - - [21296, 8192, 1, 256] - - [280, 63.224] - - - [27136, 2865, 1, 256] - - [252, 70.677] - - - [29696, 3329, 1, 256] - - [264, 70.496] - - - [30000, 256, 1, 256] - - [280, 48.765] - - - [29232, 2816, 1, 256] - - [271, 63.791] - - - [29488, 2816, 1, 256] - - [285, 63.082] - - - [30512, 2816, 1, 256] - - [280, 63.905] - - - [20480, 256, 1, 256] - - [250, 53.911] - - - [26624, 2816, 1, 256] - - [264, 73.242] - - - [27648, 256, 1, 256] - - [305, 58.412] - - - [30720, 6144, 1, 256] - - [264, 74.708] - - - [23088, 9728, 1, 256] - - [271, 63.913] - - - [33024, 6144, 1, 256] - - [270, 73.922] - - - [21504, 1280, 1, 256] - - [282, 68.976] - - - [24880, 2865, 1, 256] - - [263, 62.105] - - - [20224, 2816, 1, 256] - - [282, 71.778] - - - [25856, 3329, 1, 256] - - [252, 69.116] - - - [26880, 6144, 1, 256] - - [264, 74.125] - - - [25600, 256, 1, 256] - - [307, 55.352] - - - [20480, 10240, 1, 256] - - [254, 75.272] - - - [20272, 7168, 1, 256] - - [271, 62.663] - - - [23296, 9728, 1, 256] - - [256, 74.626] - - - [31792, 2865, 1, 256] - - [256, 62.045] - - - [23856, 2865, 1, 256] - - [263, 60.587] - - - [22016, 8960, 1, 256] - - [258, 75.62] - - - [23856, 512, 1, 256] - - [252, 57.075] - - - [22784, 9216, 1, 256] - - [255, 74.474] - - - [32000, 8192, 1, 256] - - [254, 74.349] - - - [28672, 256, 1, 256] - - [271, 51.858] - - - [23808, 6144, 1, 256] - - [264, 74.136] - - - [20736, 7680, 1, 256] - - [252, 74.828] - - - [26160, 2816, 1, 256] - - [285, 63.826] - - - [24320, 768, 1, 256] - - [299, 65.439] - - - [34560, 1024, 1, 256] - - [267, 69.938] - - - [30720, 2816, 1, 256] - - [256, 73.516] - - - [20736, 3329, 1, 256] - - [256, 68.948] - - - [21760, 8704, 1, 256] - - [254, 75.019] - - - [22784, 9472, 1, 256] - - [294, 74.935] - - - [27904, 4352, 1, 256] - - [286, 73.693] - - - [24064, 256, 1, 256] - - [267, 52.765] - - - [30464, 1280, 1, 256] - - [294, 68.821] - - - [21808, 8704, 1, 256] - - [263, 63.111] - - - [20736, 1024, 1, 256] - - [287, 68.087] - - - [24576, 1024, 1, 256] - - [254, 64.881] - - - [30976, 4864, 1, 256] - - [294, 72.422] - - - [22016, 1280, 1, 256] - - [261, 69.22] - - - [31536, 7936, 1, 256] - - [280, 64.014] - - - [28464, 2865, 1, 256] - - [280, 61.458] - - - [26624, 3328, 1, 256] - - [282, 73.411] - - - [34352, 768, 1, 256] - - [263, 61.293] - - - [34560, 10240, 1, 256] - - [264, 74.662] - - - [20480, 1281, 1, 256] - - [254, 62.783] - - - [31232, 6144, 1, 256] - - [257, 74.376] - - - [23344, 256, 1, 256] - - [254, 47.781] - - - [32000, 6144, 1, 256] - - [256, 74.002] - - - [29232, 5632, 1, 256] - - [285, 63.1] - - - [31488, 1280, 1, 256] - - [261, 70.391] - - - [29184, 2865, 1, 256] - - [284, 70.156] - - - [30720, 4608, 1, 256] - - [264, 74.138] - - - [32512, 10240, 1, 256] - - [257, 74.736] - - - [22528, 3328, 1, 256] - - [256, 73.174] - - - [31744, 2816, 1, 256] - - [261, 73.414] - - - [21552, 256, 1, 256] - - [263, 45.306] - - - [22784, 3328, 1, 256] - - [269, 72.078] - - - [30208, 2816, 1, 256] - - [318, 72.979] - - - [30256, 256, 1, 256] - - [252, 48.965] - - - [28160, 1280, 1, 256] - - [258, 70.64] - - - [34560, 3329, 1, 256] - - [259, 69.597] - - - [32000, 256, 1, 256] - - [285, 56.124] - - - [24624, 10240, 1, 256] - - [268, 63.948] - - - [33536, 9728, 1, 256] - - [251, 74.661] - - - [20480, 2865, 1, 256] - - [256, 70.005] - - - [21760, 6144, 1, 256] - - [264, 74.339] - - - [27648, 2816, 1, 256] - - [252, 73.351] - - - [26672, 3072, 1, 256] - - [252, 62.488] - - - [23552, 9728, 1, 256] - - [252, 75.537] - - - [25088, 1280, 1, 256] - - [252, 70.249] - - - [27136, 3328, 1, 256] - - [258, 73.302] - - - [20480, 3328, 1, 256] - - [262, 72.255] - - - [33328, 9728, 1, 256] - - [285, 63.695] - - - [31744, 2865, 1, 256] - - [252, 70.877] - - - [32000, 5888, 1, 256] - - [254, 73.811] - - - [34352, 2865, 1, 256] - - [271, 61.36] - - - [26112, 2048, 1, 256] - - [283, 70.991] - - - [29952, 3329, 1, 256] - - [264, 69.674] - - - [31744, 1280, 1, 256] - - [256, 71.007] - - - [31488, 2048, 1, 256] - - [288, 70.757] - - - [23040, 256, 1, 256] - - [311, 50.859] - - - [31744, 3329, 1, 256] - - [281, 70.267] - - - [29184, 1281, 1, 256] - - [301, 64.472] - - - [23808, 1792, 1, 256] - - [276, 71.018] - - - [26880, 2816, 1, 256] - - [262, 72.335] - - - [27136, 6144, 1, 256] - - [254, 74.561] - - - [33536, 1280, 1, 256] - - [261, 70.538] - - - [25088, 2865, 1, 256] - - [251, 70.235] - - - [26672, 10240, 1, 256] - - [268, 62.731] - - - [34352, 2816, 1, 256] - - [285, 63.786] - - - [28208, 4864, 1, 256] - - [271, 63.705] - - - [26624, 3072, 1, 256] - - [264, 73.536] - - - [22576, 256, 1, 256] - - [256, 46.957] - - - [29440, 5888, 1, 256] - - [258, 73.964] - - - [32000, 10240, 1, 256] - - [256, 74.514] - - - [30976, 2865, 1, 256] - - [284, 68.745] - - - [29232, 2865, 1, 256] - - [280, 61.737] - - - [30512, 2865, 1, 256] - - [280, 61.791] - - - [28416, 4352, 1, 256] - - [294, 73.602] - - - [25344, 1280, 1, 256] - - [284, 70.141] - - - [33024, 1281, 1, 256] - - [318, 63.336] - - - [28416, 4864, 1, 256] - - [258, 73.651] - - - [30464, 4352, 1, 256] - - [257, 72.6] - - - [25136, 256, 1, 256] - - [254, 49.859] - - - [27440, 2816, 1, 256] - - [271, 64.396] - - - [24320, 2816, 1, 256] - - [286, 72.422] - - - [28416, 5120, 1, 256] - - [258, 73.518] - - - [20736, 1281, 1, 256] - - [267, 62.761] - - - [30512, 10240, 1, 256] - - [268, 62.92] - - - [22784, 10240, 1, 256] - - [286, 74.611] - - - [25856, 10240, 1, 256] - - [264, 74.853] - - - [20736, 2865, 1, 256] - - [264, 69.727] - - - [31024, 10240, 1, 256] - - [288, 62.465] - - - [30000, 2865, 1, 256] - - [280, 61.012] - - - [24832, 768, 1, 256] - - [266, 65.797] - - - [24624, 1280, 1, 256] - - [249, 61.013] - - - [29952, 3328, 1, 256] - - [284, 72.644] - - - [30256, 6656, 1, 256] - - [263, 63.547] - - - [33792, 1536, 1, 256] - - [256, 71.307] - - - [25600, 5888, 1, 256] - - [254, 74.704] - - - [32256, 2816, 1, 256] - - [294, 73.279] - - - [25392, 10240, 1, 256] - - [285, 63.224] - - - [22528, 3329, 1, 256] - - [254, 70.282] - - - [33024, 2865, 1, 256] - - [281, 69.197] - - - [27392, 10240, 1, 256] - - [268, 74.262] - - - [27648, 6144, 1, 256] - - [254, 74.648] - - - [34560, 1281, 1, 256] - - [305, 64.272] - - - [32000, 1281, 1, 256] - - [264, 63.609] - - - [24880, 1536, 1, 256] - - [256, 61.665] - - - [34816, 2816, 1, 256] - - [249, 73.569] - - - [30464, 7168, 1, 256] - - [251, 72.343] - - - [32304, 10240, 1, 256] - - [264, 63.119] - - - [29696, 10240, 1, 256] - - [254, 75.448] - - - [27392, 1281, 1, 256] - - [303, 62.129] - - - [26160, 2560, 1, 256] - - [263, 63.321] - - - [23808, 256, 1, 256] - - [286, 52.087] - - - [33280, 2048, 1, 256] - - [270, 71.202] - - - [23040, 9984, 1, 256] - - [254, 75.6] - - - [33328, 9984, 1, 256] - - [263, 64.001] - - - [22832, 9728, 1, 256] - - [285, 63.859] - - - [33280, 1281, 1, 256] - - [274, 64.269] - - - [32048, 256, 1, 256] - - [254, 52.152] - - - [26880, 3329, 1, 256] - - [259, 69.209] - - - [24832, 3329, 1, 256] - - [257, 69.453] - - - [34560, 2816, 1, 256] - - [254, 72.636] - - - [23600, 256, 1, 256] - - [256, 48.338] - - - [25344, 5632, 1, 256] - - [292, 73.279] - - - [31024, 256, 1, 256] - - [262, 50.513] - - - [29184, 3072, 1, 256] - - [252, 72.996] - - - [20480, 6656, 1, 256] - - [254, 74.673] - - - [20736, 1280, 1, 256] - - [282, 69.078] - - - [29488, 10240, 1, 256] - - [264, 63.086] - - - [29440, 3329, 1, 256] - - [274, 69.625] - - - [34560, 768, 1, 256] - - [287, 68.277] - - - [29184, 5632, 1, 256] - - [254, 74.567] - - - [26112, 2865, 1, 256] - - [264, 70.579] - - - [20992, 1024, 1, 256] - - [267, 68.585] - - - [29952, 6656, 1, 256] - - [278, 74.395] - - - [34352, 256, 1, 256] - - [263, 53.882] - - - [32816, 2865, 1, 256] - - [255, 62.415] - - - [33792, 1280, 1, 256] - - [262, 71.087] - - - [34096, 256, 1, 256] - - [263, 53.881] - - - [30464, 1281, 1, 256] - - [286, 62.873] - - - [25136, 2865, 1, 256] - - [263, 61.245] - - - [24320, 4608, 1, 256] - - [284, 73.571] - - - [31744, 2048, 1, 256] - - [288, 71.217] - - - [34048, 256, 1, 256] - - [261, 58.331] - - - [22272, 10240, 1, 256] - - [274, 74.77] - - - [29744, 10240, 1, 256] - - [268, 63.34] - - - [23552, 2048, 1, 256] - - [268, 70.837] - - - [27440, 3840, 1, 256] - - [285, 64.087] - - - [21808, 256, 1, 256] - - [285, 45.844] - - - [29744, 256, 1, 256] - - [271, 48.643] - - - [33536, 3329, 1, 256] - - [252, 69.443] - - - [22272, 9216, 1, 256] - - [268, 74.35] - - - [29440, 1280, 1, 256] - - [284, 70.202] - - - [26624, 1281, 1, 256] - - [254, 63.119] - - - [24064, 3328, 1, 256] - - [284, 72.942] - - - [33280, 9728, 1, 256] - - [269, 75.157] - - - [30464, 6144, 1, 256] - - [278, 72.887] - - - [28928, 256, 1, 256] - - [252, 51.625] - - - [26672, 2865, 1, 256] - - [256, 61.236] - - - [25088, 2816, 1, 256] - - [261, 72.671] - - - [29488, 5888, 1, 256] - - [285, 63.407] - - - [21808, 10240, 1, 256] - - [280, 63.114] - - - [23808, 1280, 1, 256] - - [250, 69.635] - - - [25856, 256, 1, 256] - - [317, 55.701] - - - [29952, 2865, 1, 256] - - [252, 70.007] - - - [34304, 3328, 1, 256] - - [269, 73.509] - - - [24576, 2816, 1, 256] - - [265, 68.802] - - - [23296, 9472, 1, 256] - - [254, 75.084] - - - [20736, 2048, 1, 256] - - [290, 70.538] - - - [20480, 512, 1, 256] - - [287, 62.552] - - - [24064, 2865, 1, 256] - - [274, 70.099] - - - [20992, 256, 1, 256] - - [261, 55.358] - - - [28928, 5120, 1, 256] - - [252, 73.722] - - - [34304, 1281, 1, 256] - - [288, 64.906] - - - [33584, 2816, 1, 256] - - [280, 63.774] - - - [20224, 6912, 1, 256] - - [252, 74.537] - - - [33536, 3328, 1, 256] - - [268, 72.67] - - - [31024, 7680, 1, 256] - - [285, 63.222] - - - [25344, 1281, 1, 256] - - [266, 63.325] - - - [27952, 10240, 1, 256] - - [271, 63.351] - - - [32768, 2816, 1, 256] - - [265, 59.065] - - - [27904, 2865, 1, 256] - - [258, 69.642] - - - [22016, 1281, 1, 256] - - [299, 63.18] - - - [22528, 9216, 1, 256] - - [254, 74.957] - - - [23808, 10240, 1, 256] - - [254, 75.037] - - - [34816, 1024, 1, 256] - - [250, 70.048] - - - [30768, 2816, 1, 256] - - [252, 61.946] - - - [25600, 3328, 1, 256] - - [261, 73.271] - - - [22064, 8704, 1, 256] - - [263, 63.169] - - - [22784, 768, 1, 256] - - [267, 66.495] - - - [22528, 9472, 1, 256] - - [254, 75.892] - - - [28672, 5376, 1, 256] - - [254, 74.327] - - - [24576, 256, 1, 256] - - [252, 53.728] - - - [24576, 3328, 1, 256] - - [249, 68.462] - - - [27440, 256, 1, 256] - - [256, 53.249] - - - [30000, 2816, 1, 256] - - [280, 63.72] - - - [22064, 2816, 1, 256] - - [285, 63.866] - - - [28416, 256, 1, 256] - - [287, 51.395] - - - [32768, 512, 1, 256] - - [256, 57.901] - - - [26928, 256, 1, 256] - - [252, 52.393] - - - [26368, 512, 1, 256] - - [282, 62.805] - - - [33328, 2865, 1, 256] - - [271, 62.95] - - - [25856, 2865, 1, 256] - - [256, 69.722] - - - [28416, 2560, 1, 256] - - [284, 72.27] - - - [20480, 1280, 1, 256] - - [282, 68.503] - - - [34864, 1024, 1, 256] - - [252, 61.22] - - - [23296, 3328, 1, 256] - - [264, 72.158] - - - [32560, 2816, 1, 256] - - [271, 64.998] - - - [34608, 768, 1, 256] - - [280, 60.915] - - - [25856, 2048, 1, 256] - - [305, 70.029] - - - [30000, 6656, 1, 256] - - [271, 63.377] - - - [33280, 9216, 1, 256] - - [269, 74.887] - - - [28416, 1281, 1, 256] - - [299, 62.992] - - - [26368, 2048, 1, 256] - - [301, 70.201] - - - [32768, 2048, 1, 256] - - [249, 56.064] - - - [28672, 512, 1, 256] - - [261, 62.915] - - - [34560, 6144, 1, 256] - - [256, 74.161] - - - [29952, 2048, 1, 256] - - [250, 70.524] - - - [25600, 1792, 1, 256] - - [261, 71.951] - - - [29952, 1280, 1, 256] - - [282, 70.288] - - - [26416, 3072, 1, 256] - - [285, 62.558] - - - [34560, 2560, 1, 256] - - [256, 72.91] - - - [30464, 2865, 1, 256] - - [300, 68.347] - - - [23808, 2865, 1, 256] - - [252, 69.84] - - - [20736, 6912, 1, 256] - - [252, 74.702] - - - [33792, 256, 1, 256] - - [280, 58.332] - - - [30464, 256, 1, 256] - - [261, 54.235] - - - [20224, 512, 1, 256] - - [250, 62.059] - - - [20272, 2816, 1, 256] - - [263, 63.653] - - - [27904, 2816, 1, 256] - - [284, 72.231] - - - [24112, 10240, 1, 256] - - [263, 63.496] - - - [24064, 6144, 1, 256] - - [257, 74.387] - - - [30768, 2865, 1, 256] - - [256, 61.038] - - - [25392, 2865, 1, 256] - - [256, 61.179] - - - [25088, 10240, 1, 256] - - [257, 75.251] - - - [23552, 1792, 1, 256] - - [282, 71.439] - - - [28160, 2816, 1, 256] - - [294, 72.65] - - - [22832, 9472, 1, 256] - - [271, 64.024] - - - [26928, 3328, 1, 256] - - [280, 63.435] - - - [33024, 1280, 1, 256] - - [261, 69.175] - - - [26368, 2865, 1, 256] - - [254, 70.011] - - - [21504, 8192, 1, 256] - - [254, 75.141] - - - [31232, 2865, 1, 256] - - [254, 70.381] - - - [24368, 256, 1, 256] - - [254, 49.499] - - - [21248, 2865, 1, 256] - - [254, 69.655] - - - [24320, 3329, 1, 256] - - [278, 69.541] - - - [27648, 2865, 1, 256] - - [256, 70.739] - - - [28672, 2560, 1, 256] - - [264, 72.702] - - - [30256, 2865, 1, 256] - - [263, 61.639] - - - [30464, 4608, 1, 256] - - [274, 72.193] - - - [29440, 2816, 1, 256] - - [258, 72.647] - - - [20992, 2048, 1, 256] - - [285, 70.296] - - - [20224, 3328, 1, 256] - - [261, 71.953] - - - [25344, 6144, 1, 256] - - [255, 72.892] - - - [21248, 3329, 1, 256] - - [254, 69.165] - - - [20992, 7936, 1, 256] - - [252, 75.164] - - - [32000, 3329, 1, 256] - - [281, 69.733] - - - [24624, 1024, 1, 256] - - [255, 59.008] - - - [29440, 1281, 1, 256] - - [266, 64.108] - - - [25136, 10240, 1, 256] - - [280, 63.177] - - - [33840, 2816, 1, 256] - - [290, 63.323] - - - [34048, 6144, 1, 256] - - [278, 73.573] - - - [32048, 2865, 1, 256] - - [285, 61.728] - - - [25088, 1536, 1, 256] - - [261, 69.981] - - - [20480, 6912, 1, 256] - - [252, 74.992] - - - [22064, 256, 1, 256] - - [254, 46.283] - - - [26880, 1024, 1, 256] - - [250, 68.001] - - - [30208, 10240, 1, 256] - - [274, 75.228] - - - [26880, 10240, 1, 256] - - [264, 74.707] - - - [33536, 2048, 1, 256] - - [311, 70.838] - - - [31536, 10240, 1, 256] - - [271, 62.995] - - - [32768, 768, 1, 256] - - [249, 56.214] - - - [26112, 2560, 1, 256] - - [264, 73.095] - - - [22528, 10240, 1, 256] - - [252, 75.441] - - - [22320, 256, 1, 256] - - [271, 46.987] - - - [28464, 256, 1, 256] - - [282, 46.705] - - - [29744, 6400, 1, 256] - - [280, 63.787] - - - [22832, 256, 1, 256] - - [280, 47.124] - - - [25856, 2560, 1, 256] - - [264, 72.187] - - - [31280, 2816, 1, 256] - - [263, 63.623] - - - [26368, 1281, 1, 256] - - [287, 63.562] - - - [33536, 1536, 1, 256] - - [262, 70.649] - - - [26624, 2048, 1, 256] - - [288, 70.108] - - - [21760, 2048, 1, 256] - - [290, 70.422] - - - [29696, 1280, 1, 256] - - [261, 70.63] - - - [24832, 2865, 1, 256] - - [252, 69.873] - - - [34304, 2865, 1, 256] - - [278, 70.546] - - - [25136, 1536, 1, 256] - - [254, 61.602] - - - [24880, 1280, 1, 256] - - [280, 63.141] - - - [29184, 5888, 1, 256] - - [274, 74.176] - - - [34352, 10240, 1, 256] - - [255, 62.689] - - - [25344, 10240, 1, 256] - - [255, 73.908] - - - [26624, 512, 1, 256] - - [261, 63.853] - - - [23040, 3328, 1, 256] - - [305, 72.72] - - - [25856, 2304, 1, 256] - - [252, 71.728] - - - [33280, 2865, 1, 256] - - [278, 70.866] - - - [21504, 6144, 1, 256] - - [254, 74.607] - - - [30464, 2816, 1, 256] - - [258, 71.467] - - - [24576, 2048, 1, 256] - - [254, 64.874] - - - [32256, 1280, 1, 256] - - [294, 71.001] - - - [30976, 5120, 1, 256] - - [268, 72.377] - - - [24112, 2865, 1, 256] - - [285, 61.649] - - - [21760, 2816, 1, 256] - - [262, 72.102] - - - [32000, 1280, 1, 256] - - [261, 70.416] - - - [20016, 6656, 1, 256] - - [280, 64.154] - - - [22272, 3328, 1, 256] - - [294, 72.025] - - - [31232, 5120, 1, 256] - - [251, 74.185] - - - [30208, 1280, 1, 256] - - [282, 70.586] - - - [28160, 10240, 1, 256] - - [256, 74.999] - - - [31744, 5632, 1, 256] - - [249, 74.894] - - - [29952, 1792, 1, 256] - - [252, 71.606] - - - [25344, 1792, 1, 256] - - [294, 70.475] - - - [25088, 2048, 1, 256] - - [285, 70.772] - - - [32768, 3329, 1, 256] - - [330, 56.152] - - - [32256, 256, 1, 256] - - [258, 56.472] - - - [23296, 1280, 1, 256] - - [276, 69.319] - - - [30976, 1280, 1, 256] - - [286, 69.531] - - - [21248, 10240, 1, 256] - - [254, 74.773] - - - [22528, 768, 1, 256] - - [267, 66.36] - - - [30720, 6912, 1, 256] - - [254, 75.396] - - - [31232, 3329, 1, 256] - - [264, 69.855] - - - [30976, 7424, 1, 256] - - [268, 73.418] - - - [34048, 1280, 1, 256] - - [286, 70.127] - - - [25600, 2304, 1, 256] - - [256, 72.842] - - - [26928, 3584, 1, 256] - - [280, 63.394] - - - [24832, 4864, 1, 256] - - [284, 74.103] - - - [23808, 2816, 1, 256] - - [261, 72.377] - - - [27392, 3328, 1, 256] - - [270, 71.508] - - - [23552, 1281, 1, 256] - - [301, 63.035] - - - [22016, 10240, 1, 256] - - [274, 75.337] - - - [31536, 256, 1, 256] - - [280, 50.925] - - - [27904, 1792, 1, 256] - - [284, 71.214] - - - [23296, 2816, 1, 256] - - [254, 72.116] - - - [22784, 1280, 1, 256] - - [261, 68.723] - - - [27392, 4096, 1, 256] - - [270, 72.243] - - - [32560, 10240, 1, 256] - - [264, 64.059] - - - [28672, 2865, 1, 256] - - [252, 70.166] - - - [20992, 7424, 1, 256] - - [254, 75.243] - - - [27696, 2865, 1, 256] - - [264, 61.694] - - - [22016, 256, 1, 256] - - [285, 49.001] - - - [28976, 256, 1, 256] - - [282, 47.864] - - - [34304, 3329, 1, 256] - - [274, 69.933] - - - [31488, 3329, 1, 256] - - [281, 69.544] - - - [22016, 8448, 1, 256] - - [278, 75.105] - - - [29440, 256, 1, 256] - - [287, 52.381] - - - [20480, 2048, 1, 256] - - [254, 67.869] - - - [28416, 2304, 1, 256] - - [282, 71.87] - - - [21296, 256, 1, 256] - - [271, 44.768] - - - [27648, 1792, 1, 256] - - [264, 72.137] - - - [23040, 2865, 1, 256] - - [257, 69.926] - - - [25856, 1792, 1, 256] - - [261, 71.058] - - - [24832, 1281, 1, 256] - - [296, 63.319] - - - [32000, 2816, 1, 256] - - [254, 72.454] - - - [24576, 1281, 1, 256] - - [254, 58.687] - - - [29744, 2865, 1, 256] - - [280, 61.455] - - - [29696, 1281, 1, 256] - - [255, 64.089] - - - [31488, 2816, 1, 256] - - [264, 72.59] - - - [26624, 3329, 1, 256] - - [252, 70.492] - - - [25392, 256, 1, 256] - - [280, 50.845] - - - [26112, 6144, 1, 256] - - [252, 74.593] - - - [32304, 2865, 1, 256] - - [285, 62.87] - - - [31488, 8192, 1, 256] - - [264, 74.293] - - - [31232, 1281, 1, 256] - - [260, 64.494] - - - [24112, 768, 1, 256] - - [271, 58.71] - - - [23552, 1536, 1, 256] - - [256, 69.852] - - - [31280, 10240, 1, 256] - - [268, 63.142] - - - [21248, 8192, 1, 256] - - [256, 74.558] - - - [32304, 8960, 1, 256] - - [285, 64.13] - - - [28464, 10240, 1, 256] - - [268, 62.755] - - - [31232, 10240, 1, 256] - - [278, 75.057] - - - [27440, 10240, 1, 256] - - [264, 62.863] - - - [29696, 3584, 1, 256] - - [252, 74.288] - - - [31744, 10240, 1, 256] - - [254, 75.225] - - - [24320, 256, 1, 256] - - [271, 53.207] - - - [29696, 2865, 1, 256] - - [254, 70.8] - - - [29952, 5888, 1, 256] - - [286, 73.889] - - - [30768, 7424, 1, 256] - - [264, 62.446] - - - [24064, 4096, 1, 256] - - [318, 73.037] - - - [30976, 3329, 1, 256] - - [286, 68.061] - - - [20224, 6144, 1, 256] - - [252, 73.983] - - - [26368, 2560, 1, 256] - - [276, 72.493] - - - [28928, 3072, 1, 256] - - [254, 72.55] - - - [23552, 9984, 1, 256] - - [254, 75.93] - - - [34048, 1281, 1, 256] - - [316, 63.624] - - - [32768, 6144, 1, 256] - - [249, 58.987] - - - [31232, 1024, 1, 256] - - [287, 69.645] - - - [34864, 256, 1, 256] - - [271, 54.082] - - - [26416, 10240, 1, 256] - - [280, 62.855] - - - [23296, 2048, 1, 256] - - [280, 69.765] - - - [21760, 2865, 1, 256] - - [252, 69.561] - - - [28672, 4608, 1, 256] - - [252, 73.631] - - - [32768, 2865, 1, 256] - - [330, 56.916] - - - [32512, 6144, 1, 256] - - [257, 74.065] - - - [27392, 2816, 1, 256] - - [303, 70.787] - - - [23552, 10240, 1, 256] - - [254, 75.547] - - - [33024, 10240, 1, 256] - - [270, 74.697] - - - [20736, 3328, 1, 256] - - [256, 72.065] - - - [22528, 512, 1, 256] - - [261, 60.959] - - - [22272, 2304, 1, 256] - - [261, 71.492] - - - [31488, 7936, 1, 256] - - [264, 74.414] - - - [22784, 3329, 1, 256] - - [256, 69.258] - - - [24320, 3328, 1, 256] - - [286, 72.697] - - - [33072, 9728, 1, 256] - - [280, 64.079] - - - [30208, 6656, 1, 256] - - [257, 74.77] - - - [28416, 6144, 1, 256] - - [254, 73.683] - - - [27184, 2816, 1, 256] - - [285, 63.994] - - - [31744, 1536, 1, 256] - - [282, 71.1] - - - [24368, 10240, 1, 256] - - [256, 63.847] - - - [25344, 3328, 1, 256] - - [286, 71.605] - - - [29440, 2048, 1, 256] - - [301, 70.786] - - - [30208, 3329, 1, 256] - - [264, 69.876] - - - [29184, 2816, 1, 256] - - [251, 72.862] - - - [33328, 2816, 1, 256] - - [285, 64.013] - - - [22832, 10240, 1, 256] - - [285, 63.212] - - - [24832, 2048, 1, 256] - - [299, 70.487] - - - [33840, 10240, 1, 256] - - [268, 63.711] - - - [28672, 10240, 1, 256] - - [254, 74.999] - - - [31488, 2865, 1, 256] - - [264, 70.05] - - - [33536, 2816, 1, 256] - - [284, 72.7] - - - [20224, 1281, 1, 256] - - [287, 63.716] - - - [25600, 6144, 1, 256] - - [254, 74.742] - - - [32768, 10240, 1, 256] - - [293, 58.859] - - - [24576, 3329, 1, 256] - - [259, 65.155] - - - [30256, 2816, 1, 256] - - [280, 63.427] - - - [26880, 256, 1, 256] - - [263, 57.446] - - - [33792, 2048, 1, 256] - - [313, 71.385] - - - [20272, 256, 1, 256] - - [287, 50.146] - - - [32256, 10240, 1, 256] - - [278, 75.088] - - - [28928, 5376, 1, 256] - - [274, 73.7] - - - [20224, 2048, 1, 256] - - [250, 69.842] - - - [24064, 3329, 1, 256] - - [278, 69.727] - - - [20016, 6912, 1, 256] - - [280, 64.096] - - - [22784, 256, 1, 256] - - [261, 50.256] - - - [32048, 8704, 1, 256] - - [280, 63.329] - - - [24064, 768, 1, 256] - - [290, 65.425] - - - [32512, 3329, 1, 256] - - [300, 69.597] - - - [28160, 2304, 1, 256] - - [254, 72.19] - - - [20224, 7168, 1, 256] - - [256, 73.51] - - - [21296, 2816, 1, 256] - - [271, 63.29] - - - [26160, 2865, 1, 256] - - [263, 61.461] - - - [20480, 7424, 1, 256] - - [252, 75.186] - - - [20784, 2816, 1, 256] - - [280, 63.356] - - - [20224, 10240, 1, 256] - - [264, 74.867] - - - [29232, 256, 1, 256] - - [271, 47.674] - - - [22016, 2816, 1, 256] - - [254, 72.506] - - - [26112, 10240, 1, 256] - - [254, 75.339] - - - [31744, 8192, 1, 256] - - [254, 74.947] - - - [29952, 10240, 1, 256] - - [257, 74.674] - - - [34560, 2304, 1, 256] - - [262, 72.547] - - - [25088, 1024, 1, 256] - - [290, 68.105] - - - [25344, 5376, 1, 256] - - [269, 73.025] - - - [32768, 8960, 1, 256] - - [265, 59.008] - - - [32512, 2865, 1, 256] - - [300, 69.874] - - - [24576, 768, 1, 256] - - [261, 60.724] - - - [27952, 4608, 1, 256] - - [280, 63.343] - - - [34096, 512, 1, 256] - - [280, 59.465] - - - [21040, 2816, 1, 256] - - [263, 64.179] - - - [22784, 2816, 1, 256] - - [262, 71.985] - - - [23296, 6144, 1, 256] - - [254, 74.154] - - - [25136, 1792, 1, 256] - - [256, 62.163] - - - [23808, 1281, 1, 256] - - [287, 63.94] - - - [27648, 4352, 1, 256] - - [254, 74.703] - - - [22320, 2865, 1, 256] - - [252, 61.292] - - - [28720, 5120, 1, 256] - - [255, 61.738] - - - [31232, 3328, 1, 256] - - [268, 73.148] - - - [34048, 9984, 1, 256] - - [278, 74.355] - - - [21504, 2816, 1, 256] - - [276, 72.798] - - - [23552, 3840, 1, 256] - - [264, 74.249] - - - [25600, 1281, 1, 256] - - [268, 63.258] - - - [23808, 3329, 1, 256] - - [254, 69.505] - - - [23040, 9216, 1, 256] - - [268, 74.942] - - - [33584, 2865, 1, 256] - - [285, 62.267] - - - [27904, 1281, 1, 256] - - [267, 63.938] - - - [20992, 2865, 1, 256] - - [264, 70.053] - - - [28928, 3328, 1, 256] - - [252, 72.199] - - - [26416, 2816, 1, 256] - - [280, 63.431] - - - [26112, 1281, 1, 256] - - [260, 64.084] - - - [29440, 1536, 1, 256] - - [262, 70.113] - - - [22528, 2816, 1, 256] - - [264, 73.081] - - - [28976, 5376, 1, 256] - - [263, 63.383] - - - [22016, 2865, 1, 256] - - [254, 70.061] - - - [24624, 2865, 1, 256] - - [282, 59.672] - - - [23296, 2865, 1, 256] - - [254, 69.699] - - - [23040, 1281, 1, 256] - - [283, 63.922] - - - [30720, 2865, 1, 256] - - [254, 70.964] - - - [34864, 2816, 1, 256] - - [256, 62.188] - - - [27696, 4352, 1, 256] - - [252, 63.284] - - - [28672, 1281, 1, 256] - - [256, 63.048] - - - [28464, 5120, 1, 256] - - [271, 63.253] - - - [21248, 3328, 1, 256] - - [282, 72.083] - - - [32256, 8448, 1, 256] - - [274, 75.036] - - - [34048, 2865, 1, 256] - - [300, 69.561] - - - [28720, 256, 1, 256] - - [256, 46.969] - - - [22784, 8960, 1, 256] - - [254, 74.796] - - - [23808, 512, 1, 256] - - [267, 63.037] - - - [42624, 13824, 1, 384] - - [255, 88.278] - - - [33408, 15360, 1, 384] - - [256, 90.765] - - - [31488, 2688, 1, 384] - - [263, 89.239] - - - [31872, 5760, 1, 384] - - [264, 90.116] - - - [41856, 1152, 1, 384] - - [280, 87.469] - - - [44160, 1153, 1, 384] - - [285, 78.2] - - - [43008, 9216, 1, 384] - - [255, 88.337] - - - [32640, 7297, 1, 384] - - [281, 85.715] - - - [43776, 13441, 1, 384] - - [281, 87.285] - - - [37632, 1152, 1, 384] - - [263, 85.178] - - - [31872, 7297, 1, 384] - - [256, 88.03] - - - [39936, 7297, 1, 384] - - [254, 87.892] - - - [35712, 3072, 1, 384] - - [264, 89.547] - - - [36480, 1152, 1, 384] - - [280, 85.411] - - - [42624, 15360, 1, 384] - - [255, 87.319] - - - [32640, 1153, 1, 384] - - [305, 73.418] - - - [32640, 6912, 1, 384] - - [264, 88.737] - - - [39168, 3840, 1, 384] - - [256, 89.866] - - - [33792, 3072, 1, 384] - - [264, 88.025] - - - [38784, 7296, 1, 384] - - [263, 90.329] - - - [42240, 1152, 1, 384] - - [271, 87.93] - - - [38784, 9216, 1, 384] - - [285, 90.782] - - - [43776, 7297, 1, 384] - - [281, 86.204] - - - [43392, 8064, 1, 384] - - [254, 90.729] - - - [38400, 7296, 1, 384] - - [264, 90.343] - - - [40704, 3072, 1, 384] - - [256, 89.188] - - - [36864, 1920, 1, 384] - - [264, 87.498] - - - [43008, 13824, 1, 384] - - [256, 90.147] - - - [41472, 12288, 1, 384] - - [288, 89.607] - - - [35712, 6912, 1, 384] - - [252, 90.541] - - - [36480, 9216, 1, 384] - - [271, 90.768] - - - [36096, 15360, 1, 384] - - [257, 89.555] - - - [37632, 2688, 1, 384] - - [271, 88.713] - - - [38400, 13440, 1, 384] - - [264, 90.905] - - - [33792, 13440, 1, 384] - - [252, 90.724] - - - [31872, 3072, 1, 384] - - [252, 88.853] - - - [34560, 1153, 1, 384] - - [263, 77.466] - - - [39168, 4224, 1, 384] - - [252, 90.179] - - - [41472, 6528, 1, 384] - - [285, 90.214] - - - [36480, 13441, 1, 384] - - [254, 88.921] - - - [34560, 13440, 1, 384] - - [252, 90.784] - - - [37632, 15360, 1, 384] - - [252, 90.795] - - - [37248, 13440, 1, 384] - - [264, 90.849] - - - [41088, 3072, 1, 384] - - [259, 86.816] - - - [33792, 7296, 1, 384] - - [252, 90.356] - - - [31104, 7297, 1, 384] - - [264, 88.048] - - - [35712, 1152, 1, 384] - - [263, 86.709] - - - [37632, 2304, 1, 384] - - [280, 88.3] - - - [39552, 3072, 1, 384] - - [252, 89.481] - - - [33408, 4608, 1, 384] - - [285, 89.822] - - - [32256, 7296, 1, 384] - - [254, 90.393] - - - [38016, 9216, 1, 384] - - [263, 90.741] - - - [34944, 2688, 1, 384] - - [271, 88.484] - - - [39168, 7297, 1, 384] - - [264, 88.429] - - - [34560, 13441, 1, 384] - - [254, 89.15] - - - [36864, 1153, 1, 384] - - [252, 76.423] - - - [31488, 13440, 1, 384] - - [252, 90.896] - - - [36096, 1152, 1, 384] - - [281, 83.95] - - - [37632, 9216, 1, 384] - - [271, 90.767] - - - [34944, 5760, 1, 384] - - [254, 90.259] - - - [41088, 11904, 1, 384] - - [301, 89.957] - - - [32640, 15360, 1, 384] - - [281, 89.211] - - - [31104, 1153, 1, 384] - - [252, 75.891] - - - [42240, 13440, 1, 384] - - [285, 91.201] - - - [33408, 3072, 1, 384] - - [252, 89.593] - - - [39168, 7296, 1, 384] - - [252, 90.527] - - - [35328, 15360, 1, 384] - - [254, 90.572] - - - [37248, 3072, 1, 384] - - [254, 89.373] - - - [40704, 1152, 1, 384] - - [256, 85.429] - - - [34944, 1153, 1, 384] - - [280, 78.011] - - - [43008, 7297, 1, 384] - - [252, 87.793] - - - [34176, 1920, 1, 384] - - [252, 87.757] - - - [37632, 3072, 1, 384] - - [252, 88.886] - - - [34944, 1152, 1, 384] - - [285, 85.364] - - - [33792, 1536, 1, 384] - - [256, 86.629] - - - [37632, 7296, 1, 384] - - [264, 90.485] - - - [36096, 6912, 1, 384] - - [278, 89.39] - - - [43776, 9216, 1, 384] - - [301, 89.701] - - - [39552, 15360, 1, 384] - - [264, 90.804] - - - [33792, 1153, 1, 384] - - [254, 78.249] - - - [35328, 3072, 1, 384] - - [252, 88.672] - - - [38400, 3072, 1, 384] - - [252, 88.81] - - - [31872, 1152, 1, 384] - - [271, 84.902] - - - [38016, 15360, 1, 384] - - [256, 91.029] - - - [33024, 13441, 1, 384] - - [259, 88.236] - - - [36864, 13440, 1, 384] - - [252, 90.3] - - - [37248, 9216, 1, 384] - - [256, 90.012] - - - [31488, 3072, 1, 384] - - [254, 88.897] - - - [37248, 7297, 1, 384] - - [251, 87.954] - - - [36480, 1536, 1, 384] - - [256, 85.898] - - - [43392, 13441, 1, 384] - - [252, 89.041] - - - [36480, 7296, 1, 384] - - [285, 90.429] - - - [36096, 768, 1, 384] - - [289, 81.715] - - - [31872, 13441, 1, 384] - - [252, 88.734] - - - [33024, 1152, 1, 384] - - [280, 83.603] - - - [32640, 3840, 1, 384] - - [256, 87.392] - - - [32256, 1152, 1, 384] - - [254, 85.845] - - - [43776, 2688, 1, 384] - - [301, 87.926] - - - [32256, 6528, 1, 384] - - [252, 89.983] - - - [36096, 7296, 1, 384] - - [257, 89.202] - - - [38784, 7297, 1, 384] - - [252, 87.841] - - - [37632, 8448, 1, 384] - - [280, 90.624] - - - [41856, 13056, 1, 384] - - [280, 91.0] - - - [36864, 7680, 1, 384] - - [252, 89.447] - - - [39168, 13440, 1, 384] - - [264, 90.989] - - - [34176, 15360, 1, 384] - - [252, 90.734] - - - [33792, 4992, 1, 384] - - [254, 90.089] - - - [39168, 9984, 1, 384] - - [252, 90.689] - - - [43008, 1536, 1, 384] - - [256, 86.668] - - - [40704, 7296, 1, 384] - - [285, 90.655] - - - [39168, 13441, 1, 384] - - [256, 89.187] - - - [38400, 15360, 1, 384] - - [264, 90.558] - - - [38400, 13441, 1, 384] - - [252, 88.937] - - - [39936, 4608, 1, 384] - - [256, 88.688] - - - [34176, 13441, 1, 384] - - [257, 88.89] - - - [44160, 13441, 1, 384] - - [264, 88.866] - - - [39936, 13441, 1, 384] - - [264, 88.989] - - - [32640, 3072, 1, 384] - - [252, 86.224] - - - [34944, 15360, 1, 384] - - [264, 90.744] - - - [40320, 5376, 1, 384] - - [263, 90.308] - - - [34176, 4992, 1, 384] - - [256, 89.892] - - - [41856, 1153, 1, 384] - - [252, 78.384] - - - [40704, 2304, 1, 384] - - [280, 88.393] - - - [40704, 5760, 1, 384] - - [264, 90.322] - - - [38784, 1152, 1, 384] - - [252, 87.044] - - - [34560, 2304, 1, 384] - - [264, 88.169] - - - [35712, 13440, 1, 384] - - [256, 90.934] - - - [43392, 14208, 1, 384] - - [252, 90.644] - - - [35712, 6528, 1, 384] - - [271, 90.358] - - - [31104, 9216, 1, 384] - - [271, 90.311] - - - [39936, 11136, 1, 384] - - [252, 90.513] - - - [41856, 768, 1, 384] - - [280, 85.598] - - - [34176, 7297, 1, 384] - - [251, 87.794] - - - [36480, 7680, 1, 384] - - [285, 90.618] - - - [44160, 2688, 1, 384] - - [271, 89.787] - - - [40704, 2688, 1, 384] - - [263, 89.517] - - - [44160, 7296, 1, 384] - - [285, 90.668] - - - [38016, 3072, 1, 384] - - [256, 89.602] - - - [41856, 9216, 1, 384] - - [285, 90.762] - - - [40320, 13441, 1, 384] - - [254, 88.933] - - - [41856, 7297, 1, 384] - - [285, 88.115] - - - [33408, 1152, 1, 384] - - [263, 85.116] - - - [31104, 1920, 1, 384] - - [264, 86.98] - - - [31872, 7296, 1, 384] - - [252, 90.056] - - - [35328, 13440, 1, 384] - - [252, 90.941] - - - [34944, 3072, 1, 384] - - [252, 89.225] - - - [31104, 1152, 1, 384] - - [285, 86.317] - - - [43776, 14592, 1, 384] - - [268, 89.832] - - - [33024, 7296, 1, 384] - - [256, 89.596] - - - [31104, 7296, 1, 384] - - [252, 90.339] - - - [34944, 13441, 1, 384] - - [264, 89.08] - - - [35328, 13441, 1, 384] - - [252, 89.221] - - - [40320, 1920, 1, 384] - - [280, 89.136] - - - [39168, 1153, 1, 384] - - [280, 78.445] - - - [41088, 1152, 1, 384] - - [301, 84.162] - - - [40704, 11520, 1, 384] - - [252, 90.92] - - - [42240, 15360, 1, 384] - - [263, 91.13] - - - [31104, 13440, 1, 384] - - [256, 90.806] - - - [34176, 3072, 1, 384] - - [256, 88.904] - - - [43392, 1153, 1, 384] - - [280, 78.69] - - - [43008, 15360, 1, 384] - - [281, 89.299] - - - [43776, 7296, 1, 384] - - [268, 89.302] - - - [38400, 9216, 1, 384] - - [283, 89.62] - - - [32256, 6144, 1, 384] - - [254, 89.891] - - - [33408, 4224, 1, 384] - - [263, 89.662] - - - [38016, 7296, 1, 384] - - [256, 90.545] - - - [34944, 7297, 1, 384] - - [254, 87.974] - - - [39936, 1920, 1, 384] - - [254, 88.364] - - - [32256, 7297, 1, 384] - - [254, 88.391] - - - [41472, 3072, 1, 384] - - [264, 89.299] - - - [36864, 7296, 1, 384] - - [264, 89.855] - - - [40320, 7297, 1, 384] - - [256, 88.106] - - - [43008, 13441, 1, 384] - - [252, 88.6] - - - [35328, 9216, 1, 384] - - [301, 89.367] - - - [40320, 13440, 1, 384] - - [254, 91.065] - - - [39936, 10752, 1, 384] - - [264, 90.046] - - - [36864, 1536, 1, 384] - - [252, 85.867] - - - [31872, 2688, 1, 384] - - [280, 88.748] - - - [38400, 1153, 1, 384] - - [263, 77.123] - - - [41856, 13441, 1, 384] - - [256, 88.889] - - - [39552, 4608, 1, 384] - - [271, 89.957] - - - [42240, 3072, 1, 384] - - [256, 89.467] - - - [35712, 768, 1, 384] - - [271, 82.475] - - - [40704, 13441, 1, 384] - - [254, 89.252] - - - [34176, 8448, 1, 384] - - [320, 90.045] - - - [36864, 13441, 1, 384] - - [252, 87.818] - - - [37248, 1152, 1, 384] - - [280, 87.015] - - - [41088, 6144, 1, 384] - - [288, 88.988] - - - [44160, 15360, 1, 384] - - [271, 90.988] - - - [35328, 6144, 1, 384] - - [256, 89.727] - - - [33408, 7680, 1, 384] - - [256, 90.515] - - - [43776, 8832, 1, 384] - - [301, 89.563] - - - [32256, 9216, 1, 384] - - [254, 89.77] - - - [37248, 7296, 1, 384] - - [264, 90.367] - - - [34560, 8448, 1, 384] - - [289, 90.284] - - - [41472, 15360, 1, 384] - - [252, 90.467] - - - [41088, 13441, 1, 384] - - [281, 87.66] - - - [37248, 1920, 1, 384] - - [254, 88.403] - - - [40320, 2304, 1, 384] - - [271, 89.112] - - - [42240, 13056, 1, 384] - - [264, 90.791] - - - [40704, 7297, 1, 384] - - [280, 88.281] - - - [39936, 4992, 1, 384] - - [254, 90.119] - - - [37632, 1153, 1, 384] - - [271, 78.129] - - - [40704, 9216, 1, 384] - - [280, 90.812] - - - [43392, 13440, 1, 384] - - [252, 90.96] - - - [41088, 12288, 1, 384] - - [301, 89.769] - - - [39936, 3072, 1, 384] - - [281, 87.229] - - - [41088, 13440, 1, 384] - - [264, 90.307] - - - [33792, 8064, 1, 384] - - [264, 90.4] - - - [35328, 7297, 1, 384] - - [252, 88.144] - - - [33024, 6912, 1, 384] - - [264, 89.944] - - - [42240, 768, 1, 384] - - [280, 86.132] - - - [39552, 7297, 1, 384] - - [252, 88.337] - - - [44160, 9216, 1, 384] - - [285, 90.934] - - - [39552, 13441, 1, 384] - - [252, 88.971] - - - [43008, 7296, 1, 384] - - [256, 90.267] - - - [41472, 13440, 1, 384] - - [252, 90.964] - - - [34944, 9216, 1, 384] - - [263, 90.258] - - - [32640, 9216, 1, 384] - - [255, 88.653] - - - [39552, 1152, 1, 384] - - [263, 86.13] - - - [42624, 9216, 1, 384] - - [274, 86.819] - - - [38784, 9600, 1, 384] - - [271, 90.879] - - - [30720, 15360, 1, 384] - - [281, 89.232] - - - [34944, 8832, 1, 384] - - [252, 90.609] - - - [34560, 7296, 1, 384] - - [264, 90.444] - - - [40320, 7296, 1, 384] - - [264, 90.462] - - - [41088, 9216, 1, 384] - - [301, 89.557] - - - [39552, 4224, 1, 384] - - [271, 90.185] - - - [36480, 7297, 1, 384] - - [252, 87.916] - - - [34176, 8064, 1, 384] - - [252, 90.295] - - - [34560, 3072, 1, 384] - - [254, 89.269] - - - [32256, 13441, 1, 384] - - [252, 89.398] - - - [35328, 1153, 1, 384] - - [252, 76.639] - - - [12672, 7296, 1, 384] - - [252, 88.855] - - - [19968, 13440, 1, 384] - - [256, 90.89] - - - [19968, 9216, 1, 384] - - [280, 89.746] - - - [17280, 3072, 1, 384] - - [254, 87.102] - - - [21120, 1536, 1, 384] - - [264, 85.882] - - - [21120, 9216, 1, 384] - - [283, 90.106] - - - [12288, 4608, 1, 384] - - [254, 86.933] - - - [7296, 6912, 1, 384] - - [254, 88.378] - - - [3072, 2688, 1, 384] - - [262, 72.028] - - - [8064, 7680, 1, 384] - - [254, 87.628] - - - [24960, 8064, 1, 384] - - [256, 90.562] - - - [29568, 15360, 1, 384] - - [278, 90.385] - - - [11136, 10752, 1, 384] - - [254, 89.659] - - - [19584, 3072, 1, 384] - - [264, 87.287] - - - [15360, 7296, 1, 384] - - [252, 88.972] - - - [5376, 5377, 1, 384] - - [256, 82.103] - - - [11904, 4224, 1, 384] - - [271, 88.137] - - - [20352, 7297, 1, 384] - - [252, 88.385] - - - [17280, 13441, 1, 384] - - [256, 89.218] - - - [20352, 6528, 1, 384] - - [271, 90.017] - - - [15744, 8064, 1, 384] - - [252, 89.745] - - - [20352, 7296, 1, 384] - - [256, 89.816] - - - [16128, 1152, 1, 384] - - [254, 79.04] - - - [19200, 13441, 1, 384] - - [254, 89.246] - - - [7680, 7297, 1, 384] - - [252, 86.162] - - - [23424, 9216, 1, 384] - - [254, 89.747] - - - [28032, 2304, 1, 384] - - [263, 88.371] - - - [11904, 1152, 1, 384] - - [252, 77.734] - - - [26112, 1153, 1, 384] - - [285, 75.83] - - - [14976, 1536, 1, 384] - - [252, 82.728] - - - [5376, 5376, 1, 384] - - [252, 82.183] - - - [22656, 3072, 1, 384] - - [254, 87.879] - - - [13824, 3072, 1, 384] - - [254, 85.697] - - - [28800, 13440, 1, 384] - - [271, 90.974] - - - [28416, 13441, 1, 384] - - [254, 89.206] - - - [4992, 4608, 1, 384] - - [252, 84.18] - - - [6912, 3072, 1, 384] - - [252, 83.439] - - - [11904, 11905, 1, 384] - - [252, 88.706] - - - [25728, 2688, 1, 384] - - [263, 87.846] - - - [9984, 7297, 1, 384] - - [252, 86.231] - - - [3456, 1920, 1, 384] - - [261, 69.386] - - - [15744, 2304, 1, 384] - - [252, 84.376] - - - [3072, 3072, 1, 384] - - [264, 70.365] - - - [16128, 13440, 1, 384] - - [254, 90.31] - - - [21504, 9216, 1, 384] - - [288, 88.698] - - - [10368, 1153, 1, 384] - - [263, 68.413] - - - [8832, 1536, 1, 384] - - [252, 77.661] - - - [24192, 1153, 1, 384] - - [263, 74.563] - - - [20736, 13440, 1, 384] - - [254, 90.841] - - - [18048, 1536, 1, 384] - - [252, 82.772] - - - [13440, 13056, 1, 384] - - [252, 90.019] - - - [1920, 1920, 1, 384] - - [276, 50.157] - - - [28800, 2688, 1, 384] - - [280, 88.798] - - - [3840, 3841, 1, 384] - - [256, 76.356] - - - [21888, 2304, 1, 384] - - [281, 85.589] - - - [18816, 3072, 1, 384] - - [252, 86.239] - - - [20736, 6912, 1, 384] - - [280, 89.894] - - - [29952, 3072, 1, 384] - - [264, 88.052] - - - [15360, 8064, 1, 384] - - [256, 89.763] - - - [3456, 2304, 1, 384] - - [282, 69.753] - - - [14208, 1153, 1, 384] - - [263, 69.98] - - - [1920, 1921, 1, 384] - - [267, 49.571] - - - [8448, 3072, 1, 384] - - [256, 82.728] - - - [4992, 768, 1, 384] - - [250, 52.095] - - - [8448, 1152, 1, 384] - - [262, 72.41] - - - [19584, 1153, 1, 384] - - [254, 76.048] - - - [28416, 2688, 1, 384] - - [280, 88.293] - - - [16128, 7296, 1, 384] - - [256, 89.025] - - - [26880, 13056, 1, 384] - - [285, 90.886] - - - [20352, 13441, 1, 384] - - [252, 89.041] - - - [21120, 7680, 1, 384] - - [252, 90.502] - - - [26112, 13440, 1, 384] - - [264, 90.947] - - - [6144, 5761, 1, 384] - - [252, 82.18] - - - [9600, 9216, 1, 384] - - [263, 88.337] - - - [24960, 2304, 1, 384] - - [264, 88.378] - - - [28800, 3072, 1, 384] - - [254, 87.63] - - - [25728, 3072, 1, 384] - - [256, 88.902] - - - [18816, 7296, 1, 384] - - [263, 89.387] - - - [13440, 13440, 1, 384] - - [252, 90.071] - - - [8064, 768, 1, 384] - - [262, 63.254] - - - [26496, 13056, 1, 384] - - [252, 90.911] - - - [18048, 9216, 1, 384] - - [285, 89.946] - - - [22656, 1152, 1, 384] - - [271, 83.362] - - - [24960, 13441, 1, 384] - - [263, 89.752] - - - [11136, 3072, 1, 384] - - [254, 86.907] - - - [16896, 3456, 1, 384] - - [256, 87.275] - - - [6528, 6528, 1, 384] - - [254, 86.8] - - - [2688, 1536, 1, 384] - - [261, 56.219] - - - [6144, 5760, 1, 384] - - [254, 85.977] - - - [12288, 12288, 1, 384] - - [256, 88.86] - - - [25344, 8448, 1, 384] - - [256, 90.218] - - - [15744, 7296, 1, 384] - - [252, 89.763] - - - [15360, 1920, 1, 384] - - [252, 83.634] - - - [6912, 1152, 1, 384] - - [261, 69.651] - - - [28800, 1153, 1, 384] - - [280, 76.181] - - - [20736, 7297, 1, 384] - - [254, 88.128] - - - [8832, 7297, 1, 384] - - [263, 86.031] - - - [23424, 9600, 1, 384] - - [252, 90.657] - - - [26880, 13441, 1, 384] - - [271, 89.585] - - - [9600, 9600, 1, 384] - - [252, 88.649] - - - [29568, 3456, 1, 384] - - [280, 88.517] - - - [27648, 9216, 1, 384] - - [288, 88.785] - - - [12672, 12289, 1, 384] - - [280, 87.114] - - - [21888, 9216, 1, 384] - - [255, 87.296] - - - [22656, 7296, 1, 384] - - [256, 90.219] - - - [10752, 1153, 1, 384] - - [261, 70.09] - - - [23424, 13440, 1, 384] - - [264, 90.874] - - - [18432, 1920, 1, 384] - - [256, 85.746] - - - [18816, 1153, 1, 384] - - [280, 73.484] - - - [11520, 11521, 1, 384] - - [256, 88.63] - - - [15360, 13441, 1, 384] - - [252, 89.017] - - - [26496, 3072, 1, 384] - - [254, 87.645] - - - [18816, 13441, 1, 384] - - [271, 89.24] - - - [20352, 3072, 1, 384] - - [254, 88.042] - - - [28800, 9216, 1, 384] - - [285, 90.373] - - - [29952, 1153, 1, 384] - - [264, 76.04] - - - [19584, 5760, 1, 384] - - [264, 89.61] - - - [7296, 3072, 1, 384] - - [254, 82.095] - - - [12288, 4992, 1, 384] - - [264, 86.87] - - - [30720, 1152, 1, 384] - - [256, 85.402] - - - [21504, 7296, 1, 384] - - [254, 89.468] - - - [23808, 6912, 1, 384] - - [280, 90.007] - - - [2688, 2688, 1, 384] - - [261, 64.683] - - - [29568, 13440, 1, 384] - - [263, 90.69] - - - [2688, 2689, 1, 384] - - [276, 63.266] - - - [22272, 13440, 1, 384] - - [254, 90.599] - - - [21504, 7680, 1, 384] - - [256, 90.021] - - - [15360, 1536, 1, 384] - - [254, 80.28] - - - [11136, 3456, 1, 384] - - [254, 85.406] - - - [18048, 7297, 1, 384] - - [256, 87.754] - - - [23040, 9600, 1, 384] - - [252, 90.44] - - - [10752, 7297, 1, 384] - - [252, 86.643] - - - [29184, 9216, 1, 384] - - [283, 89.646] - - - [23808, 1152, 1, 384] - - [271, 82.731] - - - [25344, 1152, 1, 384] - - [252, 83.112] - - - [14976, 7296, 1, 384] - - [264, 89.538] - - - [7680, 7681, 1, 384] - - [252, 85.928] - - - [5760, 1152, 1, 384] - - [252, 68.56] - - - [17664, 768, 1, 384] - - [280, 77.465] - - - [9984, 2688, 1, 384] - - [263, 85.359] - - - [17664, 3072, 1, 384] - - [252, 86.289] - - - [8448, 8448, 1, 384] - - [280, 88.552] - - - [4224, 2688, 1, 384] - - [276, 74.08] - - - [29184, 1152, 1, 384] - - [252, 85.452] - - - [6912, 6913, 1, 384] - - [264, 86.541] - - - [14208, 7296, 1, 384] - - [252, 89.3] - - - [29184, 7296, 1, 384] - - [252, 90.039] - - - [2304, 1153, 1, 384] - - [261, 49.728] - - - [8448, 8065, 1, 384] - - [254, 86.48] - - - [1536, 1537, 1, 384] - - [286, 44.725] - - - [19968, 7296, 1, 384] - - [252, 90.127] - - - [4608, 1152, 1, 384] - - [262, 69.296] - - - [26880, 7297, 1, 384] - - [285, 88.499] - - - [22272, 7296, 1, 384] - - [252, 89.771] - - - [26880, 13440, 1, 384] - - [280, 90.909] - - - [23424, 13441, 1, 384] - - [274, 88.888] - - - [21120, 1152, 1, 384] - - [254, 82.711] - - - [28032, 13440, 1, 384] - - [285, 90.785] - - - [3456, 1152, 1, 384] - - [261, 54.395] - - - [7680, 3072, 1, 384] - - [252, 80.713] - - - [13056, 1153, 1, 384] - - [263, 69.844] - - - [29568, 1153, 1, 384] - - [285, 77.447] - - - [9216, 9216, 1, 384] - - [252, 87.436] - - - [8064, 1152, 1, 384] - - [252, 68.874] - - - [22272, 13441, 1, 384] - - [285, 89.718] - - - [19584, 6144, 1, 384] - - [285, 89.72] - - - [8064, 7681, 1, 384] - - [252, 86.659] - - - [24960, 7296, 1, 384] - - [254, 89.816] - - - [19968, 6528, 1, 384] - - [271, 89.457] - - - [29952, 13440, 1, 384] - - [264, 90.917] - - - [8832, 1153, 1, 384] - - [266, 65.553] - - - [2688, 2305, 1, 384] - - [261, 64.521] - - - [28032, 3072, 1, 384] - - [264, 89.145] - - - [21888, 1920, 1, 384] - - [265, 83.085] - - - [9600, 9601, 1, 384] - - [285, 88.003] - - - [19584, 2688, 1, 384] - - [263, 86.563] - - - [29568, 1152, 1, 384] - - [281, 85.277] - - - [18816, 4992, 1, 384] - - [280, 88.521] - - - [22656, 13440, 1, 384] - - [254, 90.994] - - - [3456, 1153, 1, 384] - - [262, 53.265] - - - [15744, 8448, 1, 384] - - [280, 90.086] - - - [28032, 14208, 1, 384] - - [285, 90.88] - - - [11520, 768, 1, 384] - - [280, 75.835] - - - [18432, 1153, 1, 384] - - [254, 72.221] - - - [25344, 9216, 1, 384] - - [285, 90.665] - - - [8832, 1152, 1, 384] - - [276, 75.114] - - - [29952, 7296, 1, 384] - - [254, 90.178] - - - [16896, 13440, 1, 384] - - [254, 90.773] - - - [30336, 9216, 1, 384] - - [280, 90.28] - - - [10368, 10368, 1, 384] - - [285, 89.575] - - - [24576, 1920, 1, 384] - - [256, 83.311] - - - [12672, 5376, 1, 384] - - [280, 88.794] - - - [3072, 3073, 1, 384] - - [254, 69.368] - - - [9984, 9600, 1, 384] - - [252, 88.848] - - - [26496, 7296, 1, 384] - - [252, 90.121] - - - [5760, 5761, 1, 384] - - [254, 84.571] - - - [26880, 768, 1, 384] - - [280, 81.248] - - - [13440, 7296, 1, 384] - - [252, 89.037] - - - [24960, 3072, 1, 384] - - [254, 88.551] - - - [27648, 13441, 1, 384] - - [256, 88.928] - - - [22272, 2304, 1, 384] - - [252, 87.15] - - - [24192, 13440, 1, 384] - - [254, 90.816] - - - [29952, 1152, 1, 384] - - [252, 84.039] - - - [8448, 7296, 1, 384] - - [254, 87.405] - - - [21120, 1153, 1, 384] - - [264, 76.613] - - - [28032, 13441, 1, 384] - - [254, 89.389] - - - [19584, 9216, 1, 384] - - [280, 89.89] - - - [29952, 768, 1, 384] - - [280, 83.522] - - - [9600, 9217, 1, 384] - - [289, 86.678] - - - [25344, 8832, 1, 384] - - [254, 90.401] - - - [2304, 1920, 1, 384] - - [276, 59.968] - - - [26496, 1153, 1, 384] - - [280, 77.269] - - - [19200, 3072, 1, 384] - - [264, 87.757] - - - [21504, 7297, 1, 384] - - [252, 87.803] - - - [17664, 3840, 1, 384] - - [285, 88.524] - - - [28032, 14592, 1, 384] - - [263, 91.167] - - - [11136, 1152, 1, 384] - - [254, 74.068] - - - [19968, 3456, 1, 384] - - [280, 87.779] - - - [17280, 3840, 1, 384] - - [254, 88.77] - - - [21120, 3072, 1, 384] - - [252, 88.775] - - - [6528, 6529, 1, 384] - - [254, 83.475] - - - [14592, 7296, 1, 384] - - [252, 88.764] - - - [24576, 1536, 1, 384] - - [264, 83.575] - - - [4608, 3072, 1, 384] - - [262, 80.893] - - - [17664, 1152, 1, 384] - - [263, 80.529] - - - [16896, 9216, 1, 384] - - [271, 89.757] - - - [27264, 9216, 1, 384] - - [252, 89.624] - - - [24576, 11136, 1, 384] - - [249, 87.405] - - - [23424, 7296, 1, 384] - - [252, 90.291] - - - [8832, 8832, 1, 384] - - [264, 88.346] - - - [22656, 8832, 1, 384] - - [285, 90.025] - - - [6528, 1152, 1, 384] - - [276, 65.966] - - - [28800, 15360, 1, 384] - - [263, 90.849] - - - [13440, 1152, 1, 384] - - [280, 79.284] - - - [3456, 3456, 1, 384] - - [282, 77.322] - - - [4224, 1153, 1, 384] - - [262, 63.2] - - - [25728, 7297, 1, 384] - - [285, 88.005] - - - [30720, 4992, 1, 384] - - [252, 89.461] - - - [21504, 1920, 1, 384] - - [254, 87.192] - - - [22656, 6144, 1, 384] - - [285, 89.491] - - - [9216, 7296, 1, 384] - - [256, 87.749] - - - [23040, 13441, 1, 384] - - [256, 89.079] - - - [30336, 4224, 1, 384] - - [256, 89.287] - - - [11904, 11904, 1, 384] - - [285, 90.052] - - - [11904, 1536, 1, 384] - - [264, 78.25] - - - [9984, 2304, 1, 384] - - [285, 83.947] - - - [18432, 7297, 1, 384] - - [256, 87.867] - - - [12288, 11904, 1, 384] - - [264, 89.561] - - - [3072, 2689, 1, 384] - - [261, 71.091] - - - [26496, 7297, 1, 384] - - [256, 88.044] - - - [6912, 6912, 1, 384] - - [256, 87.066] - - - [15744, 13440, 1, 384] - - [256, 90.395] - - - [8448, 7297, 1, 384] - - [254, 86.977] - - - [18432, 7296, 1, 384] - - [254, 89.726] - - - [30720, 1536, 1, 384] - - [254, 85.263] - - - [24192, 9216, 1, 384] - - [285, 90.538] - - - [24576, 1153, 1, 384] - - [256, 73.377] - - - [12672, 1920, 1, 384] - - [264, 82.978] - - - [27264, 1536, 1, 384] - - [254, 84.468] - - - [21504, 4992, 1, 384] - - [254, 89.361] - - - [18432, 3072, 1, 384] - - [264, 86.579] - - - [24192, 1536, 1, 384] - - [254, 85.442] - - - [13824, 6528, 1, 384] - - [280, 88.339] - - - [23424, 9984, 1, 384] - - [256, 90.684] - - - [23424, 768, 1, 384] - - [280, 83.094] - - - [9600, 2304, 1, 384] - - [254, 81.095] - - - [26880, 3072, 1, 384] - - [256, 88.803] - - - [23040, 6144, 1, 384] - - [285, 89.401] - - - [30336, 1536, 1, 384] - - [254, 86.95] - - - [14976, 1152, 1, 384] - - [252, 80.267] - - - [28800, 13441, 1, 384] - - [252, 88.768] - - - [16128, 2688, 1, 384] - - [285, 84.65] - - - [10368, 7296, 1, 384] - - [254, 89.3] - - - [24960, 9216, 1, 384] - - [263, 90.634] - - - [14208, 768, 1, 384] - - [267, 70.96] - - - [21888, 13440, 1, 384] - - [255, 88.568] - - - [12288, 1920, 1, 384] - - [254, 80.703] - - - [6528, 1153, 1, 384] - - [261, 64.455] - - - [7296, 1153, 1, 384] - - [287, 62.16] - - - [27264, 13441, 1, 384] - - [254, 88.624] - - - [4992, 4992, 1, 384] - - [264, 85.007] - - - [20352, 13440, 1, 384] - - [252, 90.851] - - - [13056, 7296, 1, 384] - - [256, 89.592] - - - [13824, 6144, 1, 384] - - [252, 88.162] - - - [17280, 13440, 1, 384] - - [254, 90.76] - - - [3456, 3072, 1, 384] - - [261, 78.283] - - - [24192, 7680, 1, 384] - - [252, 90.064] - - - [8832, 3072, 1, 384] - - [256, 81.695] - - - [23424, 6912, 1, 384] - - [254, 90.398] - - - [18048, 3072, 1, 384] - - [252, 87.997] - - - [11520, 3840, 1, 384] - - [256, 86.779] - - - [11136, 10753, 1, 384] - - [254, 87.819] - - - [26496, 13441, 1, 384] - - [254, 89.289] - - - [25728, 7296, 1, 384] - - [254, 89.917] - - - [18816, 9216, 1, 384] - - [263, 89.943] - - - [16896, 1153, 1, 384] - - [271, 75.524] - - - [22656, 13441, 1, 384] - - [285, 89.502] - - - [26496, 768, 1, 384] - - [263, 80.366] - - - [14592, 6912, 1, 384] - - [256, 89.845] - - - [19584, 7297, 1, 384] - - [252, 88.16] - - - [21120, 7297, 1, 384] - - [274, 88.024] - - - [12288, 7296, 1, 384] - - [256, 89.211] - - - [9984, 9601, 1, 384] - - [254, 88.172] - - - [21120, 7296, 1, 384] - - [254, 89.903] - - - [27648, 3072, 1, 384] - - [256, 86.696] - - - [28032, 1152, 1, 384] - - [285, 86.259] - - - [22272, 5760, 1, 384] - - [252, 89.478] - - - [26880, 7296, 1, 384] - - [254, 90.428] - - - [9984, 7296, 1, 384] - - [254, 88.426] - - - [10368, 9984, 1, 384] - - [256, 89.268] - - - [5376, 1152, 1, 384] - - [262, 64.654] - - - [14592, 3072, 1, 384] - - [256, 87.402] - - - [23424, 1152, 1, 384] - - [254, 85.461] - - - [13056, 2304, 1, 384] - - [256, 85.22] - - - [18048, 1152, 1, 384] - - [254, 81.926] - - - [21888, 7297, 1, 384] - - [294, 83.447] - - - [21120, 4224, 1, 384] - - [254, 88.889] - - - [3840, 2688, 1, 384] - - [282, 76.533] - - - [13824, 1153, 1, 384] - - [263, 73.584] - - - [11136, 768, 1, 384] - - [267, 73.494] - - - [24192, 7297, 1, 384] - - [256, 88.31] - - - [30720, 4608, 1, 384] - - [252, 88.764] - - - [29184, 1153, 1, 384] - - [271, 77.332] - - - [24960, 13440, 1, 384] - - [271, 90.738] - - - [15360, 13440, 1, 384] - - [254, 90.359] - - - [8832, 8448, 1, 384] - - [254, 88.439] - - - [19968, 3072, 1, 384] - - [264, 86.678] - - - [15360, 1152, 1, 384] - - [254, 81.733] - - - [14976, 1153, 1, 384] - - [280, 73.21] - - - [16512, 8832, 1, 384] - - [268, 87.877] - - - [22272, 1152, 1, 384] - - [256, 82.343] - - - [15744, 1920, 1, 384] - - [252, 85.498] - - - [29952, 13441, 1, 384] - - [254, 89.176] - - - [13440, 2688, 1, 384] - - [285, 83.727] - - - [22656, 9216, 1, 384] - - [263, 90.68] - - - [20736, 1152, 1, 384] - - [252, 81.729] - - - [15744, 7297, 1, 384] - - [252, 88.247] - - - [29952, 7297, 1, 384] - - [264, 88.269] - - - [9600, 3072, 1, 384] - - [252, 83.769] - - - [27648, 13824, 1, 384] - - [256, 90.458] - - - [13824, 13440, 1, 384] - - [254, 90.055] - - - [23808, 1153, 1, 384] - - [271, 76.937] - - - [12288, 3072, 1, 384] - - [254, 86.147] - - - [30336, 13440, 1, 384] - - [252, 90.835] - - - [23040, 9216, 1, 384] - - [285, 90.364] - - - [22656, 5760, 1, 384] - - [252, 89.68] - - - [28032, 7296, 1, 384] - - [256, 90.255] - - - [7680, 1152, 1, 384] - - [276, 75.841] - - - [13056, 5376, 1, 384] - - [254, 89.113] - - - [17280, 9216, 1, 384] - - [285, 89.968] - - - [4608, 1153, 1, 384] - - [267, 55.973] - - - [30720, 7297, 1, 384] - - [252, 88.048] - - - [9984, 3072, 1, 384] - - [264, 82.616] - - - [3840, 1152, 1, 384] - - [276, 60.113] - - - [14592, 7297, 1, 384] - - [256, 87.088] - - - [22272, 9216, 1, 384] - - [280, 90.666] - - - [1536, 1536, 1, 384] - - [286, 44.997] - - - [26112, 12288, 1, 384] - - [280, 90.415] - - - [13440, 6144, 1, 384] - - [263, 89.227] - - - [17280, 1152, 1, 384] - - [254, 78.728] - - - [2304, 1921, 1, 384] - - [262, 59.305] - - - [16896, 1152, 1, 384] - - [252, 82.505] - - - [27264, 13824, 1, 384] - - [264, 90.563] - - - [30336, 7296, 1, 384] - - [256, 90.3] - - - [14592, 1152, 1, 384] - - [254, 78.634] - - - [10752, 10752, 1, 384] - - [254, 89.21] - - - [5376, 4992, 1, 384] - - [264, 85.527] - - - [24576, 7296, 1, 384] - - [254, 87.228] - - - [19200, 7296, 1, 384] - - [256, 89.94] - - - [18048, 4224, 1, 384] - - [252, 88.169] - - - [22272, 8832, 1, 384] - - [252, 90.167] - - - [14208, 13440, 1, 384] - - [264, 90.637] - - - [24192, 1152, 1, 384] - - [252, 83.698] - - - [16128, 1153, 1, 384] - - [258, 72.093] - - - [4992, 4993, 1, 384] - - [254, 79.64] - - - [14208, 3072, 1, 384] - - [254, 85.44] - - - [24576, 8064, 1, 384] - - [249, 87.204] - - - [1920, 1537, 1, 384] - - [261, 55.069] - - - [3072, 1536, 1, 384] - - [262, 63.324] - - - [16896, 3072, 1, 384] - - [252, 87.3] - - - [13824, 1152, 1, 384] - - [256, 81.264] - - - [19968, 13441, 1, 384] - - [254, 89.301] - - - [29184, 13440, 1, 384] - - [256, 90.843] - - - [9216, 8832, 1, 384] - - [264, 88.217] - - - [19968, 1153, 1, 384] - - [263, 73.427] - - - [9216, 3072, 1, 384] - - [254, 84.6] - - - [16128, 8832, 1, 384] - - [264, 89.401] - - - [8064, 1153, 1, 384] - - [286, 67.1] - - - [25728, 13441, 1, 384] - - [263, 89.082] - - - [26496, 12672, 1, 384] - - [263, 90.847] - - - [10368, 2688, 1, 384] - - [263, 83.77] - - - [17664, 13440, 1, 384] - - [280, 90.735] - - - [25728, 13440, 1, 384] - - [263, 90.911] - - - [6144, 3072, 1, 384] - - [252, 80.518] - - - [3840, 2304, 1, 384] - - [276, 76.562] - - - [11136, 11137, 1, 384] - - [252, 88.387] - - - [29568, 3072, 1, 384] - - [278, 87.919] - - - [18816, 1920, 1, 384] - - [254, 83.941] - - - [8448, 768, 1, 384] - - [267, 67.489] - - - [5376, 3072, 1, 384] - - [254, 77.57] - - - [27648, 7296, 1, 384] - - [264, 90.303] - - - [29184, 15360, 1, 384] - - [254, 90.741] - - - [7296, 7297, 1, 384] - - [252, 84.469] - - - [16384, 3072, 1, 256] - - [249, 62.582] - - - [20992, 7168, 1, 256] - - [254, 73.905] - - - [23552, 3584, 1, 256] - - [252, 73.921] - - - [7168, 1280, 1, 256] - - [256, 57.663] - - - [18224, 256, 1, 256] - - [267, 46.948] - - - [15360, 128, 1, 384] - - [284, 37.498] - - - [14592, 128, 1, 384] - - [266, 35.35] - - - [39552, 512, 1, 384] - - [271, 79.797] - - - [39552, 23297, 1, 384] - - [264, 89.357] - - - [36864, 2048, 1, 384] - - [264, 82.377] - - - [43776, 512, 1, 384] - - [253, 79.402] - - - [42240, 4096, 1, 384] - - [263, 90.427] - - - [35328, 1024, 1, 384] - - [285, 83.966] - - - [38784, 4096, 1, 384] - - [263, 89.674] - - - [42240, 8192, 1, 384] - - [280, 91.007] - - - [38016, 4096, 1, 384] - - [285, 89.941] - - - [35328, 19457, 1, 384] - - [259, 87.338] - - - [38400, 4096, 1, 384] - - [271, 88.905] - - - [36480, 2048, 1, 384] - - [271, 87.979] - - - [36864, 20609, 1, 384] - - [264, 88.091] - - - [38016, 2048, 1, 384] - - [271, 87.947] - - - [35328, 384, 1, 384] - - [256, 76.791] - - - [39168, 512, 1, 384] - - [254, 79.257] - - - [35328, 2048, 1, 384] - - [285, 87.335] - - - [40704, 384, 1, 384] - - [252, 79.014] - - - [41472, 25217, 1, 384] - - [256, 89.363] - - - [37632, 512, 1, 384] - - [263, 81.473] - - - [42240, 26369, 1, 384] - - [254, 89.63] - - - [44160, 8192, 1, 384] - - [285, 90.909] - - - [43392, 27137, 1, 384] - - [252, 88.823] - - - [41472, 25601, 1, 384] - - [259, 87.287] - - - [43392, 8192, 1, 384] - - [285, 90.761] - - - [36480, 4096, 1, 384] - - [263, 90.149] - - - [41088, 4096, 1, 384] - - [301, 88.437] - - - [35328, 19073, 1, 384] - - [256, 89.484] - - - [36864, 8192, 1, 384] - - [259, 87.044] - - - [38016, 21761, 1, 384] - - [252, 89.316] - - - [39552, 8192, 1, 384] - - [285, 90.847] - - - [41856, 384, 1, 384] - - [263, 80.661] - - - [43008, 26753, 1, 384] - - [281, 88.861] - - - [39168, 384, 1, 384] - - [252, 76.955] - - - [44544, 4096, 1, 384] - - [301, 88.698] - - - [43008, 2048, 1, 384] - - [301, 86.187] - - - [36864, 512, 1, 384] - - [280, 80.085] - - - [43392, 4096, 1, 384] - - [263, 90.249] - - - [39936, 23681, 1, 384] - - [252, 89.105] - - - [44544, 512, 1, 384] - - [254, 82.624] - - - [41856, 25985, 1, 384] - - [254, 89.011] - - - [38400, 1024, 1, 384] - - [280, 86.389] - - - [42624, 26753, 1, 384] - - [278, 83.126] - - - [37248, 2048, 1, 384] - - [285, 87.981] - - - [43392, 27521, 1, 384] - - [252, 89.004] - - - [42624, 512, 1, 384] - - [254, 79.899] - - - [40704, 2048, 1, 384] - - [280, 88.243] - - - [39936, 4096, 1, 384] - - [288, 87.649] - - - [41088, 8192, 1, 384] - - [301, 89.419] - - - [40320, 4096, 1, 384] - - [285, 90.213] - - - [39552, 4096, 1, 384] - - [285, 90.197] - - - [39936, 24065, 1, 384] - - [264, 88.775] - - - [38016, 1024, 1, 384] - - [263, 85.94] - - - [40704, 24833, 1, 384] - - [264, 88.984] - - - [41856, 4096, 1, 384] - - [263, 89.878] - - - [39552, 2048, 1, 384] - - [285, 89.065] - - - [36480, 8192, 1, 384] - - [285, 90.778] - - - [40704, 8192, 1, 384] - - [280, 90.642] - - - [35328, 512, 1, 384] - - [271, 77.624] - - - [36096, 1024, 1, 384] - - [320, 83.824] - - - [38784, 22529, 1, 384] - - [259, 87.014] - - - [41472, 4096, 1, 384] - - [320, 88.798] - - - [38784, 2048, 1, 384] - - [280, 87.715] - - - [41088, 24833, 1, 384] - - [281, 88.119] - - - [43392, 512, 1, 384] - - [254, 81.03] - - - [42624, 4096, 1, 384] - - [252, 87.761] - - - [44160, 512, 1, 384] - - [285, 82.101] - - - [38016, 512, 1, 384] - - [252, 82.165] - - - [43776, 2048, 1, 384] - - [288, 86.412] - - - [38400, 384, 1, 384] - - [263, 75.246] - - - [35712, 19457, 1, 384] - - [259, 86.829] - - - [41472, 1024, 1, 384] - - [271, 86.378] - - - [40704, 1024, 1, 384] - - [271, 84.888] - - - [40704, 4096, 1, 384] - - [285, 90.082] - - - [39936, 8192, 1, 384] - - [301, 88.219] - - - [38784, 8192, 1, 384] - - [263, 90.697] - - - [35712, 4096, 1, 384] - - [285, 89.523] - - - [38784, 22913, 1, 384] - - [254, 89.099] - - - [37248, 384, 1, 384] - - [263, 79.868] - - - [40704, 24449, 1, 384] - - [254, 89.212] - - - [44160, 27905, 1, 384] - - [252, 88.95] - - - [36480, 384, 1, 384] - - [271, 78.501] - - - [44544, 28289, 1, 384] - - [259, 89.099] - - - [36096, 4096, 1, 384] - - [301, 87.498] - - - [35712, 384, 1, 384] - - [256, 77.148] - - - [41088, 25217, 1, 384] - - [259, 88.117] - - - [39168, 22913, 1, 384] - - [256, 89.058] - - - [41088, 512, 1, 384] - - [251, 79.633] - - - [39168, 2048, 1, 384] - - [263, 88.606] - - - [35712, 512, 1, 384] - - [256, 78.178] - - - [43008, 8192, 1, 384] - - [281, 87.95] - - - [41088, 1024, 1, 384] - - [266, 83.173] - - - [44544, 384, 1, 384] - - [263, 78.949] - - - [43776, 8192, 1, 384] - - [288, 89.499] - - - [37632, 21377, 1, 384] - - [264, 89.33] - - - [43776, 4096, 1, 384] - - [301, 88.614] - - - [39552, 1024, 1, 384] - - [285, 85.957] - - - [38400, 22529, 1, 384] - - [281, 87.23] - - - [39168, 1024, 1, 384] - - [280, 85.143] - - - [40320, 1024, 1, 384] - - [285, 87.081] - - - [24192, 2048, 1, 384] - - [285, 86.935] - - - [21888, 4096, 1, 384] - - [251, 84.551] - - - [30336, 4096, 1, 384] - - [285, 89.998] - - - [2304, 1793, 1, 384] - - [282, 55.398] - - - [9216, 1024, 1, 384] - - [267, 70.126] - - - [27264, 11393, 1, 384] - - [252, 88.608] - - - [18816, 1024, 1, 384] - - [263, 81.935] - - - [31104, 14849, 1, 384] - - [264, 89.596] - - - [11136, 1024, 1, 384] - - [280, 73.552] - - - [9216, 512, 1, 384] - - [262, 61.98] - - - [17664, 512, 1, 384] - - [280, 67.758] - - - [15744, 8065, 1, 384] - - [264, 88.222] - - - [24192, 1024, 1, 384] - - [271, 84.078] - - - [21888, 2048, 1, 384] - - [252, 80.794] - - - [13056, 1024, 1, 384] - - [263, 76.455] - - - [10752, 6785, 1, 384] - - [252, 86.162] - - - [15360, 512, 1, 384] - - [261, 68.193] - - - [30720, 14465, 1, 384] - - [252, 89.054] - - - [34176, 1024, 1, 384] - - [263, 84.779] - - - [17664, 384, 1, 384] - - [267, 69.638] - - - [18048, 10369, 1, 384] - - [264, 88.443] - - - [33408, 4096, 1, 384] - - [280, 89.946] - - - [18816, 512, 1, 384] - - [250, 71.129] - - - [13824, 2048, 1, 384] - - [271, 84.822] - - - [4992, 1024, 1, 384] - - [290, 67.094] - - - [16512, 1024, 1, 384] - - [334, 72.635] - - - [29184, 1024, 1, 384] - - [263, 84.597] - - - [6528, 384, 1, 384] - - [254, 47.136] - - - [2688, 1153, 1, 384] - - [254, 57.037] - - - [20736, 384, 1, 384] - - [252, 68.633] - - - [26112, 8192, 1, 384] - - [271, 90.445] - - - [24192, 512, 1, 384] - - [280, 78.806] - - - [32256, 8192, 1, 384] - - [264, 90.325] - - - [1920, 1409, 1, 384] - - [262, 50.649] - - - [9216, 5633, 1, 384] - - [264, 84.683] - - - [28800, 8192, 1, 384] - - [285, 90.586] - - - [23040, 15361, 1, 384] - - [281, 87.18] - - - [22272, 14209, 1, 384] - - [285, 89.567] - - - [32256, 16001, 1, 384] - - [254, 89.634] - - - [32640, 384, 1, 384] - - [254, 74.229] - - - [10368, 512, 1, 384] - - [261, 68.48] - - - [29568, 4096, 1, 384] - - [253, 88.237] - - - [8832, 384, 1, 384] - - [280, 61.953] - - - [10368, 6401, 1, 384] - - [254, 86.647] - - - [29568, 512, 1, 384] - - [271, 77.505] - - - [19200, 384, 1, 384] - - [254, 64.612] - - - [8448, 4481, 1, 384] - - [254, 83.925] - - - [26496, 10241, 1, 384] - - [285, 85.965] - - - [4992, 3457, 1, 384] - - [254, 80.037] - - - [15360, 1024, 1, 384] - - [263, 80.098] - - - [30720, 384, 1, 384] - - [254, 75.441] - - - [30720, 14849, 1, 384] - - [264, 88.592] - - - [5760, 384, 1, 384] - - [284, 42.14] - - - [12672, 384, 1, 384] - - [250, 63.627] - - - [10752, 7169, 1, 384] - - [320, 83.986] - - - [22272, 4096, 1, 384] - - [285, 89.256] - - - [12288, 512, 1, 384] - - [287, 65.525] - - - [19584, 512, 1, 384] - - [280, 73.656] - - - [5376, 3841, 1, 384] - - [254, 81.167] - - - [29952, 2048, 1, 384] - - [285, 86.831] - - - [9984, 512, 1, 384] - - [261, 66.79] - - - [24192, 8321, 1, 384] - - [285, 88.409] - - - [17280, 9601, 1, 384] - - [285, 88.482] - - - [16512, 8449, 1, 384] - - [259, 85.255] - - - [24576, 4096, 1, 384] - - [252, 82.017] - - - [33024, 16769, 1, 384] - - [259, 88.675] - - - [16128, 2048, 1, 384] - - [280, 83.245] - - - [28800, 512, 1, 384] - - [254, 75.38] - - - [27648, 11393, 1, 384] - - [254, 88.687] - - - [12672, 4096, 1, 384] - - [263, 87.969] - - - [11904, 384, 1, 384] - - [267, 60.364] - - - [24576, 8192, 1, 384] - - [259, 84.517] - - - [17664, 2048, 1, 384] - - [271, 84.228] - - - [9984, 6017, 1, 384] - - [264, 84.842] - - - [21120, 4096, 1, 384] - - [283, 88.935] - - - [19200, 11521, 1, 384] - - [264, 88.647] - - - [25728, 384, 1, 384] - - [261, 72.113] - - - [20736, 1024, 1, 384] - - [263, 83.369] - - - [34560, 8192, 1, 384] - - [252, 90.331] - - - [30336, 2048, 1, 384] - - [285, 87.643] - - - [19200, 2048, 1, 384] - - [280, 86.532] - - - [15744, 7681, 1, 384] - - [256, 87.392] - - - [10752, 384, 1, 384] - - [267, 56.107] - - - [24960, 384, 1, 384] - - [271, 70.646] - - - [8064, 512, 1, 384] - - [261, 55.111] - - - [24960, 4096, 1, 384] - - [280, 89.572] - - - [16512, 8833, 1, 384] - - [281, 85.805] - - - [23808, 1024, 1, 384] - - [271, 83.004] - - - [8448, 4865, 1, 384] - - [264, 83.744] - - - [29184, 4096, 1, 384] - - [271, 88.774] - - - [9984, 1024, 1, 384] - - [290, 75.292] - - - [14592, 6913, 1, 384] - - [254, 87.934] - - - [7296, 1024, 1, 384] - - [267, 65.639] - - - [26880, 10625, 1, 384] - - [271, 89.218] - - - [18048, 1024, 1, 384] - - [263, 79.175] - - - [12288, 2048, 1, 384] - - [254, 79.298] - - - [32256, 384, 1, 384] - - [271, 78.402] - - - [29952, 14081, 1, 384] - - [257, 89.172] - - - [19200, 11137, 1, 384] - - [274, 88.991] - - - [25728, 1024, 1, 384] - - [285, 83.763] - - - [21120, 13441, 1, 384] - - [274, 89.157] - - - [34560, 512, 1, 384] - - [254, 81.541] - - - [32640, 16769, 1, 384] - - [259, 87.83] - - - [12672, 1024, 1, 384] - - [280, 74.602] - - - [25344, 4096, 1, 384] - - [263, 89.495] - - - [24576, 1024, 1, 384] - - [259, 78.371] - - - [14208, 6145, 1, 384] - - [283, 85.307] - - - [34560, 1024, 1, 384] - - [280, 85.683] - - - [21504, 2048, 1, 384] - - [288, 84.642] - - - [14592, 4096, 1, 384] - - [285, 86.938] - - - [34176, 18305, 1, 384] - - [278, 89.147] - - - [30720, 8192, 1, 384] - - [301, 87.502] - - - [25728, 2048, 1, 384] - - [280, 86.383] - - - [27264, 2048, 1, 384] - - [280, 84.397] - - - [30336, 384, 1, 384] - - [263, 74.444] - - - [22656, 4096, 1, 384] - - [263, 89.102] - - - [33408, 384, 1, 384] - - [271, 73.364] - - - [22272, 512, 1, 384] - - [264, 73.8] - - - [32640, 512, 1, 384] - - [303, 72.894] - - - [24960, 1024, 1, 384] - - [285, 81.728] - - - [25344, 2048, 1, 384] - - [263, 87.652] - - - [28416, 12545, 1, 384] - - [256, 89.264] - - - [13824, 5761, 1, 384] - - [256, 85.86] - - - [9600, 2048, 1, 384] - - [285, 83.301] - - - [34944, 18689, 1, 384] - - [252, 89.205] - - - [26880, 384, 1, 384] - - [252, 74.9] - - - [29568, 8192, 1, 384] - - [280, 89.975] - - - [28032, 2048, 1, 384] - - [280, 88.256] - - - [6528, 2945, 1, 384] - - [254, 75.792] - - - [5376, 512, 1, 384] - - [286, 51.87] - - - [21504, 384, 1, 384] - - [271, 70.928] - - - [3840, 1921, 1, 384] - - [276, 64.843] - - - [28416, 2048, 1, 384] - - [285, 87.239] - - - [16128, 384, 1, 384] - - [299, 62.837] - - - [2688, 1793, 1, 384] - - [276, 63.259] - - - [13440, 1024, 1, 384] - - [263, 78.549] - - - [34560, 18689, 1, 384] - - [264, 89.298] - - - [13056, 2048, 1, 384] - - [280, 84.795] - - - [25344, 9089, 1, 384] - - [278, 88.616] - - - [26880, 512, 1, 384] - - [285, 78.239] - - - [21888, 1024, 1, 384] - - [252, 78.92] - - - [16896, 4096, 1, 384] - - [263, 87.888] - - - [26496, 4096, 1, 384] - - [280, 89.126] - - - [24576, 512, 1, 384] - - [264, 78.16] - - - [6528, 512, 1, 384] - - [280, 61.445] - - - [26112, 2048, 1, 384] - - [280, 87.556] - - - [15744, 2048, 1, 384] - - [285, 86.199] - - - [11136, 2048, 1, 384] - - [285, 83.043] - - - [30720, 1024, 1, 384] - - [280, 84.187] - - - [31104, 2048, 1, 384] - - [263, 87.584] - - - [25344, 384, 1, 384] - - [271, 71.672] - - - [16128, 4096, 1, 384] - - [285, 87.503] - - - [4224, 2305, 1, 384] - - [261, 71.826] - - - [6144, 512, 1, 384] - - [262, 58.515] - - - [15360, 384, 1, 384] - - [261, 61.467] - - - [33792, 384, 1, 384] - - [280, 73.841] - - - [34176, 17921, 1, 384] - - [251, 88.898] - - - [34176, 8192, 1, 384] - - [254, 90.199] - - - [7680, 4097, 1, 384] - - [271, 82.31] - - - [9984, 2048, 1, 384] - - [271, 80.888] - - - [30336, 1024, 1, 384] - - [280, 83.496] - - - [13440, 5377, 1, 384] - - [256, 85.376] - - - [31872, 4096, 1, 384] - - [280, 89.55] - - - [8064, 384, 1, 384] - - [285, 57.231] - - - [16896, 1024, 1, 384] - - [271, 80.371] - - - [33792, 4096, 1, 384] - - [301, 87.283] - - - [29952, 1024, 1, 384] - - [263, 82.722] - - - [9984, 384, 1, 384] - - [267, 52.346] - - - [11904, 7937, 1, 384] - - [254, 87.153] - - - [32640, 16385, 1, 384] - - [259, 85.468] - - - [29952, 13697, 1, 384] - - [264, 89.409] - - - [24960, 512, 1, 384] - - [280, 73.643] - - - [10752, 512, 1, 384] - - [261, 58.246] - - - [32256, 16385, 1, 384] - - [281, 85.964] - - - [27648, 512, 1, 384] - - [256, 79.871] - - - [13440, 512, 1, 384] - - [280, 71.196] - - - [29184, 2048, 1, 384] - - [280, 86.734] - - - [33408, 17153, 1, 384] - - [254, 89.285] - - - [33792, 1024, 1, 384] - - [271, 83.815] - - - [4224, 1024, 1, 384] - - [261, 58.094] - - - [9600, 5633, 1, 384] - - [285, 85.785] - - - [7680, 1024, 1, 384] - - [267, 68.745] - - - [11904, 512, 1, 384] - - [254, 63.746] - - - [16128, 8449, 1, 384] - - [256, 88.103] - - - [32256, 1024, 1, 384] - - [280, 84.016] - - - [11136, 7169, 1, 384] - - [289, 84.227] - - - [13440, 4096, 1, 384] - - [280, 87.546] - - - [27264, 1024, 1, 384] - - [254, 82.207] - - - [30720, 512, 1, 384] - - [271, 80.004] - - - [20736, 4096, 1, 384] - - [280, 88.305] - - - [24960, 2048, 1, 384] - - [271, 86.756] - - - [10368, 384, 1, 384] - - [261, 54.061] - - - [23040, 4096, 1, 384] - - [285, 88.538] - - - [20736, 512, 1, 384] - - [264, 77.072] - - - [23040, 1024, 1, 384] - - [263, 80.485] - - - [11136, 512, 1, 384] - - [261, 60.033] - - - [19584, 11521, 1, 384] - - [254, 89.051] - - - [25728, 4096, 1, 384] - - [280, 89.093] - - - [26880, 4096, 1, 384] - - [280, 88.944] - - - [2304, 1409, 1, 384] - - [328, 59.428] - - - [25344, 9473, 1, 384] - - [252, 89.006] - - - [18048, 2048, 1, 384] - - [280, 85.495] - - - [31104, 8192, 1, 384] - - [254, 90.454] - - - [7296, 512, 1, 384] - - [311, 50.843] - - - [27264, 4096, 1, 384] - - [254, 87.94] - - - [15744, 4096, 1, 384] - - [271, 88.574] - - - [5760, 2177, 1, 384] - - [262, 71.734] - - - [28416, 8192, 1, 384] - - [280, 90.413] - - - [7296, 3713, 1, 384] - - [254, 80.855] - - - [19584, 2048, 1, 384] - - [280, 85.165] - - - [33024, 4096, 1, 384] - - [280, 88.724] - - - [26496, 8192, 1, 384] - - [263, 90.525] - - - [11520, 7937, 1, 384] - - [252, 87.554] - - - [8064, 4481, 1, 384] - - [254, 82.664] - - - [31104, 512, 1, 384] - - [254, 80.719] - - - [29184, 384, 1, 384] - - [263, 72.718] - - - [26112, 512, 1, 384] - - [252, 76.532] - - - [8448, 384, 1, 384] - - [260, 59.829] - - - [4608, 3073, 1, 384] - - [254, 72.603] - - - [34944, 512, 1, 384] - - [256, 82.01] - - - [33024, 2048, 1, 384] - - [280, 87.021] - - - [17280, 384, 1, 384] - - [311, 68.205] - - - [14208, 2048, 1, 384] - - [280, 82.778] - - - [8064, 4097, 1, 384] - - [283, 81.905] - - - [33024, 384, 1, 384] - - [263, 70.222] - - - [23424, 15361, 1, 384] - - [283, 85.936] - - - [32256, 512, 1, 384] - - [254, 77.446] - - - [5760, 512, 1, 384] - - [254, 54.858] - - - [31488, 4096, 1, 384] - - [285, 89.699] - - - [18816, 11137, 1, 384] - - [252, 89.01] - - - [31872, 1024, 1, 384] - - [263, 83.409] - - - [13056, 4096, 1, 384] - - [271, 87.556] - - - [12288, 1024, 1, 384] - - [285, 78.16] - - - [21888, 512, 1, 384] - - [256, 71.776] - - - [8448, 2048, 1, 384] - - [285, 80.242] - - - [10368, 6785, 1, 384] - - [252, 86.949] - - - [31104, 1024, 1, 384] - - [280, 85.253] - - - [34176, 4096, 1, 384] - - [285, 89.32] - - - [18432, 512, 1, 384] - - [250, 69.917] - - - [9984, 6401, 1, 384] - - [254, 85.589] - - - [22656, 2048, 1, 384] - - [280, 87.419] - - - [1920, 1025, 1, 384] - - [262, 37.494] - - - [16512, 2048, 1, 384] - - [288, 80.148] - - - [32256, 4096, 1, 384] - - [253, 88.284] - - - [21120, 512, 1, 384] - - [252, 78.172] - - - [8064, 2048, 1, 384] - - [256, 76.349] - - - [23808, 2048, 1, 384] - - [271, 85.843] - - - [19968, 4096, 1, 384] - - [285, 88.078] - - - [25344, 1024, 1, 384] - - [263, 82.986] - - - [20352, 12673, 1, 384] - - [252, 89.358] - - - [32640, 8192, 1, 384] - - [259, 88.408] - - - [22656, 1024, 1, 384] - - [285, 84.226] - - - [17664, 4096, 1, 384] - - [285, 87.84] - - - [28032, 11777, 1, 384] - - [285, 89.285] - - - [33792, 512, 1, 384] - - [280, 80.242] - - - [9216, 384, 1, 384] - - [288, 64.444] - - - [14208, 1024, 1, 384] - - [263, 75.175] - - - [4992, 384, 1, 384] - - [271, 36.602] - - - [3456, 1537, 1, 384] - - [261, 56.34] - - - [24576, 8321, 1, 384] - - [252, 83.293] - - - [13440, 384, 1, 384] - - [250, 66.827] - - - [6912, 1024, 1, 384] - - [267, 73.539] - - - [3840, 2305, 1, 384] - - [261, 66.085] - - - [7680, 512, 1, 384] - - [282, 53.268] - - - [23808, 4096, 1, 384] - - [271, 88.714] - - - [29184, 512, 1, 384] - - [254, 76.749] - - - [28032, 512, 1, 384] - - [263, 80.796] - - - [18048, 9985, 1, 384] - - [254, 88.583] - - - [27648, 2048, 1, 384] - - [301, 85.281] - - - [19584, 1024, 1, 384] - - [280, 79.278] - - - [6912, 384, 1, 384] - - [284, 49.908] - - - [24960, 8192, 1, 384] - - [280, 90.373] - - - [11904, 2048, 1, 384] - - [285, 83.021] - - - [12288, 4609, 1, 384] - - [254, 83.753] - - - [29184, 8192, 1, 384] - - [256, 90.304] - - - [23424, 1024, 1, 384] - - [280, 81.746] - - - [25728, 512, 1, 384] - - [252, 75.456] - - - [9600, 1024, 1, 384] - - [280, 72.644] - - - [18048, 4096, 1, 384] - - [280, 87.568] - - - [24192, 8192, 1, 384] - - [280, 90.674] - - - [6144, 384, 1, 384] - - [290, 44.949] - - - [6528, 1024, 1, 384] - - [262, 69.453] - - - [19968, 1024, 1, 384] - - [280, 80.777] - - - [20736, 2048, 1, 384] - - [285, 86.572] - - - [3072, 1153, 1, 384] - - [261, 48.324] - - - [14592, 384, 1, 384] - - [267, 58.606] - - - [29184, 13313, 1, 384] - - [281, 86.922] - - - [21888, 13825, 1, 384] - - [286, 85.031] - - - [4608, 1024, 1, 384] - - [262, 63.474] - - - [14976, 4096, 1, 384] - - [285, 86.86] - - - [24192, 16129, 1, 384] - - [285, 89.733] - - - [27648, 4096, 1, 384] - - [288, 87.529] - - - [33024, 8192, 1, 384] - - [289, 89.435] - - - [6528, 128, 1, 256] - - [335, 37.791] - - - [7808, 128, 1, 256] - - [336, 35.357] - - - [9088, 128, 1, 384] - - [336, 46.368] - - - [7808, 128, 1, 384] - - [337, 40.491] - - - [12928, 128, 1, 384] - - [338, 54.014] - - - [12928, 128, 1, 256] - - [339, 45.988] - - - [128, 128, 1, 256] - - [340, 1.074] - - - [9088, 128, 1, 256] - - [341, 40.41] - - - [1408, 128, 1, 384] - - [342, 13.439] - - - [2688, 128, 1, 256] - - [343, 20.246] - - - [11648, 128, 1, 256] - - [344, 42.668] - - - [1408, 128, 1, 256] - - [345, 11.358] - - - [10368, 128, 1, 256] - - [336, 43.982] - - - [128, 128, 1, 384] - - [346, 1.211] - - - [5248, 128, 1, 256] - - [336, 31.75] - - - [11648, 128, 1, 384] - - [347, 50.278] - - - [384, 3072, 1, 384] - - [336, 48.474] - - - [3968, 128, 1, 384] - - [335, 28.344] - - - [5248, 128, 1, 384] - - [335, 36.66] - - - [3968, 128, 1, 256] - - [337, 24.407] - - - [384, 768, 1, 384] - - [337, 21.71] - - - [768, 1152, 1, 384] - - [336, 46.95] - - - [1536, 768, 1, 384] - - [348, 46.914] - - - [256, 512, 1, 256] - - [349, 8.398] - - - [512, 2048, 1, 256] - - [348, 37.096] - - - [768, 1024, 1, 256] - - [350, 36.135] - - - [256, 2048, 1, 256] - - [341, 25.406] - - - [3584, 257, 1, 256] - - [341, 33.197] - - - [256, 3072, 1, 256] - - [350, 36.28] - - - [512, 3072, 1, 256] - - [344, 44.68] - - - [512, 768, 1, 256] - - [335, 22.789] - - - [1792, 768, 1, 256] - - [351, 40.595] - - - [3328, 257, 1, 256] - - [337, 31.019] - - - [768, 2048, 1, 256] - - [339, 44.352] - - - [1024, 1280, 1, 256] - - [352, 43.314] - - - [13825, 128, 1, 128] - - [353, 26.713] - - - [6017, 128, 1, 256] - - [354, 30.524] - - - [2305, 128, 1, 128] - - [355, 10.976] - - - [8833, 128, 1, 128] - - [356, 23.452] - - - [641, 128, 1, 128] - - [357, 3.197] - - - [9217, 128, 1, 128] - - [358, 23.704] - - - [2177, 128, 1, 128] - - [345, 10.367] - - - [13057, 128, 1, 256] - - [347, 37.574] - - - [1793, 128, 1, 128] - - [359, 8.538] - - - [4609, 128, 1, 256] - - [360, 24.923] - - - [10625, 128, 1, 128] - - [361, 23.134] - - - [12545, 128, 1, 256] - - [362, 36.242] - - - [5633, 128, 1, 128] - - [363, 18.479] - - - [641, 128, 1, 256] - - [360, 4.731] - - - [9601, 128, 1, 128] - - [364, 24.866] - - - [13697, 128, 1, 256] - - [365, 39.415] - - - [9601, 128, 1, 256] - - [366, 35.612] - - - [9985, 128, 1, 128] - - [367, 25.327] - - - [2689, 128, 1, 256] - - [345, 18.463] - - - [4993, 128, 1, 128] - - [345, 19.814] - - - [6913, 128, 1, 256] - - [368, 33.681] - - - [6785, 128, 1, 128] - - [359, 22.258] - - - [7169, 128, 1, 256] - - [369, 27.705] - - - [11905, 128, 1, 256] - - [364, 35.504] - - - [1409, 128, 1, 128] - - [370, 7.027] - - - [12673, 128, 1, 128] - - [366, 26.276] - - - [1409, 128, 1, 256] - - [359, 10.4] - - - [7297, 128, 1, 128] - - [364, 20.023] - - - [10753, 128, 1, 256] - - [371, 32.662] - - - [7937, 128, 1, 128] - - [363, 21.46] - - - [11393, 128, 1, 128] - - [372, 24.445] - - - [12161, 128, 1, 256] - - [364, 36.415] - - - [8449, 128, 1, 128] - - [356, 22.514] - - - [10241, 128, 1, 256] - - [373, 35.572] - - - [6913, 128, 1, 128] - - [359, 21.806] - - - [4993, 128, 1, 256] - - [374, 26.999] - - - [6401, 128, 1, 256] - - [375, 32.584] - - - [13057, 128, 1, 128] - - [376, 26.404] - - - [2945, 128, 1, 128] - - [377, 13.846] - - - [3713, 128, 1, 256] - - [345, 20.452] - - - [10753, 128, 1, 128] - - [378, 23.073] - - - [3841, 128, 1, 128] - - [345, 15.578] - - - [12929, 128, 1, 128] - - [379, 26.806] - - - [12545, 128, 1, 128] - - [376, 25.865] - - - [11777, 128, 1, 256] - - [371, 34.359] - - - [11777, 128, 1, 128] - - [380, 24.695] - - - [5377, 128, 1, 128] - - [381, 17.639] - - - [8065, 128, 1, 256] - - [366, 31.167] - - - [6145, 128, 1, 128] - - [375, 19.981] - - - [5633, 128, 1, 256] - - [382, 28.675] - - - [4865, 128, 1, 128] - - [360, 19.203] - - - [385, 128, 1, 256] - - [355, 2.856] - - - [3841, 128, 1, 256] - - [374, 21.236] - - - [8833, 128, 1, 256] - - [379, 33.694] - - - [4225, 128, 1, 128] - - [355, 17.135] - - - [11009, 128, 1, 256] - - [379, 33.235] - - - [385, 128, 1, 128] - - [355, 1.973] - - - [9473, 128, 1, 256] - - [379, 35.137] - - - [5761, 128, 1, 128] - - [359, 19.155] - - - [11905, 128, 1, 128] - - [376, 25.324] - - - [4097, 128, 1, 256] - - [383, 22.568] - - - [9089, 128, 1, 256] - - [384, 34.055] - - - [10369, 128, 1, 256] - - [379, 37.425] - - - [6401, 128, 1, 128] - - [385, 20.814] - - - [7553, 128, 1, 256] - - [382, 29.265] - - - [8321, 128, 1, 128] - - [378, 22.334] - - - [1153, 128, 1, 128] - - [359, 5.75] - - - [1025, 128, 1, 128] - - [355, 5.182] - - - [4865, 128, 1, 256] - - [359, 26.404] - - - [9857, 128, 1, 256] - - [386, 36.107] - - - [11521, 128, 1, 128] - - [375, 24.366] - - - [8449, 128, 1, 256] - - [367, 32.146] - - - [4097, 128, 1, 128] - - [345, 16.708] - - - [12161, 128, 1, 128] - - [378, 25.868] - - - [1921, 128, 1, 256] - - [355, 13.314] - - - [9985, 128, 1, 256] - - [367, 36.306] - - - [7937, 128, 1, 256] - - [378, 30.434] - - - [9857, 128, 1, 128] - - [358, 25.528] - - - [13825, 128, 1, 256] - - [365, 39.476] - - - [9089, 128, 1, 128] - - [358, 24.22] - - - [6785, 128, 1, 256] - - [368, 33.387] - - - [5249, 128, 1, 256] - - [378, 27.094] - - - [7681, 128, 1, 256] - - [366, 29.376] - - - [3329, 128, 1, 128] - - [355, 15.075] - - - [11137, 128, 1, 256] - - [378, 34.039] - - - [1153, 128, 1, 256] - - [359, 8.596] - - - [13697, 128, 1, 128] - - [376, 27.251] - - - [3073, 128, 1, 128] - - [377, 14.356] - - - [257, 128, 1, 128] - - [370, 1.282] - - - [4481, 128, 1, 256] - - [387, 24.683] - - - [6017, 128, 1, 128] - - [388, 19.916] - - - [7297, 128, 1, 256] - - [354, 28.273] - - - [7553, 128, 1, 128] - - [389, 20.572] - - - [11393, 128, 1, 256] - - [364, 34.324] - - - [11521, 128, 1, 256] - - [390, 34.428] - - - [12929, 128, 1, 256] - - [386, 37.424] - - - [13313, 128, 1, 128] - - [391, 25.928] - - - [2561, 128, 1, 128] - - [392, 12.04] - - - [1537, 128, 1, 128] - - [360, 7.718] - - - [12289, 128, 1, 256] - - [371, 35.432] - - - [4225, 128, 1, 256] - - [345, 23.101] - - - [2945, 128, 1, 256] - - [345, 20.127] - - - [6529, 128, 1, 128] - - [360, 21.514] - - - [6145, 128, 1, 256] - - [378, 30.96] - - - [8705, 128, 1, 256] - - [367, 32.533] - - - [384, 128, 1, 256] - - [393, 3.132] - - - [12673, 128, 1, 256] - - [338, 36.9] - - - [8321, 128, 1, 256] - - [390, 31.823] - - - [5249, 128, 1, 128] - - [359, 17.854] - - - [13441, 128, 1, 256] - - [339, 38.454] - - - [5377, 128, 1, 256] - - [390, 27.466] - - - [11137, 128, 1, 128] - - [391, 24.107] - - - [7681, 128, 1, 128] - - [358, 20.692] - - - [7169, 128, 1, 128] - - [354, 19.671] - - - [11009, 128, 1, 128] - - [394, 23.76] - - - [13313, 128, 1, 256] - - [395, 37.292] - - - [4609, 128, 1, 128] - - [359, 18.29] - - - [5761, 128, 1, 256] - - [375, 29.529] - - - [2689, 128, 1, 128] - - [370, 12.405] - - - [8705, 128, 1, 128] - - [358, 22.785] - - - [10241, 128, 1, 128] - - [380, 25.624] - - - [3457, 128, 1, 128] - - [355, 15.372] - - - [12289, 128, 1, 128] - - [376, 25.127] - - - [2177, 128, 1, 256] - - [396, 15.018] - - - [4481, 128, 1, 128] - - [360, 18.173] - - - [8065, 128, 1, 128] - - [354, 21.886] - - - [3457, 128, 1, 256] - - [355, 22.482] - - - [6529, 128, 1, 256] - - [364, 32.895] - - - [10625, 128, 1, 256] - - [390, 32.474] - - - [9217, 128, 1, 256] - - [380, 33.105] - - - [13441, 128, 1, 128] - - [379, 27.181] - - - [3713, 128, 1, 128] - - [345, 15.058] - - - [10369, 128, 1, 128] - - [358, 26.301] - - - [1921, 128, 1, 128] - - [345, 8.975] - - - [9473, 128, 1, 128] - - [367, 24.62] - - - [1408, 897, 1, 128] - - [350, 29.971] - - - [640, 128, 1, 128] - - [355, 3.662] - - - [10496, 128, 1, 128] - - [341, 31.751] - - - [10880, 128, 1, 128] - - [341, 29.524] - - - [256, 129, 1, 128] - - [397, 1.465] - - - [128, 129, 1, 128] - - [346, 0.738] - - - [10624, 128, 1, 128] - - [336, 28.619] - - - [5248, 128, 1, 128] - - [392, 22.135] - - - [4608, 128, 1, 128] - - [343, 20.246] - - - [7936, 128, 1, 128] - - [335, 26.151] - - - [5504, 128, 1, 128] - - [350, 22.823] - - - [7040, 128, 1, 128] - - [337, 27.349] - - - [768, 512, 1, 128] - - [346, 16.196] - - - [2560, 128, 1, 128] - - [343, 13.997] - - - [2304, 128, 1, 128] - - [377, 12.504] - - - [9856, 128, 1, 128] - - [335, 30.957] - - - [1280, 128, 1, 128] - - [340, 7.324] - - - [640, 512, 1, 128] - - [340, 13.792] - - - [768, 129, 1, 128] - - [397, 4.395] - - - [13696, 128, 1, 128] - - [341, 33.364] - - - [11776, 128, 1, 128] - - [398, 29.869] - - - [3328, 512, 1, 128] - - [350, 32.644] - - - [11008, 128, 1, 128] - - [336, 29.228] - - - [1792, 512, 1, 128] - - [336, 24.27] - - - [384, 128, 1, 128] - - [343, 2.232] - - - [3968, 128, 1, 128] - - [340, 17.751] - - - [7168, 128, 1, 128] - - [399, 24.049] - - - [1664, 512, 1, 128] - - [335, 26.701] - - - [2048, 512, 1, 128] - - [400, 27.115] - - - [2304, 512, 1, 128] - - [336, 29.196] - - - [768, 257, 1, 128] - - [346, 8.688] - - - [1024, 513, 1, 128] - - [393, 18.031] - - - [3072, 512, 1, 128] - - [401, 30.642] - - - [9600, 128, 1, 128] - - [350, 30.281] - - - [13184, 128, 1, 128] - - [341, 32.655] - - - [2688, 128, 1, 128] - - [343, 14.172] - - - [1408, 512, 1, 128] - - [350, 23.224] - - - [13312, 128, 1, 128] - - [341, 31.595] - - - [6144, 128, 1, 128] - - [335, 25.055] - - - [3584, 128, 1, 128] - - [340, 16.958] - - - [4096, 128, 1, 128] - - [393, 18.547] - - - [8832, 128, 1, 128] - - [398, 28.344] - - - [8704, 128, 1, 128] - - [337, 28.178] - - - [12288, 128, 1, 128] - - [350, 30.85] - - - [13824, 128, 1, 128] - - [341, 32.393] - - - [9344, 128, 1, 128] - - [337, 29.6] - - - [9216, 128, 1, 128] - - [399, 28.702] - - - [256, 257, 1, 128] - - [393, 2.987] - - - [1280, 769, 1, 128] - - [350, 25.116] - - - [384, 257, 1, 128] - - [349, 4.378] - - - [1408, 769, 1, 128] - - [341, 27.266] - - - [1152, 513, 1, 128] - - [342, 19.814] - - - [2432, 512, 1, 128] - - [402, 30.686] - - - [3712, 128, 1, 128] - - [392, 16.913] - - - [12928, 128, 1, 128] - - [341, 32.348] - - - [1920, 512, 1, 128] - - [350, 25.535] - - - [2816, 128, 1, 128] - - [355, 14.953] - - - [256, 128, 1, 128] - - [393, 1.5] - - - [1152, 512, 1, 128] - - [393, 20.49] - - - [2688, 512, 1, 128] - - [352, 29.178] - - - [12800, 128, 1, 128] - - [348, 31.388] - - - [5760, 128, 1, 128] - - [336, 23.233] - - - [11392, 128, 1, 128] - - [341, 30.138] - - - [5632, 128, 1, 128] - - [336, 22.716] - - - [8320, 128, 1, 128] - - [336, 26.935] - - - [12672, 128, 1, 128] - - [341, 32.477] - - - [1152, 128, 1, 128] - - [346, 6.643] - - - [1024, 128, 1, 128] - - [343, 5.905] - - - [4992, 128, 1, 128] - - [340, 21.548] - - - [4352, 128, 1, 128] - - [340, 19.236] - - - [1536, 1025, 1, 128] - - [399, 29.099] - - - [7424, 128, 1, 128] - - [337, 24.684] - - - [1024, 512, 1, 128] - - [359, 18.435] - - - [640, 129, 1, 128] - - [393, 3.662] - - - [4736, 128, 1, 128] - - [343, 20.684] - - - [1280, 641, 1, 128] - - [335, 24.901] - - - [11136, 128, 1, 128] - - [335, 29.674] - - - [6400, 128, 1, 128] - - [350, 25.674] - - - [5376, 128, 1, 128] - - [335, 22.045] - - - [1024, 385, 1, 128] - - [349, 15.36] - - - [3840, 128, 1, 128] - - [342, 17.389] - - - [1664, 128, 1, 128] - - [342, 9.305] - - - [9728, 128, 1, 128] - - [336, 30.297] - - - [8448, 128, 1, 128] - - [398, 27.47] - - - [2176, 512, 1, 128] - - [348, 28.055] - - - [896, 257, 1, 128] - - [403, 9.692] - - - [13440, 128, 1, 128] - - [399, 32.958] - - - [768, 128, 1, 128] - - [343, 4.429] - - - [6528, 128, 1, 128] - - [336, 26.046] - - - [2816, 512, 1, 128] - - [335, 30.344] - - - [11520, 128, 1, 128] - - [348, 30.26] - - - [2048, 128, 1, 128] - - [343, 11.281] - - - [4480, 128, 1, 128] - - [343, 19.801] - - - [128, 128, 1, 128] - - [340, 0.75] - - - [6272, 128, 1, 128] - - [337, 25.437] - - - [2944, 128, 1, 128] - - [342, 15.411] - - - [7552, 128, 1, 128] - - [335, 24.997] - - - [7296, 128, 1, 128] - - [341, 24.703] - - - [6016, 128, 1, 128] - - [348, 24.399] - - - [512, 512, 1, 128] - - [342, 10.955] - - - [11264, 128, 1, 128] - - [404, 28.769] - - - [3456, 512, 1, 128] - - [335, 33.676] - - - [1408, 128, 1, 128] - - [393, 7.933] - - - [2560, 512, 1, 128] - - [348, 30.724] - - - [7808, 128, 1, 128] - - [350, 25.729] - - - [11904, 128, 1, 128] - - [341, 31.381] - - - [6656, 128, 1, 128] - - [337, 26.414] - - - [512, 385, 1, 128] - - [342, 8.547] - - - [13568, 128, 1, 128] - - [399, 32.943] - - - [9088, 128, 1, 128] - - [341, 29.292] - - - [12544, 128, 1, 128] - - [341, 31.709] - - - [13056, 128, 1, 128] - - [399, 32.667] - - - [6784, 128, 1, 128] - - [350, 26.778] - - - [12416, 128, 1, 128] - - [337, 31.601] - - - [1792, 128, 1, 128] - - [340, 10.02] - - - [7680, 128, 1, 128] - - [335, 25.42] - - - [3200, 128, 1, 128] - - [403, 16.517] - - - [896, 128, 1, 128] - - [340, 5.167] - - - [2944, 512, 1, 128] - - [350, 31.266] - - - [896, 385, 1, 128] - - [405, 13.814] - - - [8064, 128, 1, 128] - - [341, 26.572] - - - [10240, 128, 1, 128] - - [348, 30.977] - - - [1280, 512, 1, 128] - - [393, 22.362] - - - [13952, 128, 1, 128] - - [341, 33.545] - - - [3072, 128, 1, 128] - - [343, 16.313] - - - [9984, 128, 1, 128] - - [341, 30.964] - - - [3456, 128, 1, 128] - - [393, 17.354] - - - [1920, 128, 1, 128] - - [343, 10.498] - - - [384, 385, 1, 128] - - [406, 6.362] - - - [2432, 128, 1, 128] - - [407, 13.103] - - - [896, 512, 1, 128] - - [345, 16.743] - - - [10368, 128, 1, 128] - - [335, 31.493] - - - [3328, 128, 1, 128] - - [345, 17.059] - - - [4224, 128, 1, 128] - - [345, 18.782] - - - [5120, 128, 1, 128] - - [343, 21.595] - - - [6912, 128, 1, 128] - - [341, 27.43] - - - [11648, 128, 1, 128] - - [350, 30.816] - - - [1536, 897, 1, 128] - - [399, 27.683] - - - [8192, 128, 1, 128] - - [336, 26.405] - - - [8576, 128, 1, 128] - - [341, 27.764] - - - [2176, 128, 1, 128] - - [343, 11.724] - - - [12032, 128, 1, 128] - - [399, 30.73] - - - [1536, 128, 1, 128] - - [346, 8.857] - - - [10752, 128, 1, 128] - - [341, 29.285] - - - [1536, 512, 1, 128] - - [350, 24.918] - - - [9472, 128, 1, 128] - - [337, 29.878] - - - [4864, 128, 1, 128] - - [340, 20.996] - - - [5888, 128, 1, 128] - - [335, 24.145] - - - [8960, 128, 1, 128] - - [337, 28.755] - - - [12160, 128, 1, 128] - - [399, 31.828] - - - [512, 128, 1, 128] - - [405, 2.93] - - - [10112, 128, 1, 128] - - [337, 31.099] - - - [1152, 641, 1, 128] - - [336, 23.269] - - - [3200, 512, 1, 128] - - [350, 32.245] - - - [2816, 512, 1, 256] - - [408, 42.312] - - - [2560, 512, 1, 256] - - [352, 43.314] - - - [1280, 512, 1, 256] - - [348, 30.85] - - - [1536, 1024, 1, 256] - - [409, 44.029] - - - [1536, 512, 1, 256] - - [350, 35.992] - - - [1024, 512, 1, 256] - - [336, 24.986] - - - [2304, 512, 1, 256] - - [348, 41.352] - - - [3072, 512, 1, 256] - - [410, 44.68] - - - [1536, 768, 1, 256] - - [348, 40.855] - - - [1280, 768, 1, 256] - - [348, 35.43] - - - [512, 128, 1, 256] - - [340, 4.27] - - - [2049, 128, 1, 256] - - [407, 14.336] - - - [49, 128, 1, 256] - - [385, 0.348] - - - [1537, 128, 1, 256] - - [411, 11.345] - - - [257, 128, 1, 256] - - [385, 1.888] - - - [9728, 128, 1, 256] - - [348, 42.487] - - - [3840, 128, 1, 256] - - [341, 23.522] - - - [1280, 128, 1, 256] - - [340, 10.616] - - - [7168, 128, 1, 256] - - [335, 32.965] - - - [6656, 128, 1, 256] - - [348, 37.792] - - - [2561, 128, 1, 256] - - [345, 17.834] - - - [6912, 128, 1, 256] - - [335, 38.946] - - - [2048, 128, 1, 256] - - [343, 15.996] - - - [2304, 128, 1, 256] - - [343, 17.807] - - - [1536, 128, 1, 256] - - [393, 12.739] - - - [4864, 128, 1, 256] - - [335, 29.428] - - - [8448, 128, 1, 256] - - [335, 38.138] - - - [3072, 128, 1, 256] - - [343, 23.02] - - - [3329, 128, 1, 256] - - [345, 22.543] - - - [3328, 128, 1, 256] - - [337, 24.565] - - - [8960, 128, 1, 256] - - [336, 39.721] - - - [9216, 128, 1, 256] - - [412, 38.761] - - - [2817, 128, 1, 256] - - [377, 19.524] - - - [6400, 128, 1, 256] - - [398, 37.641] - - - [561, 128, 1, 256] - - [345, 3.87] - - - [2816, 128, 1, 256] - - [341, 21.21] - - - [3073, 128, 1, 256] - - [387, 20.073] - - - [2097, 128, 1, 256] - - [345, 14.071] - - - [768, 128, 1, 256] - - [393, 6.405] - - - [9984, 128, 1, 256] - - [341, 42.97] - - - [3584, 128, 1, 256] - - [336, 22.61] - - - [817, 128, 1, 256] - - [355, 5.636] - - - [5632, 128, 1, 256] - - [336, 33.525] - - - [9472, 128, 1, 256] - - [341, 41.248] - - - [2305, 128, 1, 256] - - [355, 15.9] - - - [1329, 128, 1, 256] - - [345, 9.168] - - - [5888, 128, 1, 256] - - [337, 34.907] - - - [7680, 128, 1, 256] - - [404, 33.944] - - - [4608, 128, 1, 256] - - [341, 28.462] - - - [2353, 128, 1, 256] - - [392, 15.789] - - - [5120, 128, 1, 256] - - [335, 30.601] - - - [769, 128, 1, 256] - - [411, 5.733] - - - [1792, 128, 1, 256] - - [345, 14.072] - - - [1073, 128, 1, 256] - - [345, 7.437] - - - [513, 128, 1, 256] - - [359, 3.806] - - - [4096, 128, 1, 256] - - [337, 25.3] - - - [7424, 128, 1, 256] - - [350, 34.356] - - - [4352, 128, 1, 256] - - [337, 26.77] - - - [1793, 128, 1, 256] - - [345, 12.725] - - - [8192, 128, 1, 256] - - [413, 35.992] - - - [1281, 128, 1, 256] - - [357, 9.551] - - - [305, 128, 1, 256] - - [396, 2.104] - - - [2560, 128, 1, 256] - - [343, 19.481] - - - [2609, 128, 1, 256] - - [345, 17.831] - - - [1585, 128, 1, 256] - - [377, 10.783] - - - [8704, 128, 1, 256] - - [336, 37.244] - - - [10240, 128, 1, 256] - - [414, 41.644] - - - [256, 128, 1, 256] - - [342, 2.135] - - - [1025, 128, 1, 256] - - [411, 7.604] - - - [2865, 128, 1, 256] - - [370, 19.051] - - - [5376, 128, 1, 256] - - [337, 32.001] - - - [1841, 128, 1, 256] - - [392, 12.41] - - - [7936, 128, 1, 256] - - [335, 36.158] - - - [6144, 128, 1, 256] - - [337, 35.292] - - - [1024, 128, 1, 256] - - [342, 8.54] - - - [5168, 256, 1, 256] - - [348, 41.239] - - - [768, 768, 1, 256] - - [335, 28.462] - - - [48, 49, 1, 256] - - [393, 0.141] - - - [2352, 256, 1, 256] - - [350, 28.229] - - - [816, 256, 1, 256] - - [342, 12.953] - - - [2096, 256, 1, 256] - - [350, 25.259] - - - [1072, 256, 1, 256] - - [343, 16.399] - - - [3376, 256, 1, 256] - - [335, 37.471] - - - [1280, 256, 1, 256] - - [343, 19.581] - - - [3072, 256, 1, 256] - - [336, 35.708] - - - [3840, 256, 1, 256] - - [415, 34.149] - - - [1024, 256, 1, 256] - - [343, 15.829] - - - [2816, 256, 1, 256] - - [336, 33.257] - - - [48, 256, 1, 256] - - [340, 0.779] - - - [5936, 256, 1, 256] - - [409, 41.044] - - - [2304, 256, 1, 256] - - [335, 28.11] - - - [560, 256, 1, 256] - - [342, 8.889] - - - [256, 49, 1, 256] - - [342, 0.778] - - - [512, 513, 1, 256] - - [343, 15.777] - - - [2608, 256, 1, 256] - - [341, 30.8] - - - [304, 49, 1, 256] - - [393, 0.914] - - - [4656, 256, 1, 256] - - [341, 39.388] - - - [4608, 256, 1, 256] - - [402, 39.897] - - - [768, 513, 1, 256] - - [335, 22.833] - - - [2048, 256, 1, 256] - - [350, 25.3] - - - [4096, 256, 1, 256] - - [413, 36.316] - - - [6400, 256, 1, 256] - - [410, 46.542] - - - [1280, 1280, 1, 256] - - [348, 47.358] - - - [304, 256, 1, 256] - - [343, 4.8] - - - [5120, 256, 1, 256] - - [416, 41.991] - - - [2048, 512, 1, 256] - - [401, 36.208] - - - [256, 257, 1, 256] - - [406, 4.192] - - - [768, 256, 1, 256] - - [342, 12.668] - - - [256, 256, 1, 256] - - [393, 4.319] - - - [4400, 256, 1, 256] - - [350, 37.764] - - - [1024, 769, 1, 256] - - [350, 35.338] - - - [5376, 256, 1, 256] - - [410, 40.594] - - - [6656, 256, 1, 256] - - [409, 47.239] - - - [3632, 256, 1, 256] - - [348, 32.495] - - - [816, 817, 1, 256] - - [348, 30.034] - - - [1024, 1025, 1, 256] - - [412, 35.087] - - - [512, 257, 1, 256] - - [393, 8.338] - - - [1024, 1024, 1, 256] - - [399, 36.317] - - - [5632, 256, 1, 256] - - [410, 42.857] - - - [560, 512, 1, 256] - - [336, 17.046] - - - [512, 305, 1, 256] - - [397, 9.788] - - - [6192, 256, 1, 256] - - [409, 42.027] - - - [2304, 768, 1, 256] - - [351, 48.821] - - - [512, 256, 1, 256] - - [393, 8.445] - - - [1328, 256, 1, 256] - - [335, 19.803] - - - [3120, 256, 1, 256] - - [350, 35.025] - - - [1024, 768, 1, 256] - - [348, 35.569] - - - [1024, 817, 1, 256] - - [348, 36.274] - - - [560, 305, 1, 256] - - [340, 10.423] - - - [1280, 1073, 1, 256] - - [417, 40.001] - - - [4352, 256, 1, 256] - - [350, 38.47] - - - [5680, 256, 1, 256] - - [409, 39.74] - - - [3328, 256, 1, 256] - - [341, 37.36] - - - [6704, 256, 1, 256] - - [408, 44.38] - - - [560, 561, 1, 256] - - [341, 18.392] - - - [1584, 256, 1, 256] - - [343, 23.152] - - - [1792, 256, 1, 256] - - [341, 22.324] - - - [512, 512, 1, 256] - - [403, 15.996] - - - [4912, 256, 1, 256] - - [336, 40.173] - - - [816, 768, 1, 256] - - [350, 29.143] - - - [5424, 256, 1, 256] - - [337, 38.221] - - - [3376, 512, 1, 256] - - [410, 44.698] - - - [6144, 256, 1, 256] - - [409, 44.901] - - - [1840, 256, 1, 256] - - [335, 22.542] - - - [768, 512, 1, 256] - - [335, 23.02] - - - [768, 561, 1, 256] - - [341, 24.478] - - - [5888, 256, 1, 256] - - [418, 43.899] - - - [1072, 817, 1, 256] - - [350, 30.424] - - - [2048, 768, 1, 256] - - [344, 44.901] - - - [4144, 256, 1, 256] - - [337, 36.091] - - - [1792, 512, 1, 256] - - [335, 33.381] - - - [1072, 1073, 1, 256] - - [335, 36.546] - - - [2864, 256, 1, 256] - - [337, 33.16] - - - [3888, 256, 1, 256] - - [337, 34.576] - - - [1536, 256, 1, 256] - - [343, 22.905] - - - [3328, 512, 1, 256] - - [408, 48.048] - - - [768, 769, 1, 256] - - [336, 28.381] - - - [304, 305, 1, 256] - - [346, 5.569] - - - [1280, 1025, 1, 256] - - [410, 38.019] - - - [6448, 256, 1, 256] - - [419, 43.664] - - - [3584, 256, 1, 256] - - [341, 33.276] - - - [6912, 256, 1, 256] - - [351, 48.589] - - - [1072, 1024, 1, 256] - - [337, 36.172] - - - [2560, 256, 1, 256] - - [350, 31.104] - - - [1328, 1073, 1, 256] - - [344, 38.13] - - - [1280, 1024, 1, 256] - - [352, 42.945] - - - [816, 561, 1, 256] - - [336, 21.462] - - - [768, 768, 1, 384] - - [335, 32.393] - - - [1152, 768, 1, 384] - - [336, 46.805] - - - [768, 769, 1, 384] - - [341, 32.23] - - - [1536, 1152, 1, 384] - - [409, 59.095] - - - [1152, 769, 1, 384] - - [398, 35.891] - - - [1152, 1153, 1, 384] - - [408, 46.328] - - - [1152, 1152, 1, 384] - - [399, 53.392] - - - [2304, 768, 1, 384] - - [409, 59.096] - - - [1920, 768, 1, 384] - - [344, 50.514] - - - [4864, 256, 1, 256] - - [420, 41.505] - - - [13441, 128, 1, 384] - - [421, 47.394] - - - [10753, 128, 1, 384] - - [422, 40.289] - - - [12289, 128, 1, 384] - - [423, 42.921] - - - [385, 128, 1, 384] - - [387, 3.33] - - - [11136, 128, 1, 384] - - [409, 48.351] - - - [13440, 128, 1, 384] - - [408, 55.421] - - - [1153, 128, 1, 384] - - [359, 10.092] - - - [6145, 128, 1, 384] - - [375, 37.382] - - - [4225, 128, 1, 384] - - [378, 26.504] - - - [1537, 128, 1, 384] - - [359, 13.779] - - - [8064, 128, 1, 384] - - [335, 42.016] - - - [3072, 128, 1, 384] - - [348, 28.226] - - - [3457, 128, 1, 384] - - [345, 26.58] - - - [5760, 128, 1, 384] - - [335, 39.611] - - - [8449, 128, 1, 384] - - [367, 38.339] - - - [2305, 128, 1, 384] - - [377, 18.904] - - - [11520, 128, 1, 384] - - [410, 49.823] - - - [11521, 128, 1, 384] - - [424, 42.238] - - - [6528, 128, 1, 384] - - [335, 44.071] - - - [768, 128, 1, 384] - - [337, 7.268] - - - [12672, 128, 1, 384] - - [338, 53.145] - - - [9216, 128, 1, 384] - - [352, 45.249] - - - [8448, 128, 1, 384] - - [348, 44.434] - - - [6144, 128, 1, 384] - - [341, 41.479] - - - [2689, 128, 1, 384] - - [345, 21.972] - - - [4224, 128, 1, 384] - - [337, 29.979] - - - [9601, 128, 1, 384] - - [367, 41.933] - - - [13056, 128, 1, 384] - - [425, 54.343] - - - [8065, 128, 1, 384] - - [354, 36.521] - - - [2304, 128, 1, 384] - - [335, 21.347] - - - [8833, 128, 1, 384] - - [378, 39.514] - - - [13824, 128, 1, 384] - - [426, 56.065] - - - [7680, 128, 1, 384] - - [352, 38.65] - - - [3840, 128, 1, 384] - - [336, 27.254] - - - [1920, 128, 1, 384] - - [341, 18.247] - - - [5761, 128, 1, 384] - - [364, 35.338] - - - [7681, 128, 1, 384] - - [427, 34.854] - - - [4608, 128, 1, 384] - - [350, 32.496] - - - [10369, 128, 1, 384] - - [386, 44.325] - - - [3841, 128, 1, 384] - - [378, 24.094] - - - [7296, 128, 1, 384] - - [337, 38.194] - - - [7297, 128, 1, 384] - - [375, 33.523] - - - [10752, 128, 1, 384] - - [371, 47.522] - - - [1536, 128, 1, 384] - - [393, 14.724] - - - [11137, 128, 1, 384] - - [428, 40.762] - - - [2688, 128, 1, 384] - - [341, 24.495] - - - [4609, 128, 1, 384] - - [390, 28.912] - - - [6529, 128, 1, 384] - - [375, 39.286] - - - [11905, 128, 1, 384] - - [428, 43.004] - - - [6912, 128, 1, 384] - - [341, 46.101] - - - [769, 128, 1, 384] - - [355, 6.811] - - - [12288, 128, 1, 384] - - [418, 52.227] - - - [9600, 128, 1, 384] - - [341, 48.868] - - - [13057, 128, 1, 384] - - [429, 46.04] - - - [10368, 128, 1, 384] - - [341, 51.018] - - - [12673, 128, 1, 384] - - [362, 45.262] - - - [9217, 128, 1, 384] - - [427, 38.95] - - - [4993, 128, 1, 384] - - [368, 31.057] - - - [9984, 128, 1, 384] - - [364, 49.458] - - - [6913, 128, 1, 384] - - [390, 40.93] - - - [8832, 128, 1, 384] - - [399, 45.694] - - - [3073, 128, 1, 384] - - [430, 23.96] - - - [384, 128, 1, 384] - - [336, 3.649] - - - [5377, 128, 1, 384] - - [375, 33.352] - - - [1152, 128, 1, 384] - - [393, 10.996] - - - [9985, 128, 1, 384] - - [386, 42.849] - - - [4992, 128, 1, 384] - - [335, 34.871] - - - [3456, 128, 1, 384] - - [399, 30.612] - - - [1921, 128, 1, 384] - - [392, 15.697] - - - [5376, 128, 1, 384] - - [336, 37.318] - - - [11904, 128, 1, 384] - - [418, 50.594] - - - [3456, 384, 1, 384] - - [335, 52.297] - - - [384, 384, 1, 384] - - [341, 10.902] - - - [2688, 512, 1, 384] - - [410, 48.391] - - - [1536, 512, 1, 384] - - [336, 41.102] - - - [1152, 257, 1, 384] - - [398, 21.611] - - - [4224, 384, 1, 384] - - [418, 55.675] - - - [768, 257, 1, 384] - - [398, 14.592] - - - [2304, 384, 1, 384] - - [337, 45.963] - - - [768, 384, 1, 384] - - [337, 21.71] - - - [3072, 384, 1, 384] - - [431, 45.05] - - - [1536, 1025, 1, 384] - - [432, 50.72] - - - [3456, 512, 1, 384] - - [410, 59.209] - - - [2688, 384, 1, 384] - - [350, 42.214] - - - [3072, 512, 1, 384] - - [408, 53.669] - - - [1920, 512, 1, 384] - - [350, 40.881] - - - [768, 512, 1, 384] - - [336, 28.582] - - - [384, 385, 1, 384] - - [341, 10.837] - - - [1536, 384, 1, 384] - - [335, 32.29] - - - [1152, 641, 1, 384] - - [341, 39.307] - - - [3840, 384, 1, 384] - - [408, 51.326] - - - [768, 385, 1, 384] - - [335, 21.492] - - - [1536, 641, 1, 384] - - [341, 39.611] - - - [1920, 384, 1, 384] - - [337, 40.235] - - - [1152, 512, 1, 384] - - [335, 32.6] - - - [4608, 384, 1, 384] - - [409, 57.433] - - - [2304, 512, 1, 384] - - [335, 47.349] - - - [1152, 384, 1, 384] - - [336, 31.364] - - - [32, 28672, 1, 32] - - [433, 6.181] - - - [32, 24576, 1, 32] - - [343, 6.63] - - - [32, 16384, 1, 32] - - [343, 5.361] - - - [32, 20480, 1, 32] - - [342, 5.761] - - - [32, 12288, 1, 32] - - [434, 4.572] - - - [32, 8192, 1, 32] - - [342, 3.499] - - - [32, 4096, 1, 32] - - [340, 1.989] - - - [32, 32768, 1, 32] - - [435, 6.934] - - - [64, 64, 1, 64] - - [436, 0.117] - - - [64, 1, 1, 64] - - [436, 0.002] - - - [1, 64, 1, 64] - - [436, 0.002] - - - [1, 1, 1, 64] - - [436, 0.0] - - - [512, 1, 1, 128] - - [436, 0.025] - - - [384, 1, 1, 384] - - [436, 0.03] - - - [256, 1, 1, 256] - - [436, 0.017] - - - [128, 1, 1, 128] - - [436, 0.006] - - - [640, 1, 1, 128] - - [436, 0.032] - - - [1, 128, 1, 256] - - [436, 0.009] - - - [15745, 128, 1, 256] - - [437, 27.345] - - - [22913, 128, 1, 256] - - [438, 36.177] - - - [28289, 128, 1, 128] - - [437, 20.694] - - - [23681, 128, 1, 256] - - [439, 36.993] - - - [24449, 128, 1, 128] - - [438, 24.486] - - - [18305, 128, 1, 256] - - [438, 30.812] - - - [21377, 128, 1, 256] - - [439, 34.302] - - - [24833, 128, 1, 256] - - [440, 38.387] - - - [17153, 128, 1, 128] - - [441, 18.729] - - - [19073, 128, 1, 128] - - [442, 20.373] - - - [27521, 128, 1, 256] - - [443, 41.204] - - - [14081, 128, 1, 128] - - [437, 15.796] - - - [22145, 128, 1, 256] - - [444, 35.265] - - - [20225, 128, 1, 256] - - [444, 33.174] - - - [14849, 128, 1, 256] - - [438, 26.003] - - - [14081, 128, 1, 256] - - [441, 24.426] - - - [16129, 128, 1, 256] - - [438, 27.59] - - - [20993, 128, 1, 128] - - [445, 21.979] - - - [14593, 128, 1, 256] - - [442, 25.525] - - - [15617, 128, 1, 128] - - [437, 17.518] - - - [25217, 128, 1, 256] - - [439, 38.617] - - - [27137, 128, 1, 256] - - [439, 41.004] - - - [24833, 128, 1, 128] - - [438, 24.538] - - - [15361, 128, 1, 128] - - [437, 17.337] - - - [21761, 128, 1, 256] - - [438, 34.842] - - - [15617, 128, 1, 256] - - [445, 27.123] - - - [14209, 128, 1, 128] - - [446, 16.135] - - - [19841, 128, 1, 256] - - [447, 32.4] - - - [28673, 128, 1, 256] - - [442, 31.898] - - - [17153, 128, 1, 256] - - [442, 29.478] - - - [14465, 128, 1, 128] - - [446, 16.375] - - - [16001, 128, 1, 128] - - [441, 17.707] - - - [22529, 128, 1, 256] - - [438, 35.8] - - - [15233, 128, 1, 256] - - [442, 26.581] - - - [16385, 128, 1, 128] - - [442, 18.492] - - - [27905, 128, 1, 128] - - [437, 26.646] - - - [20225, 128, 1, 128] - - [441, 21.175] - - - [26369, 128, 1, 128] - - [437, 25.951] - - - [26369, 128, 1, 256] - - [448, 40.172] - - - [24065, 128, 1, 128] - - [438, 24.069] - - - [20609, 128, 1, 128] - - [441, 21.547] - - - [16769, 128, 1, 128] - - [446, 18.474] - - - [19457, 128, 1, 256] - - [437, 32.31] - - - [16769, 128, 1, 256] - - [444, 28.553] - - - [16129, 128, 1, 128] - - [437, 18.011] - - - [21377, 128, 1, 128] - - [438, 21.885] - - - [23297, 128, 1, 256] - - [449, 36.822] - - - [22145, 128, 1, 128] - - [437, 22.957] - - - [20993, 128, 1, 256] - - [445, 34.018] - - - [25601, 128, 1, 256] - - [449, 38.882] - - - [14465, 128, 1, 256] - - [445, 25.211] - - - [22913, 128, 1, 128] - - [442, 23.168] - - - [14977, 128, 1, 128] - - [438, 16.955] - - - [25601, 128, 1, 128] - - [446, 24.669] - - - [18305, 128, 1, 128] - - [446, 19.753] - - - [14593, 128, 1, 128] - - [437, 16.546] - - - [15233, 128, 1, 128] - - [437, 17.088] - - - [15361, 128, 1, 256] - - [450, 26.647] - - - [18689, 128, 1, 256] - - [451, 30.86] - - - [22529, 128, 1, 128] - - [438, 22.748] - - - [14849, 128, 1, 128] - - [437, 17.019] - - - [26753, 128, 1, 256] - - [452, 40.673] - - - [15745, 128, 1, 128] - - [445, 17.609] - - - [19841, 128, 1, 128] - - [437, 21.072] - - - [17537, 128, 1, 128] - - [445, 19.148] - - - [17921, 128, 1, 256] - - [446, 30.409] - - - [19073, 128, 1, 256] - - [445, 31.601] - - - [27905, 128, 1, 256] - - [453, 41.737] - - - [16001, 128, 1, 256] - - [446, 27.499] - - - [20609, 128, 1, 256] - - [437, 33.58] - - - [23681, 128, 1, 128] - - [441, 23.717] - - - [25985, 128, 1, 128] - - [446, 25.37] - - - [23297, 128, 1, 128] - - [446, 23.719] - - - [25217, 128, 1, 128] - - [446, 24.917] - - - [28289, 128, 1, 256] - - [445, 32.474] - - - [14977, 128, 1, 256] - - [438, 26.165] - - - [17537, 128, 1, 256] - - [437, 29.553] - - - [14209, 128, 1, 256] - - [444, 24.794] - - - [16385, 128, 1, 256] - - [441, 28.591] - - - [18689, 128, 1, 128] - - [450, 19.906] - - - [26753, 128, 1, 128] - - [437, 25.881] - - - [27521, 128, 1, 128] - - [445, 26.45] - - - [21761, 128, 1, 128] - - [445, 22.216] - - - [24449, 128, 1, 256] - - [454, 37.636] - - - [25985, 128, 1, 256] - - [444, 39.546] - - - [28673, 128, 1, 128] - - [438, 20.508] - - - [19457, 128, 1, 128] - - [450, 20.546] - - - [17921, 128, 1, 128] - - [437, 19.773] - - - [24065, 128, 1, 256] - - [438, 37.356] - - - [27137, 128, 1, 128] - - [451, 26.286] -- null